app.py

print(f"Initializing..")
# Essential imports for basic video streaming
import time
import os
import cv2
import queue
import threading
import sys
import json
import math
from datetime import datetime
import argparse
from collections import deque

# Heavy imports will be loaded in background

# ============================================
# USER CONFIGURABLE VARIABLES
# ============================================

# YOLO Model and Stream Settings
model = "yolo12m"  # YOLO model to use (yolo11n, yolo12n, yolo12s, etc.)

# Parse command line arguments
parser = argparse.ArgumentParser(description="MACHINA - Video Stream Processor")
parser.add_argument(
    "--stream",
    type=str,
    default="0",
    help="RTSP stream URL or 0 for webcam (default: 0)",
)
args = parser.parse_args()

# Convert '0' string to integer 0 for webcam
if args.stream == "0":
    rtsp_stream = 0
else:
    rtsp_stream = args.stream

# Processing and Performance Settings
yolo_skip_frames = 2  # Process every Nth frame (2 = every 2nd frame)
buffer = 512  # Frame buffer size
min_confidence = 0.15  # Minimum confidence threshold for detections
min_size = 20  # Minimum size for car detections
stream_timeout_ms = 5000  # Stream timeout detection threshold (5 seconds)

# Object Tracking Settings
point_timeout = 2500  # Time before objects are considered lost (ms)
stationary_val = 16  # Movement threshold for stationary objects
idle_reset = 3000  # Time before frame skip reset (ms)
obj_max = 16  # Maximum number of tracked objects
padding = 6  # Padding around detected objects for cropping

# Display Settings
opsize = (640, 480)  # Default processing/display resolution
yolo_input_size = 640  # Square size divisible by 32 for YOLO input
snapshot_directory = "snapshots"  # Directory for snapshots
_font = cv2.FONT_HERSHEY_SIMPLEX  # Font for text overlay

# Replay and UI Settings
replay_buffer_max_size = 300  # ~10 seconds at 30fps
resolution_display_duration = 2000  # Resolution display time (ms)

# Object Detection Confidence Thresholds
class_confidence = {
    "truck": 0.35,
    "car": 0.15,
    "boat": 0.85,
    "bus": 0.5,
    "aeroplane": 0.85,
    "frisbee": 0.85,
    "pottedplant": 0.55,
    "train": 0.85,
    "chair": 0.5,
    "parking meter": 0.9,
    "fire hydrant": 0.65,
    "traffic light": 0.65,
    "backpack": 0.65,
    "bicycle": 0.55,
    "bench": 0.75,
    "zebra": 0.90,
    "tvmonitor": 0.80,
}

classlist = [
    "person",
    "car",
    "motorbike",
    "bicycle",
    "truck",
    "traffic light",
    "stop sign",
    "bench",
    "bird",
    "cat",
    "dog",
    "backpack",
    "suitcase",
    "handbag",
]

# ============================================
# SYSTEM VARIABLES (DO NOT MODIFY)
# ============================================

# Stream and processing state
frames = 0
prev_frames = 0
last_frame = 0
fps = 0
recording = False
out = None
streamsize = (0, 0)
original_opsize = None  # Will store the original stream size for Ctrl+4

# Stream timeout detection variables
last_frame_time = 0
stream_timeout_threshold = stream_timeout_ms  # Use configurable timeout from settings
max_consecutive_failures = 3  # Max consecutive read failures before restart

# Zoom and pan state
zoom_factor = 1.0
pan_x = 0
pan_y = 0
zoom_mode_active = False
stored_bounding_boxes = []

# UI state variables
hdstream = False
drawing = False
dragging = False
drag_start_x = 0
drag_start_y = 0
draw_start_x = 0
draw_start_y = 0
draw_end_x = 0
draw_end_y = 0
military_mode = False
show_info_overlay = False
show_help_text = False
help_text_start_time = 0
show_frame_skip_display = False
frame_skip_display_start_time = 0
yolo_first_processing_started = False
is_first_run = False
webcam_max_resolution = None  # Store webcam's maximum supported resolution

# Audio visualization state
audio_waveform = deque(maxlen=1024)
audio_waveform_lock = threading.Lock()
audio_visualization_enabled = False
audio_playback_enabled = False
audio_thread = None
audio_muted = False
AUDIO_WAVEFORM_HEIGHT = 60
AUDIO_WAVEFORM_TARGET_POINTS = 1024
AUDIO_WAVEFORM_SECONDS = 1.0
audio_waveform_downsample = 1
audio_waveform_max_points = 1024
audio_waveform_sample_rate = 44100


def transform(xmin, ymin, xmax, ymax, pad):
    x_scale = streamsize[0] / opsize[0]
    y_scale = streamsize[1] / opsize[1]

    new_xmin = int(xmin * x_scale) - pad
    new_ymin = int(ymin * y_scale) - pad
    new_xmax = int(xmax * x_scale) + pad
    new_ymax = int(ymax * y_scale) + pad
    return (new_xmin, new_ymin, new_xmax, new_ymax)


def resample(frame):
    global zoom_factor, pan_x, pan_y, streamsize, opsize

    zoomed_width = int(streamsize[0] / zoom_factor)
    zoomed_height = int(streamsize[1] / zoom_factor)

    center_x = streamsize[0] // 2 + pan_x
    center_y = streamsize[1] // 2 + pan_y

    start_x = max(0, min(streamsize[0] - zoomed_width, center_x - zoomed_width // 2))
    start_y = max(0, min(streamsize[1] - zoomed_height, center_y - zoomed_height // 2))

    zoomed_frame = frame[
        start_y : start_y + zoomed_height, start_x : start_x + zoomed_width
    ]

    return cv2.resize(zoomed_frame, opsize, interpolation=cv2.INTER_LINEAR_EXACT)


def rest(url, payload):
    headers = {"Content-Type": "application/json"}
    r = False
    try:
        data = json.dumps(payload)
        response = requests.post(url, data, headers=headers)
        if response.status_code == 200:
            r = json.loads(response.text)
        else:
            print(response.text)
            return False
    except Exception as e:
        print(f"-- error {e}")
    finally:
        return r


def millis():
    return round(time.perf_counter() * 1000)


def format_duration(seconds):
    """Convert seconds to human readable format like 1h20m5s, 10m, 5s"""
    if seconds < 60:
        return f"{seconds}s"
    elif seconds < 3600:
        minutes = seconds // 60
        remaining_seconds = seconds % 60
        if remaining_seconds == 0:
            return f"{minutes}m"
        else:
            return f"{minutes}m{remaining_seconds}s"
    else:
        hours = seconds // 3600
        remaining_minutes = (seconds % 3600) // 60
        remaining_seconds = seconds % 60
        if remaining_minutes == 0 and remaining_seconds == 0:
            return f"{hours}h"
        elif remaining_seconds == 0:
            return f"{hours}h{remaining_minutes}m"
        else:
            return f"{hours}h{remaining_minutes}m{remaining_seconds}s"


def toggle_fullscreen():
    """Toggle between fullscreen and windowed mode"""
    global fullscreen, window

    if fullscreen:
        # Exit fullscreen
        cv2.setWindowProperty(window, cv2.WND_PROP_FULLSCREEN, cv2.WINDOW_NORMAL)
        cv2.resizeWindow(window, original_window_size[0], original_window_size[1])
        fullscreen = False
        print("Exited fullscreen")
    else:
        # Enter fullscreen
        cv2.setWindowProperty(window, cv2.WND_PROP_FULLSCREEN, cv2.WINDOW_FULLSCREEN)
        fullscreen = True
        print("Entered fullscreen")


def window_resize_callback(val):
    """Callback for window resize events to maintain aspect ratio"""
    global window_aspect_ratio, original_window_size

    # Get current window size
    try:
        # This is a workaround since OpenCV doesn't provide direct window size callbacks
        # We'll handle aspect ratio preservation in the main loop
        pass
    except:
        pass


def reset_window_to_stream_resolution():
    """Reset window size to match stream resolution"""
    global window, streamsize, original_window_size
    reset_size = opsize

    cv2.resizeWindow(window, reset_size[0], reset_size[1])
    original_window_size = reset_size
    print(f"Reset window to {reset_size[0]}x{reset_size[1]}")


def resize_stream_dimensions(new_size):
    """Resize stream processing dimensions and OpenCV window"""
    global opsize, window, original_window_size, yolo_input_size, fullscreen, cap, rtsp_stream
    global resolution_display_active, resolution_display_text, resolution_display_start_time

    opsize = new_size

    # If using webcam, check against maximum supported resolution
    if rtsp_stream == 0 and cap and webcam_max_resolution:
        if (
            opsize[0] <= webcam_max_resolution[0]
            and opsize[1] <= webcam_max_resolution[1]
        ):
            cap.set(cv2.CAP_PROP_FRAME_WIDTH, opsize[0])
            cap.set(cv2.CAP_PROP_FRAME_HEIGHT, opsize[1])
            print(f"Set webcam resolution to {opsize[0]}x{opsize[1]}")

            # Verify what resolution was actually set
            actual_width = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
            actual_height = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
            if (actual_width, actual_height) != opsize:
                print(
                    f"Warning: Webcam set to {actual_width}x{actual_height}, not {opsize[0]}x{opsize[1]}"
                )
        else:
            print(
                f"Cannot set {opsize[0]}x{opsize[1]} - webcam max is {webcam_max_resolution[0]}x{webcam_max_resolution[1]}"
            )
            # Don't change the resolution if it exceeds webcam capability
            return

    # Adjust YOLO input size to be the larger dimension, rounded up to nearest 32
    max_dim = max(opsize[0], opsize[1])
    yolo_input_size = ((max_dim + 31) // 32) * 32  # Round up to nearest multiple of 32

    # If currently in fullscreen, maintain fullscreen after resize
    if fullscreen:
        cv2.setWindowProperty(window, cv2.WND_PROP_FULLSCREEN, cv2.WINDOW_FULLSCREEN)
    else:
        cv2.resizeWindow(window, opsize[0], opsize[1])

    original_window_size = opsize

    # Activate resolution display
    resolution_display_active = True
    resolution_display_text = f"{opsize[0]}x{opsize[1]}"
    resolution_display_start_time = millis()

    print(
        f"Resized stream to {opsize[0]}x{opsize[1]}, YOLO input: {yolo_input_size}x{yolo_input_size}"
    )


def timestamp():
    return int(time.time())


def detect_webcam_max_resolution(cap):
    """Detect the maximum resolution supported by the webcam with fingerprinting"""
    global webcam_max_resolution

    print("DEBUG: Starting webcam resolution detection with fingerprinting...")

    # Get webcam fingerprint
    webcam_id, fingerprint = get_webcam_fingerprint(cap)

    # Load existing webcam configurations
    webcam_configs = load_webcam_config()

    # Check if we already know this webcam
    if webcam_id in webcam_configs:
        print(f"DEBUG: Found existing config for webcam: {webcam_id}")
        saved_config = webcam_configs[webcam_id]
        max_width = saved_config["max_width"]
        max_height = saved_config["max_height"]

        print(f"DEBUG: Using saved resolution: {max_width}x{max_height}")

        # Set the webcam to the known good resolution
        cap.set(cv2.CAP_PROP_FRAME_WIDTH, max_width)
        cap.set(cv2.CAP_PROP_FRAME_HEIGHT, max_height)

        webcam_max_resolution = (max_width, max_height)
        print(f"Webcam resolution loaded from cache: {max_width}x{max_height}")
        return webcam_max_resolution

    # New webcam - need to test resolutions
    print(f"DEBUG: New webcam detected: {webcam_id}")
    print("DEBUG: Testing resolutions for the first time...")

    # Test resolutions from highest to lowest - stop at first working one
    test_resolutions = [
        (1920, 1080),  # 1080p
        (1280, 720),  # 720p
        (640, 480),  # 480p
    ]

    print(
        f"DEBUG: Testing resolutions in order (will stop at first working): {[f'{w}x{h}' for w, h in test_resolutions]}"
    )

    max_width = 0
    max_height = 0

    for i, (width, height) in enumerate(test_resolutions):
        print(
            f"DEBUG: Testing resolution {i+1}/{len(test_resolutions)}: {width}x{height}"
        )

        # Try to set resolution
        print(f"DEBUG: Setting width to {width}...")
        cap.set(cv2.CAP_PROP_FRAME_WIDTH, width)
        print(f"DEBUG: Setting height to {height}...")
        cap.set(cv2.CAP_PROP_FRAME_HEIGHT, height)

        # Check what was actually set
        print("DEBUG: Getting actual width...")
        actual_width = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
        print("DEBUG: Getting actual height...")
        actual_height = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))

        print(f"DEBUG: Actual resolution set: {actual_width}x{actual_height}")

        # If we got the requested resolution, this is supported
        if (
            actual_width >= width * 0.9 and actual_height >= height * 0.9
        ):  # Allow 10% tolerance
            print(f"DEBUG: Resolution {width}x{height} works! Using this resolution.")
            max_width = actual_width
            max_height = actual_height
            print(f"DEBUG: Selected resolution: {max_width}x{max_height}")
            break  # Stop testing - use the first working resolution
        else:
            print(f"DEBUG: Resolution {width}x{height} not supported, trying next...")

    print("DEBUG: Resolution detection complete")
    webcam_max_resolution = (max_width, max_height)

    # Save the results for this webcam
    webcam_configs[webcam_id] = {
        "fingerprint": fingerprint,
        "max_width": max_width,
        "max_height": max_height,
        "tested_date": datetime.now().isoformat(),
        "resolutions_tested": test_resolutions,
    }

    if save_webcam_config(webcam_configs):
        print(f"DEBUG: Saved webcam config for future use: {webcam_id}")

    print(f"Webcam maximum resolution detected and saved: {max_width}x{max_height}")
    return webcam_max_resolution


def get_available_resolutions():
    """Get available resolutions based on current stream type"""
    global webcam_max_resolution, rtsp_stream

    # Define resolution presets
    all_resolutions = [(640, 480), (800, 600), (1024, 768), (1280, 800), (1920, 1080)]

    if rtsp_stream == 0 and webcam_max_resolution:
        # Filter resolutions that fit within webcam capability
        available = []
        for res in all_resolutions:
            if (
                res[0] <= webcam_max_resolution[0]
                and res[1] <= webcam_max_resolution[1]
            ):
                available.append(res)
        return available if available else [webcam_max_resolution]
    else:
        # For RTSP streams, all resolutions are available (will be scaled)
        return all_resolutions


def cycle_resolution(direction):
    """Cycle through available resolution presets with + and - keys"""
    global opsize

    available_resolutions = get_available_resolutions()

    # Find current resolution in available list
    try:
        current_index = available_resolutions.index(opsize)
    except ValueError:
        # Current resolution not in list, start from first
        current_index = 0

    if direction > 0:  # + key - increase resolution
        current_index = (current_index + 1) % len(available_resolutions)
    else:  # - key - decrease resolution
        current_index = (current_index - 1) % len(available_resolutions)

    new_resolution = available_resolutions[current_index]

    # Show available resolutions for debugging
    if rtsp_stream == 0:
        print(f"Available webcam resolutions: {available_resolutions}")

    resize_stream_dimensions(new_resolution)
    print(f"Resolution cycled to: {new_resolution[0]}x{new_resolution[1]}")


def load_config():
    """Load configuration from config.json, create if doesn't exist"""
    global yolo_skip_frames, opsize, show_info_overlay, is_first_run

    config_file = "config.json"
    default_config = {
        "processing_nth_frame": 2,
        "screen_resolution": [640, 480],
        "window_position": [100, 100],
        "first_run": True,
    }

    try:
        if os.path.exists(config_file):
            with open(config_file, "r") as f:
                config = json.load(f)
                yolo_skip_frames = config.get("processing_nth_frame", 2)
                opsize = tuple(config.get("screen_resolution", [640, 480]))
                window_pos = config.get("window_position", [100, 100])
                is_first_run = config.get("first_run", False)
                print(
                    f"Config loaded: skip={yolo_skip_frames}, resolution={opsize}, pos={window_pos}"
                )
                return window_pos
        else:
            # First run - create config file
            with open(config_file, "w") as f:
                json.dump(default_config, f, indent=2)
            is_first_run = True
            print("First run detected - created config.json")
            return default_config["window_position"]
    except Exception as e:
        print(f"Error loading config: {e}")
        is_first_run = True
        return default_config["window_position"]


def save_config():
    """Save current configuration to config.json"""
    global yolo_skip_frames, opsize, window

    config_file = "config.json"

    try:
        # Get current window position (if possible)
        window_pos = [100, 100]  # Default fallback
        try:
            # OpenCV doesn't provide direct way to get window position
            # We'll use the stored values or defaults
            pass
        except:
            pass

        config = {
            "processing_nth_frame": yolo_skip_frames,
            "screen_resolution": list(opsize),
            "window_position": window_pos,
            "first_run": False,
        }

        with open(config_file, "w") as f:
            json.dump(config, f, indent=2)
        print(f"Configuration saved: skip={yolo_skip_frames}, resolution={opsize}")
    except Exception as e:
        print(f"Error saving config: {e}")


def get_gpu_info():
    """Get GPU name and VRAM info"""
    try:
        import torch

        if torch.cuda.is_available():
            gpu_name = torch.cuda.get_device_name(0)
            gpu_memory = torch.cuda.get_device_properties(0).total_memory
            gpu_vram_gb = round(gpu_memory / 1024**3, 1)
            return gpu_name, f"{gpu_vram_gb}GB"
        else:
            return "No CUDA GPU", "N/A"
    except:
        return "Unknown GPU", "N/A"


def get_webcam_fingerprint(cap):
    """Get webcam identifying information for fingerprinting"""
    try:
        # Try to get webcam properties that might identify it
        backend_name = cap.getBackendName()

        # Get various properties that might be unique to this webcam
        width = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
        height = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
        fps = cap.get(cv2.CAP_PROP_FPS)
        brightness = cap.get(cv2.CAP_PROP_BRIGHTNESS)
        contrast = cap.get(cv2.CAP_PROP_CONTRAST)
        saturation = cap.get(cv2.CAP_PROP_SATURATION)
        hue = cap.get(cv2.CAP_PROP_HUE)

        # Create a fingerprint from available properties
        fingerprint = {
            "backend": backend_name,
            "default_width": width,
            "default_height": height,
            "fps": fps,
            "brightness": brightness,
            "contrast": contrast,
            "saturation": saturation,
            "hue": hue,
        }

        # Create a unique ID by combining key properties
        unique_id = f"{backend_name}_{width}x{height}_{int(fps)}fps"

        print(f"DEBUG: Webcam fingerprint created: {unique_id}")
        print(f"DEBUG: Fingerprint details: {fingerprint}")

        return unique_id, fingerprint

    except Exception as e:
        print(f"DEBUG: Error getting webcam fingerprint: {e}")
        # Fallback fingerprint
        return "unknown_webcam", {"backend": "unknown"}


def load_webcam_config():
    """Load webcam configuration from webcam.json"""
    config_file = "webcam.json"

    try:
        if os.path.exists(config_file):
            with open(config_file, "r") as f:
                config = json.load(f)
                print(f"DEBUG: Loaded webcam config with {len(config)} entries")
                return config
        else:
            print("DEBUG: No webcam.json found, will create new one")
            return {}
    except Exception as e:
        print(f"DEBUG: Error loading webcam config: {e}")
        return {}


def save_webcam_config(webcam_configs):
    """Save webcam configuration to webcam.json"""
    config_file = "webcam.json"

    try:
        with open(config_file, "w") as f:
            json.dump(webcam_configs, f, indent=2)
        print(f"DEBUG: Saved webcam config with {len(webcam_configs)} entries")
        return True
    except Exception as e:
        print(f"DEBUG: Error saving webcam config: {e}")
        return False


def get_app_modified_time():
    """Get last modified time of the running script with relative time"""
    try:
        # Get the actual script filename that's being executed
        script_path = os.path.abspath(sys.argv[0])
        script_name = os.path.basename(script_path)

        modified_time = os.path.getmtime(script_path)
        modified_datetime = datetime.fromtimestamp(modified_time)
        current_datetime = datetime.now()

        # Calculate days difference
        time_diff = current_datetime - modified_datetime
        days_ago = time_diff.days

        # Format the date (handle both Linux and Windows formats)
        try:
            formatted_date = modified_datetime.strftime("%-d %b %Y")  # Linux format
        except:
            formatted_date = modified_datetime.strftime("%#d %b %Y")  # Windows format

        if days_ago == 0:
            relative_time = "today"
        elif days_ago == 1:
            relative_time = "1 day ago"
        else:
            relative_time = f"{days_ago} days ago"

        return f"{script_name}: {formatted_date} ({relative_time})"
    except:
        return "Script: Unknown"


def draw_info_overlay(img):
    """Draw the MACHINA info overlay with reduced brightness"""
    overlay = img.copy()

    # Reduce brightness by 50%
    overlay = cv2.convertScaleAbs(overlay, alpha=0.5, beta=0)

    height, width = img.shape[:2]

    # Draw MACHINA title at top
    title = "MACHINA"
    title_font_scale = 2.0
    title_thickness = 3
    title_size = cv2.getTextSize(title, _font, title_font_scale, title_thickness)[0]
    title_x = (width - title_size[0]) // 2
    title_y = 60

    # Draw title with shadow
    cv2.putText(
        overlay,
        title,
        (title_x + 2, title_y + 2),
        _font,
        title_font_scale,
        (0, 0, 0),
        title_thickness + 2,
    )
    cv2.putText(
        overlay,
        title,
        (title_x, title_y),
        _font,
        title_font_scale,
        (255, 255, 255),
        title_thickness,
    )

    # Draw separator line
    line_y = title_y + 20
    line_start_x = width // 4
    line_end_x = 3 * width // 4
    cv2.line(overlay, (line_start_x, line_y), (line_end_x, line_y), (255, 255, 255), 2)

    # Keyboard commands list
    commands = [
        "SPACE - Frame skip",
        "Q - Quit",
        "R - Toggle recording",
        "S - Take snapshot",
        "F - Reset window size",
        "M - Toggle military mode",
        "ENTER - Toggle fullscreen",
        "ESC - Exit fullscreen",
        "TAB - Toggle this info",
        "BACKSPACE - Toggle replay mode",
        "1-6 - Change resolution",
        "+ / - - Adjust frame processing frequency",
        "Mouse wheel - Zoom",
        "Right click drag - Pan",
        "Left click drag - Save selection to elements folder",
    ]

    # Draw commands
    cmd_y = line_y + 40
    cmd_font_scale = 0.6
    cmd_thickness = 1
    line_height = 25

    for i, cmd in enumerate(commands):
        y_pos = cmd_y + (i * line_height)
        if y_pos > height - 100:  # Don't go too far down
            break
        cv2.putText(
            overlay,
            cmd,
            (50, y_pos),
            _font,
            cmd_font_scale,
            (255, 255, 255),
            cmd_thickness,
        )

    # Get system info
    gpu_name, gpu_vram = get_gpu_info()
    app_modified = get_app_modified_time()

    # Draw system info at bottom
    info_y = height - 60
    cv2.putText(
        overlay, f"GPU: {gpu_name}", (20, info_y), _font, 0.5, (0, 255, 255), 1
    )  # Yellow
    cv2.putText(
        overlay, f"VRAM: {gpu_vram}", (20, info_y + 20), _font, 0.5, (0, 0, 255), 1
    )  # Red
    cv2.putText(
        overlay, app_modified, (20, info_y + 40), _font, 0.4, (255, 255, 255), 1
    )

    return overlay


labels = open("db/coco.names").read().strip().split("\n")
classlist = [labels.index(x) for x in classlist]

object_count = 0
old_count = 0
obj_break = millis()
obj_idle = 0
obj_list = []
obj_max = 16
fskip = False
last_fskip = timestamp()
obj_score = labels

bounding_boxes = []
obj_number = 1

# YOLO processing state
yolo_frame_count = 0
cached_yolo_results = None
zoom_pan_active = False
zoom_pan_pause_time = 0
last_yolo_processing_duration = 0  # Store last YOLO processing time

# Person detection cache for clustering
cached_person_detections = []
last_person_update_frame = 0

# Replay system variables
replay_buffer = []
replay_mode = False
replay_index = 0
replay_last_flash_time = 0

# Window management variables
fullscreen = False
window_aspect_ratio = 4 / 3  # Default aspect ratio
original_window_size = (640, 480)

# Resolution display variables
resolution_display_active = False
resolution_display_text = ""
resolution_display_start_time = 0

# Selection and clipboard variables
selection_complete = False
clipboard_message_active = False
clipboard_message_start_time = 0
clipboard_message_duration = 2000  # Show message for 2 seconds
selection_start_x = 0
selection_start_y = 0
selection_end_x = 0
selection_end_y = 0
clean_processed_img = None  # Store clean image before UI overlays


def center(xmin, ymin, xmax, ymax):
    center_x = (xmin + xmax) // 2
    center_y = (ymin + ymax) // 2
    return (center_x, center_y)


def _size(x1, y1, x2, y2):
    return abs(x1 - y2)


def mouse_callback(event, x, y, flags, param):
    global drawing, draw_start_x, draw_start_y, draw_end_x, draw_end_y, dragging, drag_start_x, drag_start_y, zoom_factor, pan_x, pan_y, zoom_pan_active, zoom_pan_pause_time, cached_yolo_results, zoom_mode_active, stored_bounding_boxes, bounding_boxes, fullscreen, window, selection_complete, clipboard_message_active, clipboard_message_start_time, selection_start_x, selection_start_y, selection_end_x, selection_end_y, audio_muted
    if event == cv2.EVENT_RBUTTONDOWN:
        dragging = True
        zoom_pan_active = True
        zoom_pan_pause_time = millis()
        cached_yolo_results = None  # Clear cache when pan starts
        drag_start_x = x
        drag_start_y = y

    if event == cv2.EVENT_RBUTTONUP:
        dragging = False
        zoom_pan_active = False
        cached_yolo_results = None  # Clear cache when pan ends

    if event == cv2.EVENT_LBUTTONUP:
        if drawing and draw_start_x > 0 and draw_end_x > 0:
            # Store selection coordinates before reset
            selection_start_x = draw_start_x
            selection_start_y = draw_start_y
            selection_end_x = draw_end_x
            selection_end_y = draw_end_y
            # Mark selection as complete for clipboard copying
            selection_complete = True
        drawing = False
        draw_end_x = 0
        draw_end_y = 0
        draw_start_x = 0
        draw_start_y = 0

    if event == cv2.EVENT_LBUTTONDOWN:
        # Start drawing
        drawing = True
        draw_end_x = 0
        draw_end_y = 0
        draw_start_x = x
        draw_start_y = y

    if event == cv2.EVENT_MOUSEWHEEL:
        zoom_pan_active = True
        zoom_pan_pause_time = millis()
        cached_yolo_results = None  # Clear cache on zoom in/out

        # Store objects when entering zoom mode
        if zoom_factor == 1.0 and not zoom_mode_active:
            pan_x = 0
            pan_y = 0

        old_zoom_factor = zoom_factor

        if flags > 0:
            zoom_factor = min(6.0, zoom_factor * 1.1)
        else:
            zoom_factor = max(1.0, zoom_factor / 1.1)

        # Entering zoom mode - store current tracking state
        if old_zoom_factor == 1.0 and zoom_factor > 1.0 and not zoom_mode_active:
            stored_bounding_boxes = copy.deepcopy(bounding_boxes)
            zoom_mode_active = True
            print(f"Entering zoom mode - stored {len(stored_bounding_boxes)} objects")

        # Exiting zoom mode - restore tracking state only when returning to 1.0x
        elif zoom_factor == 1.0 and zoom_mode_active:
            zoom_mode_active = False
            bounding_boxes = copy.deepcopy(stored_bounding_boxes)
            stored_bounding_boxes = []
            print(f"Exiting zoom mode - restored {len(bounding_boxes)} objects")

    if event == cv2.EVENT_MBUTTONDOWN:
        audio_muted = not audio_muted
        state = "muted" if audio_muted else "unmuted"
        print(f"Audio {state} via middle mouse button")
        with audio_waveform_lock:
            audio_waveform.clear()

    if event == cv2.EVENT_MOUSEMOVE:
        if drawing:
            draw_end_x = x
            draw_end_y = y

        if dragging:
            dx = x - drag_start_x
            dy = y - drag_start_y

            pan_x -= int(dx * zoom_factor)
            pan_y -= int(dy * zoom_factor)

            drag_start_x = x
            drag_start_y = y


class BoundingBox:
    def __init__(self, name, points, size, image):
        global obj_number
        self.nr = obj_number
        obj_number += 1
        self.x, self.y = points
        self.created = millis()
        self.timestamp = self.created
        self.size = size
        self.name = name
        self.checkin = True
        self.detections = 0
        self.idle = 0
        self.image = image
        self.desc = False
        self.state = 0
        self.seen = self.created
        self.disappeared_cycles = 0  # Track consecutive disappearance cycles

        self.init()
        print(
            "New object: " + self.name + "#" + str(self.nr) + " size:" + str(self.size)
        )

    def see(self):
        self.seen = millis()

    def ping(self):
        self.timestamp = millis()
        idle = self.timestamp - self.created
        if idle >= 1000:
            self.idle = idle // 1000
        else:
            self.idle = 0
        return self.idle

    def export(self):
        _, buffer = cv2.imencode(".png", self.image)
        base64_image = base64.b64encode(buffer.tobytes()).decode("utf-8")
        return base64_image

    def init(self):
        self.min_x = self.x - stationary_val
        self.max_x = self.x + stationary_val
        self.min_y = self.y - (stationary_val)
        self.max_y = self.y + (stationary_val)

    def contains(self, x, y, time):
        return (
            ((self.checkin == False) and self.min_x <= x <= self.max_x)
            and (self.min_y <= y <= self.max_y)
            and (time - self.seen < point_timeout)
        )

    def update(self, time, new_x, new_y):
        self.checkin = True
        self.timestamp = time
        idle = self.timestamp - self.created
        if idle >= 1000:
            self.idle = idle // 1000
        else:
            self.idle = 0
        self.x = new_x
        self.y = new_y
        self.detections += 1
        self.disappeared_cycles = (
            0  # Reset disappearance counter when object is detected
        )
        self.init()


def resetIteration():
    global bounding_boxes
    # Reset checkin status and increment disappeared_cycles for objects not seen
    for item in bounding_boxes:
        if item.checkin:
            item.disappeared_cycles = 0  # Reset counter if object was seen
        else:
            item.disappeared_cycles += 1  # Increment if object not detected
        item.checkin = False


def save_selection_to_elements(frame, start_x, start_y, end_x, end_y):
    """Save the selected rectangular region to elements folder as PNG"""
    global clipboard_message_active, clipboard_message_start_time

    try:
        # Ensure coordinates are within frame bounds
        h, w = frame.shape[:2]
        x1 = max(0, min(start_x, end_x))
        y1 = max(0, min(start_y, end_y))
        x2 = min(w, max(start_x, end_x))
        y2 = min(h, max(start_y, end_y))

        # Extract the selected region
        selected_region = frame[y1:y2, x1:x2]

        if selected_region.size > 0:
            # Create elements directory if it doesn't exist
            elements_dir = "elements"
            if not os.path.exists(elements_dir):
                os.makedirs(elements_dir)

            # Generate filename with current date-time
            timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
            filename = f"element_{timestamp}.png"
            filepath = os.path.join(elements_dir, filename)

            # Save the selected region as PNG
            success = cv2.imwrite(filepath, selected_region)

            if success:
                # Show success message
                clipboard_message_active = True
                clipboard_message_start_time = millis()
                print(f"Element saved: {filepath} ({x2-x1}x{y2-y1})")
                return True
            else:
                print("Failed to save element")
                return False

    except Exception as e:
        print(f"Error saving element: {e}")
        return False

    return False


def take_snapshot(frame):
    global snapshot_directory

    if not os.path.exists(snapshot_directory):
        os.makedirs(snapshot_directory)

    timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
    filename = f"snapshot_{timestamp}.jpg"
    filepath = os.path.join(snapshot_directory, filename)

    cv2.imwrite(filepath, frame)
    print(f"Snapshot saved: {filepath}")


def start_recording(cap):
    global recording, out
    if not recording:
        timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
        filename = f"recordings/recording_{timestamp}.mp4"

        fourcc = cv2.VideoWriter_fourcc(*"mp4v")
        out = cv2.VideoWriter(filename, fourcc, 20, opsize)

        recording = True
        print(f"Started recording: {filename}")


def stop_recording():
    global recording, out
    if recording:
        out.release()
        recording = False
        print("Stopped recording")


def add(num):
    if len(obj_list) >= obj_max:
        obj_list.pop(0)
    obj_list.append(num)


def average():
    l = len(obj_list)
    if l <= 0:
        return 0
    return round(sum(obj_list) / l)


def draw_dashed_rectangle(img, pt1, pt2, color, thickness=1, dash_length=8):
    def draw_dashed_line(img, pt1, pt2, color, thickness, dash_length):

        dist = math.sqrt((pt1[0] - pt2[0]) ** 2 + (pt1[1] - pt2[1]) ** 2)
        dashes = int(dist / dash_length)
        for i in range(dashes):
            start = [
                int(pt1[0] + (pt2[0] - pt1[0]) * i / dashes),
                int(pt1[1] + (pt2[1] - pt1[1]) * i / dashes),
            ]
            end = [
                int(pt1[0] + (pt2[0] - pt1[0]) * (i + 0.5) / dashes),
                int(pt1[1] + (pt2[1] - pt1[1]) * (i + 0.5) / dashes),
            ]
            cv2.line(img, tuple(start), tuple(end), color, thickness)

    draw_dashed_line(img, pt1, (pt2[0], pt1[1]), color, thickness, dash_length)
    draw_dashed_line(img, (pt2[0], pt1[1]), pt2, color, thickness, dash_length)
    draw_dashed_line(img, pt2, (pt1[0], pt2[1]), color, thickness, dash_length)
    draw_dashed_line(img, (pt1[0], pt2[1]), pt1, color, thickness, dash_length)
    return img


def generate_color_shades(num_classes):
    # This function will be called after numpy is imported
    colors = np.zeros((num_classes, 3), dtype=np.uint8)

    green = [0, 200, 0]
    orange = [0, 165, 255]
    yellow = [0, 200, 255]
    red = [0, 0, 255]

    base_colors = [green, orange, yellow, red]
    num_base_colors = len(base_colors)

    for i in range(num_classes):
        base_color_index = i % num_base_colors
        base_color = np.array(base_colors[base_color_index])
        shade_factor = (i // num_base_colors) / (num_classes // num_base_colors + 1)
        shade = (
            base_color * (1 - shade_factor) + np.array([128, 128, 128]) * shade_factor
        )
        colors[i] = shade.astype(np.uint8)
    return colors


# start showing the stream in this line

# Global variables for background loading
yolo_model_loaded = False
yolo_model = None
yolo_loading_complete = False
basic_stream_mode = True
colors = None
device = None

print(f"Starting up..")

# Load configuration
window_position = load_config()
print(f"First run status: {is_first_run}")

# Initialize stream immediately
print("DEBUG: Initializing video capture...")
loop = True
print(f"DEBUG: Creating VideoCapture with rtsp_stream = {rtsp_stream}")
cap = cv2.VideoCapture(rtsp_stream)

print(f"DEBUG: VideoCapture created, isOpened() = {cap.isOpened()}")

if rtsp_stream == 0:
    print("DEBUG: Webcam mode detected, starting resolution detection...")
    # Detect webcam maximum supported resolution
    max_res = detect_webcam_max_resolution(cap)

    print(f"DEBUG: Setting webcam to max resolution {max_res[0]}x{max_res[1]}...")
    # Set to maximum supported resolution
    cap.set(cv2.CAP_PROP_FRAME_WIDTH, max_res[0])
    cap.set(cv2.CAP_PROP_FRAME_HEIGHT, max_res[1])

    print("DEBUG: Verifying final webcam resolution...")
    # Verify final setting
    actual_width = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
    actual_height = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
    print(f"Webcam initialized at {actual_width}x{actual_height}")

    # Update opsize to match webcam capability if needed
    if max_res[0] < opsize[0] or max_res[1] < opsize[1]:
        print(f"Adjusting display size to webcam capability: {max_res[0]}x{max_res[1]}")
        opsize = max_res

print("DEBUG: Getting FPS property...")
fps = cap.get(cv2.CAP_PROP_FPS)
print(f"DEBUG: FPS = {fps}")

print("DEBUG: Attempting to read first frame...")
ret, img = cap.read()
print(
    f"DEBUG: First frame read result: ret={ret}, img shape={img.shape if img is not None else 'None'}"
)

# Initialize frame timeout tracking
if ret:
    last_frame_time = millis()

print("DEBUG: Setting up window and getting stream size...")
window = str(rtsp_stream)
print("DEBUG: Getting stream width...")
stream_width = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
print("DEBUG: Getting stream height...")
stream_height = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
streamsize = (stream_width, stream_height)
print(f"DEBUG: Stream size = {streamsize}")

# Store original stream size for Ctrl+4 functionality
if streamsize[0] > 0 and streamsize[1] > 0:
    original_opsize = streamsize  # Store original stream dimensions
    window_aspect_ratio = streamsize[0] / streamsize[1]
    original_window_size = streamsize
else:
    original_opsize = opsize  # Fallback to default size
    window_aspect_ratio = opsize[0] / opsize[1]
    original_window_size = opsize

if streamsize[0] > opsize[0] or streamsize[1] > opsize[1]:
    hdstream = True
    print("HD stream mode enabled")
elif rtsp_stream == 0:
    hdstream = True
    print("Webcam mode - HD stream enabled for resampling")
else:
    print("Stream resolution matches to window size")

cv2.namedWindow(window, cv2.WINDOW_NORMAL)
cv2.resizeWindow(window, opsize[0], opsize[1])
cv2.setWindowProperty(window, cv2.WND_PROP_TOPMOST, 1)
cv2.setMouseCallback(window, mouse_callback)
q = queue.Queue(maxsize=buffer)


def start_audio_capture():
    """Start background audio capture for playback and waveform visualization."""

    global audio_thread, audio_visualization_enabled, audio_playback_enabled

    if isinstance(rtsp_stream, int):
        print("Audio playback disabled: webcam sources are not supported.")
        return

    if audio_thread and audio_thread.is_alive():
        return

    sd = None

    try:
        import av  # type: ignore
        import numpy as _np  # Local import to avoid global dependency before background load
    except ImportError:
        print("Audio playback disabled: install 'av' for audio support.")
        return

    try:
        import sounddevice as sd  # type: ignore
    except ImportError:
        print("Audio playback unavailable: install 'sounddevice' to enable speakers.")

    def _audio_loop():
        nonlocal _np
        global audio_visualization_enabled, audio_playback_enabled
        global audio_waveform, audio_waveform_downsample, audio_waveform_max_points
        global audio_waveform_sample_rate, audio_muted

        playback_stream = None
        playback_ready = sd is not None

        try:
            container = av.open(rtsp_stream, options={"rtsp_transport": "tcp"})
        except Exception as exc:
            print(f"Audio playback disabled: unable to open stream ({exc}).")
            return

        audio_stream = next((s for s in container.streams if s.type == "audio"), None)

        if audio_stream is None:
            print("Audio playback disabled: no audio stream detected.")
            container.close()
            return

        audio_stream.thread_type = "AUTO"
        sample_rate = audio_stream.codec_context.sample_rate or 44100
        channels = audio_stream.codec_context.channels or 1

        audio_waveform_sample_rate = sample_rate
        downsample = max(
            1,
            int(sample_rate * AUDIO_WAVEFORM_SECONDS / AUDIO_WAVEFORM_TARGET_POINTS),
        )
        max_points = max(1, int(sample_rate * AUDIO_WAVEFORM_SECONDS / downsample))

        with audio_waveform_lock:
            audio_waveform_downsample = downsample
            audio_waveform_max_points = max_points
            audio_waveform = deque(maxlen=audio_waveform_max_points)

        if playback_ready:
            try:
                playback_stream = sd.OutputStream(
                    samplerate=sample_rate,
                    channels=channels,
                    dtype="float32",
                    latency="low",
                )
                playback_stream.start()
                audio_playback_enabled = True
            except Exception as exc:
                print(f"Audio playback disabled: unable to open output device ({exc}).")
                playback_stream = None
                audio_playback_enabled = False
        else:
            audio_playback_enabled = False

        audio_visualization_enabled = True

        try:
            for packet in container.demux(audio_stream):
                if not loop:
                    break

                if packet.dts is None:
                    continue

                try:
                    frames_iter = packet.decode()
                except Exception as exc:
                    print(f"Audio decode error: {exc}")
                    continue

                for frame in frames_iter:
                    if frame is None:
                        continue

                    # Skip processing when muted to save CPU
                    if audio_muted:
                        continue

                    try:
                        samples = frame.to_ndarray()
                    except Exception as exc:
                        print(f"Audio frame conversion error: {exc}")
                        continue

                    if samples.size == 0:
                        continue

                    if samples.dtype == _np.int16:
                        normalized = samples.astype(_np.float32) / 32768.0
                    elif samples.dtype == _np.int32:
                        normalized = samples.astype(_np.float32) / 2147483648.0
                    elif samples.dtype == _np.uint8:
                        normalized = (samples.astype(_np.float32) - 128.0) / 128.0
                    elif samples.dtype == _np.float32:
                        normalized = samples
                    else:
                        normalized = samples.astype(_np.float32)

                    normalized = _np.clip(normalized, -1.0, 1.0)

                    if normalized.ndim == 1:
                        playback_chunk = normalized
                    elif normalized.ndim == 2:
                        if normalized.shape[0] == channels:
                            playback_chunk = normalized.T
                        elif normalized.shape[1] == channels:
                            playback_chunk = normalized
                        else:
                            if channels > 1:
                                playback_chunk = normalized.reshape(-1, channels)
                            else:
                                playback_chunk = normalized.reshape(-1)
                    else:
                        if channels > 1:
                            playback_chunk = normalized.reshape(-1, channels)
                        else:
                            playback_chunk = normalized.reshape(-1)

                    playback_chunk = _np.ascontiguousarray(playback_chunk, dtype=_np.float32)
                    
                    if playback_stream is not None:
                        chunk_to_write = (
                            playback_chunk
                            if not audio_muted
                            else _np.zeros_like(playback_chunk)
                        )
                        try:
                            playback_stream.write(chunk_to_write)
                        except Exception as exc:
                            print(f"Audio playback error: {exc}")
                            playback_stream.stop()
                            playback_stream.close()
                            playback_stream = None
                            audio_playback_enabled = False

                    if playback_chunk.ndim == 2:
                        mono_samples = playback_chunk.mean(axis=1)
                    else:
                        mono_samples = playback_chunk

                    mono_samples = _np.clip(mono_samples, -1.0, 1.0)

                    downsample = max(1, audio_waveform_downsample)
                    downsampled = mono_samples[::downsample]

                    if downsampled.size == 0:
                        continue

                    with audio_waveform_lock:
                        audio_waveform.extend(downsampled.tolist())
        finally:
            audio_visualization_enabled = False
            audio_playback_enabled = False
            if playback_stream is not None:
                try:
                    playback_stream.stop()
                except Exception:
                    pass
                try:
                    playback_stream.close()
                except Exception:
                    pass
            container.close()

    audio_thread = threading.Thread(target=_audio_loop, name="audio_capture", daemon=True)
    audio_thread.start()


start_audio_capture()


def get_audio_waveform_offset(image_height):
    if not audio_visualization_enabled or audio_muted or image_height <= 0:
        return 0
    return min(AUDIO_WAVEFORM_HEIGHT, max(30, image_height // 8))


def draw_audio_waveform(img):
    if not audio_visualization_enabled or audio_muted:
        return img

    try:
        import numpy as np
    except ImportError:
        return img

    with audio_waveform_lock:
        data = list(audio_waveform)

    if len(data) < 2:
        return img

    wave_height = get_audio_waveform_offset(img.shape[0])
    if wave_height <= 0:
        return img

    base_y = img.shape[0] - 1
    top_y = max(0, base_y - wave_height + 1)

    overlay = img.copy()
    cv2.rectangle(overlay, (0, top_y), (img.shape[1], base_y), (0, 40, 0), -1)
    cv2.addWeighted(overlay, 0.35, img, 0.65, 0, img)

    arr = np.array(data, dtype=np.float32)
    if arr.size < 2:
        return img

    arr = np.clip(arr, -1.0, 1.0)

    x_coords = np.linspace(0, img.shape[1] - 1, num=img.shape[1], dtype=np.int32)
    sample_positions = np.linspace(0, arr.size - 1, num=img.shape[1])
    samples = np.interp(sample_positions, np.arange(arr.size), arr)

    center_y = top_y + wave_height // 2
    amplitude_scale = max(1, (wave_height // 2) - 2)
    y_coords = np.clip(
        (center_y - (samples * amplitude_scale)).astype(np.int32),
        top_y,
        base_y,
    )

    pts = np.column_stack((x_coords, y_coords))
    cv2.polylines(img, [pts], False, (80, 255, 80), 1, cv2.LINE_AA)

    return img


# Background loading function
def load_yolo_and_components():
    """Load YOLO model and other components in background"""
    global yolo_model_loaded, yolo_model, yolo_loading_complete, model, device, colors, bounding_boxes

    print(f"Loading heavy imports and YOLO model in background...")

    # Import heavy libraries in background
    import numpy as np
    import math
    import torch
    import base64
    import requests
    import json
    import copy
    from sklearn.cluster import DBSCAN
    from ultralytics import YOLO

    # Make imports available globally
    globals()["np"] = np
    globals()["math"] = math
    globals()["torch"] = torch
    globals()["base64"] = base64
    globals()["requests"] = requests
    globals()["json"] = json
    globals()["copy"] = copy
    globals()["DBSCAN"] = DBSCAN
    globals()["YOLO"] = YOLO

    print(torch.cuda.is_available())
    print(torch.cuda.get_device_name(0))
    print(torch.version.cuda)

    colors = generate_color_shades(len(labels))
    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    print(f"Initializing model..")
    yolo_model = YOLO("models/" + model + ".pt")
    print(f"Loading model to {device}")
    yolo_model.to(device)

    print("loading objects")
    bounding_boxes = []

    # Update global model reference
    model = yolo_model
    yolo_model_loaded = True
    yolo_loading_complete = True

    # Check if first run and show help immediately after YOLO loads
    global is_first_run, show_info_overlay
    if is_first_run:
        show_info_overlay = True
        print("First run detected - help overlay will be shown")

    print("YOLO model loaded successfully! Switching to full processing mode.")


# Start background loading thread
yolo_thread = threading.Thread(target=load_yolo_and_components)
yolo_thread.daemon = True
yolo_thread.start()


def get_clusters(detected_points, eps=50, min_samples=2):
    if not isinstance(detected_points, np.ndarray) or detected_points.size == 0:
        return {}

    db = DBSCAN(eps=eps, min_samples=min_samples).fit(detected_points)
    labels = db.labels_

    clusters = {}
    for label in set(labels):
        if label == -1:
            continue
        cluster_points = detected_points[labels == label]

        if cluster_points.size == 0:
            continue

        min_x, min_y = np.min(cluster_points, axis=0)
        max_x, max_y = np.max(cluster_points, axis=0)
        count = len(cluster_points)
        clusters[label] = (min_x, min_y, max_x, max_y, count)

    return clusters


def is_point_in_cluster(x, y, clusters):
    """Checks if (x, y) is inside any cluster bounding box."""
    for cluster_id, (min_x, min_y, max_x, max_y, count) in clusters.items():
        if min_x <= x <= max_x and min_y <= y <= max_y:
            return True
    return False


def add_to_replay_buffer(frame):
    """Add frame to circular replay buffer, maintaining max size"""
    global replay_buffer, replay_buffer_max_size

    if len(replay_buffer) >= replay_buffer_max_size:
        replay_buffer.pop(0)  # Remove oldest frame

    replay_buffer.append(frame.copy())


def stream():
    """Continuously read frames from the capture source with graceful reconnects."""
    global cap, obj_idle, last_fskip, last_frame_time, stream_timeout_threshold, max_consecutive_failures

    consecutive_failures = 0
    waiting_for_reconnect = False

    while loop:
        current_time = millis()

        # If the capture is not open, keep trying to re-establish the connection
        if cap is None or not cap.isOpened():
            if not waiting_for_reconnect:
                print("Stream disconnected. Attempting to reconnect...")
                waiting_for_reconnect = True

            if cap is not None:
                cap.release()

            cap = cv2.VideoCapture(rtsp_stream)

            if not cap.isOpened():
                time.sleep(1.0)
                continue

            print("Stream reconnected.")
            consecutive_failures = 0
            waiting_for_reconnect = False
            last_frame_time = current_time
            q.queue.clear()

            # Give the stream a short moment to produce valid frames
            time.sleep(0.2)
            continue

        try:
            ret, frame = cap.read()
        except cv2.error as exc:
            print(f"OpenCV read error: {exc}")
            ret = False
            frame = None

        if ret and frame is not None:
            # Update last successful frame time
            last_frame_time = current_time
            consecutive_failures = 0

            # Add frame to replay buffer when not in replay mode
            if not replay_mode:
                add_to_replay_buffer(frame)

            if (
                (obj_idle > 0)
                and obj_idle >= idle_reset
                and (timestamp() - last_fskip >= 30)
            ):
                last_fskip = timestamp()
                q.queue.clear()
                obj_idle = 0
            else:
                q.put(frame)
        else:
            consecutive_failures += 1

            # Check for early timeout detection
            time_since_last_frame = current_time - last_frame_time

            if (
                consecutive_failures >= max_consecutive_failures
                or time_since_last_frame > stream_timeout_threshold
            ):
                if time_since_last_frame > stream_timeout_threshold:
                    print(
                        f"Stream timeout detected ({time_since_last_frame}ms without frames), restarting video..."
                    )
                else:
                    print("Can't receive frame, restarting video...")

                cap.release()
                consecutive_failures = 0
                waiting_for_reconnect = True

                # Avoid hammering the reconnect if the source is down
                time.sleep(0.5)
            else:
                # Short delay before retry for temporary failures
                time.sleep(0.05)


def find_closest_point(points, point):
    closest_point = None
    min_distance = float("inf")

    for x, y in points:
        distance = math.sqrt((x - point[0]) ** 2 + (y - point[1]) ** 2)
        if distance < min_distance:
            min_distance = distance
            closest_point = (x, y)

    return closest_point, min_distance


def getObject(point, cname):
    global bounding_boxes
    x, y = point
    time = millis()

    for i, box in enumerate(bounding_boxes):
        if cname != box.name:
            continue

        if box.contains(x, y, time):
            box.update(time, x, y)
            bounding_boxes[i].checkin = True
            return bounding_boxes[i]

        # Remove objects that have been missing for 3+ consecutive cycles
        if box.disappeared_cycles >= 3:
            del bounding_boxes[i]
    return False


def is_point_inside(x, y, list):
    for item in list:
        if item[0] <= x <= item[2] and item[1] <= y <= item[3]:
            return True
    return False


def process_basic_stream(photo):
    """Basic stream processing without YOLO - just display the video"""
    # Simple resize if needed, avoiding resample which might use numpy
    if hdstream == True:
        img = cv2.resize(photo, opsize, interpolation=cv2.INTER_LINEAR)
    else:
        img = photo.copy()

    # Apply military mode processing
    if military_mode:
        img = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
        img = cv2.convertScaleAbs(img, alpha=1.2, beta=10)
        img = cv2.cvtColor(img, cv2.COLOR_GRAY2BGR)

    # Add loading indicator
    if not yolo_loading_complete:
        cv2.putText(img, "Loading YOLO model...", (16, 30), _font, 0.5, (0, 0, 0), 2)
        cv2.putText(
            img, "Loading YOLO model...", (16, 30), _font, 0.5, (0, 255, 255), 1
        )
    else:
        cv2.putText(
            img,
            "YOLO model ready! Switching automatically...",
            (16, 30),
            _font,
            0.5,
            (0, 0, 0),
            2,
        )
        cv2.putText(
            img,
            "YOLO model ready! Switching automatically...",
            (16, 30),
            _font,
            0.5,
            (0, 255, 0),
            1,
        )

    return img


def process(photo):
    if hdstream == True:
        img = resample(photo)
    else:
        img = photo.copy()

    # Apply military mode processing
    if military_mode:
        img = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
        img = cv2.convertScaleAbs(img, alpha=1.2, beta=10)
        img = cv2.cvtColor(img, cv2.COLOR_GRAY2BGR)

    global obj_score, bounding_boxes, yolo_frame_count, cached_yolo_results, zoom_pan_active, zoom_pan_pause_time, dragging, zoom_mode_active, cached_person_detections, last_person_update_frame

    # Check if zoom/pan operation finished recently (wait 500ms after last activity)
    current_time = millis()
    if zoom_pan_active and not dragging and (current_time - zoom_pan_pause_time > 500):
        zoom_pan_active = False
        cached_yolo_results = None  # Clear cache when zoom/pan timeout ends
        cached_person_detections = []  # Clear person cache when zoom/pan ends

    # Skip YOLO inference during zoom/pan operations or when zoomed in
    if zoom_pan_active or dragging or zoom_factor > 1.0:
        # Use last cached results during zoom/pan or when zoomed
        results = cached_yolo_results
    elif yolo_skip_frames > 5:
        # Skip YOLO processing entirely when processing every nth frame over 6
        yolo_processing_happened = False
        results = None
        yolo_frame_count += 1
    else:
        # Run YOLO inference based on skip setting (0 = all frames, >0 = every nth frame)
        yolo_processing_happened = (
            yolo_skip_frames == 0 or yolo_frame_count % (yolo_skip_frames + 1) == 0
        )
        if yolo_processing_happened:
            # Trigger help text on first YOLO processing
            global yolo_first_processing_started, show_help_text, help_text_start_time, is_first_run, show_info_overlay
            if not yolo_first_processing_started:
                yolo_first_processing_started = True

                if is_first_run:
                    # First run - show help screen immediately
                    show_info_overlay = True
                    print("First run detected - showing help overlay")
                else:
                    # Regular run - show help text 2 seconds after first YOLO processing
                    import threading

                    def delayed_help_text():
                        import time

                        time.sleep(2)
                        global show_help_text, help_text_start_time
                        show_help_text = True
                        help_text_start_time = millis()

                    help_thread = threading.Thread(target=delayed_help_text)
                    help_thread.daemon = True
                    help_thread.start()

            # Start timing YOLO processing
            global last_yolo_processing_duration
            yolo_start = time.perf_counter_ns()

            # Create square input for YOLO (required to be divisible by 32)
            yolo_img = cv2.resize(
                img, (yolo_input_size, yolo_input_size), interpolation=cv2.INTER_LINEAR
            )

            img_tensor = (
                torch.from_numpy(cv2.cvtColor(yolo_img, cv2.COLOR_BGR2RGB))
                .to(device)
                .float()
                / 255.0
            )
            img_tensor = img_tensor.permute(2, 0, 1).unsqueeze(0)

            with torch.no_grad():
                results = model(
                    img_tensor,
                    verbose=False,
                    iou=0.45,
                    half=False,
                    max_det=32,
                    conf=min_confidence,
                    classes=classlist,
                )

            # Store YOLO processing duration
            last_yolo_processing_duration = time.perf_counter_ns() - yolo_start

            # Cache the results for next 2 frames
            cached_yolo_results = results
        else:
            # Use cached results from previous YOLO inference
            results = cached_yolo_results

        yolo_frame_count += 1

    obj_score = [0 for _ in range(len(obj_score))]
    c = 0
    points = []
    boxes = [box for r in results for box in r.boxes] if results else []
    now = millis()

    # Only reset iteration when not in zoom mode
    if not zoom_mode_active:
        resetIteration()

    rawcrowd = []

    # Cache person detections after YOLO processing for clustering continuity
    if "yolo_processing_happened" in locals() and yolo_processing_happened:
        # Will update cached_person_detections after processing boxes
        pass

    for i, box in enumerate(boxes):
        class_id = int(box.cls)
        class_name = labels[class_id]
        confidence = float(box.conf)
        c = c + 1
        if (class_name in class_confidence) and (
            confidence <= class_confidence[class_name]
        ):
            continue

        if confidence <= min_confidence:
            continue

        xmin, ymin, xmax, ymax = box.xyxy[0]
        xmin, ymin, xmax, ymax = map(int, [xmin, ymin, xmax, ymax])

        # Transform coordinates from YOLO square input back to display dimensions
        scale_x = opsize[0] / yolo_input_size
        scale_y = opsize[1] / yolo_input_size
        xmin = int(xmin * scale_x)
        ymin = int(ymin * scale_y)
        xmax = int(xmax * scale_x)
        ymax = int(ymax * scale_y)
        width = xmax - xmin
        height = ymax - ymin

        if (
            xmin == 0
            or ymin == 0
            or xmax == 0
            or ymax == 0
            or xmax == img.shape[1]
            or ymax == img.shape[0]
        ):
            continue

        if class_name == "car" and (
            (width > height and (width / height) >= 2)
            or (width < min_size or height < min_size)
        ):
            continue

        if zoom_factor > 1.0:
            color = colors[class_id].tolist()
            alpha = 0.35
            color_with_alpha = color + [alpha]

            text = f"{class_name}" + " " + str(round(confidence, 6))
            text_offset_x = xmin
            text_offset_y = ymin - 5

            overlay = img[ymin : ymax + 1, xmin : xmax + 1].copy()
            cv2.rectangle(
                overlay,
                (0, 0),
                (xmax - xmin, ymax - ymin),
                color_with_alpha,
                thickness=-1,
            )
            cv2.addWeighted(
                overlay,
                alpha,
                img[ymin : ymax + 1, xmin : xmax + 1],
                1 - alpha,
                0,
                img[ymin : ymax + 1, xmin : xmax + 1],
            )
            draw_dashed_rectangle(img, (xmin, ymin), (xmax, ymax), color, 1, 8)
            continue

        idx = class_id
        obj_score[idx] = obj_score[idx] + 1

        point = center(xmin, ymin, xmax, ymax)
        if class_name == "person":
            rawcrowd.append(point)

        size = _size(xmin, ymin, xmax, ymax)
        closest, distance = find_closest_point(points, point)
        obj = getObject(point, class_name)
        if obj != False:
            if distance < 6.0:
                continue

            points.append(point)
            obj.see()
            if obj.desc != False:
                sid = obj.desc
            else:
                sid = obj.name + "#" + str(obj.nr)

            color = colors[class_id].tolist()

            if military_mode:
                cv2.rectangle(img, (xmin, ymin), (xmax, ymax), (0, 0, 255), 2)

                # Add entity label with white text on crimson background
                label_text = class_name.upper()
                font_scale = 0.5
                font_thickness = 1

                # Get text dimensions
                text_size = cv2.getTextSize(
                    label_text, _font, font_scale, font_thickness
                )[0]
                text_width = text_size[0]
                text_height = text_size[1]

                # Position label at top-left of bounding box
                label_x = xmin
                label_y = ymin - 5

                # Ensure label doesn't go off-screen
                if label_y - text_height < 0:
                    label_y = ymin + text_height + 5

                # Draw crimson background rectangle
                cv2.rectangle(
                    img,
                    (label_x, label_y - text_height - 2),
                    (label_x + text_width + 4, label_y + 2),
                    (60, 20, 220),  # Crimson in BGR
                    -1,
                )

                # Draw white text
                cv2.putText(
                    img,
                    label_text,
                    (label_x + 2, label_y),
                    _font,
                    font_scale,
                    (255, 255, 255),  # White
                    font_thickness,
                )
            else:
                cv2.circle(img, point, 1, (0, 0, 255), 2)

            idle = format_duration(obj.idle)

            cv2.putText(img, idle, (obj.x, obj.y - 6), _font, 0.35, (0, 0, 0), 2)
            cv2.putText(img, idle, (obj.x, obj.y - 6), _font, 0.35, (200, 200, 200), 1)

        else:
            text = f"{class_name}" + " " + str(round(confidence, 6))
            text_offset_x = xmin
            text_offset_y = ymin - 5

            if military_mode:
                cv2.rectangle(img, (xmin, ymin), (xmax, ymax), (0, 0, 255), 1)

                # Add entity label with white text on crimson background
                label_text = class_name.upper()
                font_scale = 0.5
                font_thickness = 1

                # Get text dimensions
                text_size = cv2.getTextSize(
                    label_text, _font, font_scale, font_thickness
                )[0]
                text_width = text_size[0]
                text_height = text_size[1]

                # Position label at top-left of bounding box
                label_x = xmin
                label_y = ymin - 5

                # Ensure label doesn't go off-screen
                if label_y - text_height < 0:
                    label_y = ymin + text_height + 5

                # Draw crimson background rectangle
                cv2.rectangle(
                    img,
                    (label_x, label_y - text_height - 2),
                    (label_x + text_width + 4, label_y + 2),
                    (60, 20, 220),  # Crimson in BGR
                    -1,
                )

                # Draw white text
                cv2.putText(
                    img,
                    label_text,
                    (label_x + 2, label_y),
                    _font,
                    font_scale,
                    (255, 255, 255),  # White
                    font_thickness,
                )
            else:
                cv2.circle(img, point, 1, (255, 255, 0), 2)
            cv2.putText(
                img,
                text,
                (text_offset_x, text_offset_y),
                _font,
                0.35,
                (0, 0, 0),
                2,
            )
            cv2.putText(
                img,
                text,
                (text_offset_x, text_offset_y),
                _font,
                0.35,
                (255, 255, 255),
                1,
            )
            qxmin, qymin, qxmax, qymax = transform(xmin, ymin, xmax, ymax, padding)
            snap = photo[qymin:qymax, qxmin:qxmax]
            item = BoundingBox(class_name, point, size, snap)
            bounding_boxes.append(item)

    # Update cached person detections if YOLO processing happened
    if "yolo_processing_happened" in locals() and yolo_processing_happened:
        cached_person_detections = rawcrowd.copy()
        last_person_update_frame = yolo_frame_count

    if zoom_factor > 1.0:
        add(c)
        return img

    # Skip object tracking operations when in zoom mode
    if zoom_mode_active:
        add(c)
        return img

    # Use cached person detections if YOLO didn't run this frame
    if (
        "yolo_processing_happened" in locals()
        and not yolo_processing_happened
        and cached_person_detections
    ):
        rawcrowd = cached_person_detections.copy()

    cenable = False
    crowds = get_clusters(np.array(rawcrowd), 50, 2)
    rect = []
    if len(crowds) > 0:
        cenable = True
        for crowd in crowds:
            min_x, min_y, max_x, max_y, count = crowds[crowd]
            rect.append((min_x, min_y, max_x, max_y, count))
            cv2.rectangle(img, (min_x, min_y), (max_x, max_y), (0, 255, 0), 1)
            cid = str(count) + " people"
            cv2.putText(img, cid, (min_x, min_y - 18), _font, 0.35, (0, 0, 0), 2)
            cv2.putText(img, cid, (min_x, min_y - 18), _font, 0.35, (255, 255, 255), 1)

    for obj in bounding_boxes:

        if (
            obj.checkin == False
            and obj.detections >= 3
            and obj.idle > 1
            and obj.idle < 8
        ):
            _, distance = find_closest_point(points, (obj.x, obj.y))
            if distance < 6.0:
                continue

            obj.ping()
            if now - obj.seen > point_timeout:
                continue

            if cenable == True and is_point_inside(obj.x, obj.y, rect) == True:
                continue

            if obj.desc != False:
                sid = obj.desc
            else:
                sid = obj.name + "#" + str(obj.nr)

            idle = format_duration(obj.idle)

            cv2.circle(img, (obj.x, obj.y), 1, (0, 255, 255), 2)

    add(c)
    return img


sthread = threading.Thread(target=stream)
print(f"Starting..")

sthread.start()

while loop:
    img = None
    original_img = None  # Store original frame for clipboard copying

    # Handle replay mode
    if replay_mode and len(replay_buffer) > 0:
        if replay_index < len(replay_buffer):
            img = replay_buffer[replay_index]
            original_img = img.copy()  # Store original for clipboard
            replay_index += 1
        else:
            # Replay finished, exit replay mode
            replay_mode = False
            replay_index = 0
            print("Replay completed")
            continue
    elif (q.empty() != True) and (fskip != True):
        img = q.get_nowait()
        original_img = img.copy()  # Store original for clipboard

        # Update main loop frame time for additional timeout detection
        current_time = millis()
        if last_frame_time > 0 and (current_time - last_frame_time) > stream_timeout_threshold:
            print(f"Main loop timeout detected ({current_time - last_frame_time}ms since last frame)")

            # Create connection lost screen with countdown
            if img is not None:
                # Convert to grayscale
                lost_screen = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
                lost_screen = cv2.cvtColor(lost_screen, cv2.COLOR_GRAY2BGR)

                # Countdown from 3 to 1
                for countdown in range(3, 0, -1):
                    display_frame = lost_screen.copy()

                    # Draw "CONNECTION LOST" text centered
                    lost_text = "CONNECTION LOST"
                    font_scale = 2.0
                    thickness = 4
                    text_size = cv2.getTextSize(lost_text, _font, font_scale, thickness)[0]
                    text_x = (display_frame.shape[1] - text_size[0]) // 2
                    text_y = (display_frame.shape[0] - text_size[1]) // 2

                    # Black outline
                    cv2.putText(display_frame, lost_text, (text_x, text_y), _font, font_scale, (0, 0, 0), thickness + 4)
                    # Red text
                    cv2.putText(display_frame, lost_text, (text_x, text_y), _font, font_scale, (0, 0, 255), thickness)

                    # Draw countdown text below
                    countdown_text = f"retrying in {countdown}..."
                    countdown_font_scale = 1.0
                    countdown_thickness = 2
                    countdown_size = cv2.getTextSize(countdown_text, _font, countdown_font_scale, countdown_thickness)[0]
                    countdown_x = (display_frame.shape[1] - countdown_size[0]) // 2
                    countdown_y = text_y + text_size[1] + 40

                    # Black outline
                    cv2.putText(display_frame, countdown_text, (countdown_x, countdown_y), _font, countdown_font_scale, (0, 0, 0), countdown_thickness + 2)
                    # White text
                    cv2.putText(display_frame, countdown_text, (countdown_x, countdown_y), _font, countdown_font_scale, (255, 255, 255), countdown_thickness)

                    cv2.imshow(str(rtsp_stream), display_frame)
                    cv2.waitKey(1000)  # Wait 1 second

                # After countdown, reconnect
                print("Reconnecting stream...")
                cap.release()
                cap = cv2.VideoCapture(rtsp_stream)
                last_frame_time = millis()
                q.queue.clear()
                time.sleep(0.5)  # Give stream time to initialize

    if img is not None:

        key = cv2.waitKey(1) & 0xFF
        if key == 32:
            print("frame skip")
            q.queue.clear()

        if key == ord("q"):
            print("Saving configuration...")
            save_config()
            loop = False
        elif key == 8:
            if not replay_mode and len(replay_buffer) > 0:
                replay_mode = True
                replay_index = 0
                q.queue.clear()
                print(f"Starting replay of {len(replay_buffer)} frames")
            elif replay_mode:
                replay_mode = False
                q.queue.clear()
                print("Exiting replay mode")
        elif key == ord("r"):
            if not recording:
                start_recording(cap)
                q.queue.clear()
            else:
                stop_recording()
                q.queue.clear()

        elif key == ord("s"):
            take_snapshot(img)
            q.queue.clear()
        elif key == ord("f"):
            reset_window_to_stream_resolution()
            q.queue.clear()
        elif key == ord("m"):
            military_mode = not military_mode
            print(f"Military mode: {'ON' if military_mode else 'OFF'}")
            yolo_skip_frames = 1 if military_mode else 2
            q.queue.clear()

        # Handle Ctrl+number combinations for resizing
        elif key == ord("1") or key == 49:  # Ctrl+1: 640x480
            resize_stream_dimensions((640, 480))
            q.queue.clear()
        elif key == ord("2") or key == 50:  # Ctrl+2: 800x600
            resize_stream_dimensions((800, 600))
            q.queue.clear()
        elif key == ord("3") or key == 51:  # Ctrl+3: 1024x768
            resize_stream_dimensions((1024, 768))
            q.queue.clear()
        elif key == ord("4") or key == 52:  # Ctrl+4: 1280x800
            resize_stream_dimensions((1280, 800))
            q.queue.clear()
        elif key == ord("5") or key == 53:  # Ctrl+5: 1920x1080
            resize_stream_dimensions((1920, 1080))
            q.queue.clear()
        elif key == ord("6") or key == 54:  # Ctrl+6: original stream size
            if original_opsize:
                resize_stream_dimensions(original_opsize)
                q.queue.clear()
        elif key == 13:  # Enter key - toggle fullscreen
            toggle_fullscreen()
            q.queue.clear()
        elif key == 27:  # Esc key - exit fullscreen only
            if fullscreen:
                toggle_fullscreen()  # Exit fullscreen
            q.queue.clear()
        elif key == ord("+") or key == ord(
            "="
        ):  # + key (decrease frame skip - more processing)
            yolo_skip_frames = max(0, yolo_skip_frames - 1)
            show_frame_skip_display = True
            frame_skip_display_start_time = millis()
            if yolo_skip_frames == 0:
                print("Frame processing: ALL frames")
            elif yolo_skip_frames <= 5:
                print(f"Frame processing: Every {yolo_skip_frames + 1} frames")
            else:
                print(f"Frame processing: Every {yolo_skip_frames + 1} frames (YOLO disabled)")
            q.queue.clear()
        elif key == ord("-"):  # - key (increase frame skip - less processing)
            yolo_skip_frames = min(5, yolo_skip_frames + 1)  # Limit to 5 (process every 6 frames max)
            show_frame_skip_display = True
            frame_skip_display_start_time = millis()
            if yolo_skip_frames <= 5:
                print(f"Frame processing: Every {yolo_skip_frames + 1} frames")
            else:
                print(f"Frame processing: Every {yolo_skip_frames + 1} frames (YOLO disabled)")
            q.queue.clear()

        elif key == 9:  # Tab key - toggle info overlay
            # On first run, any tab press marks first run as complete
            if is_first_run:
                is_first_run = False
                print("First run completed - normal tab behavior enabled")

            show_info_overlay = not show_info_overlay
            print(f"Info overlay: {'ON' if show_info_overlay else 'OFF'}")

        if key < 255 and key > 0:
            print(f"Key pressed: {key}")

        # Skip YOLO processing during replay mode
        if replay_mode:
            # Simple display processing for replay frames
            if hdstream == True:
                img = cv2.resize(img, opsize, interpolation=cv2.INTER_LINEAR)

            # Apply military mode processing if enabled
            if military_mode:
                img = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
                img = cv2.convertScaleAbs(img, alpha=1.2, beta=10)
                img = cv2.cvtColor(img, cv2.COLOR_GRAY2BGR)
        else:
            # Use basic stream processing until YOLO is loaded, then auto-switch
            if basic_stream_mode:
                img = process_basic_stream(img)
                # Automatically switch to YOLO processing when model is loaded
                if yolo_loading_complete:
                    basic_stream_mode = False
                    print("Switching to YOLO processing mode!")
            else:
                img = process(img)  # YOLO timing is handled inside this function

        # Store clean processed image before UI overlays for selection saving
        # This ensures we capture the image with zoom/pan applied but without UI elements
        clean_processed_img = img.copy()

        # Skip object count and FPS calculations during replay mode
        if not replay_mode:
            object_count = average()
            if object_count != old_count:
                obj_break = millis()
                obj_idle = 0
            else:
                obj_idle = millis() - obj_break

            old_count = object_count
            frames += 1
            if millis() - last_frame >= 1000:
                fps = (frames - prev_frames) * 1
                prev_frames = frames
                last_frame = millis()

        waveform_offset = get_audio_waveform_offset(img.shape[0])
        if waveform_offset > 0:
            img = draw_audio_waveform(img)

        if replay_mode:
            _fps = "REPLAY MODE - LAG: 0ms"
        else:
            _fps = (
                "FPS: "
                + str(fps)
                + " - LAG: "
                + str(last_yolo_processing_duration // 1000000)
                + "ms"
            )

        text_y = max(15, img.shape[0] - 5 - waveform_offset)
        cv2.putText(img, _fps, (16, text_y), _font, 0.4, (0, 0, 0), 2)
        if replay_mode:
            cv2.putText(
                img, _fps, (16, text_y), _font, 0.4, (255, 128, 0), 1
            )  # Orange for replay
        else:
            cv2.putText(img, _fps, (16, text_y), _font, 0.4, (0, 255, 0), 1)

        if recording:
            out.write(img)
            rec_y = max(38, img.shape[0] - 38 - waveform_offset)
            cv2.putText(img, "REC", (16, rec_y), _font, 0.5, (0, 0, 0), 2)
            cv2.putText(img, "REC", (16, rec_y), _font, 0.5, (0, 0, 255), 1)

        bb = str(len(bounding_boxes))

        tracking_y = max(26, img.shape[0] - 26 - waveform_offset)
        cv2.putText(
            img, "Tracking: " + bb, (16, tracking_y), _font, 0.4, (0, 0, 0), 2
        )
        cv2.putText(
            img,
            "Tracking: " + bb,
            (16, tracking_y),
            _font,
            0.4,
            (64, 255, 255),
            1,
        )

        # Show flashing REPLAY text instead of clock during replay mode
        if replay_mode:
            # Flash every 500ms
            current_time = millis()
            if current_time - replay_last_flash_time > 500:
                replay_last_flash_time = current_time

            show_replay_text = (current_time - replay_last_flash_time) < 250

            if show_replay_text:
                replay_text = "REPLAY"
                text_size = cv2.getTextSize(replay_text, _font, 0.6, 2)[0]
                text_x = img.shape[1] - text_size[0] - 10
                text_y = max(20, img.shape[0] - 8 - waveform_offset)
                cv2.putText(
                    img, replay_text, (text_x, text_y), _font, 0.6, (0, 0, 0), 3
                )
                cv2.putText(
                    img, replay_text, (text_x, text_y), _font, 0.6, (0, 0, 255), 2
                )
        else:
            clock = datetime.now().strftime("%H:%M:%S")
            text_size = cv2.getTextSize(clock, _font, 0.5, 1)[0]
            text_x = img.shape[1] - text_size[0] - 10
            text_y = max(20, img.shape[0] - 8 - waveform_offset)
            cv2.putText(img, clock, (text_x, text_y), _font, 0.4, (0, 0, 0), 2)
            cv2.putText(img, clock, (text_x, text_y), _font, 0.4, (255, 255, 255), 1)

        # Handle selection completion for element saving
        if selection_complete and clean_processed_img is not None:
            if selection_start_x > 0 and selection_end_x > 0:
                # Use the clean processed image which has zoom/pan applied but no UI overlays
                # This ensures we save exactly what's visible under the rectangle without UI elements
                save_selection_to_elements(
                    clean_processed_img,
                    selection_start_x,
                    selection_start_y,
                    selection_end_x,
                    selection_end_y,
                )

            selection_complete = False

        if drawing and draw_start_x > 0 and draw_end_x > 0:
            # Create semi-transparent green rectangle overlay
            overlay = img.copy()
            cv2.rectangle(
                overlay,
                (draw_start_x, draw_start_y),
                (draw_end_x, draw_end_y),
                (0, 255, 0),
                thickness=-1,  # Fill the rectangle
            )
            # Apply transparency
            alpha = 0.3  # 30% transparency
            cv2.addWeighted(overlay, alpha, img, 1 - alpha, 0, img)

            # Draw thick border
            cv2.rectangle(
                img,
                (draw_start_x, draw_start_y),
                (draw_end_x, draw_end_y),
                (0, 255, 0),
                thickness=2,
            )

        # Display resolution temporarily when changed
        if resolution_display_active:
            current_time = millis()
            if (
                current_time - resolution_display_start_time
                < resolution_display_duration
            ):
                # Draw resolution text in big yellow font at top left
                font_scale = 0.85
                thickness = 2
                color = (0, 200, 200)  # Yellow in BGR
                position = (16, 32)  # Top left corner with some padding

                # Draw black outline for better visibility
                cv2.putText(
                    img,
                    resolution_display_text,
                    position,
                    _font,
                    font_scale,
                    (16, 16, 16),
                    thickness + 3,
                )
                # Draw yellow text
                cv2.putText(
                    img,
                    resolution_display_text,
                    position,
                    _font,
                    font_scale,
                    color,
                    thickness,
                )
            else:
                # Timer expired, disable display
                resolution_display_active = False

        # Show help text for 2 seconds after YOLO processing starts
        if show_help_text:
            current_time = millis()
            if current_time - help_text_start_time < 2000:  # Show for 2 seconds
                help_text = "press tab for help"
                cv2.putText(
                    img, help_text, (16, 50), _font, 0.4, (0, 255, 255), 1
                )  # Small yellow text, no outline
            else:
                show_help_text = False

        # Show frame skip display for 2 seconds when changed
        if show_frame_skip_display:
            current_time = millis()
            if (
                current_time - frame_skip_display_start_time < 2000
            ):  # Show for 2 seconds
                if yolo_skip_frames == 0:
                    skip_text = "Processing: ALL FRAMES"
                elif yolo_skip_frames <= 5:
                    skip_text = f"Processing: Every {yolo_skip_frames + 1} frames"
                else:
                    skip_text = f"Processing: Every {yolo_skip_frames + 1} frames (YOLO DISABLED)"

                # Position at top center
                text_size = cv2.getTextSize(skip_text, _font, 0.7, 2)[0]
                text_x = (img.shape[1] - text_size[0]) // 2
                text_y = 40

                cv2.putText(
                    img, skip_text, (text_x, text_y), _font, 0.7, (0, 0, 0), 4
                )  # Black outline
                cv2.putText(
                    img, skip_text, (text_x, text_y), _font, 0.7, (0, 255, 255), 2
                )  # Yellow text
            else:
                show_frame_skip_display = False

        # Show element save success message at bottom center
        if clipboard_message_active:
            current_time = millis()
            if current_time - clipboard_message_start_time < clipboard_message_duration:
                message_text = "Element saved to elements folder"
                font_scale = 0.6
                thickness = 2

                # Get text dimensions for centering
                text_size = cv2.getTextSize(message_text, _font, font_scale, thickness)[
                    0
                ]
                text_x = (img.shape[1] - text_size[0]) // 2
                text_y = max(30, img.shape[0] - 15 - waveform_offset)

                # Draw black outline for better visibility
                cv2.putText(
                    img,
                    message_text,
                    (text_x, text_y),
                    _font,
                    font_scale,
                    (0, 0, 0),
                    thickness + 2,
                )
                # Draw green text
                cv2.putText(
                    img,
                    message_text,
                    (text_x, text_y),
                    _font,
                    font_scale,
                    (0, 255, 0),
                    thickness,
                )
            else:
                # Timer expired, disable message
                clipboard_message_active = False

        # Apply info overlay if enabled
        if show_info_overlay:
            img = draw_info_overlay(img)

        cv2.imshow(str(rtsp_stream), img)
    else:
        fskip = False
        time.sleep(0.01)

print("closing cv window..")
cv2.destroyAllWindows()
print("terminating..")
loop = False
sys.exit(0)