AI-Powered Search and Rescue Drone System

Prototype computer vision system for detecting humans in two disaster rescue scenarios:

• Underwater mode: low visibility, blue-green color cast, blur, poor lighting.
• Rubble mode: partial occlusion, debris, simulated thermal/infrared view.

The demo uses Python, OpenCV, YOLOv8, PyTorch, and Streamlit. It is designed for a final year project presentation where a clean live dashboard matters as much as model accuracy.

1. Folder Structure

ai-search-rescue-drone/
  app.py
  requirements.txt
  README.md
  config/
    app_config.yaml
    dataset_underwater.yaml
    dataset_rubble.yaml
  src/
    __init__.py
    alerts.py
    detector.py
    gps.py
    heatmap.py
    preprocessing.py
    utils.py
  scripts/
    train_yolo.py
    run_inference.py
  data/
    raw/
    processed/
  models/
    README.md
  outputs/

2. Installation

python -m venv .venv
.\.venv\Scripts\activate
python -m pip install --upgrade pip
pip install -r requirements.txt

YOLOv8 will automatically download yolov8n.pt the first time you run the app. For a better demo, place your trained model in:

models/best.pt

3. Run the Dashboard

streamlit run app.py

Then open the local Streamlit URL shown in the terminal.

4. Dataset Suggestions

Underwater human detection:

• Use underwater diver/swimmer videos from public sources and label humans manually with Roboflow, CVAT, or LabelImg.
• Search terms: "underwater diver detection dataset", "underwater swimmer dataset", "UFO-120 underwater object dataset", "SUIM underwater image dataset".
• If a dataset does not contain a person class, create a small custom dataset from underwater rescue/diver footage.

Rubble/disaster detection:

• Use COCO person images and augment them with occlusion patches, dust, blur, and debris overlays.
• Use disaster/search-rescue imagery from open emergency response datasets when licensing allows.
• For thermal support, use FLIR ADAS or other thermal human datasets, then map the class to person.

Recommended annotation format: YOLO format.

dataset/
  images/train
  images/val
  labels/train
  labels/val

Each label file:

class_id x_center y_center width height

Coordinates are normalized from 0 to 1.

5. Training

Edit config/dataset_underwater.yaml or config/dataset_rubble.yaml so the paths point to your dataset.

Train underwater model:

python scripts/train_yolo.py --data config/dataset_underwater.yaml --weights yolov8n.pt --epochs 50 --imgsz 640 --name underwater_human_detector

Train rubble model:

python scripts/train_yolo.py --data config/dataset_rubble.yaml --weights yolov8n.pt --epochs 50 --imgsz 640 --name rubble_human_detector

After training, copy the best checkpoint:

copy runs\detect\underwater_human_detector\weights\best.pt models\underwater_best.pt
copy runs\detect\rubble_human_detector\weights\best.pt models\rubble_best.pt

6. Inference Without UI

python scripts/run_inference.py --source 0 --mode underwater
python scripts/run_inference.py --source path\to\video.mp4 --mode rubble

7. Why YOLOv8?

YOLOv8 is a strong fit for this prototype because it is fast enough for webcam/video inference, has a simple training API, supports transfer learning, and produces bounding boxes with confidence scores. For a drone-style rescue demo, low latency matters: the system should show detections live rather than only after offline processing.

8. Main Challenges

Underwater detection is difficult because water absorbs red light, creates blue-green color distortion, reduces contrast, adds haze, and introduces blur or floating particles. The prototype handles this with white balance, CLAHE, contrast enhancement, sharpening, and optional dehazing.

Rubble detection is difficult because victims may be partially visible, covered by debris, badly lit, or viewed from unusual angles. The prototype helps by using confidence thresholding, simulated thermal visualization, heatmap accumulation, and a clean alert workflow.

9. Future Improvements

• Train separate underwater and rubble models on real rescue imagery.
• Add segmentation with YOLOv8-seg or SAM for victim outlines.
• Add pose estimation for partially visible limbs.
• Use real thermal/infrared camera input.
• Fuse RGB, thermal, GPS, IMU, and altitude data.
• Deploy on an edge device such as NVIDIA Jetson.
• Send SOS alerts to a control room through SMS, MQTT, or a cloud dashboard.

10. Professor Demo Flow

1. Start Streamlit dashboard.
2. Select Underwater Mode and show preprocessing toggle.
3. Play underwater/diver video or webcam sample.
4. Show bounding boxes, confidence, victim count, alert banner, and heatmap.
5. Switch to Rubble Mode.
6. Enable thermal simulation and show partially occluded person examples.
7. Explain how the trained model can be deployed on a drone camera feed.