Skip to content

README.en.md

Latest commit

 

History

History
148 lines (96 loc) · 4.6 KB

README.en.md

File metadata and controls

148 lines (96 loc) · 4.6 KB
Raw

yoloTracker

🌐 한국어 버전

yoloTracker extends Darknet (YOLO) with object tracking, multi-camera streaming, and OpenMP-based parallel reception. While the original YOLO only produces per-frame detection results, this repository assigns a persistent ID to each detected object and tracks it across frames.

Changes from upstream Darknet/YOLO

1. Object ID tracking

Detected bounding boxes are matched to the same object in subsequent frames. opencv_tracker/ keeps IDs stable by combining:

  • MOSSE tracker
  • ORB feature matching
  • Velocity-based position prediction
  • History-based correction

Related files: opencv_tracker/tracker_main.cpp, tracking_dot.cpp, orb.cpp, tracker.h

2. Lightweight MOSSE-only tracker

opencv_tracker_mosse/ is a simplified variant that uses only MOSSE. It shares the same out.txt / out.jpg interface but has a simpler tracking strategy.

Related files: opencv_tracker_mosse/tracker_main.cpp, tracking_dot.cpp

3. Multi-camera input

1–4 network streams are merged into a single frame with hconcat / vconcat before being passed to YOLO. The number of streams is controlled by the STREAM define.

Related files: src/image_opencv.cpp, src/demo.c, include/darknet.h

4. OpenMP parallel reception

open_video_stream_cus() receives and decodes JPEG buffers from multiple sockets in parallel. Enable with OPENMP=1 in the root Makefile.

Related files: Makefile, src/image_opencv.cpp

5. Video streaming tools

TCP-based camera send/receive utilities:

  • videoserver: encodes camera frames as JPEG and streams them
  • videoclient: receives a single stream
  • videoclient_4cam: receives four streams

Related files: video_streaming/videoserver.cpp, videoclient.cpp, videoclient_4cam.cpp

6. Miscellaneous tools

  • ORB/: ORB feature-matching experiments
  • video_maker/: tool for generating result videos
  • nano_cam_on.cpp, nano_cam_on.sh: scripts for launching the streaming server on a remote device (environment-specific)

Directory layout

Path Role
src/ Modified Darknet core — multi-stream reception and frame merging
opencv_tracker/ Main tracker combining ORB, MOSSE, and prediction
opencv_tracker_mosse/ Lightweight MOSSE-only tracker
video_streaming/ Camera send/receive utilities
ORB/ ORB matching experiments
video_maker/ Result video generation tool

Pipeline

  1. videoserver streams camera frames over the network.
  2. The modified Darknet receives multiple streams and merges them into one frame.
  3. YOLO detection results are written to out.jpg and out.txt.
  4. opencv_tracker or opencv_tracker_mosse reads those files and assigns persistent IDs to each object.
  5. Tracking results are sent to an external server or displayed on screen.

Detection results are handed off to the tracker via files (out.txt / out.jpg), not through a direct API call.

Building

Darknet core

Default Makefile flags:

GPU=1
CUDNN=1
OPENCV=1
OPENMP=1
DEBUG=1
make

Outputs: darknet, libdarknet.a, libdarknet.so

Trackers and tools

cd opencv_tracker       && make tracker
cd opencv_tracker_mosse && make tracker
cd video_streaming      && make
cd ORB                  && make tracker

Configuration checklist

CAM_NUM, STREAM — include/darknet.h

#define CAM_NUM 0
#define STREAM 4

Set CAM_NUM for local cameras or STREAM for network streams, depending on your setup.

Streaming IP / port — src/demo.c, src/image_opencv.cpp

Camera server addresses and ports are hardcoded. Update them to match your environment.

Tracking result server — opencv_tracker/tracker_main.cpp

The destination server address for tracking output is also hardcoded.

Key customization points

src/image_opencv.cpp

OpenCV I/O, multi-stream reception and merging, OpenMP parallel receive logic.

src/demo.c

Connects the custom input source to the YOLO demo loop, initializes sockets, and writes out.jpg / out.txt.

opencv_tracker/

Converts raw YOLO detections into tracked results: manages per-object tags and miss counts, handles motion prediction and screen-exit logic.

Caveats

  • IP addresses, ports, and server addresses are hardcoded throughout the source.
  • The remote camera scripts are tied to a specific development environment.
  • YOLO and the tracker communicate through files (out.txt, out.jpg), not a direct API.
  • DEBUG=1 is the default, so the build prioritizes debug convenience over performance.
  • OpenMP optimization covers only the multi-stream reception path.