This project is a real-time object detection and tracking system using the YOLOv5 model and the SORT (Simple Online and Realtime Tracking) algorithm. It captures video from a webcam, detects objects, tracks them, and outputs both a video and individual frames with bounding boxes and tracking IDs.

• Real-time object detection using a pre-trained YOLOv5 model.
• Object tracking with the SORT algorithm.
• Frame saving and video recording with bounding boxes.
• Object lifecycle tracking, including start and end times.
• Output video and individual frame images stored in a folder.

• Python 3.6 or higher
• OpenCV
• ONNX Runtime
• NumPy
• Other dependencies (listed in requirements)

Clone this repository to your local machine:

git clone https://github.com/yourusername/edge-lite-object-tracker.git
cd edge-lite-object-tracker

You can install the required dependencies using pip. It is recommended to use a virtual environment.

pip install -r requirements.txt

Here is a sample requirements.txt:

opencv-python
onnxruntime
numpy

Download the YOLOv5 ONNX model (yolov5s.onnx) from the official YOLOv5 repository or another reliable source and place it in the models/ directory.

Now you can run the program to start the object detection and tracking:

python src/main.py

This will start capturing video from your webcam, detect objects, and track them. The processed frames will be saved in the output_frames/ folder, and the video will be saved as output_video.mp4.

You can press 'q' to stop the process.

1. Detection:

   • The system uses a YOLOv5 model (converted to ONNX format) for object detection. The model detects objects and returns their bounding boxes and labels.

2. Tracking:

   • The SORT algorithm is used to track the detected objects across multiple frames. Each object is assigned a unique tracking ID.
   • The system performs tracking by comparing the detected bounding boxes using the Intersection over Union (IoU) metric to associate objects across frames.

3. Output:

   • For each detected object, bounding boxes and tracking IDs are drawn on the frames.
   • The processed frames are saved as individual images in the output_frames/ folder.
   • The processed video is saved as output_video.mp4.

4. Lifecycle Tracking:

   • Each object’s lifecycle, including the start and end times, is tracked and stored in a CSV file (tracking_data.csv).

edge-lite-object-tracker/
├── models/                  # Store your YOLOv5 model here (e.g., yolov5s.onnx)
├── output_frames/           # Folder where the processed frames will be saved
├── src/
│   ├── main.py              # Main script for running the program
│   ├── detector.py          # Object detection using YOLOv5
│   └── tracker.py           # Tracking algorithm (SORT)
└── requirements.txt         # Python dependencies

• main.py:

  • Contains the main function to run the video capture, object detection, and tracking.
  • Initializes the ObjectDetector and Sort tracker and handles video output.

• detector.py:

  • Defines the ObjectDetector class that loads and uses the YOLOv5 model to detect objects.
  • Includes helper functions like non-maximum suppression (NMS) for filtering overlapping bounding boxes.

• tracker.py:

  • Implements the SORT tracking algorithm.
  • Tracks the objects using the bounding boxes across frames and assigns unique tracking IDs.

You can adjust the following parameters to suit your needs:

• max_age: The maximum number of frames that an object can be unseen before it is removed from tracking.
• min_hits: The minimum number of frames an object must be seen to start tracking.
• iou_threshold: The Intersection over Union (IoU) threshold used for matching detected objects to tracked objects. You can adjust this threshold in the Sort class constructor if required.

1. Frames:

   • All frames with object bounding boxes will be saved as individual images in the output_frames/ folder.

2. Video:

   • A video file named output_video.mp4 will be saved with the tracked objects.

3. CSV:

   • The tracking_data.csv file contains the following columns:

     • track_id: The unique ID assigned to each tracked object.
     • label: The object label (e.g., "person", "car").
     • x1, y1, x2, y2: Coordinates of the bounding box.
     • start_time: The timestamp when the object was first detected.
     • end_time: The timestamp when the object was last seen.