Skip to content

Object Tracking using YOLO (ONNX) + ByteTrack

This example demonstrates how to perform multi-object tracking in a video using computer vision.

The system detects objects in each frame and assigns a consistent tracking ID to each object as it moves across frames.

Object tracking allows systems to understand movement patterns and trajectories of objects in video streams.

This is ideal for:

  • 🚗 Traffic monitoring systems
  • 🏬 Retail customer movement analysis
  • 🏭 Industrial object tracking
  • 🎥 Smart surveillance systems
  • 📊 Motion analytics in video streams

🚀 What This Project Does

The system performs the following steps:

  1. Detects objects using a YOLO ONNX model
  2. Tracks objects across frames using ByteTrack
  3. Assigns a unique ID to each detected object
  4. Maintains object identity across frames
  5. Draws bounding boxes and tracking IDs on the video
  6. Saves the annotated output video

Each detected object is labeled as:

#ID
Example:

#3

This means the detected object has tracking ID 3 and will keep the same ID as long as it remains visible.

📁 Project Structure

tracking/
├── yolov11x-1280_onnx.py        # Main tracking script
├── models/
│   └── yolov11x-1280.onnx       # YOLO ONNX detection model
├── data/
│   └── input.mp4                # Example input video
└── README.md                    # Example documentation

📥 Model Download

Pretrained models for swatahVision are available in the Model Zoo.

🔗 https://visionai4bharat.github.io/swatahVision/model_zoo/

📁 Project Setup

1️⃣ Clone the Repository

git clone https://github.com/VisionAI4Bharat/swatahVision.git
cd swatahVision/examples/tracking

2️⃣ (Optional) Create Virtual Environment

python3 -m venv venv
source venv/bin/activate

3️⃣ Install Dependencies

pip install -r requirements.txt

🚗 Run Object Tracking

Run the tracking script:

python yolov11x-1280_onnx.py \
    --source_video_path "data/input.mp4" \
    --source_weights_path "models/yolov11x-1280.onnx" \
    --classes 0 \
    --confidence_threshold 0.3 \
    --iou_threshold 0.7

⚙️ Command Line Arguments

Required

--source_video_path
Path to input video file.

Optional

--source_weights_path
Path to YOLO ONNX detection model .

--classes
List of class IDs to track.
Example:

--classes 0
Common class IDs: - 0 → Person - 2 → Car - 5 → Bus

If this argument is empty, all detected classes will be tracked.

--confidence_threshold Detection confidence threshold.
Default: 0.3

--iou_threshold
IoU threshold for Non-Max Suppression.
Default: 0.7

🧠 How It Works (Technical Overview)

Processing Pipeline:

  1. Load YOLO ONNX object detection model
  2. Read frames from the input video
  3. Run object detection on each frame
  4. Filter detections by selected classes
  5. Pass detections to the ByteTrack tracker
  6. Assign unique tracking IDs to objects
  7. Maintain identity across frames
  8. Draw bounding boxes and IDs on the video
  9. Save annotated output video

Multi-object tracking systems aim to estimate object locations and maintain consistent identities across frames in a video.

🧠 ByteTrack Tracking Algorithm

This example uses ByteTrack, a high-performance multi-object tracking algorithm.

Key features:

  • Tracks multiple objects simultaneously
  • Maintains consistent IDs across frames
  • Handles occlusions and missed detections
  • Works in real-time video analytics systems ByteTrack associates detection boxes across frames to maintain object identities.

📊 Output

The generated output video contains: - Bounding boxes around detected objects - Unique tracking IDs for each object - Continuous tracking across frames - Annotated output video saved to disk

Example overlay: 

#1
#4
#7
Each number represents a tracked object ID.

🖥 Model Requirements

This example requires a YOLO ONNX object detection model.

You can:

  • Export a model from Ultralytics YOLO
  • Use a pre-trained ONNX model
  • Use your own trained detection model

Example export:

pip install ultralytics
yolo export model=yolov8x.pt format=onnx imgsz=1280

🎯 Practical Applications

  • Smart surveillance systems
  • Crowd movement analysis
  • Traffic monitoring and vehicle tracking
  • Industrial object monitoring
  • Retail customer behavior analysis

📌 Notes

  • GPU recommended for real-time performance
  • Works best with stable camera footage
  • Supports tracking multiple objects simultaneously
  • Tracking accuracy depends on detection quality
  • Suitable for both indoor and outdoor video analytics