Skip to content

🔥 Heatmap & Multi-Object Tracking using YOLO (ONNX) + ByteTrack

This project demonstrates crowd movement analysis using a YOLO ONNX detection model and ByteTrack multi-object tracking.

It processes a video to:

  • Detect people
  • Track each person with a unique ID
  • Generate a dynamic heatmap of movement
  • Save an annotated output video

🚀 Features

  • YOLO ONNX inference (GPU supported)
  • Real-time multi-object tracking (ByteTrack)
  • Person-only filtering (class_id = 0)
  • Accumulated movement heatmap
  • Configurable detection & tracking thresholds
  • Annotated video export

📥 Model Download

Pretrained models for swatahVision are available in the Model Zoo.

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

📁 Project Structure

heatmap_and_track/
│
├── yolov11x-1280_onnx.py      # Main script
├── requirements.txt           # Dependencies
├── people_walking.mp4         # Sample input video
├── weight.onnx                # YOLO ONNX model (user provided)
└── README.md

📦 Requirements

Install dependencies:

pip install -r requirements.txt

Or manually:

pip install opencv-python numpy onnxruntime swatahVision

⬇️ Model Setup

1. Download YOLO ONNX Model

Download a YOLO ONNX model (example: yolov11x-1280.onnx).

Rename it to:

weight.onnx

Place it inside:

heatmap_and_track/

▶️ Run the Script

Basic run:

python yolov11x-1280_onnx.py \
    --source_weights_path weight.onnx

Run with custom video:

python yolov11x-1280_onnx.py \
    --source_weights_path weight.onnx \
    --source_video_path input_video.mp4 \
    --target_video_path output_video.mp4

⚙️ Command Line Arguments

Required

--source_weights_path
Path to YOLO ONNX weights file.

Optional

--source_video_path
Input video file.
Default: people_walking.mp4

--target_video_path
Output annotated video.
Default: output.mp4

--confidence_threshold
Detection confidence threshold.
Default: 0.35

--iou_threshold
IoU threshold for NMS.
Default: 0.5

--heatmap_alpha
Heatmap opacity (0–1).
Default: 0.5

--radius
Radius of heat circle.
Default: 25

--track_threshold
Confidence required to activate tracking.
Default: 0.35

--track_seconds
Seconds to buffer lost tracks.
Default: 5

--match_threshold
Threshold for matching tracks with detections.
Default: 0.99

🧠 How It Works

Pipeline:

  1. Load YOLO ONNX model
  2. Read video frames
  3. Run object detection
  4. Filter person class only
  5. Apply ByteTrack multi-object tracking
  6. Generate movement heatmap
  7. Annotate frame with tracking IDs
  8. Save output video

📊 Output

The output video contains:

  • Accumulated crowd heatmap
  • Persistent tracking IDs
  • Person-only tracking
  • Real-time visual annotations

🎯 Use Cases

  • Crowd density analysis
  • Retail footfall analytics
  • Smart city monitoring
  • Event crowd tracking
  • Security surveillance

📌 Notes

  • Currently tracks person class only
  • GPU recommended for real-time performance
  • Works best with stable camera footage
  • Heatmap accumulates over time