BoT-SORT: Robust Associations Multi-Pedestrian Tracking

Nir Aharon, Roy Orfaig, Ben-Zion Bobrovsky

This repository contains an unofficial C++ implementation of BoT-SORT multi-pedestrian tracking algorithm.

This implementation has been tested on NVIDIA Jetson NX and it achieves real-time performance on 1080p videos.

YOLOv8-tracking, Official BoT-SORT and Official ByteTrack GitHub repositories were used as references.

• Implement BoT-SORT tracker
• Load parameters from config file
• Implement Re-ID model for BoT-SORT tracker using TensorRT
• Dockerize the project
• Re-ID performance analysis
• Re-ID model quantization

These results demonstrate the BoT-SORT tracker, implemented in this repository, utilizing MOT provided detections.

More Result Videos

1. Install Docker and NVIDIA Container Toolkit

2. Clone the repository

   git clone [email protected]:viplix3/BoTSORT-cpp.git
   cd BoTSORT-cpp

3. Build Docker image

   ./docker_build.sh

4. Run Docker container

   ./docker_run.sh

5. Build BoT-SORT tracker

   cd /home/user/work/
   mkdir build && cd build
   cmake .. -DCMAKE_BUILD_TYPE=Release
   make -j$(nproc)

Example code for using the BoT-SORT tracker is provided in botsort_tracking_example.cpp The project by default builds the example code.

To test the example code, run the following command:

cd <project-root-dir>/build
./bin/botsort_tracking_example ../config/tracker.ini ../config/gmc.ini ../config/reid.ini ../assets/osnet_x0_25_market1501.onnx ../examples/data/MOT20-01.mp4 ../examples/data/det/det.txt ../output/

The performance of the BoT-SORT tracker, implemented in this repository, was evaluated on the MOT20 dataset. Details are provided in this document.

The accuracy of the BoT-SORT tracker, implemented in this repository, was evaluated on the MOT20 train set. Details are provided in this document.