linkinpark213/online-mot-by-detection

A simple online multi-object tracking toolbox.

Online MOT-by-Detection

A simple online multi-object tracking toolbox following the tracking-by-detection paradigm.

This repository is still under construction.

Requirements

• Python 3
• numpy
• scipy
• argparse
• opencv-python
• PyTorch 1.0+ (1.3+ if Detectron2 is used)
• Detectron2 (optional)
• mmdetection (optional)
• CenterNet (optional)
• PySOT (optional)

Note: This repository doesn't include any object detector implementations but provides interfaces for Detectron2, MMdetection, darknet and CenterNet. You can install any of them in your environment.

Note: This project is developed on Ubuntu 18.04 with Python 3.7 and NVIDIA GTX 1080Ti.

Get Started

Clone the repo

git clone git@github.com:linkinpark213/online-mot-by-detection.git --recurse-submodules

Installation

With all requirements installed, you only need to add the code directory to your PYTHONPATH. Edit the line below and run in your terminal:

export PYTHONPATH=/path/to/online-mot-by-detection:$PYTHONPATH

(Optional) If you use YOLO detector

You'll need to compile the darknet project in third_party. Please refer to darknet readme.

(Optional) If you use CenterNet detector

You'll need to replace the DCNv2 in third_party/CenterNet/src/lib/models/networks with the new version with support for PyTorch 1.0+:

rm -rf third_party/CenterNet/src/lib/models/networks/DCNv2
mv third_party/DCNv2 third_party/CenterNet/src/lib/models/networks/DCNv2
cd third_party/CenterNet/src/lib/models/networks/DCNv2
./make.sh

Download weights for components

The weights of CenterNet/DGNet/OpenReID that we used are uploaded here. For pre-trained weights of MMDetection or Detectron2, please refer to their model zoos.

Run the demo

A few example configs for building up online multi-object trackers are in the configs directory.

Here is a demo that runs a customized DeepSORT tracker with a MMDetection Faster R-CNN detector, a DG-Net re-ID encoder, a Kalman filter for target motion prediction and a cascaded bipartite matcher based on Hungarian algorithm.

MMDetection will automatically download the weight files needed but you'll have to download DG-Net weight files manually here and move the wetght file id_00100000.pt to mot/encode/DGNet.

python tools/demo.py configs/deepsort.py --demo_path /path/to/any/video