MTMCT

This is an implementation of a Multi-Target Multi-Camera Tracking (MTMCT) solution. Pipeline of our solution:

Tracking performance

Results and comparisons with FairMOT and WDA Tracker trained and tested on a 6x2-minute MTA dataset

Method	Single-Camera					Multi-Camera
Method	MOTA	IDF1	IDs	MT	ML	MOTA	IDF1	IDs	MT	ML
WDA	58.2	37.3	534.2	16.8%	17.2	46.6	19.8	563.8	6.5%	7.0%
FairMOT	64.1	48.0	588.2	34.7%	7.8%	N/A	N/A	N/A	N/A	N/A
Ours	70.8	47.8	470.2	40.5%	5.6%	65.6	31.5	494.5	31.2%	1.1%

Video demos on Multi Camera Track Auto (MTA) dataset

Installation

conda create -n mtmct python=3.7.7 -y
conda activate mtmct
pip install -r requirements.txt

Install dependencies for FairMOT:

cd trackers/fair
conda install pytorch==1.7.0 torchvision==0.8.0 cudatoolkit=10.2 -c pytorch
pip install cython
pip install -r requirements.txt
cd DCNv2
./make.sh
conda install -c conda-forge ffmpeg

Download data

Go to https://github.com/schuar-iosb/mta-dataset to download the MTA data. Or use other datasets that match the same format.

Configurations

Modify config files under tracker_configs and clustering_configs for customization. Create a work_dirs and see more instructions at ifzhang/FairMOT and koehlp/wda_tracker.

E.g. in configs/tracker_configs/fair_high_30e set the data -> source -> base_folder to your dataset location.

Tracking

Run single and the multi-camera tracking with one script:

sh start.sh fair_high_30e

Modify config files under tracker_configs and clustering_configs for customization. More instructions can be found at ifzhang/FairMOT and koehlp/wda_tracker.

Acknowledgement

A large part of the code is borrowed from ifzhang/FairMOT and koehlp/wda_tracker. The dataset used is schuar-iosb/mta-dataset

Copyright

Ruizhe Zhang is the author of this repository and the corresponding report, the copyright belongs to Wireless System Research Group (WiSeR), McMaster University.

GJTNB/mtmct