Pedestrian Tracking by DeepSORT and Hybrid Task Cascade with PyTorch
Pedestrian Tracking by DeepSORT and Hybrid Task Cascade with PyTorch.
This project is used to participate in zte algorithm contest（中兴捧月算法大赛阿尔法·勒克斯特派）, which get 77.838 on the A board.
Pedestrian detection is obtained by Hybrid Task Cascade, which implemented by MMDetection.
I choose to use DeepSORT to achieve the data association. This section is modified by other authors’ implementation.
Several other detection algorithms, such as Cascade R-CNN and EfficientDet, were also tested, but with poor results.
git clone https://github.com/FinalFlowers/pedestrian_tracking.git
cd pedestrian_tracking
pip install -r requirements.txt -i https://pypi.tuna.tsinghua.edu.cn/simple
pip install -v -e .
Note: there is a point at the end of the command.
Download detection and ReID feature extraction model parameters from Baidu Netdisk with code: bboh.
Put htc_dconv_c3-c5_mstrain_400_1400_x101_64x4d_fpn_20e.pth under pedestrian_tracking/models/
Put ckpt.t7 under pedestrian_tracking/deep_sort/deep/checkpoint/
Run the following code for pedestrian tracking:
python htc_deepsort.py /your/trackdata/
The output format is:
<frame>，<id>，<bb_left>，<bb_top>，<bb_width>，<bb_height>，<conf>，<type>
Note:
Conf and type are fixed as 0.9 and 0 respectively.
The input should be a path to images ending in /
The results will be saved under pedestrian_tracking/results/ in .txt format
Run the following code will visualize the tracking results while testing:
python htc_deepsort.py /your/trackdata/ --display
You can adjust the tracking configuration in person_tracking/configs/deep_sort.yaml and detection configuration in person_tracking/models/htc_dconv_c3-c5_mstrain_400_1400_x101_64x4d_fpn_20e.py.