Pedestrian Tracking

Pedestrian Tracking by DeepSORT and Hybrid Task Cascade with PyTorch.

Introduction

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.

Installation

1.Download the project

git clone https://github.com/FinalFlowers/pedestrian_tracking.git

2.Install the required libraries

cd pedestrian_tracking

pip install -r requirements.txt -i https://pypi.tuna.tsinghua.edu.cn/simple

3.Compile MMDetection

pip install -v -e .

Note: there is a point at the end of the command.

4.Download the weight files

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/

Testing

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

Further information

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.