DTTM-Vehicle-Counting

Introduction

In this repo, we include the submission to AICity Challenge 2020 Vehicle Counts by Class at Multiple Intersections (Didi Chuxingsubmission).

We propose a robust and fast vehicle turn-counts at intersections via an integrated solution from detection, tracking and trajectory modeling. Our team ranks 6th in Public leaderboard and models of our algorithms are not trained with any extra datasets.

Installation

Our code is tested on Tesla P40, 24G with following setting:

1. Linux
2. Python 3.6 (only test on python 3.6)
3. PyTorch 1.1 or higher
4. CUDA 10
5. NCCL 2
6. GCC 4.9 or higher

The fast way to install our code is running commond as follows:

pip3 install -r requirements.txt

Attention: If there are any errors about mmcv or mmdetection, please refer Mmdetection to install mmdetection first.

Test

Get videoes directory of Track1:

After downloading packages of AICity Challenge 2020 Track1, please unzip and $DirPath_to_Track1_AIC20_track1 is the final directory after unzip.

Download detection model:

1. Download our detection model on RetinaNetNas-FPN
2. Put the model in directory $root/detection/NAS_FPN/checkpoints
3. Then our code can be run as follows

Run test code

python  multi_process.py --video_dir=$DirPath_to_Track1_AIC20_track1

The final counting results will be stored in $root/count_nums/

Reference

Mmdetection