AIC 2020 Challenge Track 1 by Orange-Control

3rd place (with the best Efficiency Score) in Challenge Track 1: Multi-Class Multi-Movement Vehicle Counting in 2020 AI City Challenge @ CVPR.

By Team Orange-Control from ¹AI Labs, Didi Chuxing (DiDi AI Labs) and ²Beijing University of Posts and Telecommunications (BUPT).

• Team Members: Wenwei Li^2,1, Haowen Wang^2,1, Yue Shi¹, Ke Dong¹, Bo Jiang¹, Zhengping Che¹ (team leader), Jian Tang¹ (advisor), and Xiuquan Qiao² (advisor)

Requirements

• python3.6

torch==1.5.0+cu101
torchvision==0.6.0+cu101
opencv-python==4.2
pandas==0.24.2
numpy==1.17.3
sklearn==0.21.2
filterpy==1.4.5
easydict==1.9
ffmpeg==2.8.15
numba==0.46.0
tensorboard>=1.14
pycocotools==2.0.0
tqdm==4.33.0
pillow==6.1.0

Inference

Here are the steps to reproduce our results:

1. Download the corresponding model file best.pt and put it in the folder weights.
2. Make sure the raw video files and required txt files are in the folder data/Dataset_A.
3. Run inference.py to get separate result files in the folder output for all 31 videos.
4. Run result.py to combine all 31 csv files and get the single submission file track1.txt.

mkdir weights
mkdir output
python3 inference.py 1 31
python3 result.py

We use YOLOv3+sort to detect and track vehicles. To count the movement, we use a detection line (detection line) for each movement by annotating the provided training videos (Data set A), as defined in get_lines.py. If a vehicle passes the detection line, the count of the corresponding movement will increase by 1 after a short pre-defined delay calculated based on the training data.

Training the detector

We use yolov3 as our detector, which is initialized by the public COCO pre-trained model and fine-tuned with some annotated frames from the training set (which will be described later). The corresponding files are in the folder yolov3_pytorch. Below are the steps to train the detector.

1. Make sure the following files are placed correctly.
   • The training images (extracted from the raw training videos) in data/images/
   • The annotation text files in yolov3_pytorch/data/labels/
   • object.data and object.names in yolov3_pytorch/data, which describe the input data and output classes
2. Downdload the official coco pretrained model yolov3.weights from https://pjreddie.com/darknet/yolo/ and put it in yolov3_pytorch/weights.
3. Use the following train command to finetine the pretrained model. The train_best.pt file is the final model.

cd yolov3_pytorch
unzip data/labels.zip
python3 train.py --data data/object.data --cfg cfg/yolov3.cfg --epochs 200

Some of the utility scripts are borrowed from https://github.com/ultralytics/yolov3.

Annotation Data

We selected 5 videos from the provided training videos (Data set A), including cam3.mp4, cam5.mp4, cam7.mp4, cam8.mp4, cam20.mp4. A subset of 3835 frames was extracted from these videos for manual annotation.

You can use the following command to extract frames directly from the videos.

ffmpeg -i cam_x.mp4 -r 1 -f image2 yolov3_pytorch/data/images/%06d.jpg

The extracted frames should be put in the folder yolov3_pytorch/data/images.