SEED2020 - Vehicle Cross Lane Event Detection

This repo is developed for the competition of SEED2020 Vehicle Cross Lane Event Detection.

Install

$ pip install requirements.txt

Usage

$ cd project/code  
$ python main.py

Note : the default data path is /data/test/, the result is saved in /data/result as a result.zip file.

Code Intro

This repo is mainly composed of three parts as following:

Part1 Vehicle detection

Related code could be found in project/train/efficientdet.

Given a video file, EfficientDet is taken to output the size and position of vehicles via frame-by-frame.

Considering the trade-off between inference acurracy and speed, efficientdet-d5 model is used,, see more in weights/.

Thanks a lot to the efforts from zylo117 Yet-Another-EfficientDet-Pytorch.

In temp_data/, the results are saved as json files.

Part2 Lane detection

Related code could be found in project/train/PINet.

PINet is taken to detect the lane points and output the results in temp_data.

Thanks a lot to the efforts from koyeongmin PINet_new.

Part3 Cross detection

This is the most important part that is contributed to the final result. Related code could be found in project/code/crossDet.

Specifically, vehicle tracking, lane tracking ,lane type classfication, ROI mask, occlussion object are included in this part. See more in utils.py

Docker

1. build image

docker build -t {image_name} .

2. load image

docker load my_image.tar
nvidia-docker run --name {container_name} {my_image_name}