The main objective is to build an end-to-end pipeline for jointly detecting cars and lane divisions. The approaches exploit pre-trained-deep-learning-SotA models, concretely, YOLOv3 for car detection, and LaneATT for lane divisions. This repository additionally performs statistics on the number of cars on separate lanes.
- Python >= 3.5
- Pytorch == 1.6, torchvision == 0.7, cudatoolkit == 9.2
- CUDA to compile NMS code in lane detectors
- Other dependencies found in environment.yml
Conda is necessary for the installation, which might take several minutes.
conda env create -f environment.yml
conda activate carlane
cd lane_detector/lib/nms; python setup.py install; cd -Download weights of pre-trained models.
batch download_weights.shThe codes will automatically create the output directory. The detection results include lane divisions and its borders, car detection and its accuracies, a lane number to which a car belongs (X is annotated if cars are on parking or can not detect lanes for cars), number of cars per lane.
python main.py -m image -dp data/imagespython main.py -m video -dp data/video.mov --fps 20Distributed under the MIT License. See LICENSE for more information.
Khoa NGUYEN - @v18nguyen - [email protected]