Paper CDNet: link

This repository represents Crosswalk Detection Network (CDNet), which is a specific implementation of crosswalk (zebra crossing) detection and vehicle crossing behavior analysis under the vision of vehicle-mounted camera.

• A crosswalk detection and vehicle crossing behavior detection network is realized.
• The accuracy and speed exceed YOLOv5 in the specific task.
• High robustness in real complex scenarios such as in cloudy, sunny, rainy and at night is achieved.
• Real-time detection (33.1 FPS) is implemented on Jetson nano edge-computing device. +The datasets, tutorials and source codes are available on GitHub.

On the basis of YOLOv5, the following technologies are proposed to improve the speed and accuracy :

• SENet (Squeeze-and-Excitation Network)
• NST (Negative Samples Training)
• ROI (Region Of Interest)
• SSVM (Slide receptive field Short-term Vectors Memory)

• Jan 2021: 1.0.1 version of CDNet is released! With a detection speed of 33.1 FPS on Jetson nano, it obtained an average F1 score of 94.72% in the complex scenarios.

Get CDNet code and configure the environment, please check out docs/INSTALL.md

Please check out docs/MODEL_ZOO.md

Download trainsets and testsets, please check out docs/DATASETS.md

Once you get the CDNet code, configure the environment and download the dataset, juse type:

python train.py --trainset_path /path/to/trainset/folder
(such as: /home/xxx/datasets/train_data_yolov5_format) 

The training results and weights will be saved in runs/expxx/ directory.

The main optional arguments:

--device "0, 1"  # cpu or gpu id, "0, 1" means use two gpu to train.
--img-size 640 
--batch-size 32 
--epochs 100 
--not-use-SE  # use original YOLOv5 which not SE-module embedded if there is the flag

Detect the crosswalk image by image and analyze the vehicle crossing behavior.

python detect.py

The main optional arguments:

--source example/images  # images dir
--output example/output  # output dir
--img-size 640  # inference model size
--device "0"   # use cpu or gpu(gpu id)
--plot-classes ["crosswalk"]  # plot classes
--field-size 5  # the Slide receptive field size of SSVM 
--not-use-ROI  # not use roi for accelerate inference speed if there is the flag
--not-use-SSVM  # not use ssvm method for analyse vehicle crossing behavior if there is the flag

For more details, please refer to docs/INSTALL.md and docs/DATASETS.md.

The square dots denote detection size of 640, while triangle dots denote 288. The detection speed measured on the NVIDIA edge computing device Jetson nano.

CDNet is authored by Zhengde Zhang, Menglu Tan, Zhicai Lan, Haichun Liu, Ling Pei and Wenxian Yu.

Currently, it is maintained by Zheng-De Zhang ([email protected]).

xxx

CDNet is freely available for free non-commercial use, and may be redistributed under these conditions. For commercial queries, please drop an e-mail at [email protected]. We will send the detail agreement to you.