Pothole-related Traffic Safety Detection based on Deep Learning

Road traffic safety has always been a topic of close attention. Many factors, such as potholes, seriously affects the road safety and bring inconvenience to people’s lives. In recent years, with the continuous development of computer vision and machine learning, the method of image processing using deep learning has surpassed the previous methods. The driving detection technology for traffic safety provides efficient and practical methods to assist driver. This research targets to develop a traffic safety system based on pothole detection with the detection for automobile, traffic lanes and traffic sign, which are the objects usually appeared when driving, to intelligent driving system. For implementing detection, the authors use images and videos that captured from Internet and token by authors as input sources for machine learning. The labeling of detect targets is applied trans-polygon method by manual way, which adapt an image labeling software names VGG Image Annotator, to increase the accuracy. The machine learning method, which is named Mask R-CNN, is an image instance segmentation method start process by these labeled training images. Finally, the experiments and the recognition rate will be shown.