Real-Time Pothole Detection Using Deep Learning

Abstract

**— Roads are connecting lines between different places and are used in our daily life. Roads' periodic maintenance keeps them safe and functional. However, asphalt pavement distresses cause potholes which may increase the number of accidents. Detecting and reporting the existence of potholes to responsible departments can save the roads from getting worse. This study deployed and tested different deep learning architectures to detect the presence of potholes. First, several images of potholes are captured by a cellphone mounted on the car windshield. Then pothole images downloaded from the internet increased the size and the variability of our database (1087 images with more than 2000 potholes). Second, various object detection algorithms are employed to detect potholes in the road images. Real-time Deep Learning algorithms with several configurations like SSD-TensorFlow, YOLOv3-Darknet53, and YOLOv4-CSPDarknet53 are used to compare their performances on pothole detection. YOLOv4 achieved a high recall of 81%, high precision of 85% and 85.39% mean Average Precision (mAP). The speed of YOLOv4 processing is recorded around 20 frames per second (FPS) at an image resolution of 832832. Furthermore, the proposed system detected potholes at distances reaching a hundred meters. Compared with other state-of-the-art methods, our result demonstrated superior performance in real-time. The proposed method can help in reporting road potholes to government agencies, increasing the safety of drivers by detecting potholes time ahead, and improving the performance of self-driving cars to ensure safe trips for passengers in the future.*