Pothole Detection for Autonomous Vehicles in Indian Scenarios using Deep Learning

The Ministry of Road Transport and Highways of India reported that 4,775 and 3,564 road crashes in 2019 and 2020 were due to potholes. Autonomous vehicles are expected to revolutionize transportation in India, but their safe operation depends on effectively detecting potholes. Potholes are typically treated as static objects from the perspective of an autonomous vehicle, and they pose a danger to road commuters, particularly at high speeds. They usually develop during rainy seasons or continuous usage of roads by heavy vehicles like trucks. Over the years, with the advancement of image processing and deep learning, it has become feasible to detect potholes. Plenty of research was done, and several methods were proposed for pothole detection. The work proposed in this paper is improvement along those lines. Besides improving the detection accuracies, we implemented our models on an autonomous testing vehicle. Testing models in real-time made us encounter several bottlenecks in developing an end-to-end solution for pothole detection. Our approach uses Faster Region-based Convolutional Neural Network (FRCNN), and You Only Look Once (YOLOv5) object detection algorithms, which yielded noticeable results after thorough experimentation.