Deep Learning based Detection of potholes in Indian roads using YOLO

In countries like India road maintenance is a challenging task. Year after year, the accident rates are increasing due to the up-surging potholes count. As the road maintenance process is done manually in most places, it consumes enormous time, requires human labor and subjected to human errors. Thus, there is a growing need for a cost-effective automated identification of potholes. In recent trends, many approaches proved good results in applying deep learning [1] for different object detection. Convolutional Neural Networks (CNNs) have the ability to learn the art of extracting relevant features from an Image. But in countries like India, there is no potholes dataset available to train the CNN. In this paper, a new 1500 image dataset has been created on Indian roads. The dataset is annotated and trained using YOLO (You Only Look Once). The new dataset is trained on YOLOv3, YOLOv2, YOLOv3-tiny, and the results are compared. The results are evaluated based on the mAP, precision and recall. The model is tested on different pothole images and it detects with a reasonable accuracy.