Real-time pothole detection system on vehicle using improved YOLOv5 in Malaysia

Commonly in Malaysia, the road surfaces deteriorate over time which results to potholes. With the increasing number of potholes on the road, it has become a road hazard, potentially harming the safety of the Malaysian citizens. Potholes generally will cause drivers to skew away from their original direction which may result in vehicle damages and/or road accidents. Computer vision has improved significantly in the past decade to improve the technologies that require image processing, such as pothole detection. However, it has several limitations in pothole detection as potholes have inconsistent shapes, making it difficult to obtain an accurate prediction. Another limitation is the speed of the detection algorithm is not able to predict the pothole in real time. In this paper, the use of computer vision technology on vehicles is presented as a form of solution to detect the potholes in real time. A deep learning model based on Convolutional Neural Networks, YOLOv5 is found to improve the accuracy of the prediction as compared to past results. The findings on the trained YOLOv5 model have a mAP@0.5 of 80.8 %, 82.2 % and 82.5 % on the YOLOv5m6, YOLOv5s6 and YOLOv5n6 respectively. The trained YOLOv5n6 model was chosen to be the main image processing model for its robust performance to size ratio. The results were able to demonstrate the actual deployment of the trained YOLOv5n6 model on the road in real time with distance estimation.