Automatic Detection of Potholes Uses Deep Learning to Reduce Traffic Accidents

Traffic accidents are events that occur on the road that are unexpected and unplanned, involving vehicles with other road users or without other road users, resulting in human casualties or other property losses. In Indonesia in 2019, there were 116,411 traffic accidents. This number has increased by 0.06% compared to the number of cases of traffic accidents in the previous year. There are several factors that cause traffic accidents. Among them are the factors of supporting facilities and infrastructure for traffic. Infrastructure and environmental factors cause 30% of traffic accidents. The number of accidents caused by environmental factors and infrastructure requires great concern from the parties concerned. Potholes are one of the causes of accidents that are included in the category of infrastructure and environmental factors. Potholes cause 15% of traffic accidents. The development of Artificial Intelligence (AI) technology, especially Deep Learning (DL), could be one solution to help identify potholes automatically. In this research, the DL model was developed using Faster R-CNN. The proposed model is to replace the backbone convolutional neural network (CNN) from the previous model using Inception-V2 with ResNet-101. The test results with the Potholes dataset from the Electrical and Electronic Department at Stellenbosch University show that the proposed method produces better accuracy than the baseline model. The proposed method produces 73.77%, and the baseline model produces an accuracy of 67.83%.