A Comparative Evaluation of the Deep Learning Algorithms for Pothole Detection

Potholes are a menace on roads and their presence compromises the safety of both drivers and pedestrians. In most developing countries, it is one of the major reasons for road accidents and loss of human life and property. Therefore, there is a need to consistently collect and update the data on the latest road conditions, so that the drivers can be advised for alternate routes and the concerned Government department can take immediate measures to fill up the potholes for the benefit of the commuters. A simple and efficient way to detect potholes on roads is through the application of object detection algorithms on images acquired from a smartphone camera. Therefore in this paper, we focus on evaluating the performance of state-of-the-art neural network algorithms such as YOLO and Faster R-CNN with VGG16 and ResNet-18 architectures for pothole detection that is both fast and accurate. Further, an improved YOLOv2 architecture is proposed to solve the class imbalance problem of "pothole" and "normal road" classes, and its performance is compared with other object detection techniques using precision, recall, intersection over union, and the number of frames processed per second (FPS). The results showed that the modified YOLOv2 architecture outperformed all the considered models with the lowest number of parameters (35 million) and the highest FPS (28), precision (0.87), and recall (0.89). This model can be deployed in autonomous vehicles for real-time geotagged pothole detection from photographs or video streams. The pothole detection application can also suggest potential alternate eco-friendly routes and guide the commuters in low light navigation.