A Robust Solution for Pothole Detection and Mapping in Developing Countries

This paper presents an evaluation and testing of state-of-the-art object detectors, including Fast RCNN with ResNet18, Faster RCNN with VGG16 and YOLOv4 with ResNet101. They were trained using the same hyperparameters, with slight adjustments made if needed. YOLOv4 emerged as the top performer, achieving an AP0.5 of 90% and mAP of 55%, while Faster RCNN followed with an AP0.5 of 88% and mAP of 50%. The study proposes a pothole coordinate approximation technique using image and location data. A real-time application setup for pothole detection and mapping, capturing top-down image data with minimal skewing, an improved field of view and accuracy, is introduced. The positioning system components, including angular displacement, azimuth and line extension, are discussed and applied in the pseudocode for the approximation algorithm. The simulation results showed a significant increase in the accuracy of pothole coordinates, with an average improvement of 124.42 cm from the point of detection. Despite facing some errors due to an unlevelled mount and positioning system discrepancies, this technique shows promise for enhancing road distress data acquisition and facilitating further research in road distress detection and rehabilitation.