Object Detection with Detectron2 for Pothole Detection

Abstract

Damaged road surfaces can cause traffic accidents and negatively affect the tires and suspension of vehicles, potentially contributing to accidents. Potholes are one of the main causes of road damage, and rapid detection is needed to prevent it. This paper proposes a deep learning model based on Detectron2. Detectron is a training inference platform for Pytorch-based object detection and semantic segmentation. In the experiment, a total of 1334 images were trained, resulting in a class accuracy of 0.94 and a mask accuracy of 0.89, indicating good overall performance. The processing time for an image is considered sufficiently applicable in real life, with an average of 0.11 seconds per image. In the experiment, 199 test images were processed in 21.42 seconds, but the processing time varies depending on the number of instances. The proposed model can overcome the limitations of human inspection of thousands of kilometers of roads.