DEPTH-AWARE VIDEO DATASET CREATION AND PREPROCESSING FOR POTHOLE DETECTION WITH SHADOW RESISTANCE

Abstract

The timely detection and restoration of potholes is important for road safety and the maintenance of transportation infrastructure. Currently, datasets for pothole detection are mainly based on images and lack important components such as depth perception and environmental adaptability to different environments, specifically under shadowed conditions. The present study proposes an innovative video dataset acquisition and preprocessing procedure that enhances robustness by integrating techniques such as monocular depth estimation and shadow resistance. In this study, video data were collected using mobile phones and drones across various real-world scenarios. The study employed a complete preprocessing pipeline, including video stabilization, frame extraction, depth estimation, shadow detection and removal, meticulous annotation, and data augmentation. The resulting dataset, including more than 450,000 annotated samples, serves as a valuable resource for depth-aware, shadow-resistant pothole detection research using deep learning models.