Leveraging Perspective Transformation for Enhanced Pothole Detection in Autonomous Vehicles

Abstract

Poor road conditions, often resulting from inadequate maintenance or adverse weather, are a significant factor in road accidents. Additionally, human reaction time poses limitations in responding to unexpected hazards like potholes, which pose a considerable safety risk. Early detection of distant potholes is crucial to allow drivers, both human and autonomous, to react appropriately by takingavoiding actions or reducing speed, thereby minimizing vehicle damage and potential accidents. This thesis proposes a unique approach for improved pothole detection, particularly for distant ones, by leveraging the well-established You Only Look Once version 5 (YOLOv5) object detection model in conjunction with perspective transformation. This technique effectively enhancesthe visual prominence of distant potholes by virtually bringing them closer, facilitating their detection, and improving feature extraction, even under varying illumination conditions. Our approach achieves significant improvements in several metrics, exceeding 5% in terms of mean Average Precision (mAP) at various Intersection over Union (IoU) thresholds (0.5, 0.75, and 0.5–0.95) andrecall. To the best of our knowledge, this is the first work to specifically address the challenge of distant pothole detection and utilize perspective transformation as a targeted approach to address this issue.