Real-Time Wireframe Pothole Detection And Avoidance System

Engineering notes

Experimental evaluation demonstrates that the proposed system achieves 92.3% mAP@0.5, 89.7% precision, 87.5% recall, and real-time performance of 42–48 FPS on an RTX 3060-class GPU.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为端到端自动驾驶、BEV感知、3D目标检测、轨迹预测、路径规划、LiDAR感知等高价值论文补充中文说明。

Original abstract

Reliable pothole detection remains a significant challenge in autonomous driving and intelligent transportation systems because conventional two-dimensional detections provide limited information about the geometric severity of road-surface defects, while sensor-intensive solutions increase hardware cost and deployment complexity. This paper presents a cameraonly framework for real-time pothole detection and severity assessment using three coupled stages: pothole localization, monocular depth estimation, and depth-guided wireframe construction. The primary contribution of the proposed framework is the introduction of a lightweight structural modeling layer between conventional two-dimensional detection and computationally expensive dense reconstruction methods. Instead of performing full metric 3D reconstruction, the system utilizes relative depth discontinuities within detected pothole regions to generate a sparse wireframe representation that preserves boundary shape, local surface variation, and depression structure. Based on these geometric attributes, the framework computes a severity score using relative depth deviation, surface area, depth variance, and contour irregularity, enabling more informative perception than conventional boundingbox-based approaches. The framework is implemented using a lightweight YOLOv8 detector, a monocular depth estimation module, and a graphbased wireframe reconstruction pipeline operating on pothole regions of interest. Experimental evaluation demonstrates that the proposed system achieves 92.3% mAP@0.5, 89.7% precision, 87.5% recall, and real-time performance of 42–48 FPS on an RTX 3060-class GPU. In addition, the proposed depth-guided wireframe representation improves boundary approximation accuracy from 62% in the detector-only baseline to 84% in the complete framework. The results indicate that the proposed approach provides a scalable and cost-effective compromise between conventional vision-based pothole detection and expensive depth-sensor-based perception systems for real-time road-surface analysis and autonomous driving applications

Links and sources

• PDF from original source

Comments