This study proposes a computer vision-based damage inspection system for road markings. In order to evaluate the degree of damage of a marking objectively, the proposed system estimates its damage ratio according to the marking's damaged part and the marking's region. A hierarchical semantic segmentation strategy is proposed which employs a series of convolutional neural networks to recognize the 2D bounding box, damaged part and region of a marking. Specifically, this strategy can effectively identify the original region of a marking through an improved U-Net even if the marking is significantly damaged. The damage ratio estimation is enhanced by integrating information from multiple images based on object tracking and dynamic homography estimation. The experimental results confirm that the proposed system is effective in automating the inspection of road markings and producing objective damage assessments that should significantly assist road managers in prioritizing maintenance operations.

本研究提出了一种基于计算机视觉的道路标记损坏检测系统。为了客观评价标记的损坏程度，所提出的系统根据标记的损坏部分和标记的区域来估计其损坏率。提出了一种分层语义分割策略，该策略采用一系列卷积神经网络来识别标记的二维边界框、损坏部分和区域。具体来说，即使标记受到严重损坏，该策略也可以通过改进的 U-Net 有效识别标记的原始区域。基于对象跟踪和动态单应性估计，通过整合来自多个图像的信息来增强损伤率估计。