講演情報

[Data_Collection-16]Data-Centric Pavement Crack Segmentation with In-Vehicle Smartphones and Human-in-the-Loop Annotation

*CHUNLIANG XUE1, KAI XUE2, Naoto Yokozawa3, Boyu Zhao2, Tomonori Nagayama1 (1. Department of Civil Engineering, The University of Tokyo, Bunkyo-ku, Japan (Japan), 2. SmartCity Research Institute Co.Ltd, Tokyo, Japan. (Japan), 3. Public Works Research Institute (PWRI), Tsukuba, Japan (Japan))

キーワード:

pavement crack segmentation、data-centric、computer vision

Computer vision combined with in-vehicle smartphones is being explored as a low-cost, high-frequency solution for pavement inspection. Most existing approaches rely on object detection, which only provides bounding boxes around crack regions and cannot accurately capture geometric properties such as crack intersections, lengths and areas. Pixel-level crack masks can be produced by semantic segmentation, but conventional pipelines depend on large-scale, per-pixel manual annotation and are therefore difficult to deploy at engineering scale. To address this limitation, a human-in-the-loop annotation framework is proposed. Initial crack masks are generated by a segmentation model and then rapidly corrected by human annotators, so that both model performance and label quality are improved iteratively. This framework allows accurate pixel-level labels to be produced far more efficiently than by manual annotation alone, while keeping humans in control of the final annotation quality. Experiments further show that, once label quantity and quality are ensured by this process, performance differences among mainstream semantic segmentation architectures become marginal: even a basic U-Net is able to fit cracks on diverse Japanese pavement materials with high accuracy and substantially faster inference. In addition, the model exhibit high robustness in real road environments; on certain road segments in Tsukuba City, temporal increases in crack rate can be detected by our model, whereas existing open-source crack segmentation models fail to provide stable result. This finding suggests that, for smartphone-based pavement monitoring, progress is now constrained more by data than by network design. Consequently, more attention should be directed towards data-centric issues, including scalable data acquisition, annotation strategies and label quality control.