City-Level Pavement Distress Inspection Using Crowdsourced Data of Logistics Vehicles

Chenglong Liu, Hanlin Yang, Difei Wu, Yishun Li, Yuchuan Du

Abstract: Large-scale pavement distress inspection has gained attention recently. Traditional methods, involving dedicated vehicles and professionals, are effective for small-scale evaluations but are labor-intensive for extensive areas. Logistics vehicles, equipped with driving recorders, provide a vast resource of street view images covering urban roads, capturing pavement distress data. This paper proposes a novel framework leveraging logistics vehicle data for rapid distress detection and tracking. A deep convolutional neural network, enhanced with a dual-layer routing attention mechanism, improves detection precision for minor distresses. The optimized Boundary Box Regression (BBR) loss function increases accuracy for common distresses like cracks. A three-stage distress matching algorithm, based on an attentional graph neural network and adjacent-local-area matching, removes duplications and tracks distress deterioration. The Bernoulli function assesses minimal sampling frequency for road segments. Validated in Shanghai, this method achieves 79.6% mean Average Precision (mAP) and a 75.66% matching rate, enabling daily updates for timely maintenance decisions.

External IDs:doi:10.1109/tits.2025.3629112