Deep line segment detection for concrete pavement distress assessment

Abstract This study proposes a d eep l ine s egment d etection model named DLSD, for identifying four ubiquitous line segments on concrete pavements: joint, sealed joint, bridge expansion joint, and roadway boundary. DLSD associates a category with the triple‐point representation to encode a line segment. Its network employs a localization head and a classification head, attaching several auxiliary branches to integrate the line segment shape context. A novel dual‐attention mechanism further improves the line segment classification. From experiments, the structural average precision (sAP) and mean sAP of the DLSD model on class‐agnostic and class‐aware line segment detection achieve 85.0% and 73.4%, respectively. The former outperforms the existing best‐performed method by 2.7%, and the latter sets a state‐of‐the‐art performance. An automated pipeline combines the line segments with cracks to detect corner break and shattered slab on concrete pavements for an accurate distress assessment, reducing the error rate of distress ratio value from 38.7% to 11.5%.