Bridge--track relative position detection and coarse-grained attribution

Abstract

The relative geometry between railway bridges and tracks is a useful indicator for infrastructure condition screening. In field point clouds, however, recovering true structural eccentricity is often infeasible because design alignments, lower-structure references, and complete structural observations are unavailable. This paper therefore formulates the task as bridge–track relative anomaly detection and conservative coarse-grained attribution. The proposed framework constructs a moving local reference frame by locally linearizing the rail centerline in adaptive windows, estimates an apparent bridge reference from observable upper-structure boundaries, and derives continuous lateral and vertical bridge–track relative error fields. Anomaly segments are detected from robust baseline residuals, spatial continuity, confidence diagnostics, and geometric consistency constraints. Full point-cloud context is then used to assign each detected segment to bridge-related anomaly, track-related anomaly, or uncertain. Experiments on four railway bridges show that the adaptive density– curvature window provides the most balanced local-reference behavior among the tested strategies. In the main bridge2 case, nine anomaly segments were localized, including three bridge-related, one track-related, and five uncertain segments; in lower-context bridge3 and bridge4 cases, uncertain outputs dominated, reflecting the intended conservative behavior. The method produces screening-level anomaly fields without requiring explicit design models and is therefore positioned as a geometry-driven prioritization tool rather than a design-verification method.