Forward-Upper Structure Detection for Large-Component Transport Vehicles

Abstract

Large-component road transport is vulnerable to bridges, gantries, cross-road signs, and other forward-upper rigid structures, while roadside foliage can produce line-like clutter in three-dimensional light detection and ranging point clouds. This paper presents a lightweight embedded pipeline combining a fixed forward-upper region of interest, voxel and bird’s-eye-view reduction, plane/line candidate extraction, full-candidate three-dimensional validation, gated height recovery, finite line-tube context filtering, and distance-tiered warning. In a complete 434.3-s replay, the context filter reduced cumulative recovery-branch acceptances from 653 to 8 (98.8%). Single-annotator manual labeling identified 16 forward-upper rigid-structure approach events, and a corresponding accepted structure was observed in all 16. Engineering optimization reduced mean/P95 processing times from 224.63/638.67 ms to 46.06/124.10 ms. Evaluation is limited to one mainly straight route, one annotator, speeds below 30 km/h, and one vehicle-component configuration.