Dynamic Traffic Hazard Perception and Safety Impact Assessment Based on UAV LiDAR and Deep Learning

Abstract

Manual road inspections lack sufficient 3D geometric hazard data because of limited coverage and sampling frequency. This paper develops a hazard-oriented safety assessment framework integrating UAV LiDAR reconstruction, PointNet++ point-cloud segmentation, geometric hazard indicators, and traffic-mechanism analysis. Instead of merely extracting road objects, the framework converts segmented scenes into measurable metrics (lane width compression ratio, encroachment ratio, and marking degradation) and couples them with artificial-potential-field trajectory analysis, Lighthill–Whitham–Richards (LWR) traffic-wave theory, and lateral trajectory entropy. Validation with ground control points yielded a planar root mean square error (RMSE) of 0.055 m and an elevation RMSE of 0.025 m for the processed survey scenes. The five displayed class-level intersection over union (IoU) values yield an arithmetic mean of 87.88%, reported as 87.9% after rounding. The geometric hazard cases and traffic outputs are presented as analytical scenarios and sensitivity results; they are not claims of newly observed vehicle trajectories or lane-level field counts. The framework provides an interpretable basis for road maintenance, construction-zone management, and hazard early warning, while future deployment requires independent multi-site and trajectory-based validation.