Traditional crash-based intersection monitoring fails to capture the critical near-miss events that precede collisions involving vulnerable roadway users (VRUs). To provide a proactive alternative, this report investigates how emerging roadside LiDAR and AI/ML technologies can enhance intersection safety observability through three interconnected research thrusts. First, an infrastructure-based sensing survey identifies LiDAR as a premier primary modality due to its high geometric precision and robustness under varying lighting conditions. Second, a systematic study simulating vertical beam loss across six 3D detection architectures establishes a critical 20% beam-loss maintenance threshold. Beyond this point, VRU detection deteriorates rapidly, with concentrated, contiguous beam loss from sensor occlusion proving far more detrimental than dispersed loss.

Building upon these foundational insights, the third thrust introduces a real-world application via an end-to-end, auditable safety-analysis framework deployed at a signalized intersection in New York City. Utilizing a newly established, manually annotated 8,000-frame roadside dataset, the framework integrates 3D detection, tracking, stabilization, and structured human-in-the-loop quality assurance to convert raw sensor data into defensible near-miss evidence. A heavy vehicle-bicycle interaction case study demonstrates that contrasting direction-agnostic and longitudinal time-to-collision (TTC) metrics can successfully isolate lateral-intrusion conflict mechanisms that single-metric approaches miss. Together, these efforts validate roadside LiDAR as a scalable, interpretable, and highly defensible tool for proactive traffic safety management.

Resource Types: Research Report
Capabilities: Data & Information Systems, Tools & Technology
Management Processes: Monitoring & Adjustment, Performance Based Planning & Programming, Performance Reporting & Communication