Year: 2026
PIs: Yiqiao Li
Sponsor: TBD
Description: Surrogate safety measures (SSMs) have long served as the foundation of traffic conflict analysis, enabling proactive safety assessment without reliance on crash data. However, existing SSMs remain fundamentally kinematic: they quantify proximity between road users but fail to capture the behavioral intentions that give rise to safety-critical interactions. Identical time-to-collision (TTC) values may correspond to fundamentally different risk scenarios, for example, a pedestrian hesitating at a curb versus a vehicle aggressively forcing a gap, limiting the interpretability and reliability of current safety metrics.
This project introduces a new paradigm for infrastructure-based safety analysis by integrating World Models, generative neural networks capable of learning environment dynamics and predicting distributions over future states, into roadside LiDAR sensing systems. Unlike traditional deterministic pipelines, the proposed framework will produce probabilistic, intention-aware scene forecasts that explicitly encode uncertainty arising from sensing noise, occlusion, and tracking instability.
Leveraging longitudinal roadside LiDAR data collected, the project will investigate a calibrated, uncertainty-aware World Model tailored to static infrastructure-based sensing perspectives. This will enable a new class of safety metrics, probabilistic SSMs, that represent conflict risk as distributions rather than point estimates, incorporating both behavioral intent and epistemic uncertainty. The proposed framework will advance traffic safety analysis from reactive proximity-based metrics toward predictive, behavior-aware risk assessment, with direct implications for infrastructure decision-making and V2I-enabled cooperative driving systems.