TRACE-MPC: Triggered Risk Abduction and Compliance-Coupled MPC for Latent-Hazard Anticipation on Highways

In high-speed and complex highway scenarios, occlusions and rapid vehicle speeds significantly reduce available reaction time. Current autonomous driving systems primarily address visible risks, making it difficult to detect and respond to latent risks in occluded situations or scenarios that require reasoning. This paper presents TRACE-MPC (Triggered risk Abduction and Compliance-coupled Model Predictive Control), an ontology-driven risk reasoning and planning framework based on unexpected-behavior analysis that identifies unsafe, non-compliant, or unreasonable behaviors and analyzes the hidden risks they imply. We construct a seven-layer scene ontology as the foundation for reasoning. The ontology is supported by prior knowledge and traffic regulations, forming a rule base that defines and judges unexpected behaviors through interaction logic among scene elements. Potential-risk analysis uses harmful-interaction relations between traffic participants as mediating links for risk propagation. The reasoning outputs feed an MPC module to enable safe and smooth trajectory planning. Experiments in representative complex scenarios show that TRACE-MPC improves minimum time-to-collision by 172% under occlusion and avoids collision in a pedestrian-occlusion case where baselines impact at 32-35 km/h, while also improving post-merge spacing and reducing braking intensity during ramp merges, demonstrating reliable latent-risk identification and decision support.