Interactive Risk (IR): An omnidirectional safety metric of CAVs based on multimodal trajectory prediction and driving risk field

Abstract

Traffic accidents pose a significant threat to human life and property, and with the increasing presence of connected and autonomous vehicles (CAVs), effective risk assessment has become more critical. Current safety metrics, often limited to longitudinal or lateral assessments, fail to address omnidirectional risks or account for the uncertainties associated with vehicle intentions. This paper introduces a new omnidirectional safety metric, Interactive Risk (IR), which combines the concept of the driving risk field with multimodal trajectory prediction. IR captures the uncertainty of vehicle intentions, quantifies the probability and severity of potential accidents, and provides a comprehensive measure of traffic risk. Through case studies of typical collision scenarios and experiments with the simulation and real world dataset, we demonstrate that IR accurately reflects the risk levels faced by CAVs, detects collision risks earlier, and aligns more closely with human intuition compared to baseline safety metrics. Furthermore, we propose four key applications of IR, including traffic risk monitoring, ego-vehicle risk warning, driving decision-making performance evaluation, and motion and trajectory planning. The results highlight the potential of IR to enhance safety assessment in dynamic traffic environments and provide valuable insights for future research and application in autonomous vehicle systems.