Uncertainty-aware risk assessment and GRU-based risk level map generation for RSU-assisted vehicles

Detecting the surrounding environment and accurately evaluating risks to avoid collisions remain significant challenges in autonomous driving. Recent advances in artificial intelligence (AI) have enhanced trajectory prediction capabilities. However, existing risk assessment methods have not fully integrated the uncertainty inherent in multimodal prediction outputs. Furthermore, relying solely on onboard vehicle devices is challenging, whereas Roadside Units (RSUs), with their prior information, extensive sensor networks, and computing power, provide a more comprehensive risk evaluation. Thus, this paper proposes a cooperative vehicle-infrastructure risk assessment system (CVIRAS). First, an uncertainty-aware risk assessment method based on multimodal prediction results is introduced. By assuming a bivariate Gaussian uncertainty distribution, an approximate collision probability estimation approach is developed, generating risk indicators (RIs) and corresponding risk maps (RMs). Each risk map is then classified to generate a corresponding risk level map (RLM). Utilizing a Gated Recurrent Unit (GRU), an RLM recognition network model (RLMNet) is proposed. This model incorporates a pattern-matching attention mechanism, where predefined patterns with trainable parameters are established. An attention mechanism enables self-learning of the match between each preset pattern and the ego vehicle's specific state. Furthermore, we propose an end-to-end extension of RLMNet (E2E-RLMNet), which tightly couples trajectory prediction and risk assessment within a unified framework for joint training. The proposed framework is validated on real-world NGSIM and CitySim datasets. Experimental results demonstrate that RLMNet and E2E-RLMNet outperform all compared models in both three-class and five-class risk classification, while effectively mitigating the computational burden induced by multimodal and environmental complexity.