Autonomous Driving Risk Perception and Decision-Making Models: Integration and Challenges of Deep Reinforcement Learning, Imitationlearning, and Uncertainty Modeling

Despite the rapid advancement of autonomous driving technology, minimizing driving risks remains the most critical issue. This paper systematically reviews the key technologies and research progress in risk perception and decision-making models for connected and autonomous vehicles (CAVs). Focusing on the core elements of risk perception technology—environmental perception, behavior prediction, decision-making, and human-machine collaboration—it highlights four mainstream decision-making models: 1) Risk-aware reward function design based on deep reinforcement learning (DRL), which balances safety and efficiency by dynamically adjusting collision risk weights; 2) Human risk cognition transfer methods driven by imitation learning, incorporating hierarchical interpretable frameworks and personalized risk threshold modeling; 3) Multimodal perception fusion and Bayesian neural network-based uncertainty quantification techniques to improve risk assessment accuracy in complex scenarios; and 4) Hierarchical hybrid planning methods that integrate rule-driven and data-driven models to coordinate global path planning and local obstacle avoidance risk management. The study reveals three major challenges in current technologies: the fusion bottleneck of multimodal perception in extreme weather, the difficulty of adapting physiological parameters for personalized driver risk modeling, and the generalization gap between simulation environments and real-world vehicle validation. It aims to provide a reference for future research in this field.