Accident-Driven Human-Like Trajectory Planning for Permissive Left Turns

Abstract

Making a left turn at a signalized intersection without a protected green arrow presents a critical challenge for autonomous vehicles, often resulting in either overly aggressive or conservative behaviors. To resolve this conflict between safety and efficiency, this paper proposes a closed-loop adaptive trajectory planning framework incorporating Risk Homeostasis Theory. Utilizing real-world collision data from the National Automobile Accident In-Depth Investigation System (NAIS) database, we analyze failure mechanisms of rigid strategies and introduce a perception model fusing a high-order Gaussian Driving Risk Field with a Random Finite Set dynamic occupancy grid. Unlike static methods, a bio-inspired mechanism enables the vehicle to dynamically adjust its speed and safety buffers within a Frenet frame using a proportional-integral-derivative (PID) controller to maintain a target risk level. Simulations based on reconstructed accidents demonstrate that the proposed method effectively balances physical safety with ride experience. The system triggers early braking in high-threat scenarios while suppressing unnecessary hesitation in low-risk situations, successfully replicating human-like driving styles.