Trajectory planning framework for autonomous vehicles based on collision injury prediction for vulnerable road users

Abstract

Due to the escalating occurrence and high casualty rates of accidents involving Electric Two-Wheelers (E2Ws), it has become a major safety concern on the roads. Additionally, with the widespread adoption of current autonomous driving technology, a greater challenge has arisen for the safety of vulnerable road participants. Most existing trajectory planning methods primarily focus on the safety, comfort, and dynamics of autonomous vehicles themselves, often overlooking the protection of vulnerable road users (VRUs), typically E2W riders. This paper aims to investigate the kinematic response of E2Ws in vehicle collisions, including the 15 ms Head Injury Criterion (HIC15). It analyzes the impact of key collision parameters on head injuries, establishes injury prediction models for anticipated scenarios, and proposes a trajectory planning framework for autonomous vehicles based on predicting head injuries of VRUs. Firstly, a multi-rigid-body model of two-wheeler-vehicle collision was established based on a real accident database, incorporating four critical collision parameters (initial collision velocity, initial collision position, and collision angle). The accuracy of the multi-rigid-body model was validated through verifications with real fatal accidents to parameterize the collision scenario. Secondly, a large-scale effective crash dataset has been established by the multi-parameterized crash simulation automation framework combined with Monte Carlo sampling algorithm. The training and testing of the injury prediction model were implemented based on the MLP + XGBoost regression algorithm on this dataset to explore the potential relationship between the head injuries of the E2W riders and the crash variables. Finally, based on the proposed injury prediction model, this paper generated a trajectory planning framework for autonomous vehicles based on head collision injury prediction for VRUs, aiming to achieve a fair distribution of collision risks among road users. The accident reconstruction results show that the maximum error in the final relative positions of the E2W, the car, and the E2W rider compared to the real accident scene is 11 %, demonstrating the reliability of the reconstructed model. The injury prediction results indicate that the MLP + XGBoost regression prediction model used in this article achieved an R2 of 0.92 on the test set. Additionally, the effectiveness and feasibility of the proposed trajectory planning algorithm were validated in a manually designed autonomous driving traffic flow scenario.