Safety-critical trajectory segment generation for autonomous driving testing in vehicle-to-two-wheeler blind-spot...

Safety validation of autonomous driving systems (ADS) is constrained by the rarity of safety-critical interactions in naturalistic driving and the difficulty of generating realistic multi-agent trajectories from reconstructed crash data. This study focuses on a common urban safety-critical scenario in which two wheelers suddenly emerge from a driver's blind spot into the vehicle's path, putting riders at high risk of severe or fatal injuries. The objective is to generate safety-critical pre-crash trajectory segments for ADS testing in such scenarios.

A segment-level generation framework using a GRU-ODE-based generative adversarial network (GRU-ODE-GAN) is proposed, which integrates continuous-time dynamics with recurrent updates to improve temporal consistency. Using a sliding window, fixed-length 3.0 s pre-crash segments are extracted from reconstructed crash trajectories for model training. The generated segments are then embedded into counterfactual closed-loop simulations on the Baidu Apollo platform for ADS evaluation.

The generated segments closely match real segments in low-dimensional feature-space visualizations and in kinematic distributions of speed, acceleration, and jerk. In Baidu Apollo simulations, baseline replays yield a 70% crash rate, whereas generated variants reduce the crash rate to 38% while concentrating tests near the safety boundary. Despite fewer crashes, risk intensity remains high: the proportions of potential conflicts increase from 26.78% to 74.60% for $\text{PET}<3.0$ s and from 8.88% to 58.73% for $\text{TTC}<3.0$ s.

Segment-level trajectory generation can improve the efficiency of ADS safety evaluation by increasing the density of informative, high-risk interactions without sacrificing kinematic realism.