Autonomous vehicles (AVs) are expected to significantly improve road safety and traffic efficiency. However, ensuring safe decision-making in complex urban traffic scenarios remains a critical challenge due to complex interactions between AVs and surrounding vehicles. In this context, the behaviors of nearby vehicles influence the AV's actions, and vice versa, forming highly coupled bidirectional interactions. Improper handling of such interactions may lead to unsafe decisions or even traffic accidents. To this end, a game theoretic decision-making framework for interaction-aware autonomous driving is proposed in this paper, where the inter-vehicle interactions are transformed to pairwise leader-follower Stackelberg games. Specifically, a driving context-aware strategy proposal network is first developed to generate adaptive strategy spaces for surrounding vehicles by incorporating historical trajectories, road structures, and interaction features. Subsequently, a convex hull-based scanning technique is introduced to identify the most influential vehicles based on future spatiotemporal occupancy. Furthermore, a learning enhanced game-theoretic decision-making framework is formulated that incorporates predicted trajectories of vehicles into the payoff function with safety-related metrics. Finally, the effectiveness of the proposed method is validated through both qualitative and quantitative experiments in multiple highly interactive urban traffic scenarios using the CARLA simulator. Experimental results demonstrate that the proposed method could improve interaction safety while maintaining travel efficiency compared with baselines.
In autonomous driving, video-language models (VLMs) increasingly promise natural-language interpretation of driving scenes, yet whether they can reliably support safety-critical traffic risk understanding remains unclear. To address this, we benchmark seven VLMs on real-road dashcam videos in CoVLA. We propose a three-stage evaluation benchmark that decomposes driving-scene risk assessment into spatiotemporal environment perception (Q1), interaction understanding between agents (Q2), and critical risk identification (Q3), and operationalizes each stage as a scoreable multiple-choice probe derived from caption-based video labels. Using 500 dashcam clips sampled from the CoVLA dataset (1500 question items in total), we compare Qwen3VL (8B), InternVL3.5 (8B), VideoLLaMA3 (7B), BLIP3 (4B), SALMONN2+ (7B), Videochat-R1.5 (7B), and GLM 4.6v-flash (9B) under two dialogue modes: single-turn, in which each question is processed independently with a fully reset context, and multi-turn, in which questions are presented sequentially so that the model retains its own prior responses as context history. The quantitative evaluation employs overall accuracy, question-wise accuracy, mode shift (ΔOverall), and Q3 macro-F1 and balanced accuracy, while the qualitative evaluation examines context usage, visual grounding, hazard prioritization, and failure patterns across selected illustrative scenarios. In the single-turn setting, Qwen3VL achieved the highest overall accuracy at 55.6%; in the multi-turn setting it maintained the lead at 53.5%. Across both modes, Q3 was consistently the most difficult stage, and differences between dialogue modes varied sharply across models: Videochat-R1.5 improved by 3.9 percentage points, whereas SALMONN2+ collapsed to 13.5% overall with 970 unparseable outputs. These results indicate that current VLMs are not yet reliable enough for direct safety-critical decision support, but they may become useful for post-hoc risk analysis, failure triage, and human-supervised scenario review as model tuning, multi-turn design, output-format control, and broader expert adjudication improve. The practical deployment of traffic risk understanding systems will require not only high accuracy but also strong visual grounding, stable context retention, accurate hazard prioritization, and robustness against error propagation.
