Risk Assessment-based collision avoidance for autonomous vehicles with multi-driving-style interaction

Car-following and cut-in maneuvers are among the most frequent causes of road traffic accidents in mixed traffic environments involving human-driven and autonomous vehicles, often resulting in significant casualties. Autonomous driving technology has significant potential to reduce these traffic accidents. This paper proposes a risk assessment method that comprehensively considers autonomous vehicle (AV) driving style preference and surrounding disturbance vehicles (DVs) driving style. In addition, as a precursor to risk assessment, we have improved lane-changing intention recognition by integrating both frequency-domain and time-domain feature analysis. We design an Adaptive Threshold Wavelet Attention (ATWA) block integrated into a Transformer-based classifier. This block employs multi-level wavelet decomposition to extract frequency-domain features from vehicle trajectories, utilizing adaptive thresholds to suppress noise in low/high-frequency components while amplifying critical signals. Experimental results demonstrate that our model achieves encouraging accuracy on the NGSIM and HighD datasets, with an average inference time of 0.0043 s under 32 TOPS computational constraints, meeting real-time computing requirements. For risk assessment, we propose a driving style-aware decision-making strategy based on Conditional Random Fields (CRF). This method enhances interactive risks between the AV and the DV using a potential risk coefficient for the DV, dynamically adjusting braking responses according to the driving styles of the AV. PanoSim simulations validate that our method effectively prevents collisions in multi-scenarios while balancing safety and traffic efficiency. This work provides a comprehensive solution for autonomous systems, integrating high-precision intention recognition with personalized risk-responsive behaviors, thereby enhancing both safety and drivers’ acceptance in complex traffic environments.