An ADAS with better driver satisfaction under rear-end near-crash scenarios: A spatio-temporal graph transformer-based prediction framework of evasive behavior and collision risk

Current advanced driver assistance systems (ADASs) do not consider drivers’ preferences of evasive behavior types and risk levels under rear-end near-crash scenarios, which undermines driver satisfaction, trust, and use of ADASs. Additionally, spatio-temporal interactions between vehicles are not fully involved in current evasive behavior prediction models, and the influence of evasive behavior is ignored while predicting collision risk. To address these issues, this study aims to propose an ADAS with better driver satisfaction under rear-end near-crash scenarios by establishing a spatio-temporal graph transformer-based prediction framework of evasive behavior and collision risk. A total of 822 evasive events are extracted from 108,000 real vehicle trajectories on highways, and variables from three sources (i.e., road environment features, evading vehicle features, and interactive behavior features) are used to construct rear-end near-crash scenario knowledge graphs (RNSKGs). By utilizing RNSKGs embedding and multi-head self-attention mechanism, spatio-temporal graph transformer networks can effectively capture the spatio-temporal interactions between vehicles. The results show that the prediction accuracy of evasive behavior (i.e., braking-only or braking and steering) and collision risk (lower, medium, or higher risk) is 96.34% and 92.12%, respectively, superior to other commonly-used methods. After including the selected evasive behavior in predicting collision risk, the overall accuracy increases by 10.91%. Then, an autonomous evasive takeover system (AET) based on the prediction framework is developed, and its effectiveness and satisfaction are verified by driving simulation experiments. According to the self-reported data of participants, the safety, comfort, usability, and acceptability of AET proposed in this study all significantly outperform existing autonomous takeover systems (i.e., autonomous emergency braking and autonomous emergency steering). The findings of this study might contribute to the optimization of ADASs, the enhancement of mutual understanding between ADASs and human drivers, and the improvement of active driving safety.