Car-Following Safety Control Framework Based on Short-Term Safety Evolution Pattern Prediction

Abstract

Predicting and controlling car-following risks is a crucial component of advanced driver-assistance systems (ADAS). The traditional control framework activates only on reaching a specific risk threshold, which may not be timely enough to prevent collisions. This research proposes an active safety control framework based on prediction of safety evolution patterns. First, car-following samples are extracted from natural driving data, and safety and risk boundaries determined. The car-following patterns are conceptually divided into Safety, Disturbance, and Resilience based on the safety evolution trend. Subsequently, a proposed convolutional neural network (CNN) with gated recurrent unit (GRU) and with feature and temporal attention mechanisms is employed to predict the safety evolution pattern. Finally, a weight-optimized model predictive control (MPC) is applied for the safety control of samples predicted as Resilience. The results indicate that: (1) the proposed risk pattern prediction model achieves an average F1 score of 0.931, outperforming other baseline models; (2) the optimal weight combination for MPC is determined through grid search, resulting in an overall driving safety improvement of 29.31%, with the safety of all controlled samples remaining above the risk boundary and the proposed framework achieving significant improvements over traditional threshold-based solutions; and (3) there are distribution differences in the variables across safety evolution patterns, and the falsely predicted samples also exhibit feature distribution similarities to other patterns. The trade-off exists between safety and efficiency in the control algorithm. This study validates the effectiveness of the risk pattern prediction and safety control framework, contributing to the advancement of ADAS technology.