任务(项目管理)
碰撞
毒物控制
计算机科学
伤害预防
运输工程
模拟
工程类
航空学
计算机安全
物理医学与康复
医疗急救
医学
系统工程
作者
Liu Yang,Jike Zhang,Nengchao Lyu,Qianxi Zhao
标识
DOI:10.1016/j.aap.2024.107830
摘要
The lane-changing (LC) maneuver of vehicles significantly impacts highway traffic safety. Therefore, proactively predicting LC maneuver and associated collision risk is of paramount importance. However, most of the previous LC risk prediction research overlooks the prediction of LC maneuver, limiting its practical utility. Furthermore, the effectiveness of LC maneuver recognition tends to be moderate as the prediction horizon extends. To fill the gaps, this paper proposes a multi-task learning model that simultaneously predicts the probability of LC maneuver, LC risk level, and time-to-lane-change (TTLC), while further analyzing the intrinsic correlation between LC maneuver and LC risk. The model consists of a Convolutional Neural Network (CNN) and two Long Short-Term Memory networks (LSTM). The CNN is employed to extract and fuse shared features from the dynamic driving environment, while one LSTM is dedicated to estimating the probability of LC maneuver and TTLC, and the other LSTM focuses on estimating the LC risk level. Evaluation of the proposed method on the HighD dataset demonstrates its excellent performance. It can almost predict all LC maneuvers within 2 s before the vehicle crosses lane boundaries, with an 80% recall rate for high-risk LC levels. Even 3.6 s before crossing lane boundaries, the model can still predict approximately 95% of LC maneuvers. The use of the multi-task learning strategy enhances the model's understanding of traffic scenarios and its prediction robustness. LC risk analysis based on the HighD dataset shows that the risk distribution and influencing factors for left and right lane changes differ. In right lane changes, collision risks primarily arise from the leading and following vehicles in the current lane, while in left lane changes, collision risks mainly stem from the leading vehicle in the current lane and the following vehicle in the target lane. The proposed approach can be applied to advanced driver assistance systems (ADAS) to reliably and early identify LC during highway driving, while correcting potentially dangerous LC maneuvers, ensuring driving safety.
科研通智能强力驱动
Strongly Powered by AbleSci AI