重庆理工大学学报2025，Vol.39Issue(23)：245-254,10.DOI:10.3969/j.issn.1674-8425(z).2025.12.030
多源数据融合的城市道路换道风险识别研究
Multi-source data fusion for lane change risk identification on urban roads
摘要
Abstract
To further explore the relationships among driver behavior,physiological changes,and the evolving risk states during lane-changing maneuvers on urban roads,driver behavior data and physiological indicators are collected through real-vehicle experiments.A multi-source fused driving dataset is then built accordingly.A change-point detection algorithm based on steering-wheel angle data and lateral gaze position data is employed to extract lane-change events.It is futher employed to identify lane-change characteristics and determine the corresponding time window. Traffic risk indicators,including TTC,DRAC,and MTC,are employed to evaluate the safety level of each lane-change event.K-means clustering is then applied to classify the lane-change events into different risk levels.A CNN-BiLSTM-Transformer model is built to identify lane-change risks.The model has a Convolutional Neural Network(CNN)to extract local spatial features,a Bidirectional Long Short-Term Memory(BiLSTM)network to learn temporal driving patterns,and a Transformer module to capture global dependencies in the lane-change process. The SHAP method is employed to analyze the influence of each feature variable on the model's prediction outcomes,allowing for an interpretable assessment of how different features contribute to the model's decisions.The differences in driver behavior and physiological responses are compared for lane-change maneuvers with different risk levels.These differences are examined separately as in on-ramp merging scenarios and in off-ramp diverging scenarios to better understand how drivers behave in the two types of ramp environments. Results show a lane-change maneuver is defined as highly risky when the Time to Collision(TTC)is less than 3.90s,the Margin to Collision(MTC)is less than 0.55,and the Deceleration Rate to Avoid Crash(DRAC)is greater than 1.10 m/s2.Under these conditions,the driving scenario is considered to involve a significantly higher level of potential conflict.The model effectively identifies lane-change risks by integrating multi-source data features.It achieves an accuracy of 98.0％.Among all the input variables,the steering-wheel angle and the lateral acceleration are the key features that exert the biggest influence on the model's recognition results. The values of the relevant feature indicators increase markedly as the lane-change risk level becomes higher.This trend shows these indicators respond more readily when the driving involves greater risks.In ramp areas,the β-wave amplitude of the driver's EEG increases during high-risk lane-change maneuvers.On average,the β-wave amplitude is about 13％higher in high-risk conditions compared with low-risk lane-change situations.During high-risk lane-change maneuvers,drivers show wider variations in lateral gaze position.In these situations,the driver's gaze shifts more noticeably from side to side compared with low-risk lane-change conditions.The proposed method distinguishes different types of lane-change risks with finer granularity.The findings also reveal the characteristics of lane-change risks on urban roads,providing data support and theoretical guidance for traffic safety management and risk warning applications.关键词
城市道路/换道风险/深度学习/多源数据融合/匝道区域Key words
urban road/Lane change risk/deep learning/multi-source data fusion/ramp area分类
交通工程引用本文复制引用
ZHU Xinglin,WU Junchao,LIU Hongjun,WANG Nuo,WANG Guangdong,GUO Rui..多源数据融合的城市道路换道风险识别研究[J].重庆理工大学学报,2025,39(23):245-254,10.基金项目
新疆维吾尔自治区自然科学基金项目(2024D01A64) （2024D01A64）
新疆维吾尔自治区教育厅业务项目(XJEDU2024P040) （XJEDU2024P040）
