异常驾驶人行为识别与异常度量化研究  

Abnormal Driver Behavior Recognition and Quantification Based on Mix-up Hybrid Contrastive Learning

作　　者：刘卓凡 魏东 LIU Zhuo-fan;WEI Dong(Modern Postal College,Xi'an University of Posts&Telecommunications,Xi'an 710061,Shaanxi,China)

机构地区：[1]西安邮电大学现代邮政学院,陕西西安710061

出　　处：《中国公路学报》2025年第3期113-124,共12页China Journal of Highway and Transport

基　　金：国家自然科学基金项目(52002319,52402419)。

摘　　要：异常驾驶人行为(ADB)是一种常见且易引发交通事故的不安全驾驶人行为,严重危害交通安全。为了全面、准确、快速地检测ADB,提出了一种基于混合对比学习的异常驾驶人行为识别与量化模型(ADB-HCL)。模型创新性地将Mix-up数据增强策略融入对比学习框架,生成介于正常与异常驾驶人行为之间的混合样本,扩展驾驶人行为特征空间的覆盖范围,以提升模型识别未知或罕见异常行为的能力。同时,通过生成代表正常驾驶人行为的模板特征,计算待测驾驶人行为与模板特征之间的距离,量化异常程度,克服传统方法仅能提供已知异常驾驶人行为离散输出的局限性。基于DAD和AIDE数据集的试验结果显示,ADB-HCL方法在识别未知异常驾驶人行为方面表现出色,在推理时间仅为10.75 ms的情况下,准确率达到86.73%,相较于现有方法提高了6%~15%。驾驶人行为异常程度量化结果显示,该方法可以实现对异常驾驶人行为的细粒度量化。研究结果表明,ADB-HCL在识别异常驾驶人行为的全面性、准确性、细致性和检测速度方面具有显著优势,展示了其在车辆主动安全技术中应用的潜力。Abnormal driver behavior(ADB)is a common and unsafe driver behavior that can easily lead to traffic accidents and poses a serious threat to road safety.To comprehensively,accurately,and swiftly detect ADB,this study proposes an abnormal driver behavior identification and quantification model based on hybrid contrastive learning(ADB-HCL).The proposed model innovatively incorporates a Mix-up data augmentation strategy into the contrastive learning framework and thereby generates mixed samples of normal and abnormal driver behaviors.This expands the coverage of the driver behavior feature space,thus enhancing the ability of the model to identify unknown or rare abnormal behaviors.In addition,by generating a template feature representing normal driver behavior and calculating the distance between the test driver behavior and the template feature,the model quantifies the degree of abnormality and hence overcomes the limitations of traditional methods that provide only discrete outputs for known abnormal driver behaviors.Experimental results based on the DAD(Driver Anomaly Detection)and AIDE(AssIstive Driving pErception)datasets show that the ADB-HCL method excels at identifying unknown abnormal driver behaviors and achieves an accuracy of 86.73%with an inference time of only 10.75 ms,which represents a 6%to 15%improvement over existing methods.The quantification results of driver behavior abnormalities indicate that this method enables the fine-grained quantification of abnormal driver behaviors.The findings demonstrate that ADB-HCL has significant advantages in terms of the comprehensiveness,accuracy,granularity,and speed of detecting abnormal driver behaviors,this showcasing the potential applicability of ADB-HCL in vehicle active safety technologies.

关 键 词：交通工程 异常度量化 对比学习 异常驾驶人行为 Mix-up技术 交通安全 

分 类 号：U491.254[交通运输工程—交通运输规划与管理]