摘要
针对隧道环境中监控图像分辨率低与车辆运动轨迹特征异质性减弱导致的驾驶员识别准确率偏低问题,本文提出一种融合卷积与多头注意力机制的驾驶员识别方法(Multi-scale CNN with Multi Attention),通过充分利用驾驶过程人—车—路—环境多源信息的协同耦合关系提升识别精度。首先,设计开展实车驾驶试验,构建针对隧道路段的人—车—路多源驾驶数据库并设计特征集合;其次,搭建驾驶员识别模型框架,该框架通过多尺度卷积神经网络学习驾驶过程中的局部波动,并通过并行的多头自注意力层结构捕捉驾驶时间序列的长期依赖性,实现局部信息与全局信息的有效整合,从而提升隧道场景的驾驶员识别效果。结果显示,与其他先进的算法相比,所提出的模型在驾驶员身份识别任务中的准确率高达99.07%,调和F_(1)分数达到99.03%,充分证明了所提方法的有效性。此外,通过特征贡献度评估方法对隧道场景下驾驶员身份识别任务中的特征重要性进行深入探究发现,相较于车辆历史运动数据,驾驶员心理、生理及视觉特征显示出更高的贡献度。研究结果可为隧道场景多源数据应用提供支持,并对隧道安全监管提供技术支撑。
In response to the problem of low driver recognition accuracy caused by the low resolution of monitoring images and the weakening of vehicle motion trajectory features in tunnel environments,this paper proposes a driver recognition method(Multi-scale Convolutional Neural Network with Multi Attention)that integrates convolution with multi-head attention mechanism.The collaborative coupling relationship of multi-source information of human-vehicle-road-environment during the driving process was utilized to improve recognition accuracy.First,real vehicle driving experiments are designed and conducted to establish a multi-source driving database for tunnel sections and design feature sets.Second,a driver recognition model framework is built.This framework learns local fluctuations in the driving process through a multi-scale convolutional neural network and captures the long-term dependency of driving time series through parallel multi-head self-attention layer structures,effectively integrating local and global information to enhance the driver recognition effect in tunnel scenarios.The results show that compared with other advanced algorithms,the proposed model achieves an accuracy of 99.07%and a macro F_(1) score of 99.03%in driver identification tasks,fully demonstrating the effectiveness of the proposed method.In addition,through the feature contribution evaluation method,the importance of features in driver identification tasks in tunnel scenarios is explored in depth.It is found that compared with historical vehicle motion data,driver psychological,physiological,and visual features show higher contribution.The research results can provide support for the application of multi-source data in tunnel scenarios and provide technical support for tunnel safety supervision.
作者
金盛
周梦涛
白聪聪
JIN Sheng;ZHOU Mengtao;BAI Congcong(Institute of Intelligent Transportation Systems,Zhejiang University,Hangzhou 310058,China;Zhongyuan Institute,Zhejiang University,Zhengzhou 450000,China)
出处
《交通运输系统工程与信息》
EI
CSCD
北大核心
2024年第4期81-93,共13页
Journal of Transportation Systems Engineering and Information Technology
基金
国家自然科学基金(72361137006)
浙江省“尖兵”“领雁”研发攻关计划项目(2022C01042)
浙江省杰出青年基金(LR23E080002)。
关键词
交通工程
驾驶员识别
深度学习
隧道场景
多头注意力机制
多尺度卷积
traffic engineering
driver recognition
deep learning
tunnel scenarios
multi-head attention mechanism
multi-scale convolution