摘要
针对车辆行驶过程中复杂多变的车内环境和光照条件及驾驶员行为姿态多样性对驾驶员异常行为检测与识别的影响,提出一种基于对比学习的驾驶员异常驾驶行为检测算法。首先,将驾驶员驾驶行为检测视为二分类任务,使用对比学习的方法,将驾驶员正常驾驶与异常驾驶样本进行对比,并通过对比损失函数提高模型的性能。其次,以驾驶员正前方和上方的深度图像作为输入,通过提供驾驶员的深度信息解决车内环境复杂、光照强度变化及视角盲区问题。最后,在轻量化网络MobileNetV2基础上引入3D卷积,并在每个瓶颈结构的卷积层中加入通道混洗操作,提高识别的准确性。试验结果表明,提出的算法在驾驶员异常检测(DAD)数据集测试集中的准确率达到94.18%,受试者操作特征(ROC)曲线下面积(AUC)达到0.962,该算法在驾驶员异常行为检测方面具有有效性。
In the process of driving a vehicle,the complex and changing environment inside the vehicle,the change of lighting conditions and the diversity of drivers’behavioral postures affect the detection and recognition of abnormal driver behavior.To address this issue,this paper proposes a driver abnormal driving behavior detection algorithm based on contrast learning.The paper firstly considers driver’s driving behavior detection as a binary classification task,and utilizes a contrast learning approach to compare driver’s normal driving with abnormal driving samples and to improve the performance of the model by contrasting loss functions.Secondly,the depth images right ahead and above the driver serves as inputs to solve the problems of complex in-vehicle environment to change the light intensity and blind spots in viewpoint by providing the depth information of the driver.Finally,3D convolution is introduced in the lightweight network MobileNetV2,and the operation of channel blending is added to the convolution layer of each bottleneck structure to improve the accuracy of recognition.Test results show that accuracy of the proposed algorithm reaches 94.18%in the Driver’s Abnormality Detection(DAD)dataset and ROC AUC reaches 0.962,which shows the effectiveness of the algorithm in driver’s abnormal behavior detection.
作者
李仲伦
于光达
杨帅
邹世野
张鹤群
汪春雨
Li Zhonglun;Yu Guangda;Yang Shuai;Zou Shiye;Zhang Hequn;Wang Chunyu(Jilin Province Product Quality Supervision and Inspection Institute,Changchun 130103)
出处
《汽车工程师》
2025年第8期29-36,共8页
Automotive Engineer
基金
吉林省市场监督管理厅科技计划项目(2023MK008)。
关键词
异常驾驶行为检测
对比学习
二分类
3D卷积神经网络
Abnormal driving behavior detection
Contrast learning
Second classification
3D Convolution Neural Networks(CNN)
