基于孪生网络改进算法的分心驾驶行为识别  

Distracted Driving Behavior Recognition Based on Siamese Network Improvement Algorithm

作　　者：田竞 曹金璇[2] 王嘉宁 欧晓放 TIAN Jing;CAO Jinxuan;WANG Jianing;OU Xiaofang(School of Traffic Management,People's Public Security University of China,Beijing 100038,China;School of Information and Cyber Security,People's Public Security University of China,Beijing 100038,China)

机构地区：[1]中国人民公安大学交通管理学院,北京100038 [2]中国人民公安大学信息网络安全学院,北京100038

出　　处：《中国人民公安大学学报(自然科学版)》2025年第1期46-53,共8页Journal of People’s Public Security University of China(Science and Technology)

基　　金：公安交管科研数据资料库建设与管理(2022SJKJS06)。

摘　　要：分心驾驶行为是驾驶过程中普遍存在的问题,对交通安全造成了显著影响。当前分心驾驶识别模型参数量大,识别任务使用数据集多为国外开源数据集。针对上述问题,首先模拟6种驾驶行为构建了数据集。再以孪生网络(Siamese network)为框架,将特征提取层改为轻量化的LightCNN9,并融合注意力机制以更好提取图像特征。针对孪生网络特点修改损失函数为二类分交叉熵损失函数(BCE Loss function)。通过自建数据集进行模型训练、测试后,使用开源数据集验证模型可靠性,实验结果表明:检测准确度达到97.2%,模型参数大小6.26 M,验证准确度达到96.4%。消融实验结果表明:map_0.5达到96.6%,说明模型在分心驾驶识别任务上有着较好可靠性与效果。Distracted driving is a common issue that has a significant impact on traffic safety.The current model used to recognize distracted driving has many parameters and foreign open-source datasets are mostly used for recognition tasks.To address these issues,a dataset was constructed to simulate 6 kinds of driving scenarios first.Using the Siamese network as a framework,the feature extraction layer was replaced with the lightweight LightCNN9 and an attention mechanism was integrated to improve feature extraction.Moreover,the loss function was modified to the BCE loss function for the characters of Siamese network.After training and testing the model with a self-built dataset,the reliability of model was verified by using an open-source dataset.The experimental results demonstrated a detection accuracy of 97.2% and a verification accuracy of 96.4%,with a model parameter size of 6.26 M.The ablation experiment results indicated that the proposed model had good reliability and performed well in the distracted driving recognition task,with a map_0.5 of 96.6%.

关 键 词：交通安全 分心驾驶 孪生网络 深度学习 

分 类 号：D035.37[政治法律—政治学]