摘要
选取辅助驾驶中驾驶人行为识别作为研究对象,设计基于融合时间和通道注意力机制的功能模块,构建驾驶人行为识别模型,提升识别准确率。探究多角度数据对模型性能提升情况,自建多视角驾驶人行为数据集,包括4种拍摄视角、10种驾驶行为、1148个视频数据。构建TCAM-R(2+1)D驾驶人行为识别模型,以(2+1)D卷积模块为基础,结合ResNET主干网络,提出融合时间和通道注意力机制的功能模块,增强模型提取时序信息的能力。使用Adabound优化器训练模型,提高模型的识别准确率和泛化能力。实验结果表明:通过增加模型的注意力机制,相较于R(2+1)D模型,自建数据集驾驶人行为识别准确率提高3.03%。采用大型人体运动数据集(HMBD51)进行消融实验,增加融合注意力机制功能模块准确率至59.60%(提高了1.93%),验证融合时间和通道注意力机制的增益效能。
Taking driver behavior recognition in assisted driving as the research object,this paper designs a functional module based on the fusion of time and channel attention mechanism.A driver behavior recognition model is built to improve recognition accuracy.First,a self-built multi-perspective driver behavior data-set is built with 4 shooting perspectives,10 driving behaviors,and 1148 video data.Then,a TCAM-R(2+1)D driver behavior recognition model is built based on the(2+1)D convolution module.With the ResNET backbone network,a functional module integrating time and channel attention mechanisms is proposed,enhancing the model’s ability to extract temporal information.Finally,the Adabond optimizer is employed to train the model and improve its recognition accuracy and generalization ability.Experimental results show by increasing the attention mechanism of the model,the self-built data-set improves the accuracy of driver behavior recognition by 3.03%compared to the R(2+1)D model.A large-scale human motion data-set(HMBD51)is employed for ablation experiments.The fusion attention mechanism functional module’s accuracy rises to 59.60%(an improvement of 1.93%),demonstrating its superior performances in fusion time and channel attention mechanism.
作者
徐慧智
张原铭
XU Huizhi;ZHANG Yuanming(School of Civil Engineering and Transportation,Northeast Forestry University,Harbin 150040,China)
出处
《重庆理工大学学报(自然科学)》
北大核心
2025年第10期1-12,共12页
Journal of Chongqing University of Technology:Natural Science
基金
国家自然科学基金项目(62371170)。
