基于4D卷积和混合注意力的驾驶行为检测方法  

Driving behavior detection method based on 4D convolution and mixed attention

作　　者：高军 易建钢[1] GAO Jun;YI Jian-gang(School of Smart Manufacturing,Jianghan University,Wuhan 430056,China)

机构地区：[1]江汉大学智能制造学院,湖北武汉430056

出　　处：《计算机工程与设计》2023年第5期1427-1433,共7页Computer Engineering and Design

基　　金：湖北省教育厅科学研究计划指导性基金项目(B2021055)。

摘　　要：交通状况的复杂性和不确定性往往会影响驾驶行为模型的检测性能,为准确从前方道路视频中检测出本车的驾驶行为,提出一种基于混合注意力的4D卷积神经网络(MA4DN)模型。结合4D残差卷积块,MA4DN模型具备同时建模短程和长程时空表征的能力;为MA4DN设计一种混合注意力模块(MAM),联合学习时空感知特征、通道依赖特征和运动上下文特征等多类型信息;设计一种融合focal loss的损失函数,增强MA4DN模型对难样本的学习能力。在一个自然驾驶数据集上的实验结果表明,MA4DN的检测精度和泛化能力明显优于其它先进的3D卷积神经网络模型。The performance of driving behavior detection is usually influenced by the complexity and uncertainty of traffic status.To accurately detect the driving behavior from the video of road ahead,a mixed attention-based 4D convolutional neural network(MA4DN)was proposed.Combined with the 4D residual block,the MA4DN was able to model both the short-term and the long-term spatio-temporal representations simultaneously.A mixed attention module(MAM)was designed for MA4DN to jointly learn multi-branch information including spatio-temporal perception,channel interdependencies and motion context.A loss function integrated with a focal loss(FL)item was designed to enhance the learning ability of hard samples.The experiments were conducted on a naturalistic driving data set.MA4DN achieves excellent results,outperforming recent 3D CNNs by a large margin.

关 键 词：驾驶行为检测 深度学习 4D卷积网络 混合注意力 高级驾驶辅助系统 计算机视觉 时空感知 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]