基于压缩视频的驾驶行为识别方法  

Driving Action Recognition Method Based on Compressed Videos

作　　者：帅真 杨会成 胡耀聪 林园园 李雯婷 SHUAI Zhen;YANG Huicheng;HU Yaocong;LIN Yuanyuan;LI Wenting(College of Electrical Engineering,Anhui Polytechnic University,Wuhu Anhui 241000,China)

机构地区：[1]安徽工程大学电气工程学院,安徽芜湖241000

出　　处：《传感技术学报》2025年第3期504-510,共7页Chinese Journal of Sensors and Actuators

基　　金：国家自然科学基金资助项目(62203012);安徽省重点实验室开放课题项目(JCKJ2022A07)。

摘　　要：针对在识别速度快的同时无法达到较高的精度这一问题,提出了一种基于压缩视频的多分支轻量化驾驶行为识别框架。首先,直接将压缩视频处理成为三个分支,为了降低计算成本,采用轻量化卷积神经网络进行时空建模。一个分支通过使用轻量级的二维卷积网络捕获外观线索,另外两个分支采用轻量级三维卷积网络分别从运动向量和残差帧中学习时间信息,结合三个分支的测试结果,得到最终的准确度。另外引入教师模型引导轻量化模型,从高容量的时空深度学习模型中提炼出互补的知识,将其迁移到所提出的多分支轻量化模型中进一步改善识别精度。实验结果表明,所提出的框架对驾驶行为识别具有较好的可行性和有效性。Targeting at the problem of not being able to achieve high accuracy while having fast recognition speed,a multi-branch light-weight driving action recognition framework based on compressed video is proposed.First,the compressed video is direct processed into three branches,and in order to reduce the computational cost,a lightweight convolutional neural network is used for spatio-temporal modeling.One branch captures appearance cues by using a lightweight 2D convolutional network,and the other two branches employ a lightweight 3D convolutional network to learn temporal information from motion vectors and residual frames,respectively.The test results of the three branches are combined to obtain the final accuracy.In addition,a teacher model is introduced to guide the lightweight mod-el,and complementary knowledge is extracted from the high-capacity spatio-temporal deep learning model,which is migrated to the pro-posed multi-branch lightweight model to further improve the recognition accuracy.The experimental results show that the proposed framework has good feasibility and effectiveness for driving action recognition.

关 键 词：行为识别 压缩视频 轻量化网络 时空建模 知识蒸馏 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]