MEC-Net:基于运动捕捉和通道注意力的行为识别方法  

MEC-Net:Video action recognition method based on motion extract with channel attention

作　　者：郭志鑫 冯秀芳[1] GUO Zhi-xin;FENG Xiu-fang(College of Software,Taiyuan University of Technology,Jinzhong 030600,China)

机构地区：[1]太原理工大学软件学院,山西晋中030600

出　　处：《计算机工程与设计》2024年第6期1805-1811,共7页Computer Engineering and Design

基　　金：山西省重点研发计划基金项目(202102020101007)。

摘　　要：针对视频行为识别过程中面临的特征信息利用不充分、计算量过大的问题,提出一种基于运动捕捉和通道注意力的行为识别模型。模型利用卷积和池化层提高关键特征的利用率,利用空间通道注意力模块在通道维度利用自适应学习参数聚合信息,降低背景冗余信息的影响,引入时空注意力机制融合特征信息获得分类结果。所提模型在公开数据集UCF101、Kinetics-400以及HMDB51上分别获得了94.5%、80.2%和61.9%的精确度,对比其它模型具有更加精准的识别结果以及更少的计算量,验证了模型的有效性。An action recognition model based on motion capture and channel attention was proposed to solve the problems of insufficient use of feature information and excessive calculation in the process of video action recognition.The convolution and pooling layers were used to improve the utilization of key features,and the spatial channel attention module was used to aggregate information using adaptive learning parameters in the channel dimension,effectively reducing the influence of redundant background information,and a spatiotemporal attention mechanism was introduced to fuse feature information to get classification results.The proposed model achieves accuracy of 94.5%,80.2%and 61.9%respectively on public datasets UCF101,Kinetics-400 and HMDB51.Compared with other models,it has more accurate recognition results and less calculation,which verifies the model’s effectiveness.

关 键 词：行为识别 视频 混合模型 注意力机制 时空特征 自适应 通道 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]