基于运动历史图像与卷积神经网络的行为识别  

Action Recognition Based on Motion History Image and Convolution Neural Network

作　　者：石英[1] 孙明军 李之达 罗佳齐 杨明东 SHI Ying;SUN Ming-jun;LI Zhi-da;LUO Jia-qi;YANG Ming-dong(School of Automation,Wuhan University of Technology,Wuhan 430070 China)

机构地区：[1]武汉理工大学自动化学院

出　　处：《湘潭大学学报（自然科学版）》2019年第2期109-117,共9页Journal of Xiangtan University(Natural Science Edition)

基　　金：江苏省重点研发计划项目(BE2016155);国家自然科学基金项目(61673306)资助

摘　　要：针对人体行为识别难于兼顾速度与精度的问题,提出了一种结合运动历史图像(MHI)与卷积神经网络的行为识别算法.该算法首先从原始视频序列中计算MHI,不仅减少了待处理的信息量,还提取了行为识别中的关键时空信息;接着以MHI作为输入,搭建了深度卷积神经网络,可以更好地表达时空信息;最后利用随机梯度下降法与dropout策略训练网络,实现行为类别分类.对比不同卷积神经网络训练与测试实验,该算法在Weizmann行为识别数据集上取得了95%的平均识别率,相较于未改进的网络结构提升了1.2%;对于持续时间为1.6s的行为动作,该算法的识别时间为1.56s.实验结果表明,所提算法在维持较高识别准确率的同时,实现了人体行为的在线实时识别与分类.Considering the speed and accuracy of human behavior recognition, a compromise scheme was given on the base of motion history image (MHI) and convolution neural network. Firstly, MHIs were extracted from the original video sequence in order to reduce the amount of information to be processed and preserve the key spatial-temporal feature.Secondly, a convolution neural network was constructed to comprehensively describe the MHI feature. Finally, by using the network trained with stochastic gradient descent (SGD) and dropout strategy, different kinds of behaviors could be successfully classified. In the training and testing experiments, different networks were compared, and the proposed algorithm obtained recognition rate of 95% on Weizmann behavior identification dataset, which was 1.2% higher than the other two networks. Moreover, take the human behavior lasting 1.6 s for example, the recognition time of the proposed algorithm was 1.56 s. The experiment results show that the proposed algorithm can realize the real-time classification of human behaviors at a high recognition rate.

关 键 词：深度学习 运动历史图像 卷积神经网络 行为识别 Weizmann行为识别数据集 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]