基于卷积神经网络的多人行为识别方法

摘要

为了解决多人行为识别中人物角色多且难以区分、图片增加的特征维数难以表达和学习以及行为背景复杂且容易产生干扰等问题,提出了一种基于卷积神经网络的多人行为识别方法.考虑到多人行为识别的复杂性,选择较为容易的两人交互行为作为研究对象,对实验中需要的图像数据库进行了初步的收集与预处理;然后选用在特征提取中不受拍摄角度、光照强度影响的Dense-sift算法来对原始图像进行初步的特征提取.由于人体行为图片相对手写数字图片更为复杂,因此为了使该网络能够很好地识别人体行为,针对该网络在其输入、网络层数、滤波器核数、学习率、输出等方面进行了修改.实验结果表明,提出的方法对拳击、拥抱、接吻3类交互行为的识别是有效的.%In order to solve the problems in multi-person behavior recognition,for example,it is difficult to distinguish many characters,it is difficult to express and learn increased feature dimension of image,the behavior background is complex and it is easy to cause interference,this paper proposed a method of multiplayer behavior recognition based on convolutional neural network.At first,considering the complexity of multi-person behavior recognition,the simple twoperson interactive behavior is choosen as the research object and the picture database is collected.Then,because multiplayer behavior recognition has complicated background and many features in the recognition progress,a method using the Dense-sift algorithm for feature pretreatment mode is proposed.Against the complexity of the multiplayer behavior recognition,this network makes various modifications,such as input dimensions which is expanded to include layer convolution,convolution kernel increasing,output reduction,etc.Experimental results show that the proposed method can recognize simple multi-person behavior recognition,such as boxing,hug and kissing effectively.