一种结合深度信息的人体行为识别方法  

Human Behavior Recognition Method Based on Depth Information

作　　者：赵欣[1] 周海英[1] 

机构地区：[1]中北大学计算机与控制工程学院,太原030051

出　　处：《科学技术与工程》2017年第1期244-249,259,共7页Science Technology and Engineering

基　　金：山西省自然科学基金项目(2013011017-6)资助

摘　　要：针对现有的人体行为识别方法中易受到噪声、光照以及复杂背景等因素的影响,同时,未充分考虑到人体交互区域的作用。提出一种结合普通彩色视频和深度信息的人体行为识别方法。首先,对于识别中提取人体运动目标时,利用深度图中物体表面法向量提取运动目标的边缘;同时结合加权累计帧差法获取运动模板。其次,结合深度连续性提取非人体区域(人体与动作的交互区域)并进行描述,作为人体行为表示的一部分。最后利用支持向量机(support vector machine)进行训练和识别。实验部分在CAD-120数据集中测试,通过与一些现有的人体行为识别方法相比较,动作识别准确率提高了5%左右。In view of the influence of noise,illumination and complex background,at the same time,not fully taking into account the role of the human body interaction region,a method of human behavior recognition based on common color video and depth information is proposed. Firstly,in the process of extracting the human body＇s moving object,the edge of the moving object is extracted by the normal vector of the object surface in depth maps,and combined with the weighted cumulative frame difference method to obtain the motion template. Secondly,combined with the depth of the continuous to extract the non human body region（ the interaction between the body and the region） and described as part of the human behavior representation. Finally,support vector machine（ vector machine support） is used for training and recognition. In the experimental part,the dataset is CAD-120. The accuracy of motion recognition is increased by about 5% compared with the existing methods of human behavior recognition.

关 键 词：人体动作识别 时空图像分割 交互区域 支持向量机 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]