深度运动图耦合正则化表示的行为识别算法  

Action recognition algorithm based on depth motion maps and regularized representation

作　　者：李贤阳[1,2] 阳建中 杨竣辉[3] 陆安山 

机构地区：[1]钦州学院电子与信息工程学院 [2]钦州市电子产品检测重点实验室 [3]江西理工大学信息工程学院

出　　处：《电子测量与仪器学报》2018年第1期119-128,共10页Journal of Electronic Measurement and Instrumentation

基　　金：国家自然科学基金(61273328);广西高校中青年教师科研基础能力提升项目(2017KY0795);钦州市科技攻关项目(20164410);钦州市电子产品检测重点实验室开放项目资助

摘　　要：为了提高图像行为的识别精度,使其能够准确判别行为识别中的微小变化以及遮挡问题,提出了基于深度运动图(depth motion maps,DMM)与正则化协同表示的行为识别算法。首先,将深度图像序列投射到3个正交平面上,得到了3个方向的投射图。对于不同的投射图,通过测量两个连续映射之间的绝对差值来表示运动能量,并将所有深度图像序列中运动能量进行叠加,获得了3个方向的深度运动图。随后,根据这些投射图,DMM能从多个方向获取更多具有判别力的运动信息。再引入Hough变换(Hough transform,HT)算子,提取DMM中3个方向的HT特征,并其进行归一化融合,获取DMM-HT特征。最后,引入Tikhonov正则化计算系数向量,构建正则化协同表示分类器,对每个位置样本的分类标签完成深度行为分类学习,实现人体行为的准确识别。实验数据表明,与当前行为识别技术相比,算法具有更强的鲁棒性,能完成各种行为的识别,在遮挡、噪声等干扰条件下具有更高的识别精度。所提算法能够较好地适应复杂环境下的人体动作准确识别,在智能家居、视频监测、人机交互等领域具有良好的参考价值。In order to improve the recognition accuracy of image behavior to accurately distinguish small change and occlusion in behavior recognition,a behavior recognition algorithm based on DMM and regularized cooperative representation was proposed. Firstly,three projected projections map was obtained by projecting the depth image sequence into three orthogonal planes. Then the motion energy was represented by measuring the absolute difference between two successive mappings for different projective maps,and the depth motion maps in three directions were obtained by superimposing the motion energy in all the depth image sequences. Subsequently,DMM can obtain more discriminative motion information from multiple directions according to these projection maps. Then the HT features of three directions in DMM were extracted by introducing the Hough transform,and DMM-HT features were obtained by normalized fusion.Finally,a regularized cooperative representation classifier was constructed by introducing the Tikhonov regularization to compute the coefficient vectors to perform deep behavior classification learning for classified label of each location sample for achieving accurate recognition of human behavior. The experimental data show that this algorithm is more robust to complete the recognition of various behaviors,and has higher recognition accuracy under the interference of occlusion,noise and other interference conditions compared with the current behavior recognition technology. This algorithm can well adapt to the accurate recognition of human movements in complex environment,and has a good reference value in smart home,video surveillance,human-computer interaction and other fields.

关 键 词：深度运动图 行为识别 运动能量 正则化协同表示 HOUGH变换 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]