DMM-SIFT算子耦合SVM的深度图动作识别算法

摘要

Aiming at poor performance of human action recognition for deep images,a recognition method based on scale-invariant feature transform feature with multi directions was proposed.The depth sequence diagram was mapped to three mutually orthogonal planes in turn,the absolute values of the difference between adjacent frames were accumulated,and the depth motion map was obtained.SIFT features on DMM were extracted,the SIFT characteristics of three planes were calculated after obtaining the DMM on the three plane,and they were normalized.The support vector machine (SVM) was introduced,and the normalized feature description was sent to the SVM to study and test the depth action model.Experimental results show that the proposed method is effective,and it can effectively extract the depth chart information with high precision and strong robustness compared with the current commonly used motion recognition algorithms.%针对目前深度图动作识别的低效性问题,提出基于多方向的尺度不变特征转换(scale invariant feature transform,SIFT)算子的深度图识别方法.将深度序列图依次映射到3个相互正交的平面上,累加相邻帧之间差的绝对值,得到深度运动图(depth motion map,DMM);在DMM上提取SIFT特征,得到3个平面上的DMM后,分别计算3个平面的SIFT特征,对其进行归一化处理;引入支持向量机(support vector machine,SVM),将归一化的特征描述嵌入到SVM中,进行深度动作模型的学习与测试.实验结果表明,相对于当前常用的动作识别算法,所提动作识别技术具有更高的检测精度与更强的鲁棒性,能够更有效地提取出深度图里的动作信息.