为了提高人体动作识别的准确率和实时性,提出了一种基于关键帧的人体行为识别新方法。用Kinect提取人体骨架信息(各关节点的3D坐标),将中心点(人体基准参考点)分别与其他各关节点作结构向量,根据空间不变性选取中心向量,计算各个结构向量和中心向量之间的夹角,并将夹角的角速度作为一种新的姿态描述特征,利用AP(Affinity Propagation)聚类算法提取关键帧,利用SVM将得到的关键帧进行动作序列的分类。在Cornell Activity Dataset-60(CAD-60)数据库实验结果表明,新方法具有良好的识别能力。
In order to improve the accuracy and real time of human action recognition,a new method about recognition of human behavior based on key frame is proposed.The information(coordinates of all human joints)of human skeleton is extracted and the structural vectors based on the center(body baseline reference point)and each other joints are constituted.According to the invariance of spatial,the angular velocities of joint angles are calculated between the respective structural vectors.The center vector is considered as a new feature.The AP clustering algorithm is used to extract key frames.The key frames are classified by SVM.The results of experiment at Cornell Activity Dataset-60(CAD-60)show that the algorithm has a good recognition rate.