基于新投影策略的人体行为识别方法研究

A human action behavior recognition method based on new projection strategy

为解决微小动作识别率低的问题,提出一种结合新投影策略和能量均匀化视频分割的多层深度运动图的人体行为识别方法。首先,提出一种新的投影策略,将深度图像投影到三个正交笛卡尔平面,以保留更多的行为信息;其次,基于整个视频的多层深度运动图图像虽然可反映整体运动信息,但却忽略了很多细节,采用基于能量均匀化的视频分割方法,将视频划分为多个子视频序列,可以更加全面地刻画动作细节信息;最后,为描述多层深度运动图图像纹理细节,采用局部二值模式作为动作特征描述子,结合核极端学习机分类器进行动作识别。实验结果表明:在公开动作识别库MSRAction3D和手势识别库MSRGesture3D上,本文算法准确率分别达94.55%和95.67%,与现存许多算法相比,有更高的识别率。

To solve the problem of low recognition rate of micro motion,this paper proposes a multi-layer depth motion maps human action recognition method based on new projection strategy and energy homogeneous video segmentation.Firstly,the paper proposes a new projection strategy that projects the depth image into three orthogonal Cartesian planes,so as to retain more behavioral information.Secondly,considering that the image of multi-layer depth motion maps based on the whole video can reflect the whole motion information of the video but ignores a lot of detail information,the paper adopts a video segmentation method based on energy homogenization to divide an action video into multiple sub-video sequences,which can more sufficiently depict the detail information.Lastly,this paper uses a local binary pattern feature descriptor to describe the detail texture features of depth motion maps and adopts kernel extreme learning machine classifier to recognize actions.Experimental results on MSRAction3D and MSRGesture3D show that the proposed algorithm can achieve the accuracy of 94.14%and 95.67%,respectively,and has higher accuracy than the existing algorithms.