3D局部特征耦合回归森林的图像动作识别算法

Motion recognition algorithm based on 3D local feature coupling regression forest

为提高人体动作的识别与理解能力,设计3D局部特征耦合回归森林学习的动作识别方案。利用Gaussian滤波器从深度图像中提取人体轮廓的空间点,将获得的轮廓点映射到3D直方图中,形成3D局部特征;根据3D局部特征,利用关节点与边缘形成人体的图形模型,获取其结构约束Φ(x_i)和空间约束Φ(x_(ij));引入回归森林(regression forests,RF)定义Φ(x_i)、Φ(x_(ij))的回归系数,利用Gaussian密度函数计算Φ(xi)、Φ(x_(ij))的相互分布关系,对其进行分类学习,完成人体动作识别与理解。实验结果表明,与当前方法相比,所提方法具有更高的动作识别准确率,可有效学习人体结构和定位关节。

To improve the recognition and understanding of human action,an action recognition scheme for 3D local feature coupling regression forest learning was presented.The spatial point of human contour was extracted from depth image using Gaussian filter.The obtained contour points were mapped to the 3D histogram to form the 3D local feature.According to the extracted 3 Dlocal features,agraphical model of human body was formed by joint and edge,the structural constraints Φ（xi） and spatial constraints Φ（x（ij））were obtained.The regression coefficient（regression forests,RF）was introduced intoΦ（xi）,Φ（x（ij））to calculate the regression coefficient,and Gaussian density function was used to calculate theΦ（xi）,Φ（x（ij））distribution of each other for learning to complete human action recognition and understanding.Experimental results show the proposed method significantly improves the accuracy of motion recognition for depth images to effectively study the human body structure and positioning joints compared with the current methods.