RGBD人体行为识别中的自适应特征选择方法

目前在RGBD视频的行为识别中，为了提高识别准确率，许多方法采用多特征融合的方式。通过实验分析发现，行为在特定特征上的分类效果好，但是多特征融合并不能体现个别特征的分类优势，同时融合后的特征维度很高，时空开销大。为了解决这个问题，提出了RGBD人体行为识别中的自适应特征选择方法，通过随机森林和信息熵分析人体关节点判别力，以高判别力的人体关节点的数量作为特征选择的标准。通过该数量阈值的筛选，选择关节点特征或者关节点相对位置作为行为识别特征。实验结果表明，该方法相比于特征融合的算法，行为识别的准确率有了较大提高，超过了大部分算法的识别结果。

Adaptive feature selection method for action recognition of human body in RGBD data

Many methods adopt the technique of multi-feature fusion to improve the recognition accuracy of RGBD video. Experimental analyses revealed that the classification effect of certain behavior in some features is good; however, multi-feature fusion cannot reflect the classification superiority of certain features. Moreover, multi-feature fusion is highly dimensional and considerably expensive in terms of time and space. This research proposes an adaptive feature selection method for RGBD human-action recognition to solve this problem. First, random forest and information entropy were used to analyze the judgment ability of the human joints, whereas the number of human joints with high judgment ability were chosen as the feature selection criterion. By screening the threshold number, either the joint feature or the relative positions of the joints was used as the recognition feature of action. Experimental results show that compared with multi-feature fusion, the method significantly improved the accuracy of action recognition and outperformed most other algorithms.