In this paper,a new method of action recognition is proposed,which is based on the previous one.In order to better improve the action recognition accuracy and generalization,based on bilinear method for action recognition of heterogeneous information RGB features and characteristics of the depth of the previous modal,a modal skeletal features increased,so that the three characteristic modes after compression is mapped to a shared learning space,and it also has to reduce the noise and identify useful information capture.In this paper,we use the same method as the previous-matrix to express three kinds of modal characteristics in order to obtain the complex motion and time information.Using the row and column parameters of the matrix to form a low dimensional multi linear model,and to minimize the dimension of the model,a low dimension classifier is established to realize the action recognition.This method has been evaluated on RGB-D and bone two public data sets,and has achieved good results.Even if the three modal data is lost in training or test,it can be realized by other modes.%本文提出了一种新的动作识别的方法,该方法是在前人的基础上增加了一个模态特征.为了能更好的提高动作识别的准确率和泛化能力,在前人的RGB特征和深度特征模态的双线性异构信息的动作识别的方法的基础上,增加了一个骨骼特征模态,使三个特征模态经过压缩映射到一个共享学习空间上,同时这样也有利于减少噪声和捕捉有用的识别信息.本文采用和前人一样的方法—矩阵形式来表示三种模态特征以便获得复杂的动作时空信息.用矩阵的行和列参数组成低维的多线性模型,最小化模型维度后建立一个低维分类器实现动作识别.该方法在RGB-D和骨骼两个公共数据集上进行了评估,取得了不错的效果.即使三种模态数据在训练或者测试中部分丢失也能通过其他模态实现识别.
展开▼
