利用骨架模型和格拉斯曼流形的3D人体动作识别

3D Human action recognition method using joint point model and Grassmann manifold

为了通过几何特征的有效方法描述人体骨骼运动,构建3D人体动作识别系统,提出一种基于3D骨骼关节空间建模方法。首先,使用自回归和移动平均模型(ARMA)描述每个随着时间变化的运动轨迹,成功捕捉了时空动态运动信息。同时,将该模型的观察矩阵生成的子空间作为格拉斯曼流形中一个点;然后,通过学习控制切线(CT)描述每个类的均值,映射学习过程中的观察变量到所有CT形成局部切丛(LTB),LTB流形数据点可直接在分类器上完成分类;最后,提出的方法使用SVM分类器完成训练和分类。MSR-action 3D、Weizmann和UCF-Kinect三个数据库的实验结果验证了该方法的有效性,与几种基于深度数据的算法相比,该方法获得了最高的识别率,在延迟性方面的性能也表现最优,当帧数为30时,识别率达到97.91%,在延迟较高时,可达到期望识别率。

In order to effectively describe human skeleton movement by geometrical characteristics and build 3D humanaction recognition system, a modeling algorithm based on 3D bone joint is proposed. Firstly, Auto Regressive & MovingAverage(ARMA)is used to describe each trajectory over time, which successfully captures the information of temporaland spatial motion. Meanwhile, the model view matrix generated subspace generated from view matrix of the model is beingas a point on Grassmann manifold. Then, the mean of each class is described by learning Control Tangent(CT), duringthe mapping learning process, observed variables to all the control tangents is being formed Local Tangent Bundle(LTB).And the LTB manifold data points can be directly used to classify in the classifier. Finally, the proposed method uses theSVM classifier to complete training and classification. The effectiveness of the proposed algorithm is verified by experimentalresults on three databases MSR-action 3D, Weizmann and UCF-Kinect. Compared with several algorithms based on depthdata, proposed algorithm not only has achieved the highest recognition rate, but also performs best in terms of latency, andthe recognition rate is 97.91% when the number of frames is 30, so it achieves the desired recognition rate when the delayis high.