基于Kinect和ROS的骨骼轨迹人体姿态识别研究

Research on human body attitude recognition based on Kinect and ROS

为了解决不完整人体动作识别的问题,在机器人操作系统(robot operating system,ROS)下提出利用骨骼节点向量的角度累计变化作为特征向量,并采用自适应能量的方法划分视频人体动作。在人体解剖学的基础上建立投影坐标平面,进行空间位置和骨骼角度的计算。通过时间金字塔方法对不同时间间隔的骨骼角度数据编码,形成多级特征向量更好地表示人体动作。在人体受遮挡情况下,使用扩展卡尔曼滤波预测骨骼节点坐标,提高骨骼坐标的准确性。该方法具有旋转、平移不变性,识别4种不完整人体动作的正确率达到了92.25%。

In order to solve the problem of incomplete human motion recognition, the angular cumulative change of the skeleton node vector is proposed as the feature vector under the robot operating system(ROS), and the adaptive human energy method is used to divide the video of human body motion. The projection coordinate plane is established on the basis of human anatomy, and the calculation of the spatial position and the bone angle is performed. The time pyramid method is used to encode the bone angle data of different time intervals, and the multi-level feature vector is formed to better represent the human body motion. In the case of human occlusion, the extended Kalman filter is used to predict the coordinates of the bone nodes and improve the accuracy of the bone coordinates. The method has rotation and translation invariance, and the correct rate of identifying 4 incomplete human movements reaches 92.25%.