视觉遮挡下的人体姿态鲁棒估计

Robust estimation of human pose under visual occlusion

针对人体姿态估计中的视觉遮挡问题,提出了一种带有鲁棒卡尔曼滤波(RKF)的人体姿态估计方法。首先,采用随机森林方法(RFM)从深度图像中识别出人体各部件,并计算出人体各关节点在相机坐标系下的3D位置;其次,考虑到视觉自遮挡或遮挡引起的人体部件误识别,设计了一种鲁棒的卡尔曼滤波器,利用假设检验的方法对视觉遮挡造成人体姿态信息中包含的复杂噪声进行识别和分类,以提高人体姿态估计对视觉遮挡的鲁棒性。最后,通过仿真结果以及多个人体关节点的3D位置估计实验表明,所提方法可有效提高人体姿态估计的精度和鲁棒性。

A human pose estimation method with robust Kalman filter(RKF)is presented aiming at the problem of visual occlusion in human pose estimation.Firstly,the random forest method is used to identify each part of the human body from the depth image and calculate the 3D position of each node of the human body in the camera coordinate system.Secondly,considering the false recognition of human body parts caused by visual self-occlusion or occlusion,a robust Kalman filter is designed to identify and classify the complex noises in the human pose information caused by visual occlusion by using the method of hypothesis testing,so as to improve the robustness of human pose estimation against visual occlusion.Finally,by the simulation results and 3D position estimation experiments of multiple human body joints,it demonstrates that improved accuracy and robustness of the human pose estimation can be achieved by the proposed method.