关键词: 动作捕捉, 3D姿态估计, 3D模型驱动, 骨骼动画, 姿态平滑

Abstract: At present, various 3D human pose estimation models have been applied to monocular visual motion capture. However, the results of pose estimation models only contain spatial coordinate information of points and cannot calculate joint rotation, so they cannot be directly used to drive human models. the proposed algorithm aims to map the results of pose estimation algorithms to the skeletal nodes of a 3D human body model skeleton and calculating the rotation of the bones, in order to drive the 3D human body model. Firstly, interpolation is performed using the keypoint coordinates generated from pose estimation results to locate the unpredictable keypoint information. Then, an intermediate matrix is used to align the skeletons between different models, ensuring the applicability of this method on different models. Finally, Kalman filter and low-pass filter are used to smooth the posture of the human body, in order to eliminate the random noise and high-frequency noise caused by the asynchronous action input and 3D scene frame rate. Under the testing of the BlazePose pose estimation model and other human pose estimation datasets such as Human3.6M, the proposed algorithm has achieved good performance in the common 3D human body model driving process from three head to nine head, and in the video of the same action sequence, compared with directly outputting coordinate information to the 3D scene, the mean square error(MSE) of the output frame sequence through Kalman filter and low-pass filter with has been reduced from 2.871 to 0.831, which means the shaking and flicker problems of the model have been significantly improved.

Key words: Motion capture, 3D pose estimation, 3D model driven, Skeletal animation, Smooth posture