摘要：Aiming at the high computational complexity of the 3D pose estimation algorithm in the deep learning field, we propose a fast human poseestimation algorithm based on optical flow and particle filter. The temporal correlation between video frames is applied to the algorithm. The first frame of the video is defined as a keyframe, which will serve as the output of the 3D pose estimate. Then the next frame is determined by the key frame algorithm whether it is a key frame. The key frame is estimated by the 3D human pose estimation algorithm, and theoutput result of the key frame is propagated to the non-key frame through the optical flow mechanism.Non-key frames are subjected to pose estimation through particle filter. In the 3D human pose estimation problem, we propose a unified equation for 3D human pose estimation from the RGB image, combining 2D joint estimation and 3D posereconstruction. Theproposed approach outperforms all state-of-the-art methods on Human3.6m achieving a relative error reduction greater than 30% on average. Our method significantly improves detection performance compared to the original algorithm, and the detection speed can be increased by an average of 43.75%.
