A Denoising Method of Human Pose Estimation Based on Reconstruction from Keyframes
Wei Gao, Feng Qiu, Huanhuan Zhang · 2022 7th International Conference on Image, Vision and Computing (ICIVC) · 2022
In the field of 3D human motion reconstruction, it is always the case when jitters on 2D key points corrupt 3D human pose visualization. Hence, denoising is necessary dealing with 3D key points generated by neural networks. Contrary to traditional methods, a post-processing denoising method based on reconstruction from keyframes is proposed, giving consideration to both smoothness and reconstruction accuracy. This approach converts key points to bones’ rotation angles to avoid bone length skewing in post-processing. Afterwards, it selects keyframes with acceleration constraint, which is insensitive to jitters, ensuring the accuracy of keyframe extraction. Finally, Squad spline interpolation is implemented within keyframes in order to smooth human motion curves. Extensive experiment results demonstrate that this method can, guaranteeing accuracy of human poses, remove jitters within motions, increasing smoothness of motion curves, and eliminate bone length skewing in produced by filters completely. For complicated motion data, such as dances, and different noise patterns, this method works stably as well.