Human Pose Denoising Via Keyframe-Based Reconstruction

In recent years,with the development of computer vision techniques,human pose estimation has been more and more widely used in practice,in fields with promising prospects,such as virtual reality,human-computer interaction,automatic pilot,motion identification and so on.In a sentence,human pose estimation is to estimate the positions of human’s skeleton and joints from pictures and videos,to construct human body models.Nevertheless,it is of significant challenge to reconstruct 3D human pose from 2D data like pictures and videos.It is always the case that noises and jitters are within the reconstruction results.These noises may be resulted from the limitation of models’ fitting ability,or introduced within reconstruction pipelines.These flaws cannot be eliminated by human pose estimation models merely,and have nonnegligible impact on final effect of human body models.Thus,it is usually an inalienable step to perform denoising upon the results of human pose estimation in actual projections.There are often many problems when applying traditional denoising techniques to post-processing of human pose estimation.For instance,the human skeleton is a whole,bone lengths of which must be constant within movements.Otherwise,the final output will look like certain limbs are stretched or contracted.Hence,bone length deviation needs special consideration.For another example,actual human motions are natural and coherent,which means that the speed will not change swiftly,so that there will be few vacillations in visual effect.All these problems should be taken into consideration when design denoising method for human pose estimation.On that account,we introduce a whole pipeline for denoising of human pose estimation based on reconstruction from keyframes.First of all,we transform human pose estimation’s output from key points to rotation angles.By this mean can we control bone lengths and eliminate bone length deviation.Secondly,we promote a RamerDouglas-Peucker algorithm with acceleration constraint for keyframe selection.We perform this to minimize jitters within trajectory of human motions while controlling approximation error.Finally,based on the results of keyframe selection,we perform quaternion Squad spline interpolation for keyframes,smoothly connecting keyframes under a relatively low computational cost,to achieve the final denoising results.To demonstrate the effect of this article,we perform human pose estimation and post-processing methods on different human motion trajectories,and compare our results with benchmark methods.We look into a series of indicators like bone length deviation,estimation error and acceleration and analysis our advantages.The experiment results show that our method can completely eliminate bone length deviation.Meanwhile,it has the best reconstruction error within the methods compared,and motion trajectories of which is most natural,according with laws of human movements.