Robust human motion estimation using bidirectional motion prior model and spatiotemporal progressive motion optimization

Abstract

We propose a bidirectional conditional variational autoencoder (CVAE) motion prior model and a spatiotemporal progressive motion Optimization approach for robust human motion estimation. Priors have been playing an important role in handling pose estimation problems with occlusion and noise. The bidirectional motion prior model(BMP) we proposed can generate better motion priors by adjusting the initial states compared to the unidirectional generative model. To achieve plausible optimization results, we introduce a spatiotemporal progressive motion optimization (SPMO) strategy. The SPMO guides the optimization process towards a feasible and optimal solution. Qualitative and quantitative results on both 2D and 3D datasets have demonstrated the importance of the bidirectional motion model and the effectiveness of our optimization approach