Seq2seq motion continuity and limb invariance constrained spatial–temporal encoder for 3D human pose estimation

Recent years have seen deep neural network-based methods for 3D human pose estimation from 2D poses, but most directly predict key point positions without leveraging limb-joint correlations during human movement. In this article, we propose a limb-invariance-constrained hybrid spatiotemporal encoder, which extracts spatiotemporal and motion features of joints and limbs from continuous frames and uses limb invariance to boost key point estimation accuracy. Experiments on common datasets show it achieves state-of-the-art performance, supporting broader applications such as human-computer interaction and medical rehabilitation.