Abstract: In the field of 3D human pose estimation, occlusion leads to inaccurate extraction of human joint points. To address this problem, this paper proposes a 3D human pose estimation algorithm that combines spatio-temporal attention and channel attention. Firstly, a feature filtering module is proposed, which further captures the feature information of human joint points by introducing the position embedding module. Then, a mobile vision transformer temporal attention module is proposed, which can obtain more details of pose features by introducing the SiLU activation function. Finally, a channel attention module is proposed, which adjusts the weights of the output channel features by introducing a parallel branch processing architecture and adding normalization layers, so that the algorithm can focus on human pose features while reducing the influence of background features. Experiments are conducted on the Human3.6M dataset. Compared with the baseline model Strided Transformer, the mean per joint position error (MPJPE) and the procrustes-aligned mean per joint position error (P-MPJPE) decrease by 2.5% and 2.3%, respectively, when the 2D joint points extracted from the cascaded pyramid network (CPN) are used as input. The MPJPE decrease by 6.7% when the annotated 2D joint points of the Human3.6M dataset are used as input. Experimental results show that the proposed algorithm has high accuracy.

Key words: 3D human pose estimation, occlusion, spatio-temporal attention, channel attention, Transformer