Bi-Directional Human Pose Completion Based on RNN and Attention Mechanism

To tackle the occlusion problems between pedestrians in surveillance videos, a human pose sequence completion method is proposed. The method can generate the missing poses in the middle when poses before and after them are visible. The method includes the following steps:(1) A sequence-to-sequence model based on attention mechanism is used to generate the target pose sequence in the middle by taking the visible poses at both ends as input. (2) The same sequence-to-sequence model based on attention mechanism is used to generate the target pose sequence again, but in a reverse direction. (3) The two-direction prediction results are mixed together to obtain the final target pose sequence. The proposed method handles well the poses recovery problem due to occlusions when the poses before and after the target ones are known. The proposed method is tested on Human3.6M, CASIA datasets, etc. The L₂ norm distance between the generated poses and their ground truth is used as the evaluation metric. Compared with previous approach, the average poses error is reduced from 81.6 to 42.5, and the performance is increased by 47.9%.