Improved 3D Human Pose Real-time Estimation based on Monocular Video
Jing Zhang, Chunguang Wang, Liang Chen, Longhua Ren · 2021
Aiming at the ambiguity of real-time estimation of human posture depth of criminal behavior in the field of monocular visual surveillance, an improved 3D human posture real-time estimation method based on video is proposed. Firstly, the 2D joint point coordinates of each frame of characters in unmarked video are detected and used as the input of constrained generation countermeasure network to generate 3D human posture with human joint constraints, Combined with the one-dimensional motion information back projected onto the two-dimensional plane, the result is consistent with the corresponding input in the same frame. The information combination method used in this method can not only rely on a small amount of data to obtain a more ideal accuracy, but also correct and optimize the character motion in the three-dimensional space. At the same time, the constrained generation confrontation network is used to solve the ambiguous motion posture, The experimental results on human 3.6m public data set and self collected criminal behavior video set show the accuracy and practical applicability of this method, and the comparison results are also better than the previous benchmark algorithms.