A Robust 3D Pose Estimation Algorithm Based on 2D Camera Arrays
Ziting Xiao, Wei Fan, Tingting Zhao · 2024
The implementation of 3D human posture estimation is one of the classic tasks in computer vision, which is the basis of posture recognition, behavior tracking, and human body tracking technologies, and has a wide range of potential applications in the fields of rehabilitation, video surveillance, and advanced human-computer interaction. Aiming at the problems of expensive traditional cameras used for 3D human posture estimation, complicated operation process, facing occlusion as well as depth uncertainty, an algorithm for multicamera soft synchronization based on a two-layer point matching strategy is proposed, and a probabilistic graph model based on the length of the skeleton is also established to optimize and fuse the results of 3D human posture estimation from multiple RGB cameras. First, a novel digital clock video is designed and a combination of global point matching based on deep learning and local point matching based on SIFT features is used as a strategy, where each camera array can obtain accurate image capture time simultaneously; on this basis, the least squares method is used to model the camera capture time (PC side) and shutter time (sensor side); 2D body pose after synchronization, 3D body pose obtained using triangulation. Finally, a probabilistic graph model based on bone length is established to decompose the graph dependency of the a posteriori probability to infer the global optimal solution of the human joint tree map, and the optimal 3D human pose is solved based on the process of non-parametric confidence propagation. Experiments show that the camera array algorithm of this novel soft synchronization algorithm can effectively synchronize almost all types of cameras, and the average absolute difference can reach below 11 ms. The probabilistic graphical model based on the bone length can provide accurate estimation of the 3D human body pose.