Real-Time Multi-view 3D Pose Estimation System with Constant Frame Speed

Abstract

Vision-based 3D human pose estimation is a key technique in recognizing human behavior and is widely applied to various fields dealing with human-computer interactions. In particular, the multi-view-based 3D pose recognition method is a method of predicting hypothetical accurate 3D poses that solve problems such as rotation and obscuration by compensating for the shortcomings of viewpoint-dependent 2D pose recognition method and single-view-based 3D pose recognition method. The multi-view-based 3D pose recognition method has excellent prediction performance, but there are difficulties that come because it uses multiple cameras. It is a difficulty in synchronization to simultaneously control an excessive amount of computation at the central server and multiple cameras. In this paper, we propose a distributed real-time 3D pose estimation framework based on asynchronous multi-cameras. The proposed framework consists of a central server and a number of edge devices, which utilize timestamp techniques to output 3d pose estimation results at constant frame speed. Finally, we implement and demonstrate that we successfully estimate a 3D human pose of 30 fps in real time by constructing the proposed framework as a demo platform.