3D Human Pose Estimation From Multi Person Stereo 360 Scenes
Abstract: This paper presents a human tracking and 3D pose estimation algorithm for use with a pair of 360 cameras. We identify and track an individual throughout complex, multi-person scenes in both indoor and outdoor environments using appearance models and positional data, and produce a temporally consistent 3D skeleton by optimising a skeleton of realistic joint lengths over joint positions produce by Convolutional Pose Machines (CPMs). Our results show an average improvement of 22.67% over state of the art deep learning approaches for tracking, as well as reasonable estimates for pose using just two cameras.
Loading