Multi-View Auto-Calibration Method Based on Human Pose Estimation

Abstract

This paper contributes a novel automatic calibration method of a multi-camera system based on human joints. The main difficulty of this problem is the lack of enough known world coordinate points and corresponding two-dimensional image points. Most previous methods address this difficulty by using customized calibration tools, which is tedious due to a lot of manual intervention. In this work, we use human joints as the corresponding points between cameras to deal with this problem. In addition, the proposed method does not need a customized calibration tool, but only requires a person to walk in the calibrated scene, which can reduce the calibration cost greatly. Moreover, the experimental results of binocular and four-camera system indicate that the proposed method outperforms the state of-the-art methods in the case of small public market for both binocular cameras and multi-camera cameras.