Human posture probability density estimation based on actual motion measurement and eigenpostures

Abstract

In this paper, we construct the human posture probability density that is based on the actual human motion measurement. Human motions in the daily-life were measured for two days by wearing a mechanical motion capture. The accumulated postures were converted to quaternion for a guarantee of the uniqueness of the posture representation. In order to represent the probability density effectively, we propose the eigenpostures for the posture compression and use the kernel based reduced set density estimator (RSDE) for reduction of data samples. By applying the constructed human posture probability density for unprecedented posture detection and human motion segmentation, we verify its effectiveness for many kinds of application.