3D Human Animation from 2D Monocular Data Based on Motion Trend Prediction

Abstract

A model-based method is proposed in this paper for 3-dimensional human motion recovery, taking un-calibrated monocular data as input. The proposed method is able to generate smooth human motions that resemble the original motion from the same viewpoint the sequence was taken, and look continuous from any other viewpoint. The core of the proposed system is the motion trend prediction for reconstruction. To focus the research effort on motion reconstruction, “synthesized” input is first employed to ensure that the reconstruction algorithm is developed and evaluated accurately. Experiment results on real video data indicate that the proposed method is able to recover human motion from un-calibrated 2D monocular images with very high accuracy.