Three Dimentional Pose Reconstruction Of Human Upper Body From Mono-camera Images

Vision based 3D human pose reconstruction is the process of reconstructing 3D information with images taken from single view or multi view. Reconstruction from monocular image firstly extracts the target’s color, shape, contour and some other clues extracted from the given monocular image, and then obtains the target person’s 3D spatial information with a few known conditions. For the lack of depth information, reconstruction from monocular image is quite challenging.In this paper, we aim at 3D pose reconstruction of human upper body from monocular images or image sequences taken from single view. At first, we review and make a conclusion of the related algorithms about human 2D pose estimation and human 3D pose reconstruction. On the basis of recent researches, we develop a novel 2D limb detection algorithm. In addition, according to the result of 2D limb detection, we implement the 2D-3D pose transformation and reconstruct human upper body’s 3D pose with the representation of parameters. The main research contents and contributions are as follows:1) We propose a 2D limb detection algorithm of human upper body with given monocular image.It utilizes the improved shape context descriptor to model human upper body’s appearance, and trains corresponding classifiers for each human part. Then the input image is detected by the trained classifiers, and the final 2D pose is estimated after optimization.2) A simple approach is proposed to solve human model mapping problem.It makes a simplification to solve the camera model parameters, and decreases the parameters to be determined from 7 to only 1. This simplifying method reduces the algorithm complexity and improves the computational efficiency.3) We develop a joints’ 2D position based 3D pose reconstruction algorithm, and establish a 2D to 3D pose transformation system.It combines the 3D pose with 2D landmarks. At the same time, reprojection error based annealed particle filter is applied to reconstruct human upper body’s 3D pose. In addition, we establish a 2D to 3D pose reconstruction system, which integrates 3D pose reconstruction from monocular image and 3D pose reconstruction from image sequences.4) A large amount of experiments are carried out on four different datasets.The result of experiment shows that the approach we propose in this paper could estimate 2D pose of human upper body efficiently, and then reconstruct the 3D pose rapidly.