Loading [a11y]/accessibility-menu.js
Single view image based — 3D human pose reconstruction | IEEE Conference Publication | IEEE Xplore

Single view image based — 3D human pose reconstruction


Abstract:

In this paper, we propose an improved framework for estimating 3D human pose sequences from 2D videos. Prior methods were declared to solve this problem; however, they so...Show More

Abstract:

In this paper, we propose an improved framework for estimating 3D human pose sequences from 2D videos. Prior methods were declared to solve this problem; however, they sometimes produced invalid poses because the constraint of result poses was not robust enough to satisfy anthropometric regular body pose. Our goal is to eliminate this drawback, provide more accurate estimated poses. The proposed framework consists of three stages. First, we start by generating a set of class-independent body part candidates for each video frame and constructing a fully connected graph from these candidates. Integer linear programming is used to label each candidate by a part class and assign it to individuals to form accurate and consistent pose. We then make the outputs from the first stage suitable for the next stage by inferring the 2D joint configuration. Second, from 2D poses resulted from the first stage, 3D poses are reconstructed by a matching pursuit algorithm operating on camera projection, all the output must adapt the anthropometric regular body pose constraint such as limb length and joint angle constraint. Third, we exploit the temporal and spatial information in videos to obtain more accurate 3D poses. Our experiments on CMU MOCAP dataset show that the proposed framework produces better results compared with prior works.
Date of Conference: 19-21 October 2017
Date Added to IEEE Xplore: 23 November 2017
ISBN Information:
Conference Location: Hue, Vietnam

References

References is not available for this document.