Model based full body human motion reconstruction from video data

Abstract

This paper introduces a novel framework for full body human motion reconstruction from 2D video data using a motion capture database as knowledge base containing information on how people move. By extracting suitable twodimensional features from both, the input video sequence and the motion capture database, we are able to employ an efficient retrieval technique to run a data-driven optimization. Only little preprocessing is needed by our method, the reconstruction process runs close to real time. We evaluate the proposed techniques on synthetic two-dimensional input data obtained from motion capture data and on real video data.

Retrieval, Recognition and Reconstruction of Quadruped Motions

“…Rosenhahn et al employ geometric prior information about the movement pattern in markerless pose tracking process (Rosenhahn et al, 2008). Most of the work regarding reconstruction from video sequences has been done on human motion like (Yasin et al, 2013).…”

“…In case of video data, 2D feature sets and SURF feature detection techniques. We have used the same feature detection technique as in (Yasin et al, 2013) and refer to this work for further details. In this paper, we only deal with intrinsic camera parameters and have discarded the extrinsic camera parameters, i.e.…”

Motion Tracking, Retrieval and 3D Reconstruction from Video

“…Wei and Chai in [11] reconstruct human motion form video input data by employing physics-based modeling and minimal user interaction to annotate key intermediate frames for tracking error correction. Yasin et al [14] first detect and track video features by constructing a dictionary of features (DoF) using MSER and SURF feature detection techniques and make data-driven 3D reconstruction from these 2D detected feature sets. Dantone et al [15] estimate 2D human pose from still images by non-linear body part dependent joint regressor using two layered random forests.…”

“…For database sampling and construction of our knowledge base, we first extract 3D feature sets F 15 3D based on positions of hands, feet and head like Krüger et al have done in [17]. We project these 3D feature sets F 15 3D onto 2D plane at different viewing directions similar to Yasin et al [14] and as a consequence, we get 2D feature sets F 10 2D . To achieve previously mentioned goal, we have specified azimuth angles from 0 to 350 degrees with 10 degree step size and elevation angles from 0 to 90 degrees with step size 15 degree, as shown in Figure 2.…”

“…First, we extract features by utilizing these MSER and SURF feature detection techniques and develop a dictionary of features (DoF) along the lines of Yasin et al [14]. We stabilize feature detection process by 3D prior information retrieved from MoCap database by developing a kd-tree and the online lazy neighbourhood graph.…”