Human body orientation estimation in multiview scenarios

Estimation of human body orientation is an important cue to study and understand human behaviour, for different tasks such as video surveillance or human-robot interaction. In this paper, we propose an approach to simultaneously estimate the body orientation of multiple people in multi-view scenarios, which combines a 3D human body shape and appearance model with a 2D template matching approach. In particular, the 3D model is composed of a generic shape made up of elliptic cylinders, and a 3D colored point cloud (appearance model), obtained by back-projecting pixels from foreground images onto the geometric surfaces. In order to match the reconstructed appearance to target images in arbitrary poses, the appearance is re-projected onto each of the different views, by generating multiple templates that are pixel-wise, robustly matched to the respective foreground images. The effectiveness of the proposed approach is demonstrated through experiments in indoor sequences with manually-labeled ground truth, using a calibrated multi-camera setup.