2D human pose estimation and tracking in non-overlapping cameras

Abstract

This chapter will discuss approaches to 2D human pose estimation and tracking in a non-overlapping camera network. It will demonstrate the limitations of current approaches and suggest strategies to overcome them. In particular, computational intractability due to high dimensional limb space, violation of articulation constraints, and view-point dependence. The chapter is divided into threemajor components; namely, search space reduction, pose validation, and view-invariant pose tracking in a non-overlapping camera network. Firstly, we present approaches for search space reduction, such as Kinematic Tree based sub-region selection for each limb, Mean-Shift based maxima search on the likelihood surface, and temporal based reduction of search in parameter space. Secondly, we devise a PCA based Pose Validation strategy to prune out anatomically incorrect hypotheses. Thirdly, we propose to incorporate articulation constraints while keeping the problem tractable. Finally, we enable view-invariance through the fusion of only two pose detectors and an articulated skeleton tracker.