View-Invariance in Visual Human Motion Analysis

Abstract

This thesis makes contributions towards the solutions to two problems in the area of visual human motion analysis: human action recognition and human body pose estimation. Although there has been a substantial amount of research addressing these two problems in the past, the important issue of viewpoint invariance in the representation and recognition of poses and actions has received relatively scarce attention, and forms a key goal of this thesis. Drawing on results from 2D projective invariance theory and 3D mutual invariants, we present three different approaches of varying degrees of generality, for human action representation and recognition. A detailed analysis of the approaches reveals key challenges, which are circumvented by enforcing spatial and temporal coherency constraints. An extensive performance evaluation of the approaches on 2D projections of motion capture data and manually segmented real image sequences demonstrates that in addition to viewpoint changes, the approaches are able to handle well, varying speeds of execution of actions (and hence different frame rates of the video), different subjects and minor variabilities in the spatiotemporal dynamics of the action. Next, we present a method for recovering the body-centric coordinates of key joints and parts of a canonically scaled human body, given an image of the body and the point correspondences of specific body joints in an image. This problem is difficult to solve because of body articulation and perspective effects. To make the problem tractable, previous researchers have resorted to restricting the camera model or requiring an unrealistic number of point correspondences, both of which are more restrictive than necessary. We present a solution for the general case of a perspective uncalibrated camera. Our method requires that the torso does not twist considerably, an assumption that is usually satisfied for many poses of the body. We evaluate the quantitative performance of the method on synthetic data and the qualitative performance of the method on real images taken with unknown cameras and viewpoints. Both these evaluations show the effectiveness of the method at recovering the pose of the human body.