Video segmentation for markerless motion capture in unconstrained environments

Abstract

Segmentation is an important first step in many computer vision applications. The identification of key regions within an image or video allows for a higher level analysis of the media content. This thesis explores the application of this low level process to the monitoring of human performance. In such a context, a proposed segmentation algorithm would be required to impose a minimum of constraints in order to assure the integrity of the performance and the proper transfer of key data to higher level analysis components. Classical approaches to the segmentation problem either make assumptions on the content of the media or impose unreasonable constraints on their targets and environments. In doing so, the integrity of performance measurements cannot be assured and semantic interpretation therefore becomes skewed. The method presented within this thesis allows for unconstrained environments by using a spatiotemporal colour-texture segmentation routine that represents the media content as a set of homogenous texture regions. The routine is assisted by a non-parametric clustering algorithm in order to produce an initial colour-texture representation. The regions obtained from this algorithm undergo a merging and tracking process in order to produce a final segmented representation of a target. Experimental results reveal that the system is robust for complex environments and provides several advantages over current segmentation processes.