Inferring Human Pose and Motion from Images
Abstract
As optical gesture recognition technology advances, touchless human computer interfaces of the future will soon become a reality. One particular technology, markerless motion capture, has gained a large amount of attention, with widespread application in diverse disciplines, including medical science, sports analysis, advanced user interfaces, and virtual arts. However, the complexity of human anatomy makes markerless motion capture a non-trivial problem: I) parameterised pose configuration exhibits high dimensionality, and II) there is considerable ambiguity in surjective inverse mapping from observation to pose configuration spaces with a limited number of camera views. These factors together lead to multimodality in high dimensional space, making markerless motion capture an ill-posed problem. This study challenges these difficulties by introducing a new framework. It begins with automatically modelling specific subject template models and calibrating posture at the initial stage. Subsequent tracking is accomplished by embedding naturally-inspired global optimisation into the sequential Bayesian filtering framework. Tracking is enhanced by several robust evaluation improvements. Sparsity of images is managed by compressive evaluation, further accelerating computational efficiency in high dimensional space.