Computer vision-based human motion capture using kinematics constraints / Computer vision registrering af humane bevægelser vha. kinematiske begrænsninger

In man-machine-interfaces there is a great need for interaction methods more natural to humans, e.g. via speech and body language. The latter is based on a computer capturing the movements of the individual body parts and recognizing their meaning. In this project computer vision is utilized to investigate the capturing problem. The key approach is to have a geometric model of the articulated body parts. The model is used to predict possible "next" configurations given the past configurations. The predicted configurations are compared with the image measurements, and the true configuration of the human body is captured. To optimize this process detailed kinematics constraints related to the articulated body parts are introduced to limit the search space of possible configurations. . The limited search space is, however, still too large for a brute force search and therefore a probabilistic approach, in the form of a sequential Monte Carlo method, is adapted to identify the most likely configuration. In 2003 the project was concluded with a Ph.D. degree being awarded to Thomas B. Moeslund. (Thomas B. Moeslund, Erik Granum)