Tracking human motion with multiple cameras using articulated ICP with hard constraints

This thesis proposed a new ICP-based algorithm for tracking articulated skeletal model of a human body. The proposed algorithm takes as input multiple calibrated views of the subject, computes a volumetric reconstruction and the centerlines of the body and fits the skeletal body model in each frame using a hierarchic tree traversal version of the ICP algorithm that preserves the connection of the segments at the joints. The proposed approach uses the kinematic constraints and an Extended Kalman Filter to track the body pose. The first contribution is a new algorithm to find the skeletal points of a 3D volume. The algorithm using a slicing technique find the medial axis of a volume in a fast way using the graphic card processor and the texture units. This algorithm produce good results in quality and performance compared to other works in literature. Another contribution is the introduction of a new tracking strategy based on a hierarchical application of the ICP standard algorithm to find the match between a stick body model and a set of 3D points. The algorithm use a traversing version of ICP where also all the 3D points are weighted in such a way every limbs of the model can best fit on the right portion of the body. The application of these techniques shown the feasibility of the method and the performances obtained in terms of quality of estimate pose are comparable with other works in literature. The results presented here demonstrate the feasibility of the approach, which is is intended to be used in complete system for vision-based markerless human body tracking. Future work will aimed at optimizing the implementation, in order to achieve real-time performances.