Kinematic Tracking and Activity Recognition Using Motion Primitives

In recent years, there has been an increasing interest in monocular human tracking and activity recognition systems, due to the large amount of applications where those features can be used. Standard algorithms are not practical to employ for human tracking due to the computational cost that arises from the high number of degrees of freedom of the human body and from the ambiguity of the images obtained from a single camera. Constraints in the configuration of the human body can be used to reduce its complexity. The constraints can be deduced from demonstration, based on the human performance of different activities. A human tracking system is developed using this kind of constraints and then evaluated. The fact that the constraints are based on activities allows, while doing the tracking, the inference of the activity the human is performing.

This thesis was done at KTH and in collaboration with the Computer Science Department at Brown University, in Providence, Rhode Island, USA.