Effective Multi-Model Tracking Using Team Actuation Models

Extended Abstract Robots need to track objects. We identify two kinds of tracking problems in which: (i) the tracker is static or does not actuate on the tracked object and (ii) the tracker actu-ates on the object. This thesis focuses on the latter problem. Tracking is performed by a robot executing specific tasks acting over the object being tracked, such as a Segway RMP soccer robot grabbing and kicking a ball. Object tracking efficiency completely depends on the accuracy of the motion model and of the sensory information. The motion model of the target becomes particulary complex in our case and highly dependent on the robot's actions. In addition, this thesis also considers the challenging environment of multiple team members actuating the object being tracked. In this case, the motion can become highly discontinuous and nonlinear. We assume that robots know their own actions, and robots in a team are collaborating according to the coordination plans. We claim that the knowledge in terms of the single robot control strategy and the multi-robot coordination plan can be a valuable source of information for tracking. Communication between robots is another interesting information source. This thesis contributes an approach to incorporate prior and dynamic knowledge into tracking. The expected outcome is a considerable improvement in tracking performance. There are several approaches incorporating some kind of prior knowledge related to the general problem of tracking under no actuation. For example, hard constraints on target position, speed or acceleration have been considered in tracking problems to improve tracking performance. This kind of information is in general simple and easy to represent as a truncated density. Another example is the situation where a number of targets are moving in formation, and there is a strong dependency between the individual sensor measurements, which provides valuable information on target behavior. In terrain-aided tracking, using the ground moving target indicator (GMTI), one may have some prior information of the terrain, road maps, and visibility conditions. A related tracking approach concerned with our problem of actuation over tracked objects, considers that a Sony AIBO robot actuates on a ball and models the motion using a fixed transition model between different single models. In my current research, I have successfully introduced a preliminary tracking filter that dynamically switches among target motion models as a function of one robot's actions. This tracker, implemented using a particle filter and a dynamic Bayes network …