Abstract

companion system is supposed to adapt its behavior to the current situation. The necessary dynamic model of the environment is provided by a multi object tracking algorithm which estimates the number of persons in the environment and their positions. The tracking algorithm can be divided into two parts, the prediction step and the innovation step. In the prediction step, the expected state of an object at the next measurement time is predicted based on a motion model. In the innovation step, the received measurements are used to update the predicted states according to Bayes formulas. In order to achieve closed-form expressions, multi object tracking algorithms assume that the objects move independent of each other. Especially in case of low measurement rates or short-time occlusions combined with a high object density, it is obvious that this assumption is not fulfilled any more. Thus, the dependence between the persons has to be integrated into the tracking algorithms.

In this talk the advantages and disadvantages of several standard multi object tracking algorithms will be discussed. Afterwards an approach to adapt the uncertainty about an objects state to the current situation as well as an improved motion model for occluded persons are presented.