Conditional particle filters for simultaneous mobile robot localization and people-tracking

Abstract

Presents a probabilistic algorithm for simultaneously estimating the pose of a mobile robot and the positions of nearby people in a previously mapped environment. This approach, called the conditional particle filter, tracks a large distribution of person locations conditioned upon a smaller distribution of robot poses over time. This method is robust to sensor noise, occlusion, and uncertainty in robot localization. In fact, conditional particle filters can accurately track people in situations with global uncertainty over robot pose. The number of samples required by this filter scales linearly with the number of people being tracked, making the algorithm feasible to implement in real-time in environments with large numbers of people. Experimental results illustrate the accuracy of tracking and model selection, as well as the performance of an active following behavior based on this algorithm.