Range-Based People Detection and Tracking for Socially Enabled Service Robots

With a growing number of robots deployed in populated environments, the ability to detect and track humans, recognize their activities, attributes and social relations are key components for future service robots. In this article we will consider fundamentals towards these goals and present several results using 2D range data.We first propose a learning method to detect people in sensory data based on a set of boosted features. The method largely outperforms the state of the art that typically relies on hand-tuned classifiers. Then, we present a person tracking approach based on the detection and fusion of leg tracks. To deal with the frequent occlusion and self-occlusion of legs, we extend a Multi-Hypothesis Tracking (MHT) approach by the ability to explicitly reason about and deal with adaptive occlusion probabilities. Finally, we address the problem of tracking groups of people, a first step towards the recognition of social relations. We further extend the MHT approach by a multiple model hypothesis stage able to reflect split/merge events in group formation processes. The proposed extension is mathematically elegant, runs in real-time and further allows to accurately estimate the number of people in each group. The article concludes with prospects and suggestions for future research.