Spatially Grounded Multi-Hypothesis Tracking of People

Abstract

People tracking is an important yet chal- lenging task for mobile robots operating in populated environments and interacting with humans. What makes this problem dicult is that human behavior is complex and hard to predict. However, motion of peo- ple, the rate at which people appear and where they appear are not random but strongly place-dependent and follow patterns that are engendered by the envi- ronment. In this paper we make use of such informa- tion for the purpose of people tracking. Concretely, we learn a probabilistic representation, called spatial aordance map , to spatially ground activity events acquired by observing people in the environment. This representation is a non-homogeneous spatial Poisson process for which we derive expressions for life-long Bayesian learning. We show how the spatial aor- dance map can be used to compute refined probability distributions over hypotheses in a multi-hypothesis tracker and to make better, place-dependent predic- tions of human motion. In experiments with real data from a laser range finder, we demonstrate how both extensions lead to more accurate tracking behavior. The system runs in real-time on a typical desktop computer. I. Introduction As robots enter more domains in which they interact and cooperate closely with humans, people tracking is becoming a key technology for several areas in robotics such as human-robot interaction, intelligent cars or hu- man activity understanding. In this paper we pursue the approach to learn and represent human spatial behavior for improved people tracking. Human activity is strongly place-dependent. By learning a spatial model that represents activity events in a global reference frame and on large time scales, the robot acquires place-dependent priors on human behavior. As we will demonstrate, such priors can be used to better hypothesize about the state of the world (that is, the state of people in the world), and to make place- dependent predictions of human motion that better re- flect how people are using space. Concretely, we propose a non-homogeneous spatial Poisson process to represent the spatially varying distribution over relevant human activity events for people tracking. The representation, called spatial aordance map , holds space-dependent Poisson rates for the occurrence of track events such as creation, confirmation or false alarm. The map is then incorporated into a multi-hypothesis tracking framework using data from a laser range finder.