People Tracking with Heterogeneous Sensors using JPDAF with Entropy Based Track Management

In this paper we study the problem of tracking an arbitrary number of people with multiple heterogeneous sensors. To solve the problem, we start with a Bayesian derivation of the multiple-hypothesis tracking (MHT), and, under certain assumptions, we arrive to the joint probabilistic data association filter (JPDAF). In their original derivation, both the MHT and JPDAF assume a multiple sensor scenario which enables us to fuse the sensors measurements by asynchronously updating the tracking filters. To solve the data association problem, instead of using the optimal MHT with complex hypothesis branching, we choose the JPDAF since we are interested only in local observations by a mobile robot for people detection, tracking, and avoidance. However, the JPDAF assumes a constant and known number of objects in the scene, and therefore, we propose to extend it with an entropy based track management scheme. The benefits of the proposed approach are that all the required data come from a running filter, and that it can be readily utilized for an arbitrary type of filter, as long as such a strong mathematical principle like entropy is tractable for the underlying distribution. The proposed algorithm is implemented for the Kalman and particle filter, and the performance is verified by simulation and experiment. For the simulation purposes, we analyze two generic sensors, a location and a bearing sensor, while in the experiments we use a laser range scanner, a microphone array and an RGB-D camera.