An ultrasonic/RF GP-based sensor model robotic solution for indoors/outdoors person tracking

An non-linear Bayesian regression engine for robotic tracking based on an ultrasonic/RF sensor unit is presented in this paper. The proposed system is able to maintain systematic tracking of a leading human in indoor/outdoor settings with minimalistic instrumentation. Compared to popular camera based localization system the sonar array/RF based system has the advantage of being insensitive to background light intensity changes, a primary concern in outdoor environments. In contrast to single-plane laser range finder based tracking the proposed scheme is able to better adapt to small terrain variations, while at the same time being a significantly more affordable proposition for tracking with a robotic unit. A key novelty in this work is the utilisation of Gaussian Process Regression (GPR) to build a model for the sensor unit, which is shown to compare favourably against traditional linear triangulation approaches. The covariance function yield by the GPR sensor model also provides the additional benefit of outlier rejection. We present experimental results of indoors and outdoors tracking by mounting the sensor unit on a Garden Utility Transportation System (GUTS) robot and compare the proposed approach with linear triangulation which clearly show the inference engine capability to generalise relative localisation of human and a marked improvement in tracking accuracy and robustness.