Skin-color based human tracking using a probabilistic noise model combined with neural network

Abstract

We develop a simple and fast human tracking system based on skin-color using Kalman filter for humanoid robots. For our human tracking system we propose a fuzzy and probabilistic model of observation noise, which is important in Kalman filter implementation. The uncertainty of the observed candidate region is estimated by neural network. Neural network is also used for the verification of face-like regions obtained from skin-color information. Then the probability of observation noise is controlled based on the uncertainty value of the observation. Through the real-human tracking experiments we compare the performance of the proposed model with the conventional Gaussian noise model. The experimental results show that the proposed model enhances the tracking performance and also can compensate the biased estimations of the baseline system. © Springer-Verlag Berlin Heidelberg 2006.