Object Tracking Incorporating Transfer Learning into an Unscented Kalman Filter

Abstract

We present a novel algorithm designed to address the challenges posed by mismatched intensity of the noise in sensors performing object tracking. Our objective is to enhance the accuracy of estimation in the tracking domain, particularly in scenarios where reliable measurements are difficult to obtain due to environmental conditions affecting a specific sensor. To accomplish this, we propose a framework that integrates transfer learning techniques into an unscented Kalman filter (UKF). We introduce an additional step to model and learn the parameters of predicted observations in a learning domain at each time step. By incorporating the learned knowledge from the learning domain into the filtering process of the tracking domain, our approach demonstrates significant improvements in tracking accuracy. Through extensive simulations, we validate the effectiveness of our proposed algorithm in terms of tracking accuracy, comparing its performance to that of the traditional isolated UKF.