Mutual information based data selection in Gaussian Processes for 2D laser range finder based people tracking

In general, a model to describe human motion patterns should have a capability of enhancing tracking performance even with long term occlusions. One way of effectively learn these patterns is to apply Gaussian Processes (GP). However, with the increase of the amount of training data with time, the GP becomes computationally expensive. In this work, we have proposed a Mutual Information (MI) based technique along with the Mahalanobis Distance (MD) measure to keep the most informative data while discarding the least informative data. The algorithm is tested with data collected in an office environment with a Segway robot equipped with a laser range finder. It leads to more than 90% data reduction while keeping the limit of Average Route Mean Square Error (ARMSE). We have also implemented a GP based Particle filter tracker for long term people tracking with occlusions. The comparison results with Extended Kalman Filter (EKF) based tracker shows the superiority of the proposed approach.