Non-parametric data optimization for 2D laser based people tracking

Abstract

Generally, a model on describing human motion patterns should have an ability to enhance tracking performance particularly when dealing with long term occlusions. These patterns can be efficiently learned by applying Gaussian Processes (GPs). However, the GPs can become computationally expensive with increasing training data with time. Thus, with the proposed data selection and management using Mutual Information (MI) and Mahalanobis Distance (MD)approach, we have be able to keep the necessary portion of informative data and discard the others. This approach is then experimented by using the measurements of horizontal 2D scan of public area of our research centre with a stationary laser range finder. Experimental results show that even 90% reduction of data did not contribute to significantly increased Root Mean Square Error (RMSE). Implementation of Gaussian Process — Particle filter tracker for people tracking with long term occlusions produces a remarkable tracking performance when compared to Extended Kalman Filter (EKF) tracker.