Gaussian Processes-BayesFilters with Non-Parametric Data Optimization for Efficient 2D LiDAR Based People Tracking

A model for expressing and describing human motion patterns must be able to improve tracking accuracy. However, Conventional Bayesian Filters such as Kalman Filter (KF) and Particle Filter (PF) are vulnerable to failure when dealing with highly maneuverable targets and long-term occlusions. Gaussian Processes (GP) is then used to adapt human motion patterns and integrate the model with Bayesian Filters. In GP, all samples in training phase need to be included and periodically, new samples will be added into training samples whenever it is available. Larger amount of data will increase the computational time to produce the learned GP models due to data redundancies. As a result, Mutual Information (MI) based technique with Mahalanobis Distance (MD) is developed to keep only the informative data. This method is used to process data which is collected by a robot equipped with a LiDAR. Experiments have demonstrated that reducing data does not raise Average Root Mean Square Error (ARMSE) considerably. EKF, PF, GP-EKF and GP-PF are utilised as a tool for tracking people and all techniques have been analyzed in order to distinguish which method is more efficient. The performance of GP-EKF and GP-PF are then compared to EKF and PF where it proved that GP-BayesFilters performs better than Conventional Bayesian Filters. The proposed approach has reduced data points up to more than 90\% while keeping the ARMSE within acceptable limits. This data optimization technique will save computational time especially when deal with periodically accumulative data sets. Comparing on four tracking methods, both GP-PF and GP-EKF have achieved higher tracking performance when dealing with highly maneuverable targets and occlusions.