Trajectory Prediction Using Gaussian Process Regression : Estimating Three Dynamical States Using Two Parameters

Sammanfattning

In this thesis a Gaussian process regression (GPR) model and a Kalman filter (KF) model were developed and applied to a trajectory prediction problem. The main subject of the thesis is GPR, where the intended purpose of the KF is to compare it to the GPR model. The input data for the models consists of two noisy spherical angle coordinates of a moving target relative to a moving guided projectile. In order to perform trajectory predictions the models need to estimate the distance between the target and guided projectile since there are only two coordinates available and an estimation of three coordinates is desired. The distance estimation was done by a Low Speed Approximation. The trajectories investigated were harmonic-exponential, exponential-spiral and linear. The results showed issues with the hyperparameters of the GPR model which may be related to the preprocessing of the trajectory data. However, the GPR model did outperform the KF model when there was acceleration, despite the issues with the hyperparameters. The KF model outperformed the GPR model when the target trajectory behaved linearly. The results indicate that GPR has potential as a trajectory prediction algorithm.