Adaptive Sampling and Machine learning methods for Efficient Meta-model and application to Real-Time Target Tracking

Abstract

This paper explores two main aspects of metamodeling. In the first part, it introduces a method of creating accurate surrogate models using the least number of data points that contain necessary information, as opposed to the entire dataset. This is achieved through an adaptive sampling technique known as Multi-Point Multi-Objective Infill Sampling Criteria (MPMO ISC). Based on the Expected Improvement (EI) formula from Efficient Global Optimization (EGO), this method focuses on two objectives: one aimed at the exploration of minimum points and the other at enhancing the global accuracy considering uncertainties. The second part delves into research on future position prediction models, an integral component of target tracking. This research proposes future position prediction models for target tracking using databased techniques like Gaussian Process Regression (GPR), Long Short Term Memory (LSTM), and Convolutional Neural Network (CNN), and compares their accuracy against traditional motion model-based methods. The significance of this study lies in the ability of the MPMO ISC adaptive sampling technique to create high-accuracy surrogate models with a reduced amount of data, thereby enhancing data utilization efficiency in the industry, leading to cost and time savings. Additionally, the use of data-based methods not only predicts future trajectories from past trajectory history but also learns from operational environment and obstacle information, aspects typically not considered in conventional target tracking methods, thereby improving the accuracy of future predictions.