Non-line-of-sight object localization using multipath wave propagation - a machine learning approach

Abstract

Abstract This thesis report examines the application of machine learning models for detecting and positioning of non-line-of-sight (NLOS) and line-of-sight (LOS) targets using multi-path wave propagation. The report commences with exploring the data simulation process, emphasizing the distribution of scenarios and the incorporation of NLOS timeframes. The simulation aims to include variations in the dataset, enabling the models to understand the problem comprehensively. Next, the report focuses on machine-learning positioning, comparing the performance of two models: the U-Net and a Convolutional Neural Network (CNN). Both models achieve similar accuracy, but U-Net outperforms CNN in terms of mean squared error (MSE) and average Euclidean distance (AED). The report evaluates the models’ performance on different scenario labels, scenario types, and numbers of targets. Finally, the report examines object tracking using machine learning measurements. The tracking algorithm demonstrates high performance, achieving low average Euclidean distance (AED) and absolute errors in position, velocity, heading, and turn-rate state estimation. The report concludes that machine learning models show promise in identifying and positioning NLOS and LOS targets, and object tracking using machine learning measurements yields satisfactory results.