چكيده به لاتين
The prolonged stay of users in indoor environments for object tracking or person localization has turned navigation and guidance within these enclosed spaces into a challenging process. According to conducted research, humans spend approximately 70-90% of their life and work time in indoor environments; therefore, developing systems to provide suitable services to users in confined spaces is deemed essential. User and device localization has extensive applications in various sectors, including healthcare, industry, crisis management, building management, surveillance, and other diverse fields.
Various technologies exist in indoor positioning systems. In this thesis, Ultra-Wideband (UWB) technology has been investigated due to its high precision in indoor positioning. UWB technology has been recognized as having significant potential in constructing accurate Indoor Positioning Systems (IPS). However, indoor environments are filled with objects and individuals, which may cause signals to be reflected by obstacles. In comparison to Line of Sight (LOS) signals, the signal path delay in Non-Line of Sight (NLOS) signals introduces positive distance errors and consequently positioning inaccuracies.
To mitigate the impact of NLOS conditions on positioning, this thesis aims to initially use deep learning networks with channel impulse response data as input, without any prior knowledge of the environment, to accurately distinguish between LOS and NLOS conditions. The classification results are also compared with two other references using the same dataset. After identifying NLOS conditions, different regression networks are presented to estimate positions under two scenarios: 1) considering NLOS conditions, and 2) disregarding NLOS conditions.
Simulation results in the first part, the classification of NLOS/LOS signals, indicate that proposed Convolutional Neural Networks (CNN) and Inception networks outperform conventional neural network methods (advanced neural networks with different hidden layers and neurons) in accurately distinguishing between these two environments. Additionally, in the positioning section using regression algorithms, in both scenarios, the Support Vector Machine (SVM) machine learning network exhibits lower error in estimating tag positions compared to itself without considering NLOS conditions.
