Machine learning based approach to Ultra-WideBand (UWB) indoor localization

Providing an accurate, reliable and low-cost estimate of indoor positioning remains an active area of research, despite the availability of popular localization techniques like Acoustic Systems, Infrared Systems, etc. The aim of this thesis is to introduce a new methodology for indoor localization combining Ultra-WideBand (UWB) technology with Artificial Intelligence (AI). Albeit UWB is not a new technology, it is now being revitalized and employed for wireless connections over short distances. Many companies such as Intel, Xiaomi, Sony, Samsung, Apple and Bosh claim that this technology could prove more successful than Bluetooth as it is faster, cheaper, less power consuming and more secure. UWB is a short-range wireless communication protocol (like Wi-Fi or Bluetooth) using short radio pulses with large bandwidth. The resulting radio waves can pass through walls and other obstacles and do not interfere with different radio signals, such as those from cellular telephones. The only main limitation could be the short range, which could be easily overcome using multiple well-positioned receivers. The localization that has been performed in this thesis uses the Channel Impulse Response (CIR) shape to understand in which subarea of the environment the antenna is positioned. The tracking is achieved by classification using Machine Learning (ML). Indeed, when the two antennas communicate with each other, what the receiver gets is a composition of the direct signal with all its reflections. The process consist in analyzing at which time these signals reach the target (or Tag), and, based on the reflections delay (which are dependent on the surrounding), estimating the position in the indoor environment. The main steps performed in the thesis development are the collection of multiple datasets, the analysis and post-processing of the collected data, and the identification of the neural network able to offer the best classification performances. The final selected network shows high validation accuracy when all the datasets have been combined with each other and successively split into training and validation samples. Conversely, performance deteriorates when the datasets are kept separate and used individually as training and validation sets.