Deep learning and data augmentation techniques for indoor environment characterization via UWB technology

Ultra-wideband is a radio-based communication technology for short-range use and fast and stable transmission of data. Its main characteristics are extremely large bandwidth, very low signal power density, robustness against fading, and low cost. These features make ultra-wideband suitable for indoor localization applications. Referring to the positioning accuracy, it is around five and twenty-five times better than adopting Bluetooth Low Energy (BLE) beacons or Wi-Fi, respectively. Despite these promising results, ultra-wideband localization accuracy robustly degrades when moving to non-ideal conditions, including Non-Line-Of-Sight and the presence of dynamic environmental factors. This work aims at characterizing an indoor environment via channel impulse responses retrieved by ultra-wideband technology, opening the doors to future research in terms of sensor fusion techniques for improving indoor localization accuracy and indoor channel characterization. As a first step of this work, a sufficiently large dataset is collected in a typical indoor scenario, including static, dynamic, Line-Of-Sight and Non-Line-Of-Sight conditions. In this regard, data augmentation techniques are exploited to enlarge and enrich the dataset. Then, a deep neural network in the context of deep learning is chosen to classify a certain number of positions from the channel impulse responses. In parallel, a class activation map algorithm is considered to provide model explainability, a crucial missing aspect in many deep learning projects. Results show channel impulse responses carry enough information about the multipath delays in an indoor environment, at least in the static case. When a realistic dynamic environment is considered, indoor environment characterization becomes harder. As well known, a very large dataset is needed when dealing with deep learning models. In this regard, data augmentation comes in hand, leading to better results in terms of network generalization. The last chapter draws the conclusion of the presented work, including suggestions for future research and objectives to be pursued.