On-device Classification of Human-induced Non-Line-of-Sight on Chest-Worn Ultra-wideband Wearables

Accurate indoor localization has become increasingly essential across domains that play a crucial role in modern daily life, such as IoT smart buildings, healthcare and safety, industrial operations, and emergency response, where real-time information about a person’s location or movement is fundamental to system effectiveness. Traditional positioning technologies, however, exhibit notable limitations when applied in such demanding contexts. Satellite-based systems (e.g., GPS, Galileo) suffer from severe signal attenuation and high latency indoors, while short-range options such as Bluetooth Low Energy and Wi-Fi generally provide insufficient accuracy. Ultra-wideband (UWB) has instead emerged as a leading candidate for precise indoor positioning, offering decimeter-level accuracy with energy efficiency and timely updates. Still, despite these advantages, practical UWB deployments—which are inherently human-centric and typically involve low-power devices worn on or carried by the user—frequently suffer significant performance degradation when the body obstructs the line-of-sight (LOS) radio propagation. These human-induced occlusions, or human non-line-of-sight (HNLOS) conditions, may lead to meter-scale estimation errors, heavily reducing the reliability of UWB systems in day-to-day deployment. Although UWB performance has been widely studied, the specific effects of HNLOS on the technology remain inadequately characterized, and the efficiency of existing mitigation techniques is not well established. In practice, a common strategy is to first classify the channel condition—presuming the link state (LOS vs. HNLOS)—before applying subsequent correction or filtering methods. Despite its importance, however, trade-offs of state-of-the-art classification approaches remain insufficiently examined. Particularly, approaches suitable for low-power wearables are often under-evaluated, leaving a considerable gap between the demands of real-world deployments and current research outcomes. The overarching goal of this thesis is to enhance the robustness of human-centric UWB positioning systems, with an emphasis on battery-powered chest-worn devices, a simple and non-intrusive arrangement that closely reflects everyday usage yet presents stronger HNLOS challenges than other body placements. To achieve this, the thesis systematically investigates how human-induced occlusion affects UWB distance estimation (ranging) and derives practical, lightweight countermeasures suitable for on-device implementation, ultimately supporting dependable human-centric ranging and localization across diverse scenarios. Motivated by this objective, the thesis consists of three stages. First, we revisit the largely neglected manufacturer-provided link indicators (LIs), which perform basic NLOS detection. Establishing the capabilities of these intrinsic indicators is a necessary first step toward understanding how far lightweight, radio-native mechanisms alone can support HNLOS detection in chest-worn UWB systems. Experimental results demonstrate that, in controlled setups, LIs can reliably detect human occlusions and thus enable more effective ranging strategies. Next, we broaden the analysis to dynamic and realistic scenarios, where we conduct the first systematic side-by-side evaluation of intrinsic LI-based methods, lightweight machine learning (ML) classifiers, and newly introduced hybrid LI–ML approaches. By jointly exploiting the indicators and learning-based models, these hybrid schemes represent a novel contribution to the literature. Their evaluation enables a detailed comparison of strengths, limitations, and complementarities across multiple performance dimensions. The resulting analysis yields new empirical insights and actionable recommendations that can be adapted to a wide range of practical requirements in evolving UWB-based applications operating with low-power chest-worn devices. Finally, we extend the analysis to a newer generation of UWB radios (DW3000), which incorporates enhanced physical-layer mechanisms relevant to operation under HNLOS conditions. Through a focused experimental study, we examine the behavior of LIs, lightweight ML classifiers, and their sensitivity to configuration choices on DW3000. Prior DW1000-based results are referenced to provide qualitative context for assessing cross-platform relevance. Together, these findings offer practical guidance for designing robust, HNLOS-aware UWB ranging systems on contemporary wearable hardware. Overall, this thesis offers an in-depth characterization of HNLOS effects with chestworn UWB wearables and introduces a set of lightweight, systematically validated strategies to mitigate their impact. Beyond addressing current gaps, the findings establish the groundwork for developing more resilient UWB ranging and localization schemes, and classification pipelines suited to realistic conditions, varied platform architectures, and the growing variety of human-centric applications.