PhySynth: A Physics-Based Synthetic UWB/IMU Data Generator for Training of ML-based Tracking

Abstract

Deep learning-based localization and tracking systems require large amounts of labeled training data, which are often difficult to acquire in challenging outdoor environments such as avalanche monitoring. To address this limitation, we present PhySynth, a physics-based synthetic data generator for synchronized ultra-wideband (UWB) ranging and inertial measurement unit (IMU) measurements. PhySynth is calibrated using real-world data collected during a cable-car experiment and models trajectory dynamics, sensor noise, measurement availability, and age of information effects within a unified framework. The generated traces preserve the physical characteristics of the underlying localization system while enabling systematic variation of UWB anchor deployments and channel conditions. Validation against independent descending and ascending trajectories demonstrates close agreement between real and synthetic measurements, achieving Kolmogorov–Smirnov statistics of 0.147 and 0.057, respectively. A comprehensive study further reveals that channel conditions have a greater impact on data fidelity than anchor density, with diminishing returns observed beyond three anchors. Finally, we evaluate the usefulness of the generated data through a case study using AoI-FusionNet, a deep learning-based tracking framework. The results show that PhySynth provides a practical and reproducible approach for generating realistic multimodal datasets to support the development, training, and benchmarking of machine learning-based localization and tracking systems.