ScholarWorks@성균관대학교

초록

Indoor localization for Unmanned Aerial Vehicles (UAV) is a fundamental prerequisite for autonomous operations in environments where GPS is unavailable. Although Ultra-Wideband (UWB) with Time Difference of Arrival (TDoA) provides a reasonable localization service to those UAVs, its accuracy is often limited by conventional filtering algorithms like the Kalman Filter (KF). The performance of these filtering algorithms is fundamentally limited by their dependency on simplified, predefined motion models, which are inadequate for tracking the agile and nonlinear movements of a UAV. To overcome this limitation, this paper introduces a deep learning framework centered on a Long Short-Term Memory (LSTM) network as a direct replacement for the Kalman Filter. The model is trained on a large-scale synthetic dataset, which includes sequences of TDoA and Received Signal Strength Indicator (RSSI) values from a UAV performing diverse flight patterns. By learning directly from the raw sensor stream, our model maps the complex temporal dependency to the UAV's positional coordinates without reliance on an explicit physical model. The empirical results from our simulations confirm that our data-driven approach yields a substantial improvement in localization accuracy and robustness over the traditional KF approach, presenting a viable and superior alternative for challenging indoor navigation tasks.

키워드