Ultrawideband Localization Improvement Based on LSTM Neural Networks

Barbara Morawska, Piotr Lipinski, Krzysztof Lichy

Abstract: This study investigates the enhancement of localization accuracy through the application of the Kalman filter and LSTM neural network. Experimental results reveal the successful elimination of outliers originating from noise and signal reflections, leading to a improvement from an initial error of 2.33 meters in raw samples to about 0.7 meters with LSTM and Kalman. The proposed method uses an LSTM neural network to fuse data from accelerometer, gyroscope, and ultrawideband 2D position measurements for position estimation. Despite offering comparable results, the Kalman filter sensitivity in selecting covariance matrix parameters is highlighted because minor changes are leading to significant declines in performance The LSTM filter, in contrast, demonstrates resilience to parameter tuning and provides accurate results without requiring specific domain knowledge or equations. However, the LSTM model applicability is limited to systems with accessible training and testing data. In contrast, the Kalman filter adaptability to various navigation problems featuring Gaussian error distributions remains a distinct advantage.