Navigation with low-sampling-rate GPS and smartphone sensors: A data-driven learning-based approach

摘要

For vehicle navigation, an inertial navigation system (INS) is often used to assist the global positioning system (GPS); but the positioning accuracy is well known to be sensitive to both the GPS sampling rate and inertial sensor error aggregation. As such, many practical navigation systems adopt both highend INS and high GPS sampling rate, which costs more for hardware on the one hand and demands excessive energy consumption and communications on the other hand. In this work, we propose a learning-based approach that is capable of providing the same level of positioning accuracy but using merely cheap, low-end mobile inertial sensors and integrated GPS with very low sampling rate. The proposed approach consists of a Kalman filter and two multilayer neural networks to realize 3-dimensional, high-precision vehicle navigation. The key to the impressive outcome is the combination of the classic statespace Kalman filter with a new data-driven machine learning mechanism. The proposed approach is evaluated with real trajectory data collected on campus of CUHKSZ, and the results show that it can achieve a positioning error around 8m when the GPS sampling interval is as large as 30s (30 times of nowadays regular sampling interval). Apart from this, the proposed approach also demonstrates good adaptiveness to changing trajectories.