UAV Localization with Multipath Fingerprints and Machine Learning in Urban NLOS Scenario

Abstract

With the increasing number and application of unmanned aerial vehicle (UAV) in urban areas, positioning of UAV has become one of the key technologies for maintaining city security and managing airspace resources. Time of arrival (TOA) based location technology is widely used for its high precision, but its performance may suffer from strong multipath and NLOS propagation in urban scenario. However, the NLOS components may also be useful for positioning if the propagation path can be analyzed in a map model. In this paper, a multipath fingerprint dataset for an urban area is built by ray tracing simulation. Based on this dataset, we propose a two-stage localization method on machine learning framework. Firstly, in the stage of coarse positioning, the Random Forest (RF) algorithm is applied to determine which region the UAV is located in. Then, in the fine positioning stage, a neural network is trained to predict the specific location within the region. The simulation results in a $600 \times 600\ m^{2}$ region show that 90% of the positioning error of this method is less than 16m.