Deep Learning for Robotic RFID-Localization

摘要

This paper presents different deep learning architectures that successfully solve the problem of localization of RFID tags by a single antenna on top of a robot in 2D space. Phase measurements, collected by an RFID reader on top of a moving robot, combined with the corresponding antenna-positions, are properly structured, as proposed herein, to form the input vector of different Multilayer Machine Learning Networks. The proposed architectures are originally tested in simulated data, suffering by zero-mean Gaussian noise, achieving centimeter-level accuracy, verifying the soundness of the proposed approach. Subsequently, the models are tested on experimental data involving hundreds of RFID tags and experiments, dividing the dataset into two disjoint sets, the training set and the test set. The proposed deep learning solutions outperformed a maximum-likelihood estimator, since the latter assumes only the effects of the Line-Of-Sight link, while Neural Networks (NNs) identify patterns resulting from all contributions. To the best of our knowledge, this is the first paper that proposes a way to restructure phase measurements collected by a moving robot in a manner that can then be solved by different Machine Learning architectures. The proposed methods provide a scalable and computationally efficient alternative for real-time RFID localization tasks, which can be expanded in 3D space.