Institutional Repository
Technical University of Crete
EN  |  EL



My Space

RFID localization with multistatic interrogation and neural networks

Papadopoulos Georgios

Full record

Year 2024
Type of Item Diploma Work
Bibliographic Citation Georgios Papadopoulos, "RFID localization with multistatic interrogation and neural networks", Diploma Work, School of Electrical and Computer Engineering, Technical University of Crete, Chania, Greece, 2024
Appears in Collections


The use of low-cost, batteryless radio frequency identification (RFID) tags for efficient communication and localization via signal backscattering, has been recently established in research and industry scenarios. This work draws inspiration from a recently proposed multistatic localization technique, that exploits the form of ellipses in the plane, using phase information and potentially, (elliptical) direction-of-arrival.A 2D localization scheme is introduced, that exploits neural networks (NNs) for improved accuracy. Specifically, a Deep Feed Forward Neural Network is developed that performs regression for tag location estimation from phase-based measurements as input, offered from a custom multistatic RFID interrogation setup.In the first part, traditional techniques are contrasted with neural networks trained with 1,000,000 samples and, through simulation comparisons across an 8x4 m2 area, a pronounced advantage for the latter is demonstrated for random tag positions. Then, a substantially improved setup is presented which, by utilizing only one extra antenna, yields a remarkable 76% reduction in mean absolute error (MAE), in the order of 2.48 cm, while median absolute error is below 1 cm. Contrary to bibliography, it is proven that this outcome is not due to phase ambiguity, but the determining factor is the increased difference in the values of the phase-based inputs within the tested area, thus giving the neural network more diverse information to be trained upon.In the second part, experimental data are used to evaluate the efficiency of the neural networks in two real-world scenarios, with tags being positioned up to 1.2 meters away from the antennas and 44 samples taken in total. It is remarked that the system was trained with simulated data, while the experimental data mentioned above were only used for testing. This time, the NNs perform in a comparable fashion to the traditional methods, in both experimental scenarios, with MAE ranging from 19 cm (in the first) to 27 cm (in the second scenario) and the median within a centimeter of each other, at approximately 8.5 cm.Two major conclusions can be drawn from this thesis; firstly that NNs trained exclusively with simulated data can perform reasonably well with experimental (real-world) measurements, comparably to state-of-the-art methods; secondly, there is plenty room for improvement, by using experimental data for training, perhaps with a tag in a moving robotic platform, and carefully selected reader antenna placements, as already shown in simulations.

Available Files