RFID Localization with Multistatic Interrogation and Deep Neural Networks

Deep neural networks (DNN) have become a widely used tool for improving efficiency in various problems; however, the time-consuming training process and the need for large amounts of labeled data often limit their practicality. This work presents a deep feed-forward neural network, designed for real-time tag localization, which addresses these challenges by lever-aging easily-generated simulation data for training. The network uses phase-based feature inputs, and performance is evaluated using simulation data, as well as real-world measurements from a custom multistatic RFID interrogation testbed. In the first part, state-of-the-art techniques are contrasted to neural networks in a simulated environment, and 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. This outcome is attributed to the increased distance between antenna pairs and not to phase ambiguity. In the second part, experimental data are used to evaluate the efficiency of DNNs in two real-world immobile scenarios, using a custom multistatic experimental setup with software-defined radios (SDR). Even though DNNs are trained exclusively on simulated data, they are found to perform in a comparable fashion to the conventional methods, with MAE ranging from 19 to 27 cm for the first and second scenarios, respectively; similarly, the median absolute error is found within 1 cm between the two scenarios, at approximately 8.5 cm. These results suggest that DNNs offer tremendous potential for further localization improvements, especially if more real-world measurements, possibly through future automated (e.g., robotic) processes, are blended into the training process.