On the Implementation of Temporal Fusion Transformers for Target Recognition and Tracking

Time series forecasting is advancing across many domains, offering promising performance. However, efforts to apply forecasting techniques to direction of arrival estimation and array signal processing are limited. This paper explores the implementation of temporal fusion transformers (TFTs) for target localization and tracking, addressing the shortcomings of traditional methods and demonstrating the potential of time series forecasting in wireless communication challenges. The TFT architecture is adapted to forecast the positions of multiple targets over various horizons, utilizing its advanced temporal modeling and dynamic feature selection capabilities. Simulation results underscore the effectiveness of TFT in enhancing target localization, providing a robust, interpretable, and scalable solution to critical issues in array signal processing. Comparative evaluations with state-of-the-art neural networks prove that TFT combines both accuracy and computational efficiency.