MCST: An Adaptive Tracking Algorithm for High-Speed and Highly Maneuverable Targets Based on Bidirectional LSTM Network

Abstract

The abrupt and intricate dynamics of high-speed and highly maneuverable targets pose significant challenges in radar target tracking. The precise monitoring of such targets has long been a daunting task in radar technology. Traditional tracking algorithms are hindered by their dependence on precise prior assumptions regarding target motion states. Any inaccuracies in these assumptions can result in diminished tracking accuracy or even divergence. In contrast, neural network-based methods offer a promising alternative, as they circumvent the need for numerous prior assumptions and possess inherent advantages for tracking targets. In this article, we extensively study the variation characteristics of residuals when the target maneuvers and incorporate the concept of unscented transformation into the algorithm. Building upon this foundation, we introduce an adaptive tracking model for high-speed and highly maneuverable targets using neural networks. To begin with, a comprehensive trajectory dataset, tailored specifically to the characteristics of swift and agile targets, has been meticulously crafted. Subsequently, a neural network named maneuver compensation strong tracker (MCST), built upon bidirectional long short-term memory (Bi-LSTM), has been devised to track these targets effectively. The MCST consists of two primary components: the predictor and the updater. At its heart lie the maneuver compensation unit (MCU), Bi-LSTM, and a dual-level attention module. This model excels at adapting the state vector to accommodate target maneuvers, effectively capturing uncertainties inherent in both the state and observation vectors. To validate its effectiveness, we simulate multiple scenarios involving swift and agile target movements, conducting a comparative analysis against several cutting-edge algorithms. Our findings unequivocally underscore the superiority of our proposed approach.