Highly Maneuvering Target Tracking with a Transformer Network

The aim of this paper is the tracking of highly maneuverable radar targets using Deep Networks. Numerous statistical methods are used in the literature to guarantee good results in tracking moving objects, such as the Extended Kalman Filter (EKF) and Interacting Multiple Models (IMM). These tracking tools are based on a priori models of the target’s behavior in its environment, with a finite number of elementary movements. The use of Deep Learning makes it possible to push back these modeling limits by training neural networks with very large numbers of parameters, using huge sets of practical trajectories. Neural networks are thus able to approach the target’s actual kinematics more precisely, particularly in the case of abrupt maneuvers that are difficult for traditional methods to track, and on which this paper focuses. Therefore, it uses the latest Deep Learning concept, named Transformer, based on attention mechanism which enables a thorough consideration of the target’s temporal sequence and maneuvers. More specifically, the TrNET network which has performed well on the popular LAST dataset [1], is applied and extended, with a new formulation of the input space, and an application to prediction estimates. For this, a more realistic database is generated; and performances are evaluated for a set of practical trajectories based on GPS measurements.