A cross-and-dot-product neural network based filtering for maneuvering-target tracking

Maneuvering-target tracking is a critical topic for target tracking. However, suffering from unknown and changeable target movements, tracking performance of common algorithms still leaves large rooms for improvement. Recently, neural network-based algorithms are proposed to greatly improve the tracking performance, while their generalization ability remains a problem. That is, when the target states distribute beyond the range of training set, their tracking errors will dramatically increase. To conquer this problem, this paper proposes a Cross-and-Dot-Product neural network based filtering algorithm. Specifically, a network is designed to estimate transition matrices with turn rate information of moving targets by calculating the cross-product and dot-product data. Thereby, our network can correctly understand the maneuvering-target movements as real-time constant turn models, instead of a database-limited end-to-end mapping. Furthermore, an adaptive probability allocation method is designed to form a double-channel filtering algorithm. The proposed algorithm derives the final tracking results by fusing the states from two unscented Kalman filters together: one filter is based on the constant velocity model, and the other is based on the model with the transition matrices estimated by network. The simulation results verify that the proposed algorithm outperforms other state-of-the-art algorithms both in tracking performance and generalization ability.