A Deep Learning-Based Multiple-State Fusion Approach for Maneuvering Target Tracking

With the rapid development of sensor communication technologies in recent years， the ability to obtain high-precision data in real time has been greatly enhanced. This facilitates the development of data-driven approaches for solving maneuvering target tracking problems， thereby overcoming the challenges of traditional model-based filters in dealing with poor a-priori model information and high-speed target maneuvering. In particular， deep learning approaches estimate the target state by constructing end-to-end neural networks， which eliminate the reliance on a-priori information at the expense of reliance on training data and lack of physical interpretability. To solve these deficiencies， we combine the advantages of the model-driven interactive multiple model （IMM） algorithm and data-driven long short-term memory （LSTM） network， to propose a deep learning-based multiple-state fusion approach for target track maneuvering.This approach comprises multiple LSTM-based trackers and a single LSTM-based classifier. Each tracker is assigned a separate target motion state， and the trackers are executed in parallel to obtain state estimates. A classifier is used to determine the weight for each motion/tracker， and the final estimate is the weighted average of the estimates of all trackers. Simulation results demonstrated that the proposed algorithm outperforms both the IMM and end-to-end LSTM network-based tracking algorithms in terms of tracking accuracy.