Maneuvering Multi-target Tracking Based on DLSTM and Dynamic Association Gate

To address the problem that the model-driven algorithms have the limited expression ability and are difficult to deal with the target maneuvering in time, this paper presents a maneuvering multi-target intelligent tracking (M-MTIT) algorithm based on deep long short term memory (DLSTM) and dynamic association gate, called the M-MTIT-DLSTM algorithm. Firstly, to track multiple targets, the proposed algorithm uses a prediction network and an update network based DLSTM, and defines the association gate by using the residual and corresponding covariance. Secondly, the valid measurements are extracted, and then the global nearest neighbor (GNN) algorithm is adopted to associate the data. Finally, the cardinal distribution theory is used to design a dynamic association gate, which can clear the accumulated historical states and historical measurements. Subsequently, the tracking process of the target is restarted when a maneuver is determined. Simulation experiments demonstrate that the proposed M-MTIT-DLSTM algorithm is more robust near moments of the target maneuver, and the tracking accuracy is close to that of the interacting multiple model (IMM) algorithm with high operational efficiency compared with the current widely used IMM algorithm.