Maneuvering Target Tracking with Recurrent Neural Networks for Radar Application

Maneuvering target tracking is a problem of state estimation where the system undergoes abrupt changes. Traditional model-based approaches to this problem are threatened by the performance degradation caused by the model mismatch and the estimator usually has limited statistical precision in practice. Developed from what the machine sees as mathematically optimal in the data, deep neural network-based methods are not sensitive to variant models as long as the interested motions are fully contained in the training data. Besides, having access to the true state while training a network can make it possible to break the limit of traditional estimation precision. To coincide with the sequential manner of target tracking, the recurrent neural network is proposed to estimate the true state. Simulation results show that the proposed network handles the target motion uncertainty problem well, meanwhile, the states are estimated more accurately.