A Mnemonic Kalman Filter For Non-Linear Systems With Extensive Temporal Dependencies

Analytic dynamic models for target estimation are often approximations of the potentially complex behaviour of the object of interest. Its true motion might depend on hundreds of parameters and can involve long-term temporal correlation. However, conventional models keep the degrees of freedom low and they usually assume the Markov property to reduce computational complexity. In particular, the Kalman Filter assumes prior and posterior Gaussian densities and is hence restricted to linear transition functions which are often insufficient to reflect the behaviour of a real object. In this paper, a Mnemonic Kalman Filter is introduced which overcomes the Markov property and the linearity restriction by learning to predict a full transition probability density with Long Short-Term Memory networks.