Adaptive Filtering based on Recurrent Neural Networks

Kalman filter is an optimal filtering solution in certain cases, however, it is more often than not, regarded as a non-robust filter. The slight mismatch in noise statistics or process model may lead to large performance deterioration and the loss of optimality. This research paper proposes an alternative method for robust adaptive filtering concerning lack of information of noise statistics. The method is based on the application of recurrent neural networks trained by a dynamic identity observer. The method is explained in details and tested in the case analysis of object tracking model. Performance evaluation is made for cases of the standard Kalman filter, a noise-adaptive Kalman filter, the adaptive filter with a recurrent neural network trained by a static identity observer, and the adaptive filter with recurrent neural network trained by a dynamic identity observer. The results for different noise statistics as well as noise statistics mismatches are compared and presented. It is shown that in cases with a lack of knowledge of the noise statistics it is beneficial to use the filtering method proposed in this research work.