Unscented Kalman Filters Integrated with Deep Learning Approaches for Active Sonar Based 2D Underwater Target Tracking

Abstract

This manuscript proposes a new approach to track 2D targets using a combination of machine learning algorithms and the Unscented Kalman filter (UKF). The approach makes use of active sonar sensors to measure range and bearing, which are used to predict the target’s course and speed. So far in the literature of target tracking, researchers assumed covariance matrix of the noise in sonar measurements. In this manuscript, it is tried to estimate the same using deep learning algorithms. The Machine Learning algorithms, such as multilayer perceptron, convolutional neural network, long-short term memory, and gated recurrent unit, are employed to approximate the covariance of the noise in the input measurements. Simultaneously, the Unscented Kalman Filter (UKF) is utilised to mitigate the noise in the measurements and to estimate the position and speed of the target. The results are quantified through Monte Carlo simulations in a simulated underwater environment. The measurements are assumed to conform to a normal Gaussian distribution with a mean of zero. The findings indicate that LSTM has superior performance compared to the other models. Nevertheless, it is important to note that the results are constrained in their applicability due to the restricted set of variables employed for training the machine learning models.