An Improved Kalman Filter Based on Long Short-Memory Recurrent Neural Network for Nonlinear Radar Target Tracking

Abstract

The target tracking of nonlinear maneuvering radar in dense clutter environments is still an important but difficult problem to be solved effectively. Traditional solutions often rely on motion models and prior distributions. This paper presents a novel improved architecture of Kalman filter based on a recursive neural network, which combines the sequence learning of recurrent neural networks with the precise prediction of Kalman filter in an end-to-end manner. We employ three LSTM networks to model nonlinear motion equation, motion noise, and measurement noise, respectively, and learn their long-term dependence from a large amount of training data. They are then applied to the prediction and update process of Kalman filter to calculate the estimated target state. Our approach is able to address the tracking problem of nonlinear maneuvering radar target online end-to-end and does not require the motion models and prior distributions. Experimental results show that our method is more effective and faster than the traditional methods and more accurate than the method with LSTM network alone.

Trainable Quaternion Extended Kalman Filter with Multi-Head Attention for Dead Reckoning in Autonomous Ground Vehicles

“…Deep neural networks were used to update the parameters of an invariant EKF dynamically. A recent study also explored the use of long short-term memory (LSTM), a type of recurrent neural network (RNN), to model the nonlinear noises for KF [ 19 ] to address target tracking problems. Another approach that uses reinforcement learning to adaptively estimate the process-noise covariance matrix was proposed by Gao et al [ 20 ], in which their algorithm used the deep deterministic policy gradient (DDPG) to extract the optimal process-noise covariance matrix estimation from the continuous action space, using an integrated navigation system as the environment and the reverse of the current positioning error as the reward.…”

Long Short-Term Memory-Based Neural Networks for Missile Maneuvers Trajectories Prediction⋆

“…Since the tracking of a moving target is performed by processing the measurements of the available sensors, such as radar, sonar and camera, corruption generated by random noise is unavoidable. Under the quite restrictive assumption of regular target motion and white Gaussian distributions for the process and the measurement noise, most solutions proposed for this problem are based on the Kalman Filter (KF) theory [2]. However, when target trajectories have been characterized by great complexity and diversity and vary unexpectedly, classical KF approaches, which are based on a single dynamical model, do not achieve satisfactory performance [3], [4].…”

“…[21]. Since the missile target tracking problem can be seen as a sequence problem [14], RNNs could be employed to handle this task [22] and, unlike conventional model-based methods, allows to learn the correct behavior from the available training data in a model-free fashion, facing both the issues of measurement noise and model uncertainties, and without any a priori knowledge on the probabilistic noise distribution [2]. However, the training of standard RNNs suffers for well-known problems of vanishing gradient and exploding gradient, due to the difficulties for the gradients to propagate far in a lot of time steps consistently with an acceptable range [23], thus considerably limiting the applicability of these standard nets.…”