Deep Learning Based Kalman Filter Tracking Algorithm For Manoeuvring Targets

With the rapid development of new generation information technology,maneuvering target tracking has been more and more widely used in military and civilian fields.However,for the current increasing number of maneuvering target tracking,the classical Kalman filter tracking algorithm and some of its traditional improvement algorithms will face the challenges of nonlinearity of the equation of motion and uncertainty of the observation equation.At the same time,deep learning,as a way of machine learning,has started to receive attention for its application in the field of target tracking.Based on this,this paper carries out the research of Kalman filter tracking algorithm for maneuvering targets based on deep learning,and the specific work carried out is as follows.(1)The theory and algorithms related to maneuvering target tracking are discussed.First,the principles of target tracking are introduced followed by the analysis of the commonly used CV,CA,and CT motion models.Secondly,the most classical linear tracking algorithm Kalman filtering and the nonlinear tracking algorithm EKF algorithm in the field of target tracking are analyzed and discussed.(2)Study of adaptive multi-incremental Kalman filter tracking algorithm based on the Marxian distance.To address the problems that the traditional adaptive asymptotic Kalman filter requires certain a priori knowledge in estimating the new interest covariance matrix and cannot be adjusted in real time when tracking highly maneuverable targets,an adaptive multi asymptotic Kalman filter based on the Marcian distance is proposed.The algorithm firstly establishes the martingale distance by the difference between the new interest vector and the zero vector,and uses the hypothesis testing method to determine whether the system is anomalous according to the feature that the square of its martingale distance obeys the chi-square distribution;secondly,it uses the gradient descent method to solve the asymptotic factor,and adaptively changes the iteration step size according to the hyperbolic tangent excitation function adjusted by the martingale distance of the new interest vector;finally,it assigns the multiple asymptotic factors by the feature that different channel values of the error covariance matrix represent Finally,the multiple asymptotic factors are assigned by the feature that different channel values of the error covariance matrix represent different tracking accuracy.The simulation results show that the improved method using the Marcian distance has faster computational speed and higher tracking accuracy.(3)A study of long and short term memory neural network maneuvering target tracking algorithm based on motion model classification.Since the Interacting Multiple Model(IMM)algorithm can achieve reasonable performance only when the motion state of the maneuvering target can be approximated by a small number of models.In addition,using a fixed Markov probability matrix to switch models causes problems such as slow model switching and poor tracking accuracy.With the guidance of machine learning,we propose a new deep long and short-term memory neural network based on motion model classification(C-DLSTM)for maneuvering target tracking.With sufficient data,the network can track targets with different motion models.Specifically,the proposed C-DLSTM neural network adopts a novel layered structure.The proposed network consists of two layers of deep-length short-term memory neural nets(DLSTM),with the function of the first layer being motion model classification and the function of the second layer being target tracking.Simulation results demonstrate that the tracking accuracy of the proposed C-DLSTM neural network algorithm outperforms the conventional maneuvering target tracking algorithms including the proposed adaptive fading factor Kalman filter tracking algorithm based on the Marxian distance.