Research On Tracking Algorithm Based On Machine Learning And Multi-source Fusion

T Target tracking technology is widely used in military and civil fields,and has always been a research hotspot.With the increasing mobility and complexity of tracking targets,it is difficult for traditional algorithms to achieve high-precision target tracking.Therefore,this paper focuses on the target tracking algorithm,and the content of this paper is also the research result of the author’s participation in the Natural Science Foundation Project(No.62171060).The main research contents include:(1)The maneuvering target tracking algorithm based on machine learning is studied.In order to improve the tracking accuracy of nonlinear and complex moving targets with limited measurement information,a long short-term memory network-Kalman filter(LSTM-KF)algorithm is proposed.Firstly,the error sources of Kalman filter applied to nonlinear motion and complex motion tracking are analyzed,and LSTM is introduced to learn the tracking target motion information to reduce the error of target motion prediction.Secondly,through the analysis of correlation characteristics,a scheme based on average speed and instantaneous speed is proposed to solve the problem of poor generalization when LSTM predicts absolute coordinates.Finally,LSTM and Kalman filter are combined to realize tracking.(2)The target tracking algorithm based on multi-source information fusion is studied.In order to make full use of the redundant measurement information in multi-sensors for high-precision tracking,an adaptive federated Kalman filter(RDAFKF,Redundant Data Adaptive Federated Kalman Filter)algorithm is proposed based on the traditional federated filter.Firstly,in the stage of information allocation,an adaptive information factor allocation scheme is designed for the scene of redundant information fusion tracking.Secondly,in the stage of information fusion,firstly,a probability-based outlier detection scheme is proposed to reduce the influence of error data.For the data with correlation and Gaussian distribution,the outlier probability of the data to be evaluated is calculated by using confidence interval,and all outlier probabilities are combined by D-S evidence theory to determine whether the data to be evaluated is outlier.Then,the linear variance least square method is used to fuse the correlated data,and the final estimation result is obtained.(3)Simulation results show that the tracking accuracy and robustness of the proposed two algorithms are improved.