Research On Radar Target Tracking Algorithm Based On Supervised Learning

In the field of radar target tracking,tracking algorithms can be divided into point data sources and echo data sources according to data sources.The former includes Kalman filter(KF)based on Bayesian estimation theory.The latter is mainly a multiframe joint detection and tracking algorithm based on dynamic programming-based tracking before detection(DP-TBD).Both occupy an important position in the research of radar target tracking.However,both types of traditional tracking frameworks need to model the prior model and set its prior parameters.Therefore,in the case of prior model mismatch or parameter estimation errors,the tracking performance of traditional algorithms will drop significantly.With the wide application and development of supervised learning,it has end-to-end data-driven and high-dimensional feature mining,which can effectively improve the robustness of traditional radar target tracking algorithms.However,the current target tracking based on supervised learning is mainly concentrated in the field of visual tracking,which has limitations in the application of radar targets.For example,only supervised learning is used to assist and correct target tracking,the fusion framework of supervised learning and radar target tracking is not fully constructed;the target motion characteristics are not extracted,resulting in insufficient general applicability of supervised learning;there is no effective use of multiple supervised learning algorithms to construct an integrated radar target tracking system.Therefore,thesis studies the radar target tracking algorithm based on supervised learning.The main work and contributions are as follows:1.For the traditional radar target tracking system with different data sources,study the filtering algorithm and DP-TBD algorithm based on the Bayesian estimation theory,so as to analyze the deficiencies of the prior model dependence of the traditional modelbased algorithm.2.For the problem that the traditional point filter relies on model matching and prior parameters,a supervised learning tracking filtering(T-SLTF)algorithm framework based on tree embedding is proposed.At the same time,the XGBoost algorithm is used for specific implementation,which effectively realizes the filtering algorithm.Robust estimation in unknown complex noise environments.3.Aiming at the problem that the traditional DP-TBD algorithm of echo data source relies on the assumption of background noise and lacks adaptability to the fluctuating environment,a DBU-Net-based supervised learning multi-frame joint detection(DSLMD)algorithm is proposed,which solves the problem of the traditional algorithm in fluctuating environment and improves the detection performance of weak targets in this environment.4.Aiming at the double dependence of the DP-TBD algorithm on the assumption of target motion characteristics and the assumption of environmental noise,as well as the single auxiliary limitation of supervised learning in the traditional target tracking field,combining the above T-SLTF idea and D-SLMD algorithm,a Light GBM-based Supervised Learning Multi-Frame Integrated Tracking(L-SLMIT)algorithm is proposed,which achieves robust tracking of radar targets under complex motion.The above proposed algorithms have been verified by simulation,which proves their effectiveness in the field of radar target detection and tracking.