Research On Radar Multiple Target Tracking Algorithm Based On Deep Learning

Multi-target tracking technology has always been a hot research topic in the field of radar data processing.Traditional multi-target tracking methods are prone to wrong tracking or decreased tracking performance at target intersection points under background clutter and environmental noise,hence the need for research into multi-target tracking algorithms that can adapt to complex scenarios.In recent years,deep learning technology has gained more attention in the field of radar target tracking and has shown better performance potential than traditional methods.Therefore,this thesis focuses on the following research on deep learning-based multi-target tracking algorithms:(1)In response to the limitations of classic data association algorithms that require prior information such as target motion models and clutter densities,a data association algorithm that combines temporal convolutional networks and long short-term memory networks(TCN-LSTM)is proposed.The proposed algorithm comprehensively considers false alarms and missed detections and uses data-driven learning to map the relationship between each target and measurement,effectively solving the multi-target data association problem in cluttered environments.Simulation experiments show that the proposed algorithm has a smaller Optimal Sub-Pattern Assignment(OSPA)error,and in tracking environments with different detection probabilities,it has better tracking performance for multiple targets with crossing motions.(2)In view of the poor tracking effect of the classical tracking filtering algorithm and the large error between the measurement and the real state of the target,a space target state prediction algorithm embedded with self-attention mechanism is proposed,and a loss function with regular term is designed to smooth the tracking trajectory.The experimental results of the measured data show that,compared with the classical tracking filtering algorithm and the state prediction algorithm based on long-term memory network,the proposed algorithm has less root mean square error for strong maneuvering targets.and it has better prediction robustness under different detection probabilities.(3)A deep learning-based multi-object tracking algorithm was designed,which includes two main modules: data association and tracking filter.The experimental results show that the designed modular algorithm can effectively solve the problem of knownnumber maneuvering multi-object tracking in dense clutter environment.Compared with traditional multi-object tracking algorithms and deep learning-based data association algorithms combined with classical filtering algorithms,the proposed algorithm has smaller tracking errors at different detection probabilities.