Radar Multi-target Tracking Based On Deep Learning

Nowadays multi-target tracking has a wide application.In terms of civil applications,it is one of the core technologies in weather monitoring,video surveillance,autonomous driving and air space monitoring.In terms of military use,especially radar and sonar,where it plays a vital role,a great many researchers have been studying multi-target tracking for long time.Traditional algorithms perform poorly in complex scenarios where the number of targets are unknown,motion patterns are complicated,clutter rate is high,detection rate is low and so on.Deep learning,which has been rapidly developing in recent years,may provide another approach to multi-target tracking problem.Firstly,deep learning is able to learn latent pattern from data,which may become a substitution for handcrafted system model.Secondly,the high-level feature that provided by deep learning may solve the problem of target-measurement association,increase data association accuracy,therefore boost the performance of multi-target tracking algorithm.This thesis mainly studies detecting and tracking algorithms on deep learning,the outline of work is as follows:1)Traditional tracking algorithms are reviewed.Kalman filter,3 single-object filtering algorithms(NN,PDA,GSF),3 multi-object filtering algorithms(GNN,JPDA,MHT)are implemented and tested in different scenarios and analyses of their merits and disadvantages are reported.2)Poisson Multi-Bernoulli Mixture(PMBM)Filter is studied and its revised version-PMBM filter based on adaptive clutter density estimation algorithm is proposed.The revised algorithm calculates the distance from a measurement to its N nearest neighbors to estimate clutter density for that measurement and passes to PMBM filter in order to suppress false alarms.By careful analysis,k-d tree structure is used to search N nearest neighbors to boost speed of algorithm.3)A full convolutional network is designed for semantic segmentation for radar B-scope image and a detecting-tracking algorithm based on semantic segmentation is proposed.For overlapping or false bounding boxes from detecting network,a revised NMS(non-maximal suppression)algorithm which utilizes semantic segmentation results enclosed by bounding boxes is used to suppress false alarms.For false negatives and ID switches incurred by crossing targets,the first-order Markov model in PMBM algorithm is adaptively tuned,which greatly reduces the number of false alarms and ID switches.