Research On Detection And Tracking Methods Of Air Moving Targets

In recent years,with the emergence and wide application of various aircraft,the detection and tracking of moving targets in the air has become an increasingly noteworthy research topic.At the time when the general vision algorithm is in the ascendant,the algorithm design based on the characteristics of moving targets in the air is also an important research direction that cannot be ignored.Air moving targets have the characteristics of fast speed and huge changes in shape and scale,which creates difficulties in detection and tracking algorithms.Specifically,when the target moves too much in adjacent frames,the target tracking algorithm based on correlation filtering will either lose the target or only sacrifice efficiency to expand the search range,while the detection algorithm will also reduce accuracy in the face of the rich changes in shape and scale of the target caused by rapid movement.Aiming at the above problems,this paper proposes an improvement on the general detection and tracking algorithm based on the characteristics of air moving targets to achieve better detection and tracking results.The specific work is as follows:(1)A special detection and tracking dataset for air moving targets has been produced.In the detection dataset,specific categories of samples in the general detection dataset are collected,and the label box is corrected;In the tracking dataset,set the target trajectory and crop the target for synthesis.(2)A fast target tracking algorithm based on trajectory prediction is designed.By analyzing the motion characteristics of moving targets in the air and modeling the displacement based on the assumption that the acceleration of the target is constant within a short time series,the position of the target in the next frame is roughly estimated,which can obtain sufficient search field of view without expanding the search area,thereby accurately tracking the fast target.Experimental results show that our algorithm outperforms traditional algorithms in fast target tracking datasets.(3)A multi-target tracking algorithm based on tracker list is designed.A tracker list is created by initializing the target detection results of the first frame.Its element is a tracking component.Each tracking component is responsible for tracking a target and carrying the target number.In the subsequent frames,through the combination of detection and tracking,accurate target number and target boundary box can be obtained,so as to realize adaptive tracking of all targets.At the same time,a bimodal effect suppression method based on template scale reduction is designed,which can eliminate the bimodal effect caused by similar targets in tracking.Experiments show that the algorithm achieves 58.9% MOTA,64.7% IDF1,and 42.4% tracking frame rate on the produced flying target multitarget dataset.