Research On Anti-Jamming Tracking Algorithm Of Infrared Small Target

Infrared small target tracking is one of the key technologies of infrared target search and tracking system,which is widely used in air and marine surveillance system and early warning system.It uses the infrared radiation of flying targets for identification and can accurately locate and track moving targets in rainy,dark,electromagnetic interference and other complex environments,and obtain relevant data.Furthermore,it provides protection for precision strikes in military system.However,due to the long range and small size of air targets,and the existence of various interference factors,the effect of the existing infrared small target tracking algorithms is seriously affected.Therefore,the research on the anti-jamming tracking algorithm of an infrared small targets has important practical significance.In order to solve the influence of three main interference factors(occlusion,scale transformation,deformation)on the infrared small target tracking algorithm,this thesis conducts indepth research on infrared small target anti-jamming tracking combined with infrared image characteristics and target tracking algorithm.For the edge scenes with limited computing resources,an anti-occlusion tracking algorithm for small infrared targets based on correlation filtering is proposed,which is based on the gradient characteristics of small infrared targets and the change of the response matrix under occlusion.For the server scenes with higher accuracy requirements,an optimization strategy of the twin network tracking algorithm based on attention mechanism and feature fusion is proposed,which achieves high-precision and high-stability anti-jamming tracking of small infrared targets.The main research contents of this thesis are as follows:(1)45 groups of infrared flight target datasets with three interference factors in multiple scenarios are collected and produced.At present,there is a extremely lack of published infrared target datasets in the air.In order to provide sufficient data support for the research of infrared target anti-jamming technology,35 groups of small infrared flying targets in real scenes are photographed using the infrared camera.The average single set of data is 342 frames,including 4 types of background and three main interference factors of target scale transformation,deformation,and occlusion.Then,the collected data is cleaned and labeled.In addition,10 groups of video sequences of flying targets being occluded by interference and strike targets are simulated and produced using MATLAB software programming.(2)The three main interference factors of occlusion,scale transformation,and deformation in the tracking scene are quantitatively analyzed in detail.This thesis divides and defines the occlusion of the target in detail.According to the change of target scale and pixels,the scale transformation and target deformation are divided and defined.In addition,the gradient characteristics of the target,the response distribution map,and the accuracy of the algorithm under the interference environment are analyzed.(3)For hardware platforms with limited computing resources,the traditional kernel correlation filtering algorithm is improved into an anti-occlusion tracking algorithm that supports adaptive template update.The flux density feature based on the gradient vector field of the infrared image is proposed to quickly and accurately distinguish the small target from the background.The template update mechanism is improved,and an occlusion detection mechanism is proposed.When the target is occluded,the tracker directly enters the next frame tracking task without template update.Finally,the hand-made infrared small target dataset is migrated to the OTB platform,and the algorithm is evaluated.The accuracy of this algorithm is 0.656,and the anti-occlusion ability is significantly improved.(4)For the server scenes with higher accuracy requirements,and anti-jamming tracking algorithm based on attention mechanism and Siamese neural network is proposed.The Res Net50 is used as the feature extraction network of the twin network with two branches to obtain the most effective feature representation of small infrared targets.An attention mechanism module is designed to calibrate the importance of the latter three-stage features of the template branch network and the latter three-stage features of the dual-branch network are fused to obtain highly robust discriminative features.Afterwards,the correlation filter block is used to obtain the target template,combining temporal information to obtain saliency target features.use the correlation filter block to obtain the target template,and combine the temporal information to obtain the salient target features.Experiments show that the accuracy of this algorithm is 0.817 on the OTB platform,which is 10.7% higher than the accuracy of the original algorithm.