基于MSA特征的遥感图像多目标关联算法

遥感图像中多目标关联存在以下两个问题:一是低时间分辨率观测使得目标状态信息无法准确估计,基于Kalman滤波的多目标关联算法不再适用;二是基于图像特征的目标关联算法又无法处理大场景观测中多个目标关联引起的模糊性.针对上述问题,提出一种基于多尺度自卷积特征匹配和关联代价矩阵最优化的多目标关联算法.实验表明该算法对遥感图像中多目标关联问题具有一定的适用性.

A MSA Feature-based Multiple Targets Association Algorithm in Remote Sensing Images

Target identification fusion based onmulti-source remote sensing mi ages canmake fulluse of the redundancyand complementary information from all sensors, acquiring more accurate result of target recognition. One of the pre-condition of identification fusion is targetassociation, which is to determine if the information from two ormore mi ages arerelated to the same target and should be fused together. Due to different performance of sensor and diverse targetdistribution, the extracted information of targets generally has some uncertainty, which results in the difficulty in judgingwhether the information from two mi ages is originated from the same target. Therefore, how to utilize the information ofremote sensing mi ages to distinguishmulti-target association has become an urgen problem.There are two kinds ofmethods concerning target association when using mi age data: one isKalman filtering baseddata association and tracking, which utilizes accumulated kinematic information ofmulti-frame mi ages to estmi ate andtrack. Typicalmethods areNearestNeighbor (NN), JointProbabilistic Data Association (JPDA), Multiple-HypothesisTracking (MHT) and so on. Thesemethodsneed dense sampling ofobserved data, and the targetmotionmodel should besmi ple. The otherone uses mi agematch in computervision for reference. Typicalmethods are cross correlationmatching,featurematching and so on. Thesemethods usuallywork on condition thatonly single target is concerned.For remote sensing mi ages, there are two problemswhen associatingmultiple targets in them. Firstly, it is incapableto acquire a seriesofmulti-temporal remote sensing mi ageson the same region atpresent, so the kinematic state ofa targetcannotbe estmi ated accuratelywith low temporal resolution data and the classicalKalman filtering association algorithmsare nomore applicable. Wemust seek for other tmi e-independent information as the associatingmeasurement, which canbe mi age invariant feature. Secondly, there are two uncertainties lying in mi age feature extraction of a target. Oneuncertainty lies in determining invariant features due to various mi age distortions such as rotation, scaling and so on. Theother lies in establishing feature correspondencesbetween any two consecutive mi ages. So, it isdifficultto discrmi inate theambiguity ofmultiple targets correspondenceswhen using mi agematching-based associationmethod.In order to solve above problems, a novel multiple targets association method based on mi age invariant featurematching andAssociation CostMatrix (ACM) global optmi ization is proposed. At first, theMulti-scaleAutoconvolution(MSA) transform of a target is computed based on affine invariant theory and is used as associationmeasurement, whichcan overcome the negative influence of changes in target s pose, mi aging viewpoint and so on. Secondly, the associationcostmatrix is constructed based on the dissmi ilarities ofMSA featurematching of any two target pairs from two mi agesrespectively, representing the correspondence illegibility oftwo targets. Finally, theminmi alenergy ofACM is found usingsmi ulated annealing (SA) algorithm, and the global optmi al association result is achieved.From the smi ulation expermi ents, some conclusions can be drawn as follows: (1) Using mi age invariant features toperform target association is a validateway, overcoming the bottleneck that the tmi e-dependentkinematic feature cannotbeestmi ated from sparse remote sensing mi age series. (2) Compared with the NN local algorithm, the optmi ization ofassociation costmatrix is a globaloptmi al algorithm and has excellentperformance in dense targets circumstance. (3)Theapproxmi ate algorithms such as SA can greatly mi prove the search of optmi al association costmatrix, and then makecomplex associationmethod practicable.