Menzil profili modunda çalışan radarla otomatik hedef sınıflama

Abstract

ÖZET Bu tezde uzaktan algılama aracı olarak mikrodalga radarları algılanacak `hedef olarak da su-üstü gemileri gözönüne alınarak radarla otomatik sınıflama incelemekte, bu amaçla algoritmalar önerilerek bunlarla elde edilen performans belirlenmektedir. Sınıflama, yalnızca hedefin varlığının saptanması değil aynı zamanda radar işaretlerini uygun biçimde değerlendirerek `algılanan` gemilerin gözlem uzayında daha önceden saptanan kriterlere göre belirlenen alt gruplardan hangisine girdiği konusunda bir hipotez oluşturulmasıdır. Yüksek menzil çözünürlüğüne sahip radarlar ile aydınlatılan gemi gibi büyük hedeflerden saçılan işaret, hedefin bakış doğrultusu boyunca `dilimlenmiş` örneklerini içerir. Bu dilimlerden saçılan güçlerin menzil boyunca dizilmesi ile elde edilen menzil profili geminin hangi sınıfa ait olduğuna ilişkin bir hipotez oluşturmakta, yani sınıflamada kullanılabilir. Yüksek çözünürlüklü radar kullanarak hedef menzil profillerinin oluşturulması frekans adımlı sentetik metodla açıklanmış, hedef sınıflamada temel kavram ve yöntemler verilmiş ve menzil profili yardımı ile sınıflama incelenmiştir. Hedef sınıflama aşamalarının en önemli adımlan olan nitelik çıkarma ve sınıflama adımlan için kullanılmak üzere geliştirilmiş olan belli başlı algoritmalar karşılaştırmalı olarak incelenmiştir. Tezde kullanılan yaklaşım, nitelik çıkarma adımı için, hedefin elektromagnetik özelliklerinin hedef üzerindeki saçıcı merkezleri yardımıyla temsilini sağlayan modele dayalı yöntemlere karşı düşmektedir. Bu çalışmada söz konusu modelleme için ayrı bir simülasyon programı ile elde edilen sonuçlardan yararlanılmıştır. Otomatik sınıflama algoritması temelde En Yakın Komşu (EYK) metodu kullanılarak oluşturulmuş bir menzil profili sınıflama algoritmasıdır. Gerçek verilere, yani gemilere ait MP ölçme sonuçlarına ulaşmak olanağı bulunamadığı için sınıflama performansının belirlenmesinde bir simülasyon yazılımından yararlanılmıştır. Tezde bu yazılımın bazı önemli özellikleri tanıtılmakta, gemilerin saçıcı merkezleri yardımıyla modellenmesi ve ölçü sonucu ve referans menzil profillerine karşı düşen verilerin oluşturulması için ve otomatik sınıflama amacıyla kullanılan algoritmalar verilmektedir. Daha sonra, söz konusu algoritmalar, iki farklı senaryo yardımıyla test edilmekte ve elde edilen sayısal sonuçlar yardımıyla sınıflama performansı ve bu performansı etkileyen faktörler belirlenmektedir. Operasyonel durumda ve gerçek ölçme verileri kullanılarak otomatik sınıflama algoritmaları ile elde edilmiş olan sınıflama başarımı değerleri ile bu testler sonucunda elde edilen sınıflama performansının yakın bir uyum içinde olması, tezde kullanılan yazılımın yaklaşımının ve sonuçlan elde etmek için yararlanılan simülasyon yazılımının gerçeği yeterli doğrulukta yansıttığını göstermektedir. XI

SUMMARY AUTOMATIC TARGET CLASSIFICATION WITH RADARS OPERATING IN RANGE PROFILING Radar researchers have long believed that through detection and utilization of all the target information present in an electromagnetic signal scattered from an unknown target, significant information about the nature (i.e., point scatterer versus area scatterer) and class (i.e., tanker or warship) of that target could be derived. However, this presents a difficult task both from the standpoint of the design of suitable radars and also from the standpoint of developing sufficiently robust and efficient algorithms. Classification of sea targets is extremely important for many applications, ranging from law enforcement, traffic monitoring, search and rescue to military tasks such as maritime situation assessment. Target classification is one of the critical functions that must be performed by many military systems since, failure to correctly classify potential targets may lead to undesirable consequences. Historically, a human being has performed this function using information obtained by his own ( possibly electronically augmented ) senses. As required reaction times decrease, target density increases, the ability of the human to make accurate decisions is severely degraded. Automatic target classification techniques attempt to use computer resources to process the sensor information to enhance ( if not replace ) the human decision. The degree of technological maturity that has been achieved by recent advances in radars and the availability of fast and efficient signal processing hardware has resulted in renewed interest in automatic target classification. Accordingly, in recent years, a considerable amount of attention has been focused on this problem. Radar techniques applicable to the development and deployment of this technology include Doppler analysis, for determination and characterization of target dynamic characteristics, and synthetic aperture and inverse synthetic aperture radar ( SAR and ISAR, respectively ), for determination and characterization of target shape and structure. Radar-based target classification can be attempted based on various types of radar signatures. One natural choice for such a signature domain is that of range profiles. These profiles provide a one-dimensional 'map' of the target scatterers in the range dimension i.e., with respect to the radar line of sight. In this thesis, we investigate the feasibility of utilizing automatic classification procedures for the classifications of sea targets utilizing a high resolution radar capable of generating target range profiles. XUA schematic range profile ( RP ) is shown in Fig.l. The aim is to use some sort of an algorithm which when fed with RP data can automatically discriminate between target classes. The main building blocks of an automatic classifier are that of feature extraction and classification. The reason for extracting target features from the received signal is threefold: 1) to optimize performance, 2) to reduce the amount of information to be processed, 3) to ensure the robustness of the system. RANGE CELLS ENERGY WITHIN A RANGE CELL ^HX. 0 20 40 60 80 100 1Z0 140 160 180 200 220 240 SCALE ( m ) Fig.l A schematic range profile The first consideration, optimization of performance, is a common goal of most target classification systems, where performance is usually interpreted as classification accuracy or probability of making correct decisions. The second consideration, dealing with the amount of available information and sometimes referred to as dimensionality reduction, implies that reducing the amount of information to be processed also reduces the demands placed on the signal processor. In general, as the dimensionality of a data set increases it becomes more difficult to implement a successful classification scheme. Even though dimensionality reduction may degrade the achievable final performance, it is often used because having a tractable nonoptimal solution is better than having no solution at all. XlllThe third criterion, robustness, means that an attempt should be made to ensure that at least some (expected) changes in the target or its environment will not alter classification performance. Hence, a certain degree of robustness is viewed as a basic requirement to be fulfilled by the selected features. Several factors may be used to generate the classification decision. In this thesis we concentrate on Nearest Neighbor ( NN ) type criteria. Denoting Feature Vectors ( FV ) derived from the measured data by xm and reference or prototype FV for the k.th ship class by Xk we define a squared Euclidean distance d(k)as d(k) = xm - xk 2 In cases where the data base contains more than one prototype corresponding to a particular ship type and to a particular aspect angle we define the reference FV as an ensemble average. In the direct utilization of the NN algorithm the hypothesis generated for the ship class is based on minimizing dk. This approach will lead to a certain correct classification performance, which in general will be a function of the scenario under consideration. It will also depend quite critically on the amount of the type of the data contained in the FV. On the other hand, in most practical cases one has access to a number of sequential RP measurements which are generated for the same target and under essentially identical conditions. This extra amount of data can be used at either the feature extraction or the classification stages to improve the classification accuracy. In this thesis we choose the second approach and generalize the simple NN algorithm by the introduction of a set of Bayesian fusion rules. We begin by defining classification likelihood vectors ( CLV ). Assuming that L measurements are available and the scenario contains M ship classes we first calculate the Euclidean distances for each ship class and for each measurement as ke 1,M, le 1,L Next we normalize these distance via, Pk'=M - kel,M;<sl,L v=l XIVCLV are then constructed as o« = [Pl' £ £ f11 P2 P3 PM The fusion algorithm that has been implemented combines CLV resulting from successive measurements through the Bayesian rule : PCH^fjPCEqlHi) P(HiIEIE2...Em)=- q=1 JrP(Hr)nP(EqIHr) q=l where P (Hi) : denotes the probability that a certain hypothesis concerning for the class of the ship is true Ej xorrespond to different measurements, and P(Hi,EiE2...) :denotes the conditional probability corresponding to P(Hj) subject to the availability of the data set Ej. In this thesis we utilized a model based approach for feature extraction. In this approach, with the aid of certain simplifying assumptions [8] one tries to extract from the measured data the spatial distribution and scattering amplitudes of the main scattering centers which make up the backscattered signal for the particular aspect angle. Since measured data was not available we used scattering center data via an auxiliary modeling program. Finally, in order to test the performance of the range profiling classification algorithm suggested in this thesis we utilized a simulation program in order to obtain the FV corresponding both to the measured and also to the reference range profiles of the ships represented in the scenario. The generation for a particular ship and aspect angle the calculation of the RLV was performed utilizing following steps : - Electromagnetic properties of the ships are defined using a number of scattering centers (~ 100 ) of given amplitude placed at certain spatial positions on the ship. - Backscattered amplitudes are calculated via vectorial summation and assuming Swerling 2 type fading characteristics. Measurement errors in the aspect angle are taken into account. XV- Reference data is calculated in the same way as the measured data but assuming that sufficient amount of smoothing has been performed ( as is the case in the construction of actual target signature databases ) with the neglect of the error terms and via a scalar summation on the bases of the decorelated powers. - Although the simulation program is sufficiently general in that noise, clutter and various system losses which may effect the detection process can be taken properly into account, for the purposes of this research. We assumed that sufficient signal can be made available at the receiver input and neglected these effects. Assuming the resulting RCS values in dB corresponding to the i.th range cell are given as RCSi,k we define RCS., -RCS., f N s = 1£ imq_ and i,k g k flrj i,k where the index k refers to the ship type, the index ref denotes the data base value, the factor 6 is included to suppress the glint effect and N corresponds to the range extend of the target. We then calculate the elements of the RLV as -1/2A, e K pk~ M -1/2A, 5> K k=l RLV ?[ Pi Pi pM. where M denotes the number of ship types in the scenario under consideration. We have put our approach to test using a fairly realistic scenario which contain 10 ships each from 5 different classes which are allowed to move along quasi - random paths while remaining within the coverage area of the radar. The simulation was run for 20 hours, which is sufficient for generation of 20 RLV's on the average for each ship in the scenario. The test results are depicted in graphical form in Fig.2. Here the percentage value of correctly classified ships is plotted versus the number of detected ships as a function of time. It is seen that the final correct classification probability is approximately 95 %. This value can be compared quite favorably with the probability of correct classification for a single measurement which is obtained to be 80 %. XVIThe classification performances obtained in this thesis are in quite close agreements with those obtained using real data and under operational conditions and reported in the open literature. This is another indication that both the modeling of the ships and the simulation of the generation and evaluation of the RP data as presented in this thesis can be considered to the sufficiently realistic. 100.00 -, 75.00 : 50.00 25.00- 0.00 l- r-i-ny i-,- u r -I 1, 1 1 1 1 r- tOOOO. 20000. 30000. 40000. 50000. `i ? i 1- - i 1 1 ? i ?} r- 60000. 70000. (sn) Figure 2 Percentage of correctly classified ships XVII