Target Initialization And Tracking During Electro-optical Imaging Terminal Guidance

The precision-guided weapon plays a decisive role in modern warfares. Targetinitialization and tracking techniques are the core technologies in electro-opticalimaging terminal guidance weapon systems, which are the representatives of theprecision-guided weapon. Therefore, two key issues, namely electro-optical imagingterminal guidance target initialization and target tracking, are investigated in thisdissertation. Several new terminal guidance target initialization and target trackingapproaches are proposed, which form a relatively complete terminal guidance workflow.Target initialization in the electro-optical imaging terminal guidance requires thealgorithm to provide robust wide baseline image matching across a substantial range ofimage scaling, image rotation, addition of noise, change in illumination and3D cameraviewpoint difference between the reference image and the run-time image. In order toachieve this goal, several contributions for target initialization have been proposed:A new target initialization approach based on speeded up robust features (SURF)and a simple, yet efficient3D viewpoint compensation method is proposed. SURFfeatures are invariant to image scale and rotation, robust to illumination change andrapid to be extracted. They can be used to solve the problem of image differences inscale, rotation, gray-scale and noise between the reference image and the run-timeimage. By using the3D viewpoint compensation method, the robustness of SURF in3Dviewpoint difference is improved substantially. Hence, a reliable matching over a wider3D viewpoint angle between reference image and run-time image may be achieved.Wide baseline image matching based on affine invariant features is introduced intoterminal guidance target initialization. Two new gray-scale and rotation invariant featuredescriptors are constructed based on two modified version of Local Binary Patterns(LBP) respectively which are gray-scale and rotation invariant local binary patterns andLocal Binary Pattern Histogram Fourier features (LBPHF). These novel approaches arenot only capable of doing local image regions invariant description, but also skippingorientation assignment and local region rotation, which are necessary in the traditionalalgorithms. Based on the traditional wide baseline image matching algorithmframework, the stable and efficient target initialization has been achieved in thisdissertation with the usages of the Maximally Stable Extremal Regions (MSER) andthose two new feature descriptors.Wide baseline image matching based on random Ferns classifier is also introduced into terminal guidance target initialization. By converting the wide baseline imagematching problem to a more generic classification problem, the high-dimension featurevectors matching in traditional wide baseline matching framework which results in ahigh computational overhead is therefore transferred to a classifier offline training phase,thus yielding a fast run-time performance. However, using random Ferns classifier maylead to an amount of false matching candidate region pairs. In order to solve thisproblem, Scale Invariant Feature Transform (SIFT) descriptors of corresponding regionsin each candidate matches are computed. False matching region pairs rejecting isadopted based on Mahalanobis distance criterion. The proposed method is capable ofimproving the stability of random Ferns classifier when running wide baseline imagematching in severe conditions. Additionally, the work flow, potential problems andsolutions during the implementation of our algorithm on hardware in real systems areproposed.Terminal guidance target tracking requires the algorithms to overcome the potentialdefault when rapid change of the target size and small range of image vibration duringthe flight of the missile. In order to solve these problems, two new terminal guidancetarget tracking approaches are proposed in this dissertation.A new terminal guidance target tracking approach based on the Gaussianscale-space is proposed. Two factors which lead to the failure of terminal guidancetracking are found via the analysis of large amounts of terminal guidance imagesequences. First, the rapid change of target size leads to the tracking centre location drift.Second, as the missile flying forwards, the target image is becoming clearer. Thus theappearance of the new target texture may lead to a wrong tracking localization anduncertain template refreshing occasion. A multi-scale representation of run-time imageis acquired based on the Gaussian scale-space theory. The scale space image sizes areadjusted according to respective scale parameters to estimate the target sizes differenceat different distances between the missile and the target. This eases the problem ofuncertainty of the template size and the appearance of the new target texture issuppressed by applying the Gaussian smoothing when constructing the scale-space. Thebest matching image and location are selected based on the Mean Absolute Difference(MAD) similarity measure. The final target location and size are acquired according tothe best matching image’s zoom rate. The simulated results showed that the proposedmethod is capable of providing robust terminal guidance tracking in both city andcountryside environments, no matter using CCD camera or Infrared camera.The traditional image matching method based on Fourier-Mellin Transform (FMT) only allowed a certain degree of image scaling difference. In order to improve thestability of the FMT, an improvement based on the Gaussian scale-space theory isproposed. The new approach called Multi-Scale Fourier-Mellin Transform (MSFMT) isconstructed to evaluate image transform parameters. The scale and rotation parameterevaluation between the template and run-time images is performed based on MSFMT.The template is refreshed based on bilinear interpolation according to the scale androtation parameters, to ensure the precision and stability of target tracking. Experimentresults showed that the proposed method is capable of providing robust terminalguidance tracking in city environments using CCD camera, but unsuitable forcountryside environment or using Infrared camera.