Bobkov A.V. - Image registration in the real time applications, страница 4

Inthe simplest cases, the task can be reduced either to localisation or recognition. Forexample, if the identification procedure is applied to every image area, the objectposition can be found. Or, looking for all possible samples on the image, it can befound which of them are present on the image. The task of sample localisation is asubject of the image registration theory, whereas identification task is subject ofthe image recognition theory.This research deals with sample localisation task only.

However, solutionsthat can be found can be applied to the image recognition task as well, since bothtasks are tightly related and have its inner likelihood.1.1.3. Aims and objectivesOur primary task in this work is developing and researching methods ofimage registration that can be used in real environment and in a real time. Theprimary domain of application is the visual navigation task; however, otherapplications are also can be taken into account.15The basic idea of this work is to use image representation as a set of linesegments.

Line segments are universal feature that are usually present on most ofnatural images in enough amounts, and they can be easily extracted and matched.This makes the approach very promising and popular for many researchers.However, the results already achieved are far from optimal. The approach has bothadvantages and disadvantages, obvious and hidden. Our objective is to determinethe primary characteristics of this approach, discover the factors affected on them,research their behaviour in a different conditions and for different data, comparethem with the characteristics of other methods used in practice and understand thepractical use of the approach.We have selected the Generalized Hough Transform as the key strategy ofline segments matching.

Our next objective is analysing the possible methods ofline matching, investigate its advantages and disadvantages, measure itscharacteristics, and make conclusion about practical usage.In order to represent the image as a set of line segments, we also used theHough Transform. Existing schemes of HT for line segment detection have goodreliability, but provides insufficient accuracy. So the next objective is producingthe extended scheme of HT that would allow increasing the accuracy of linesegment detection up to one pixel without loss of performance. The accuracy isquite important factor because it allows reaching high reliability in further linematching.

We also need to research properties and characteristics of obtained linesegment detector.To implement the line segment detection, we need to extract the contour ofthe image – the line segments can be found only on the object edges. Contourdetection task is not so interesting because it is deeply researched; howeverobtaining the contour detector of proper characteristics is still difficult task. Ourobjective would be review existing approaches to the contour detection, select theproper components and combine them. Then we need to measure characteristics ofthe contour detector in order to be sure that it provides suitable characteristics forline detection.The final objective is testing of proposed methods on statisticallyrepresentative data sets in order to ensure that the characteristics are suitable forsolving practical tasks.

The direct testing on the flying vehicle is very expensiveand hard (imagine the software debugging process when every mistake can causeloss of equipment), so another approach is used. We reproduced the task conditionsin the robotic laboratory. We used a manipulator, which carried the camera overthe surface model, emulating the flying vehicle behaviour in the proper scale. Thisallows implementing image analysis in the conditions that are near to natural.Appropriate software would be also developed.1.1.4. Thesis outlineThe thesis is organized as follows.The rest of the chapter is analysis of image registration task and review ofthe literature on the image registration and the related topics – shape16representation, line detection and edge extraction.

The objectives is to reviewexisting image analysis research, to examine existing solution methodology, and todetermine a common approach to developing a practical image analysis systemthat can process a wide range of images in real time.Second part is devoted to the contour extraction.

Here we select the structureof the contour extraction algorithm and select proper components for itsimplementation. We compared direction gradient and statistical methods to choosethe best edge detector, suitable for further line segment extraction and matching.Then the proper thresholding algorithm are chosen and tested. Finally, the edgethinning method will be added, and whole contour detection algorithm will betested and tuned in order to produce good contour for further line detection.Third chapter is devoted to the line segment detector development.

We usedwell known Hough transform approach. Its disadvantages were analysed and moreaccurate and fast computation scheme were proposed. Then the properties of theproposed algorithm were tested in order to select proper parameters.Fourth chapter is devoted to the line matching. The main approach used forline matching in this work is Generalized Hough transform. We then came backagain to the Hough transform and describe possible approaches of line matching.Then we select the best of them and test them thoroughly in order to determine theperformance, accuracy and reliability in the different conditions. The measurementuses statistics collected on large data sets.Fifth chapter is the discussion of image registration methods proposed in theprevious chapters.

We analysed the performance, accuracy and reliability of themethods in different conditions and environment, showed the negative factors thatcan be affected to characteristics, and compared the method behaviour with theother popular methods commonly used. It is shown that proposed methods aresuitable for the task of image registration for visual navigation purposes, have highcharacteristics, and can be applied in many other areas.In the appendix reader can find the description of our equipment andsoftware used to provide experiment materials for the work.

This is roboticequipment for laboratory emulation of conditions of the visual navigation ofunmanned flying vehicle with a camera. The software we developed are used toprocess the obtained data, to test the contour extraction, line detection andmatching algorithms, to collect and process statistics on algorithm behaviour indifferent environment and different parameters.1.2.

Image registration problemThe Image registration problem is one of the fundamental and importantproblems in computer vision. At the current time, a problem of image matching issolved only for several particular applications and still remains an actual problem.On the one hand, researchers try to build reliable and totally autonomous systems17that can operate without human aid. Such systems have a wide practical interest,but they are developed only for a very narrow class of tasks. On the other hand,researchers try to build universal task-independent systems to check hypotheses ofimage interpretation area and to model biological systems of image acceptance.1.2.1. Image registration: statement of the problemThere are two images I1 and I2.

The registration task is to find a geometrictransformation g(I) and brightness transformation f(I), which transform image I1 insuch a way that corresponding pixels of both image match to each other:I1(X)=f(I2(g(X)))(1.1)The co-ordinate system can be different due to a different position of a videocamera, camera rotation or motion of an observed object itself. So the primary taskis to transform both images into a common co-ordinate system.

The brightnesstransformation can be also taken into account due to a possible difference in lightconditions, different season or time of the day. Beside this, images in the commoncase can be obtained by different kinds of sensors, and also require both space andbrightness transformation.Since the equation (1.1) is inaccurate on the practice, the quality functionalC1 is introduced as a matching quality criterion for selected transformations ofspace and brightness:C1 ( g , f ) = ∫ [ I1 ( X ) − f ( I 2 ( g ( X )))]dXX(1.2)The following definition will be given.Control points, or reference points are points, which positions are knownon both images.Since the control points must be transformed exactly to their pairs onanother image, this points play role of limitations to required space transformation:X i′ = g ( X i ), i = 1..N ,(1.3)′where X i and X i are control points of first and second image, N is amount ofcontrol point pairs.The error C2 in the control point mapping can be used to estimate theaccuracy of the selected space transformation:C2 ( g ) =N∑i= 1|| X i′ − g ( X i ) || 2(1.4)In such a way, the task of image registration is to find a space transformationg and brightness transformation f, which provides a minimum of target function(1.2) or (1.4).181.2.2.

Classification of image registration methodsThere are a large number of image registration methods that have beendeveloped for different limitations of source data. These limitations arise when areal practical task is investigated, and they can change from task to task. Butmethods can have significant differences even for the same limitations. So in orderto compare image registration methods and to determine their similar featuresthese methods must be classified by their several characteristics.The most popular classification scheme that uniformly describes all knownimage registration methods is the scheme which includes three classificationcharacteristics:1.Search space – a possible space transformation and brightnesstransformation between processing images.2.Type of elements used.3.The strategy used for optimal solution searching.These characteristics are not independent of each other.

For example, theselection of a typical image element shrinks the possible search space, and boththese characteristics determine, in most of cases, the search strategy.1.2.2.1. Space transformationAccording to the task conditions, it is required to find a geometrictransformation g and brightness transformation f, which provides a minimum of atarget function. It is impossible to find this minimum among all possible functionsg and f, so the class of possible transformation must be defined.In most of the work on image registration, the search of brightnesstransformation is not performed: either the brightness of both images isapproximately the same, or the method developed is invariable to a brightnesstransformation. In this case the brightness transformation can be found (if it isrequired) after the image registration.The space transformation can be described in three forms: the globaltransformation, the local transformation and the optic flow.1.2.2.1.1 Global transformationThe global transformation defines a transformation that is uniquely appliedto all image pixels or elements.

It transforms all the area of one image to anotherimage. This can be a transformation from the motion group, a homothetictransformation, affine transformation or projective transformation [Brown, 1992].The projective transformation is non-linear transformation. It is defined byfollowing equations:19x′ =a0 x + a1 y + a2,a3 x + a 4 y + 1y′ =a5 x + a 6 y + a 7a3 x + a 4 y + 1(1.5)The affine transformation can be produced from projective transformation, ifa3 and a4 is zero:x′ = a0 x + a1 y + a 2 , y ′ = a5 x + a6 y + a7(1.6)The affine transformation is a particular case of the projectivetransformation, and it is a polynomial transformation of the first degree.The homothetic transformation is a particular case of the affinetransformation, and it forms a transformation group.