6763

6305

كارشناسي ارشد

هوش مصنوعي و رباتيك

اصفهان: دانشگاه صنعتي اصفهان، دانشكده برق و كامپيوتر

1390

ده، 111ص.: مصور، جدول، نمودار

ص.ع.به: فارسي و انگليسي

مازيار پالهنگ

نيلوفر قيصري

22/3/91

محمدرضا احمدزاده، رسول امير فتاحي

1396/09/14

كتابنامه

برق و كامپيوتر

مهندسي برق و كامپيوتر

ID6305

به فارسي و انگليسي: قابل رويت در نسخه ديجيتالي

Improving contour based object detection and application of Longest Common Subsequence LCS problem in finding feature correspondence and hypothesis verification Maedeh Ahmadi m ahmadi@ec iut ac ir Date of Submission 2011 7 10 Department of Electrical and Computer Engineering Isfahan University of Technology Isfahan 84156 83111 Iran Degree M Sc Language FarsiSupervisor Maziar Palhang palhang@cc iut ac irAbstractAn important research area in computer vision is object detection Object detection means detecting objectsbelonging to a specific class like bottles humen airplanes etc in an image Its aim is designing a systemthat gets training images containing specific classes or object models as inputs and afterwards detectingobjects in new images Object detection has various applications such as security systems driver assistantsystems classification and organization of huge number of images and videos blind person assistance andsearch engines Many methods have been introduced for object detection but the performance of thesemethods is far away human ability One of the popular and successful approaches is a method based onGeneralized Hough Transform GHT This method has two main steps 1 applying Hough transform voting in 3D location scale space and generating initial hypothesis for object location and scale 2 hypothesis verification and selecting valid hypothesis Applying Hough transform is based on local featuresof the image In this thesis we use a family of local contour based features namely k adjacent segment kAS to represent and detect objects Contour based features are invariant to color texture andillumination changes They are appropriate to represent object s shape The first part of this thesis presents acontour based object detection method invariant to in plane rotation To this end dissimilarity criteria of2AS features way of applying Hough transform and Hough voting space have been changed in such a waythat object s rotation degree can be estimated as the fourth dimension of voting space along with locationand space In order to improve detection results a method based on Max Margin Hough transform hasbeen proposed To study the performance and efficiency of the proposed method rotated TUD cows reference dataset images have been used The achieved results show that the proposed method can estimateobject s location and rotation degree promisingly The second part of the thesis introduces a method for finding corresponding features between model andhypothesis resulted from Hough transform Hough transform models the object s structure by consideringeach feature location with respect to its center The shortcoming of this method is that it considers thelocation of each feature independent of others and ignores relative location of features To overcome thisproblem we formulate the feature correspondence problem between model and hypothesis as a LongestCommon Subsequence LCS problem Model and hypothesis images are represented as feature strings andlongest common substring between them is calculated So by considering features order correspondingfeatures between model and hypothesis are obtained In the final step similarity of model and thecorresponding features is calculated using shape context We applied our method on two subsets of standardETHZ shape dataset The achieved results show that the proposed method improves Hough transformperformance considerably and has comparable or better results in comparison with previous methods Keywords Contour based object detection In plane rotation invariance 4D hough space Hypothesisverification Longest Common Subsequence

مازيار پالهنگ

نيلوفر قيصري

محمدرضا احمدزاده، رسول امير فتاحي