تعيين محل هاي مناسب براي فرود اضطراري هواپيماهاي بدون سرنشين با استفاده از بينايي ماشين

چكيده انگليسي :

Landing Site Selection for UAV Forced Landing Using Machine Vision Mahdie Rezaeian m rezaeian@ec iut ac ir Date of Submission 2011 4 27 Department of Electrical and Computer Engineering Isfahan University of Technology Isfahan 84156 83111 Iran Degree M Sc Language PersianSupervisor Saeid Sadri sadri@cc iut ac ir Rasoul Amirfattahi Fattahi@cc iut ac ir Assoc Supervisor Niloofar Ghaisari n gheissari@cc iut ac ir AbstractIn this thesis a method for aerial image segmentation has been proposed to select an appropriate site forunmanned aerial vehicle UAV emergency landing Up to now the design and development of the UAVshave been mostly for military requirements Technology development UAV benefits and large number ofpotential UAV applications in civilian sector such as commercial security and scientific leading to theUAVs entrance into residential areas in the near future To gain public acceptance and to develop UAV usein the civilian branches concerns about the safety issues should be resolved and the UAV should be able toreact in a similar manner to pilots in emergency situations including that UAV must be capable of finding asuitable landing location automatically The objective of this research is Landing Site Selection for UAVForced Using Machine Vision For this different parts of aerial images should be segmented and classified Due to permanent decrease in altitude segmentation algorithm must be rapid and also accurate One of thesegmentation methods is pixel classification Important advantage of this method is that segmentation andclassification of different image areas can be done simultaneously and no extra recognition step is needed There are two main categories that can be used to achieve this outcome and they are called Supervised andUnsupervised Classification techniques Due to the different number of classes and combination variety indifferent aerial images using unsupervised classification techniques may cause over or under segmentation Therefore in this thesis supervised approach is used for pixel classification In this method different featuresof pixels are compared with a database of a certain number of classes and most likely label for each pixel isselected Due to the nature of aerial images features must be rotation and scale invariant In this thesis twoseparate KNN classifiers classify image texture and color histograms Gray image local texture characteristicand local color histograms are calculated in distinct windows and classified individually Final segmentationis obtained by evaluating the certainty with which each classifier color or texture alone would make adecision Thus it is possible to set the parameters of each classifier independently for achieving the bestresult On the other hand it is unlikely that both classifiers make mistake in the same case so it is possible tocorrect each other errors In this way the segmentation error decreases In aerial images like most naturalimages neighboring pixels usually have similar characteristics Based on this quality calculations can bereduced and better segmentation can be obtained Accordingly classification result of each feature vector isassigned to a group of adjacent pixels a patch instead of labeling each pixel Proposed method which isapplicable for any type of color texture images is simple and rapid The error rate of the proposed algorithmis low and segmentation results for aerial images are acceptable visually Keywords Automatic Landing Site Selection Aerial Image Supervised Segmentation PixelClassification Color Texture Segmentation Feature Distribution LBP Color Histogram K NN Classifier Semantic Segmentation

رضائيان، مهديه

تعيين محل هاي مناسب براي فرود اضطراري هواپيماهاي بدون سرنشين با استفاده از بينايي ماشين

قطعه بندي تحت نظارت تصاوير هوايي , بافت رنگي , طبقه بندي پيكسل ها , كلاسيناير KNN , طبقه بندي پوشش زمين , توزيع محلي مشخصه ها