On-line Boosting forCarDetection fromAerial
Barbara Gruber · 2007
Inthispaper, we present a newapproach for Building anefficient androbust framework forobject de- automatic cardetection fromaerial images. Thesystem exploits a tection frominaerial imageshasdrawntheattention of robust machine learning methodknownasboosting forefficient research community incomputer vision foryears, e.g.(32), cardetection fromhighresolution aerial images. We propose . ' .' touseon-line boosting withinteractive training framework to (29), (44), (12), (3). Theproblem ofcardetecton fromaerial efficiently train andimprove thedetector. Weuseintegral imagesimages hasavariety ofcivil andmilitary applications, for forfastcomputation offeatures. Thisalsoallows toperformexample transportation control, roadverification tocomplete exhaustive search fordetection ofcarsafter training. Forpostlanduseclassification problem forurban planning, ormilitary processing, weemploy ameanshift clustering method, whichreconnaissance, etc. improves thedetection ratesignificantly. Incontrast torelated work, ourframework doesnotrely on anypriori knowledgeA ntaeia altake obyeairplan ialarg-cale image oftheimagelike asite-model orcontextual information, butif whichcontains alotofobjects withacomplicated background necessary this information canbeincorporated. Anextensive set oftheurbanscene. Forexample, UltraCamD camerafrom ofexperiments onhighresolution aerial images using thenew Microsoft-Vexcel candeliver large format panchromatic im- UltraCamD showsthesuperiority ofourapproach.