Marker Detection in Aerial Images

Handle URI:
http://hdl.handle.net/10754/623123
Title:
Marker Detection in Aerial Images
Authors:
Alharbi, Yazeed ( 0000-0002-8073-1959 )
Abstract:
The problem that the thesis is trying to solve is the detection of small markers in high-resolution aerial images. Given a high-resolution image, the goal is to return the pixel coordinates corresponding to the center of the marker in the image. The marker has the shape of two triangles sharing a vertex in the middle, and it occupies no more than 0.01% of the image size. An improvement on the Histogram of Oriented Gradients (HOG) is proposed, eliminating the majority of baseline HOG false positives for marker detection. The improvement is guided by the observation that standard HOG description struggles to separate markers from negatives patches containing an X shape. The proposed method alters intensities with the aim of altering gradients. The intensity-dependent gradient alteration leads to more separation between filled and unfilled shapes. The improvement is used in a two-stage algorithm to achieve high recall and high precision in detection of markers in aerial images. In the first stage, two classifiers are used: one to quickly eliminate most of the uninteresting parts of the image, and one to carefully select the marker among the remaining interesting regions. Interesting regions are selected by scanning the image with a fast classifier trained on the HOG features of markers in all rotations and scales. The next classifier is more precise and uses our method to eliminate the majority of the false positives of standard HOG. In the second stage, detected markers are tracked forward and backward in time. Tracking is needed to detect extremely blurred or distorted markers that are missed by the previous stage. The algorithm achieves 94% recall with minimal user guidance. An average of 30 guesses are given per image; the user verifies for each whether it is a marker or not. The brute force approach would return 100,000 guesses per image.
Advisors:
Wonka, Peter ( 0000-0003-0627-9746 )
Committee Member:
Heidrich, Wolfgang ( 0000-0002-4227-8508 ) ; Ghanem, Bernard ( 0000-0002-5534-587X )
KAUST Department:
Computer, Electrical and Mathematical Sciences and Engineering (CEMSE) Division
Program:
Computer Science
Issue Date:
9-Apr-2017
Type:
Thesis
Appears in Collections:
Theses

Full metadata record

DC FieldValue Language
dc.contributor.advisorWonka, Peteren
dc.contributor.authorAlharbi, Yazeeden
dc.date.accessioned2017-04-10T14:27:06Z-
dc.date.available2017-04-10T14:27:06Z-
dc.date.issued2017-04-09-
dc.identifier.urihttp://hdl.handle.net/10754/623123-
dc.description.abstractThe problem that the thesis is trying to solve is the detection of small markers in high-resolution aerial images. Given a high-resolution image, the goal is to return the pixel coordinates corresponding to the center of the marker in the image. The marker has the shape of two triangles sharing a vertex in the middle, and it occupies no more than 0.01% of the image size. An improvement on the Histogram of Oriented Gradients (HOG) is proposed, eliminating the majority of baseline HOG false positives for marker detection. The improvement is guided by the observation that standard HOG description struggles to separate markers from negatives patches containing an X shape. The proposed method alters intensities with the aim of altering gradients. The intensity-dependent gradient alteration leads to more separation between filled and unfilled shapes. The improvement is used in a two-stage algorithm to achieve high recall and high precision in detection of markers in aerial images. In the first stage, two classifiers are used: one to quickly eliminate most of the uninteresting parts of the image, and one to carefully select the marker among the remaining interesting regions. Interesting regions are selected by scanning the image with a fast classifier trained on the HOG features of markers in all rotations and scales. The next classifier is more precise and uses our method to eliminate the majority of the false positives of standard HOG. In the second stage, detected markers are tracked forward and backward in time. Tracking is needed to detect extremely blurred or distorted markers that are missed by the previous stage. The algorithm achieves 94% recall with minimal user guidance. An average of 30 guesses are given per image; the user verifies for each whether it is a marker or not. The brute force approach would return 100,000 guesses per image.en
dc.language.isoenen
dc.subjectvisionen
dc.subjectcomputeren
dc.subjectdetectionen
dc.subjecttrackingen
dc.subjectobjecten
dc.subjectdescriptionen
dc.titleMarker Detection in Aerial Imagesen
dc.typeThesisen
dc.contributor.departmentComputer, Electrical and Mathematical Sciences and Engineering (CEMSE) Divisionen
thesis.degree.grantorKing Abdullah University of Science and Technologyen_GB
dc.contributor.committeememberHeidrich, Wolfgangen
dc.contributor.committeememberGhanem, Bernarden
thesis.degree.disciplineComputer Scienceen
thesis.degree.nameMaster of Scienceen
dc.person.id133407en
All Items in KAUST are protected by copyright, with all rights reserved, unless otherwise indicated.