Bild

OSMDeepOD - Object Detection on Orthophotos with and for VGI

    Samuel Kurath, Raphael Das Gupta, Stefan Keller

GI_Forum 2017, Volume 5, Issue 2, pp. 173-188, 2017/12/13

Journal for Geographic Information Science

doi: 10.1553/giscience2017_02_s173

doi: 10.1553/giscience2017_02_s173


PDF
X
BibTEX-Export:

X
EndNote/Zotero-Export:

X
RIS-Export:

X 
Researchgate-Export (COinS)

Permanent QR-Code

doi:10.1553/giscience2017_02_s173



doi:10.1553/giscience2017_02_s173

Abstract

A great deal of the interesting information captured by aerial imagery is as yet unused, even though it could help to enrich maps and improve navigation. For this information to be made available, objects such as buildings or roads need to be recognized on images. This is laborious to do entirely manually, but non-trivial to perform computationally. In this paper, we present an automated method for detecting objects of a chosen class (pedestrian crosswalks) on orthophotos, a method which can be adapted for various classes of objects. The method uses a supervised machine-learning approach with a deep convolutional neural network. We re-trained the final layer of a pre-trained neural network using specific imagery and crowdsourced geographic information from the OpenStreetMap (OSM) project. The result is an easily enhanceable and scalable application which is able to search for objects in aerial imagery. We achieved an accuracy of well over 95% for crosswalks and promising preliminary results for roundabouts.

Keywords: visual recognition, deep convolutional neural networks, aerial imagery, VGI, parallelism

Published Online: 2017/12/13 12:19:14

Object Identifier: 0xc1aa5572 0x00373589

Document viewed: Calculating...

Rights: https://creativecommons.org/licenses/by-nd/4.0/