Fraunhofer-Gesellschaft

Publica

Hier finden Sie wissenschaftliche Publikationen aus den Fraunhofer-Instituten.

Deep learning based vehicle detection in aerial imagery

 
: Sommer, L.

:
Volltext urn:nbn:de:0011-n-5069128 (4.5 MByte PDF)
MD5 Fingerprint: 9e01c3e544266eaead78b118c0ff7bb2
Erstellt am: 28.8.2018


Beyerer, Jürgen (Ed.); Pak, Alexey (Ed.); Taphanel, Miro (Ed.):
Joint Workshop of Fraunhofer IOSB and Institute for Anthropomatics, Vision and Fusion Laboratory 2017. Proceedings : July 30 - August 5, 2017, Triberg-Nussbach, Germany
Karlsruhe: KIT Scientific Publishing, 2018 (Karlsruher Schriften zur Anthropomatik 34)
ISBN: 978-3-7315-0779-6
ISBN: 3-7315-0779-X
DOI: 10.5445/KSP/1000081314
S.83-97
Fraunhofer Institute of Optronics, System Technologies and Image Exploitation and Institute for Anthropomatics, Vision and Fusion Laboratory (Joint Workshop) <2017, Triberg-Nussbach>
Englisch
Konferenzbeitrag, Elektronische Publikation
Fraunhofer IOSB ()

Abstract
Detecting vehicles in aerial images is an important task for many applications like traffic monitoring or search and rescue work. In recent years, several deep learning based frameworks have been proposed for object detection. However, these detection frameworks were developed and optimized for datasets that exhibit considerably differing characteristics compared to aerial images, e.g. size of objects to detect. In this report, we demonstrate the potential of Faster R-CNN, which is one of the state-of-the-art detection frameworks, for vehicle detection in aerial images. Therefore, we systematically investigate the impact of adapting relevant parameters. Due to the small size of vehicles in aerial images, the most improvement in performance is achieved by using features of shallower layers to localize vehicles. However, these features offer less semantic and contextual information compared to features of deeper layers. This results in more false alarms due to objects with similar shapes as vehicles. To account for that, we further propose a deconvolutional module that up-samples features of deeper layers and combines these features with features of shallower layers.

: http://publica.fraunhofer.de/dokumente/N-506912.html