Hier finden Sie wissenschaftliche Publikationen aus den Fraunhofer-Instituten.

Change detection in UAV video mosaics combining a feature based approach and extended image differencing

: Saur, Günter; Krüger, Wolfgang

Volltext urn:nbn:de:0011-n-4324726 (1.2 MByte PDF)
MD5 Fingerprint: f5ab1ad04fa03c9352a83713cf46cff5
Erstellt am: 7.2.2017

Halounova, L. ; International Society for Photogrammetry and Remote Sensing -ISPRS-:
XXIII ISPRS Congress 2016. Commission VII : 12-19 July 2016, Prague, Czech Republic
Istanbul: ISPRS, 2016 (ISPRS Archives XLI-B7)
International Society for Photogrammetry and Remote Sensing (ISPRS Congress) <23, 2016, Prague>
Konferenzbeitrag, Elektronische Publikation
Fraunhofer IOSB ()
change detection; image differencing; directed change mask; video mosaicking; image registration; UAV

Change detection is an important task when using unmanned aerial vehicles (UAV) for video surveillance. We address changes of short time scale using observations in time distances of a few hours. Each observation (previous and current) is a short video sequence acquired by UAV in near-Nadir view. Relevant changes are, e.g., recently parked or moved vehicles. Examples for non-relevant changes are parallaxes caused by 3D structures of the scene, shadow and illumination changes, and compression or transmission artifacts. In this paper we present (1) a new feature based approach to change detection, (2) a combination with extended image differencing (Saur et al., 2014), and (3) the application to video sequences using temporal filtering. In the feature based approach, information about local image features, e.g., corners, is extracted in both images. The label new object" is generated at image points, where features occur in the current image and no or weaker features are present in the previous image. The label “vanished object” corresponds to missing or weaker features in the current image and present features in the previous image. This leads to two “directed” change masks and differs from image differencing where only one “undirected” change mask is extracted which combines both label types to the single label “changed object”. The combination of both algorithms is performed by merging the change masks of both approaches. A color mask showing the different contributions is used for visual inspection by a human image interpreter.