Page - 109 - in Proceedings of the OAGM&ARW Joint Workshop - Vision, Automation and Robotics

Image Registration andObject Detection for Assessing Unexploded Ordnance Risks - A Status Report of the DeVisORProject* Simon Brenner1, Sebastian Zambanini1 and Robert Sablatnig1 I. INTRODUCTION Although the last acts of war in Central Europe date back to the times of World War II, Unexploded Ordnance (UXO) from that period still poses a serious hazard for population and construction projects [3]. For a preliminary estimation of UXO risks, specialized companies retrieve and interpret aerial images from WWII surveillance flights over the area of interest. This process includes the registration of historic aerial images to modern satellite images, and the detection andmappingofcertainobjects that indicate increasedcombat activity in the surveyed area. Currently, these tasks are performed in time-consuming manual work. The DeVisOR project, which was started in 2016 as a cooperation between the Computer Vision Lab and the Information Engineering Group (TU Wien), as well as the Luftbilddatenbank Dr. Carls GmbH as an industrial project partner, aims at supporting the above named tasks with computer vision and visualization techniques. This paper gives a half-time status update of the project achievements as well as an outlook for the final year. II. IMAGE REGISTRATION The registration of WWII aerial images to modern satellite images is particularly challenging because the landscape has changed drastically in the course of seventy years. Not only buildings and roads, but also vegetation, agricultural use and the courses of rivers may have changed, so that it becomes difficult to find reliable common features [4], [5]. Additionally, the available images are partly in suboptimal condition.We thereforeproposea semi-automatic framework for the registration process, in which first the easier task of registering the historical images among each other is performed automatically. Due to the varying conditions even among the historical images (seasonal changes, weather, destruction, image noise) and the absence of a priori in- formation about their relative rotation and translation, only feature-based registration methods, such as SIFT [2], are applicable. We found that automatic scale space feature detection is too unstable for the given image data; however, for each image the approximate aircraft altitude and the focal length of the camera is known. We can therefore normalize the scales of the images and perform a dense *This work is supported by Austrian Research Promotion Agency (FFG) under project grant 850695 1Simon Brenner, Sebastian Zambanini and Robert Sablat- nig are with Faculty of Informatics, Institute of Computer Aided Automation, Computer Vision Lab, TU Wien, 1040 Vienna, Austria sbrenner@caa.tuwien.ac.at, zamba@caa.tuwien.ac.at,sab@caa.tuwien.ac.at (a) Scale space extrema (b) Densely sampled features Fig. 1: Comparison of feature matching stability sampling of features at a fixed scale, which significantly improves the matching stability. Figure 1 shows an example. To refine the resulting registration and account for parallax effects resulting from uneven terrain and different capturing angles,wesuccessfullyappliedadeformablefine registration approach, that was originally designed for the registration of multi-modal medical data [1]. Guided by an interactive visualization of the registration results, the user can then select the most suitable historical image and manually georeference it; all the other images are then registered transitively. We are also working on a novel registration algorithm that is currently able to register about a third of the WWII images in our test data set directly to modern satellite images and thus supplement the above named framework. 109