DENSE MULTIPLE STEREO MATCHING OF HIGHLY OVERLAPPING UAV IMAGERY

International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XXXIX-B1, 2012 XXII ISPRS Congress, 25 August – 01 September 2012, Melbourne, Australia DENSE MULTIPLE STEREO MATCHING OF HIGHLY OVERLAPPING UAV IMAGERY Norbert Haala *, Mathias Rothermel Institute for Photogrammetry, University of Stuttgart [firstname.lastname]@ifp.uni-stuttgart.de Commission I, WG I/V KEY WORDS: Platforms, Point Cloud, Matching, DEM/DTM, Surface, High resolution, Three-dimensional, Performance ABSTRACT: UAVs are becoming standard platforms for applications aiming at photogrammetric data capture. Since these systems can be completely built-up at very reasonable prices, their use can be very cost effective. This is especially true while aiming at large scale aerial mapping of areas at limited extent. In principle, the photogrammetric evaluation of UAV-based imagery is feasible by of-theshelf commercial software products. Thus, standard steps like aerial triangulation, the generation of Digital Surface Models and ortho image computation can be performed effectively. However, this processing pipeline can be hindered due to the limited quality of UAV data. This is especially true if low-cost sensor components are applied. To overcome potential problems in AAT, UAV imagery is frequently captured at considerable overlaps. As it will be discussed in the paper, such highly overlapping image blocks are not only beneficial during georeferencing, but are especially advantageous while aiming at a dense and accurate image based 3D surface reconstruction. consequence, standard assumptions and implications used during standard AAT do not hold true anymore. 1. INTRODUCTION UAVs are establishing as serious alternative for traditional photogrammetric data capture. Since these systems can be completely built-up at very reasonable prices, photogrammetric data collection can be very cost effective. This is especially true while aiming at large scale aerial mapping of areas at limited extent. In principle, the photogrammetric evaluation of UAVbased imagery is feasible by of-the-shelf commercial software products. Thus, standard steps like aerial triangulation, the generation of Digital Surface Models and ortho image computation can be performed effectively. However, this processing pipeline can be hindered due to the limited quality of UAV data. This is especially true if low-cost sensor components are applied. Due to these reasons, alternative structure-and-motion algorithms are additionally integrated to guarantee successful block configuration of UAV imagery. Within this context, an increased redundancy of image measurements is used to overcome potential problems in AAT. Thus, UAV imagery is frequently captured at considerable overlaps. As it will be discussed in the paper, such highly overlapping image blocks are not only beneficial during georeferencing, but are especially advantageous while aiming at a dense and accurate image based 3D surface reconstruction. Within our investigations dense 3D point clouds are generated using the Semi-Global Matching (SGM) stereo method. The potential of the SGM algorithm was already demonstrated for different applications and data sets, including aerial images, satellite data or video sequences. This was our motivation to implement and use SGM for dense image matching from multiple overlapping UAV imagery. Since this approach aims at a pixel-wise matching, geometric surface reconstruction is feasible at a resolution similar to the available ground sapling distance of the captured imagery. For our investigations, data from flight campaigns with a low-cost fixed-wing UAV, equipped with a consumer grade digital camera and a low-cost GPS are used. The low cruising speed of the UAV easily allows an image acquisition at large overlaps in flight direction. Additionally, rather short distances between the respective flight lines were selected. This results in overlaps, which allow for a typical visibility of each object point in more than 20 images. This is a considerable increase in redundancy if compared to standard stereo capture by two image rays as available from traditional photogrammetric blocks. Standard aerial image collection is usually combined with differential GNSS measurement to provide the camera stations at centimeter accuracies. Furthermore, digital airborne cameras are frequently mounted to stabilized platforms, which guarantees camera views close to nadir. Finally, integrated GNSS/inertial systems can be used for direct georeferencing. These sensor components are beneficial for automatic block configuration and allow an aerial triangulation at considerable accuracies. In contrast, the evaluation of UAV imagery frequently rests on low-cost navigation grade GNSS measurements, which substantially limits the accuracy of the measured camera station. Automatic aerial triangulation (AAT) is additionally hindered by the high flight dynamics of UAV platforms. These can easily result in significant deviations of the respective images from nadir view. Furthermore, the use of consumer cameras for UAV image collection results in small footprints if compared to digital airborne cameras. In combination with the high flight dynamics this causes considerable deviations in mutual image overlaps. As a Within our implementation, this redundancy is utilized for image based surface reconstruction by linking corresponding pixels from multiple images. For this purpose, a base image is matched against the surrounding images. This provides * Corresponding author. 387 International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XXXIX-B1, 2012 XXII ISPRS Congress, 25 August – 01 September 2012, Melbourne, Australia corresponding measurements, which consist of pixel coordinates of the base image and the multiple pixel coordinates in the respective match images. As demonstrated within the paper, the following point determination from multiple image rays can efficiently eliminate potential mismatches by suitable filter processes. Thus, the accuracy and reliability of image based 3D point cloud generation is increased considerably. By these means high quality surface reconstruction is feasible even from imagery of limited quality as available from UAV-borne consumer grade digital cameras. Our investigations will be verified by for data sets captured at a test site already used for evaluation of digital photogrammetric camera systems. Thus, a comparison of our dense multiple stereo matching of highly overlapping UAV imagery to standard photogrammetric data collection by digital airborne cameras is feasible. difficult landings. Sensors on board are 3-axis gyros, accelerometers, 3-axis magnetometer, barometer and a single channel GPS receiver. During the flight, the information of the sensor readings is used to estimate flight path angles, velocity and position of the airplane. In order to improve the (...truncated)