Seite - 88 - in Joint Austrian Computer Vision and Robotics Workshop 2020

Imageacquisition.Acquisition of a set of SAR im- ages of the area of interest, in optimal case from as- cending and descending orbital direction (cf. Fig- ure 1). In case images are gathered from one orbital direction the stereo intersection angle has to be rea- sonably large (i.e., larger than 10◦). After import each image consists of complex valued pixels plus the according sensor model (i.e., the cocircular ge- ometry based on the Range and Doppler equations [2,10]). SARdelay correction. Adjustment of sensor mod- els, in specific the SAR internal delays in range di- rection, for the following effects (cf. [5]): (1) Iono- sphericsignalpropagationdelaycausedbyelectrons; (2) tropospheric signal propagation delay caused by air conditions, e.g., water vapor; (3) solid earth tides caused by gravity of moon and sun; and (4) plate tectonics, i.e., continental drift. For each image the range correction grid is updated, whereas the under- lying information is gathered from weather and GPS services. Point extraction. Metal objects appear as points or rather bright blobs on dark background (cf. Fig- ure 2). For detection the image is upsampled based on complex FFT oversampling with a factor of 2. Then a matched filter is applied on the amplitude to localize blobs using a spike-shaped template ker- nel (cf. [14]). Results are thresholded and the best matching2Dblob locationsare retrievedbysubpixel interpolation [6,12]. Matching of points. For each stereo pair epipo- lar rectified images using a coarse digital elevation model based on the method [13] are generated, also transferring theextractedpoints. Thismethodundis- torts the images in rangedirectionand thus increases their geometric and radiometric similarity. Those points are then matched by means of normalized cross-correlation (kernel size depends on resolution of the input images). The resulting homologous points are then transformed back into the input im- ages. Retrieval of 3D coordinates. GCPs are calculated by a multi-image least squares spatial intersection of SARrangecirclesyieldingapointcloud. Duetoover determination incorrect points can be detected and rejected. 3.ResultsandConclusion The presented workflow was applied on a multi- tudeofmulti-beamscenesdistributedover thewhole Figure 2. Roundabout traffic as perceived from an air- bornedigitalcamera(top)andfromtheSARsatellite(bot- tom). Thebrightblobs in theSARamplitudecorresponds mainly to light poles. An exemplary pole is highlighted togetherwith its extracted3Dlocation. globe, acquired with various imaging modes (i.e., Stripmap, Spotlight, HS Spotlight, Staring Spotlight [3]). Referencecoordinatesofmetalpolesweremea- sured in-situ with differential GPS with an absolute 3D accuracy of±5 cm such that inaccuracies of the cadastresystemdonotpropagate into theevaluation. Table 1 gives exemplary 3D accuracies (defined as root mean square (rms) values) as can be ex- pected from the proposed methodology. In planime- try around 15 cm are achieved and in height around 20cm,whichare impressivenumbers taking intoac- count thealtitudeof the satellite’sorbit at 514km. East [m] North [m] Height [m] rms 0.14 0.14 0.21 mean 0.04 0.09 0.07 std 0.13 0.10 0.20 min -0.38 -0.26 -0.44 max 0.32 0.26 0.40 Table 1. 3D accuracy evaluation w.r.t. in-situ measure- ments given in meters based on 26 reference points and twooppositeorbitStaringSpotlight images. Future work will deal with automatic transfer of those SAR-based GCPs to optical images by means of multi-modal image matching. Most promising re- cent works use deep learning to tackle this ill-posed issue, for instance, [7,8,1]. 88