Page - 80 - in Proceedings - OAGM & ARW Joint Workshop 2016 on "Computer Vision and Robotics“

Weobtain several noisy estimates for thepositiontLRF of theLRFthrough thecorrespondence tLRF,i= l3D,i−di · lLRF, (2) where lLRF has been normalized to unit length. We obtain the final estimate for the positiontLRF of theLRFbytakingthemedianofallnoisyestimates. TherotationRLRF isgivenbytheanglebetween theviewingdirection lLRF of the laser rangefinderand thecamerasoptical axis in theplanespanned by theoptical axisof thecameraand lLRF. 4. SparsePoseEstimationand3DSceneReconstruction The proposed approach is structured in steps typical to SfM pipelines: image recording, preprocess- ing, relative pose and motion estimation between views and ultimately 3D reconstruction. Since it is aimed at the reconstruction of building facades, which can to a large extent be modeled as a set of flat surfaces, it is sufficient to reconstruct the building as a wire-frame model using surface vertices together with a few supporting points on the walls. We compute SIFT matches to estimate homo- graphies between image pairs (Ii,Ij), which can be used to establish correspondences based on the known laserpoint l3D,i inIi and its respective2Dposition l2D,i,j inIj. Using an initial set of 4 images with full correspondences and the laser measurements, we are able to initialize and calculate an early estimate for our model and the relative camera poses. Then we iteratively add the remaining cameras and distance measurements and finally refine the poses with a global bundle adjustment. Since we know the respective distance information to each camera pose, this estimation is accurate in its scale. 4.1. ImageRecording Since we perform sparse camera pose estimation and reconstruction, the accuracy of the solution depends upon a few, yet highly significant features which are easily found in images taken from different perspectives. For a good reconstruction, the significant features should be chosen in a way such that they lie on the facade and are well-distributed over its surface including the corner points, e.g. vertices of walls and corners of windows. Figure 1 depicts the data recording process, where we take RGB images from various view points while measuring the distance of a single point in the respective image with the LRF. 4.2. PreprocessingandFeatureExtraction To keep our approach as flexible as possible and to reduce the complexity during manual data ac- quisition, we assume no particular order of the images. Initially all possible image pairs are added to a working set. We extract SIFT features [9] from gray-scale versions of the images and establish pointcorrespondencesusingaFLANN-basedmatcher [11] followedbyLowe’s ratio test tofilterout- liers. Withthesecorrespondences,werobustlyestimateahomographybetweeneachimagepairusing RANSAC[5]witha thresholdof1px. Weonlywant tokeep imagepairswithacertainoverlap in the working set, thus we filter out all with less thann= 10 inliers according to the RANSAC estimate and a ratio of inliers to number of matches of< 50%. As a measure for the quality of the remain- ing image correspondences, we define an errorEi,j for an image pair (Ii,Ij) using the 2D positions PSIFT,i andPSIFT,j of theirmatched featuresas follows: Ei,j =mean(||PSIFT,i−PSIFT,j||2),∀i,j∈N,i 6= j. (3) 80