A novel approach for positioning using smartphones and image processing techniques is developed. Using structure\nfrom motion, 3D reconstructions of given tracks are created and stored as sparse point clouds. Query images are\nmatched later to these 3D models. High computational costs of image matching and limited storage require\ncompressing point clouds without loss of positioning performance. In this work, localization is improved and memory\nand storage requirements are minimized. We assumed that the computational speed and, at the same time, storage\nrequirements benefit from reducing the number of points with appropriate outlier detection. In particular, our\nhypothesis was that positioning accuracy is maintained while reducing outliers in a reconstructed model. To evaluate\nthe hypothesis, three methods were compared: (i) density-based (Sotoodeh, International Archives of\nPhotogrammetry, Remote Sensing and Spatial Information Sciences XXXVI-5, 2006), (ii) connectivity-based (Wang et\nal. Comput Graph Forum 32(5):207ââ?¬â??10, 2013), and (iii) our distance-based approach. In tenfold cross-validation,\napplied to a pre-reconstructed reference 3D model, localization accuracy was measured. In each new model, the\npositions of test images were identified and compared to the according positions in the reference model. We\nobserved that outlier removal has a positive impact on matching run-time and storage requirements, while there are\nno significant differences in the localization error within the methods. That confirmed our initial hypothesis and allows\nmobile application of image-based positioning.
Loading....