Visual Localization Using Voting Based Image Retrieval and Particle Filtering in Indoor Scenes

We propose an image retrieval based method for visual localization in indoor scenes, provided that a geotagged image database in indoor environments is given. For image retrieval, we introduce a voting based image similarity which is robust to geometric image transformations and occlusions. In order to further improve the performance of image retrieval, we introduce two additional procedures: multiple voting and a ratio test. These two procedures are effective in increasing the true positives and in decreasing the false positives, respectively. In addition, we introduce a particle filter to smoothly estimate the trajectory of a moving camera used for visual localization. In experiments with real images captured at an university library, we show that the proposed method outperforms a structure-from-motion based method.