Place Recognition Based Visual Localization Using LBP Feature and SVM

This paper presents a visual localization method based on HOG-LBP and disparity information using stereo images. The method supposes the availability of a database composed with geo-referenced images of the traveling environment. Given an image, the method consists in searching the similar image in the geo-referenced database using SVM (support vector machine) image recognition model. To perform that, a global descriptor obtained by concatenating LBP (Local Binary Pattern) descriptors and HOG features built from the gray-scale image and its disparity map is constructed. Then, a SVM recognition model built on the global descriptors was used to identify the top best similar images. The matched image (from the reference database) to the given image is finally determined using a probability threshold. If no candidate can be selected, the current position is estimated by extrapolating the previous known positions. The integration of disparity information into HOG-LBP is valuable to decrease perceptual aliasing problems in case of bidirectional trajectory situation. To show its effectiveness, the proposed method is tested and evaluated using real data sets acquired in outdoor environments.