An accurate indoor localization approach using cellphone camera

Indoor localization has attracted more and more attention for its wide applications in location-based services. This paper proposes a new indoor localization mechanism using photos taken by cellphone camera. By scenario selection and image matching in an already built image database, we try to get the “closest” image and then utilize its reserved information to achieve the accurate indoor localization. We take advantage of convolutional neural networks (CNNs) and unsupervised learning for scenario selection which can effectively reduce the size of candidate image set, and then exploit ASIFT (Affine Scale Invariant Feature Transform) with epipolar geometry constraint for image matching within selected scenario. Our system can not only achieve sub-meter localization result but also obtain the orientation information of users, without any infrastructure but a mobile camera module. The proposed system is tested in a teaching building and achieved the promising results.