Indoor visual positioning using stationary semantic distribution registration and building information modeling

Abstract

Indoor Visual Positioning (IVP) is a prerequisite for applications like indoor location-based services in smart buildings. Building Information Modeling (BIM), representing physical and functional characteristics of buildings, is widely used in IVP. Existing BIM-based IVP methods register visual features from sensed images to BIM but suffer inaccuracies caused by dramatic disturbances from unstable objects like chairs. Stationary objects like walls may address this issue and provide a more reliable IVP scheme, yet it remains to be explored. This paper proposes an IVP scheme leveraging stationary object registration from sequential images to BIM, termed Stationary Semantic Distribution-driven Visual Positioning (S2VP). In the offline phase, S2VP generates “stationary semantic distribution-positions” datasets from BIM. During positioning, the stationary semantic distribution of sensed images is first estimated, and the indoor position is computed via a particle filter model. Experiments show that S2VP achieves an average positioning error of 0.37 m, outperforming existing methods.