VPS-based Book Retrieval System Using Smart Devices

This paper proposes a smart device book retrieval system using VPS (visual positioning system) based on camera images in indoor library environments where GPS signals are unavailable. To overcome the infrastructure dependency and radio interference vulnerability of existing indoor positioning technologies such as Wi-Fi, BLE beacons, and UWB, we build a deep learning pipeline integrating DINOv2 feature extraction, FAISS image retrieval, and Reloc3r pose estimation. To address the visual aliasing problem in repetitive bookshelf environments, we propose a spatial graph-based search restriction with 54 clusters and 98 edges. We built a COLMAP 3D map from 14,652 reference images at Sogang University’s Loyola Library and evaluated with 839 query images. DINOv2 reduces feature vector dimensions by 97.6% compared to NetVLAD while improving Top-3 accuracy by 5%p. Cluster-based search achieves 95.7% accuracy within 3m and up to 46% reduction in processing time. A usability study with 11 participants confirmed an SUS score of 81.82 (grade A) and a 25.2% reduction in book retrieval time.