What's in an Embedding? Harnessing Deep Learning to Fuse Multi-Sensor Embeddings for Room Identification
Gabriel A. Morales, Rocky Slavin · 2025
Localization technologies are prevalent in use through our daily lives ranging from map-based navigation, location-based automation, or locating friends. However, while these technologies are well-suited for outdoor applications, transferring them to granular indoor applications lacks the same level of precision. Indoor localization enables useful applications such as object and person tracking, or room-to-room navigation. As part of these applications, an underpinning concept, known as room identification, is understanding exactly which room people or objects are in. While existing techniques use single modalities to achieve this kind of task, they can lack the depth necessary to understand the context available from multiple devices and signals in the area that makes each room unique. In this work, we study room identification using embeddings from sensor fusion via deep learning methods. In addition, we use a vector database for embedding storage and flexibility to perform large-scale similarity searches for the task. Experimenting with different fusions from three modalities (3D meshes, altitude, and link layer RSSI), we demonstrate the possibility of performing room identification with the highest overall F1-Score reaching 89%.