SEMAP: Automatic semantic mapping for large-scale indoor spaces

Abstract

Due to the rapid development of indoor location-based services, semantic mapping becomes a highly promising technique for ubiquitous applications. Semantic mapping is challenging due to semantic diversity and indoor environment variations. This paper presents a novel semantic mapping method (namely SeMap) that uses crowdsourced images and motion traces. It first constructs a scene graph, which serves as the foundation for the semantics localization stage. It then extracts semantics, estimates the local pose of each image using crowdsourced images, and calculates the global pose of each image using motion traces of users. Furthermore, by combining multiple viewpoints, a semantic localization algorithm is proposed to label semantics of general entities in an indoor map. Extensive experiments are conducted on a real dataset. The experimental results indicate that our method achieves an average location error of 1.05m. It demonstrates the advantage of our proposed method.