Visual indoor localization in complex scenes based on image semantic and spatial features

In complex and high precision indoor scenarios without pre-deployed facilities, traditional visual indoor localization methods urgently need improvement due to the low retrieval efficiency of offline database images and reduced positioning accuracy under dynamic environments and lighting changes. This paper aims to propose an efficient and accurate indoor localization method. To this end, a visual localization method combining semantic and spatial features is adopted. First, common indoor infrastructure objects are taken as semantic features, and an offline database is constructed by using a hybrid semantic segmentation algorithm based on K-Net + UperNet. Then, coarse semantic retrieval is carried out to determine the positioning range, reducing retrieval time and avoiding accuracy issues caused by structural similarities. Finally, a generative adversarial network is utilized to extract spatial features as descriptors for fine retrieval, and accurate positioning is achieved by combining with prior position information. Experiments on the self-built dataset of Shandong Jianzhu University show that the hybrid semantic segmentation model performs excellently. The introduction of the dynamic convolutional kernel module of K-Net improves the model performance, and the retrieval method based on spatial and semantic features significantly enhances the retrieval accuracy. The newly proposed method in this study effectively improves the efficiency and accuracy of indoor positioning, providing a new solution for localization in complex environments.