A real-time image retrieval and localization method based on 360-degree panoramic visual feature maps

Accurate and reliable localization in large indoor environments without satellite signals remains a significant challenge. In recent years, visual localization has emerged as a popular indoor localization method. Its core idea is to pre-built a 3D sparse feature map database and estimate the 6-DoF pose of query images for precise localization. This technology holds great potential for applications such as augmented reality (AR) and AR navigation in large indoor scenes. However, the presence of weak textures and repetitive textures poses substantial challenges to the pre-built feature map database and image retrieval, severely affecting the accuracy and robustness of localization. In this paper, we propose a real-time image retrieval and localization method based on a 360-degree panoramic visual feature global map. The proposed method consists of three main components: 360° panoramic sparse feature map construction (PGFC); an image retrieval strategy based on point cloud overlap (PCO-IR); visual localization method enhanced by PCO-IR. Extensive experiments demonstrate that our approach surpasses both state-of-the-art research methods and commercial software (e.g., COLMAP, Metashape) in weak-texture and repetitive-texture regions. Across three distinct indoor scenarios, the PCO-IR enhancement yields significant accuracy gains: after optimization, PixLoc and HLOC achieve localization success rates of 95% and 97%, respectively, with mean pose errors reduced to 72% and 37% of their original values. The code for our proposed method can be found at https://github.com/ouwenwu/pco_ir.