<scp>BIM</scp> ‐Based Indoor Navigation Using End‐to‐End Visual Localization and <scp>ARCore</scp>

摘要

ABSTRACT Traditional localization methods based on sensors such as Bluetooth and Wi‐Fi struggle to achieve the required accuracy for fine‐grained applications like AR navigation or robot localization. Accurate global position estimation and seamless AR navigation in dynamic indoor environments can significantly enhance the intelligence of indoor scene applications. This paper proposes an as‐built BIM‐based indoor navigation using End‐to‐End visual localization and ARCore. First, based on pre‐captured RGB image sequences, a sparse point cloud of the indoor space can be constructed using Structure from Motion (SFM) methods, and the camera pose of each RGB image can be obtained. At the same time, image features are extracted in a multi‐scale manner using end‐to‐end deep learning, and an image database is constructed for later scene retrieval and matching. Second, Building Information Modeling (BIM) model is constructed for a specific indoor scene, and an automated reconstruction of the indoor navigation topology map is performed based on this model. Finally, by giving a query image, the optimal matching image in the image database is obtained through scene retrieval methods. And the camera pose of the query image is iteratively optimized by multiple feature maps. Then path planning is performed based on the indoor navigation topology map with ARCore. In order to assess the effectiveness of the proposed method, Experimental are conducted for the evaluation of the global positioning estimation module and AR navigation module using both public datasets and our own collected data. The experimental results demonstrate that in publicly accessible datasets, the global positioning error consistently remains within the 20‐cm range, with the majority of these datasets achieving a positioning error of 10 cm or less. Furthermore, the experimental results of AR navigation show that the method enables an approximate positioning accuracy of 10 centimeters in AR navigation within a length of 50 m.