A Lightweight Image-Based Indoor Localization System with BIM

Indoor localization based on images is a crucial technology for realizing intelligent building services. However, existing methods face significant challenges. The traditional process of acquiring indoor image data is time-consuming and labor-intensive. Additionally, it is difficult for positioning schemes to achieve both high precision and high efficiency simultaneously. To address these issues, this study presents a lightweight indoor localization method based on building information modeling (BIM) cross-domain image retrieval. Firstly, the Revit application programming interface is used to automatically generate BIM-rendered images and label location information, effectively solving the problem of low data acquisition efficiency. Secondly, by integrating the convolutional block attention module (CBAM) attention mechanism and spectrum normalization techniques, an improved CycleGAN model is proposed. This model can effectively reduce the domain differences between BIM images and real indoor images, enhancing the cross-domain retrieval accuracy by 23%. Finally, MobileNetV3-small and k-dimensional (k-d) trees are employed to achieve lightweight feature extraction and fast matching, with a single-frame processing speed of 6.1 ms achieved on embedded devices. Experiments conducted on the public University of Melbourne corridor data set and the self-built data set of Shandong Jianzhu University demonstrate that the localization accuracy reaches 95% and 88% respectively. This represents increases of 62% and 23% compared to non-cross-domain direct retrieval methods. The indoor localization method based on BIM cross-domain image retrieval proposed in this study, through the improvement of the CycleGAN-MobileNetV3 small and k-d tree architecture, realizes a high-precision and high-efficiency lightweight positioning solution. It significantly cuts down labor costs and system energy consumption, providing a new scalable real-time positioning.