Image retrieval using BIM and features from pretrained VGG network for indoor localization. (August 2018)

• Record Type: Journal Article
• Authors: Ha, Inhae; Kim, Hongjo; Park, Somin; Kim, Hyoungkwan
• Is Part Of: Building and environment. Volume 140(2018)
• Page Start: 23
• Page End: 31
• Publication Date: 2018-08
• Subjects: Image-based indoor localization -- Cross-domain image retrieval -- BIM -- Feature extraction -- Global descriptor; Buildings -- Environmental engineering -- Periodicals; Building -- Research -- Periodicals; Constructions -- Technique de l'environnement -- Périodiques; Electronic journals; 696
• DOI: 10.1016/j.buildenv.2018.05.026
• Languages: English
• ISSNs: 0360-1323

Abstract

Abstract: Various devices that are used indoors require information regarding the user's position and orientation. This information enables the devices to offer the user customized and more relevant information. This study presents a new image-based indoor localization method using building information modeling (BIM) and convolutional neural networks (CNNs). This method constructs a dataset with rendered BIM images and searches the dataset for images most similar to indoor photographs, thereby estimating the indoor position and orientation of the photograph. A pretrained CNN (the VGG network) is used for image feature extraction for the similarity evaluation of two different types of images (BIM rendered and real images). Experiments were performed in real buildings to verify the method, and the matching accuracy is 91.61% for a total of 143 images. The results also confirm that pooling layer 4 in the VGG network is best suited for feature selection. Highlights: A method utilizing CNN and BIM for image-based indoor localization is proposed. The image most similar to the photo is retrieved from the dataset built from BIM. Features are extracted by the VGG for evaluating similarity between images. The most suitable layer in the VGG for cross-domain image retrieval is identified. A high performance in the experiment verified the proposed method.