Indoor localization system using deep learning based scene recognition

• Published in: Multimedia tools and applications 2022-08, Vol.81 (20), p.28405-28429
• Main Authors: Labinghisa, Boney A., Lee, Dong Myung
• Format: Article
• Language: English
• Subjects: Algorithms; Artificial neural networks; Clustering; Computer Communication Networks; Computer Science; Data Structures and Information Theory; Deep learning; Electronic devices; Errors; Fingerprinting; Image segmentation; Indoor environments; Machine learning; Modules; Multimedia Information Systems; Object recognition; Signal strength; Special Purpose and Application-Based Systems
• ISSN: 1380-7501; 1573-7721
• DOI: 10.1007/s11042-022-12481-3

One of the major problems of indoor positioning using Wi-Fi is the unstable wireless signal in a large area inside the building. In some buildings, signal path loss occurs due to the partition separating the space. In this case, since it is difficult to identify a specific partition when indoor positioning is performed using only received signal strength indicator (RSSI), a partitioning or segmentation algorithm is required to solve this problem. The proposed system in this paper consists of a scene recognition algorithm suitable for a given environment, an image-based indoor location awareness algorithm (IILAA), and a clustering algorithm that connects these modules to identify not only the exact spatial location but also the scene cluster of the user. The system is composed of a scene recognition module which identifies an image taken by the user. A convolutional neural network trained on resnet50 architecture will identify the scene and forward the classified scene result to another module name mapping algorithm module where KNN is used to process the RSSI received by the user mobile device. As a result of the experiment, first, it was confirmed that the proposed system can recognize an average of 93.07% of a given indoor environment. In addition, it was confirmed that the average error distance of the proposed system is 1.31 m, which is about 6.2 times more accurate than 8.15 m, which is the average error distance of fingerprinting using RSSI. The main factor of this performance result is that the proposed system appropriately selects the classified cluster for each scene in the indoor space. Second, it was confirmed that the proposed IILAA’s error distance performance was 1.75 m in an environment in which both clustering/non-clustering fingerprint maps were combined, which is about 3.90 times more accurate than the 6.90 m average error distance of conventional fingerprinting. Finally, the trained model was used on MIT67 dataset to compare its 87. 6% accuracy with other approaches.