Detecting Building Entrances on Street View Images Using Deep Learning for Supporting Indoor-Outdoor Seamless Services

Spatial data is important for virtually representing the real world and is essential in developing applications for making informed decisions. With the growing interest in seamless indoor-outdoor environments, spatial data from different sources exists in various formats for use in LBS (Location-Based Services). Previous research has utilized deep learning for indoor omnidirectional images to generate NRS (Node-Relation Structure), a network-based topological data, for supporting spatial analysis for navigation while providing visualization. This study proposes an approach to detect building entrances in street view omnidirectional images through a deep learning-based object detection algorithm for supporting indoor-outdoor LBS. This paper focuses on formulating refinement conditions for constructing an image training dataset that combines both an open dataset and directly captured omnidirectional images to address the challenge of establishing a huge volume of images for training the object detection model. By applying the conditions, the mAP (mean Average Precision) of 61.20% obtained from training with open data increased to 85.72%, and applying image augmentation methods improved the mAP to 87.42%. These results show that the proposed conditions can be used as a framework for constructing generalized training data that results in accurate entrance detection in street view images, regardless of the study area.