A METHOD FOR REGIONAL ANALYSIS USING DEEP LEARNING BASED ON BIG DATA OF OMNIDIRECTIONAL IMAGES OF STREETS
- 1Tokyo Institute of Technology, 2-12-1-M1-27 Ookayama, Meguro-ku, Tokyo 152-8550, Japan
- 2The University of Tokyo, 5-1-5 Kashiwanoha, Kashiwa-shi, Chiba 277-8568, Japan
Keywords: Omnidirectional image, Deep learning, Big data, Semantic segmentation, Clustering, Computer vision
Abstract. In this paper, we propose a method for regional analysis using image recognition technology based on deep learning and big data of street images captured by omnidirectional cameras on vehicles. Specifically, we first construct a classification method of regions based on street images using a pretrained deep learning model (VGG16) for image recognition as a feature extractor. Next, we develop a method to evaluate the landscape and safety of streets based on the ratio of street components (such as buildings, roads, fences, vegetations, sky, street lights) at each shooting point, which is calculated by semantic segmentation.