Geotagging Urban Elements Using Deep-Learning and Street-Level Video

摘要

There are several works using computer vision to create databases with geotagged elements from urban areas, however, most of them are based on existing images provided by a third-party. In order to maintain up-to-date information that helps governmental, educational, economical, and private sectors decide how to invest resources, it is essential to be capable of acquiring the data on-demand. This research proposes building a prototype to capture street-level videos, using deep learning to detect the relevant objects, employing computer vision algorithms to track them, and applying least-squares to calculate their geotags. Experimental results on the object detection task using a Faster-RCNN model are carried out on a dataset of real urban imagery taken by the prototype in the city of Culiacan, showing positive results despite the high variety of appearances within objects of the same class and the challenging visual conditions of the urban scenery.