Landmark Recognition using Deep Learning in a Virtual Space

Abstract

It is a very important issue to simulate human behavior in a virtual space for emergency evacuation in the real world. Humans take actions using their own eyes and memory. Then, the identification of the scene that virtual humans are looking at in a town is one of the key elements of the behavior, and image-based pattern matching is usually used; however, the accuracy is affected by the view angle and the length between the target object and the position at which the image is taken. This paper proposes a method to identify the images of landmarks that are placed at the corners in an intersection in a virtual space using a deep learning method and reports the relationship between the accuracy and the area rate that the landmark object occupies in the image.