Visual localization on NTU campus

Abstract

The advancement of technology in localization system has been growing in demand among the industries in the modern civilisation. Localization system technology enables location identification of an environment depending on the location of the user or device. Numerous data can be extracted from location information of an environment such as set of images, 3D scene models and points cloud. With the extracted data, it can be used in a localization framework of Visual Localization. Visual localization estimates the 6 Degree-of-Freedom (DoF) camera pose from an image relative to a reference scene representation, which allows it to be feasible for indoor and outdoor environments. The objective of this project is to focus on improving and implementing a visual localization framework that can detect the accurate position of a user with a picture taken from gadgets such as smartphones. The project application aids the user to navigate and locate oneself within the NTU Campus. However, visual localization framework has limitations that has yet to overcome such as dynamic scenes with moving objects, changes in the lighting and shadow from day to night. As such, this report explores the existing research methods, mainly scene coordinate regression module of visual localization framework, and improvements that can be implemented to the framework.