Change Detection with Global Viewpoint Localization

Abstract

In this study, we address the problem of change detection in robotic mapping and localization (e.g., SLAM) applications from the novel perspective of global viewpoint uncertainty. Our goal is to develop a generic framework that can be used to detect changes under multi-modal viewpoint estimation by global viewpoint localization. To achieve this goal, we introduce a global viewpoint localization module powered by deep local convolutional features with bag-of-local-features scene model. We then study the effects of global viewpoint uncertainty in change detection tasks and develop two complementary approaches: feature matching (i.e., no-change detection) and anomaly detection (i.e., change detection). We evaluate the effectiveness of individual techniques as well as their combinations using highly dynamic scenes obtained from the publicly available Malaga dataset.