An Efficient Visual Place Recognition System by Predicting Unique Features

Abstract

Visual place recognition (VPR) is the challenge of determining whether a sensor is visiting a previously recorded location or exploring new locations using visual inputs. VPR approaches typically presume that the appearance remains the same as the moment the map (reference) was produced. This presents a significant challenge, since the premise of static appearance is invalid. Instead, the environment is constantly changing because of weather, time of day, building sites, the upgrading of facades and billboards, and so on. A prominent way to deal with the resilience of environmental change is to demand the selection of features from unique and non-unique objects. By doing this, a method can properly discriminate the image, but it is computationally expensive. In this paper, we seek to recognize a place efficiently by reducing the number of its features. In particular, we predict unique features and avoid using features from non-unique objects by taking advantage of geo-tags. Our method provides increased accuracy with lower computational costs compared with other state-of-the-art methods.