Local patch descriptor using deep convolutional generative adversarial network for loop closure detection in SLAM

Abstract

Recently, the augmented reality and virtual reality fields have been actively researched and a lot of major companies have been aggressively investing the fields. On the core of the research field, the simultaneous localization and mapping (SLAM) algorithm which estimates the camera's position in a global coordinate and simultaneously constructs a 3D environment map firmly settled. Among typical components of modern SLAM framework, we are focusing on a loop closure detection for determining whether the current position of a robot agent was visited previously. The conventional algorithms for the loop closure detection relied on clustering hand-crafted features like SIFT, SURF, and ORB which appear a weakness to handle variations in the image such as a viewpoint change, illumination change, deformation, and occlusion. In this paper, we propose a local patch descriptor using a deep convolutional generative adversarial network to deal with the variations. The experiment result displays the proposed method well clusters the image patches with similar appearances better than the hand-craft features.