Loop closure detection for visual SLAM systems using convolutional neural network

Abstract

This paper is concerned of the loop closure detection problem, which is one of the most critical parts for visual Simultaneous Localization and Mapping (SLAM) systems. Most of state-of-the-art methods use hand-crafted features and bag-of-visual-words (BoVW) to tackle this problem. Recent development in deep learning indicates that CNN features significantly outperform hand-crafted features for image representation. This advanced technology has not been fully exploited in robotics, especially in visual SLAM systems. We propose a loop closure detection method based on convolutional neural networks (CNNs). Images are fed into a pre-trained CNN model to extract features. We pre-process CNN features instead of using them directly as most of the presented approaches did before they are used to detect loops. The workflow of extracting CNN features, processing data, computing similarity score and detecting loops is presented. Finally the performance of proposed method is evaluated on several open datasets by comparing it with Fab-Map using precision-recall metric.