Visual Place Recognition Using an Unsupervised CNN Approach

Visual place recognition (VPR) plays a key role in many applications, such as mobile robot navigation and localization, location intelligence in smart warehouse etc. Convolutional neural network (CNN) based deep learning (DL) solutions have been demonstrated to outperform traditional bag-of-words (BoWs) solutions for VPR. However, CNN-based DL solutions often require a large amount of high-quality annotated data, where manual annotation is time-consuming and even difficult in some scenarios. In this paper, we propose an unsupervised CNN approach for VPR. First, we use Oriented FAST and Rotated BRIEF (ORB) feature detector, bag-of-words (BoWs) and Hough space verification methods to automatically generate image clusters. Each image cluster is assigned with a numbered keyframe representing one place. Images relating to its keyframes will be stored as a visual experience with relevant available information, such as pose of current keyframe, in the corresponding cluster. Next, the image clusters are used as training datasets to train various CNN models for VPR. Experiments have been conducted on a public dataset and competitive results have been achieved using the proposed unsupervised CNN method.