Real-Time Visual Place Recognition Based on Analyzing Distribution of Multi-scale CNN Landmarks

What makes visual place recognition difficult to solve is the variation of the real-world places. In this work, an effective similarity measurement is proposed for visual place recognition in changing environments, based on Convolutional Neural Networks (CNNs) and content-based multi-scale landmarks. The image is firstly segmented into multi-scale landmarks with content information in order to adapt variations of viewpoint, then highly representative features of landmarks are derived from Convolutional Neural Networks (CNNs), which are robust against appearance variations. In the similarity measurement, the similarity between images is determined by analyzing both spatial and scale distributions of matched landmarks. Moreover, an efficient feature extraction and reduction strategy are proposed to generate all features of landmarks at one time. The efficiency of the proposed method makes it suitable for real-time applications. The proposed method is evaluated on two widespread datasets with varied viewpoint and appearance conditions and achieves superior performance against four other state-of-the-art methods, such as the bag-of-words model DBoW3 and the CNN-based Edge Boxes landmarks. Extensive experimentation demonstrates that integrating global and local information can provide more invariance in severe appearance changes, and considering the spatial distribution of landmarks can improve the robustness against viewpoint changes.