Mining DCNN landmarks for long-term visual SLAM

Abstract

Long-term visual SLAM, in familiar, semi-dynamic, and partially changing environments is an important area of research in robotics. The main problem we faced is the question of how to describe a scene discriminatively and compactly-both of which are necessary in order to cope with changes in appearance and a large amount of visual information. In this study, we address the above issues by mining visual experience. Our strategy is to mine a library of raw visual images, termed visual experience, to find the relevant visual patterns to effectively explain the input scene. From a practical point of view, our work offers three main contributions over the previous work. First, it is the first application of discriminative visual features from deep convolutional neural networks (DCNN) to the task of visual landmark mining. Second, we show how to interpret a high-dimensional DCNN feature to a compact semantic representation of visual word. Third, we show that our approach can turn the scene description task with any feature (including the DCNN feature) into the task of mining visual experience. Experiments on a challenging cross-domain visual place recognition validate efficacy of the proposed approach.