Robust Topological Navigation via Convolutional Neural Network Feature and Sharpness Measure

Visual navigation for mobile robots has emerged in recent years. Among the various methods, topological navigation using visual information provides a scalable map representation for large-scale mapping and navigation. A topological map is essentially a graph with keyframes as its nodes and adjacency relations as its edges. Previous topological mapping uses local feature descriptors, such as scale-invariant feature transform or Speeded-Up Robust Features, to select keyframes in mapping, localization, and estimate relative pose. In practice, local features are not robust for severe motion blur or large illumination change. In this paper, we improve topological mapping to make it more efficient and robust. First, we use a convolutional neural network (CNN) feature as the holistic image representation. The CNN feature can be used to effectively retrieve keyframes that have similar appearance from a topological map, and it is robust to motion blur and illumination change. Thus, it improves the performance for place recognition and robot relocalization. Second, we use sharpness measure to select high-quality keyframes and avoid selecting blurry ones. Third, an efficient and robust non-rigid matching method, vector field consensus, is used for efficient geometric verification and to retrieve the most similar keyframe. The qualitative and quantitative experimental results demonstrate that our method is satisfactory.