Graph-based visual SLAM and visual odometry using an RGB-D camera

In this paper we present a real-time graph-based visual SLAM approach. The presented visual SLAM algorithm can be separated into three parts: feature extraction, data association, and SLAM back-end. We use FAST for feature detection and the Binary Robust Independent Elementary Features (BRIEF) as feature descriptor, which together provide a fast and stable feature extraction. The data association is solved using Locality Sensitive Hashing (LSH), which uses local hash tables and profits from binary feature descriptors. As SLAM back-end we use the general graph optimization framework g²o, which is designed to provide solutions to several SLAM variants. We further provide a novel approach to visual odometry by combining recent sensor measurements into a small pose graph and optimizing it using g²o. For finding potential neighbour nodes and loop closures we introduce the Global Feature Repository (GFR). GFR searches for loop closures and potential neighbours independent of their position in the graph. Finally, we show the accuracy and real-time ability of our algorithm by comparing it to a recently published benchmark dataset. We further provide some large-scale datasets using state-of-the-art laser localization as ground-truth.