Improved Seq SLAM for Real-Time Place Recognition and Navigation Error Correction

摘要

Place recognition plays an important role in long term navigation in challenging environment and Seq SLAM has achieved quite remarkable results. In this paper, we mainly adopt three strategies to improve the original Seq SLAM algorithm: integrating Seq SLAM with odometry, optimizing sequence searching strategy and multi-scale sequence matching. The improved algorithm is evaluated using the KITTI dataset. The template library is created online using navigation information from the sliding-window visual-inertial odometer. When a place is recognized, the corresponding information is used as observation of the filter. The result shows the superiority of the proposed method in real-time place recognition. The optimized sequence searching strategy performs much better in minor deviations. Meanwhile, the advantages of longer sequence match (higher recall rate) and short sequence match (precise location) are combined together. At last, the navigation errors are greatly reduced by close-loop detection. The overall position error of odometer with Seq SLAM is 20.3m (0.55% of the trajectory), which is much smaller than the navigation errors of the single odometer (32.0m, 0.86%).