A SLAM Pose Graph Optimization Method Using Dual Visual Odometry

Abstract

<p indent=0mm>Backend trajectory optimization is an important part of the visual simultaneous localization and mapping system, which can significantly improve localization accuracy. However, the existing optimization methods based on the bundle adjustment have a large amount of calculation in large scenes and cannot be applied to end-to-end visual odometries. To solve this problem, a universal backend pose graph optimization algorithm with two visual odometries at the front end is proposed, which can be applied to end-to-end visual odometries. This method uses a high-speed but low-precision end-to-end visual odometry to run at high frequency, while a low-speed but high-precision visual odometry runs at a low frequency. Local optimization uses Gauss-Newton method iterative optimization through the constraints provided by two odometries. Global optimization is performed simultaneously which based on key frames scene matching. Experiments show that the simultaneous localization and mapping system which apply this optimization method can run in real-time on the KITTI dataset. Compared with the two visual odometries, the accuracy has been greatly improved. And compared with several well-known open source simultaneous localization and mapping methods that apply backend trajectory optimization, low errors have been achieved in trajectory error, absolute translational error, rotation error and relative pose error, taking into account the advantages of the accuracy of traditional methods and the advantages of high speed end-to-end methods. In addition, the optimization framework can also be applied to other more visual odometries.