An Efficient Bundle Adjustment Approach for Stereo Visual Odometry with Pose Consensus

In simultaneous localization and mapping (SLAM) and structure from motion (SFM), bundle adjustment (BA) plays a vital role in adjusting the pose of the 6-DOF camera and the location of the 3D landmark points. However, the usefulness of the BA is affected by the number of optimization variables. Thus, an efficient bundle adjustment (EBA) based on a fast iteration technique can effectively estimate the pose and feature points of the camera. Firstly, for the Levenberg-Marquardt (LM) solution method in the BA algorithm, we use a variable block to add constraints when solving the increment, replacing the original Hessian matrix of the LM to estimate the descending direction and step length faster during the solution process, and obtain the optimal solution of the objective function. Secondly, based on the pose consensus, the manifold alternating direction method of multipliers (MADMM) algorithm is presented in the bundle adjustment problem to speed up the optimization of variables. Finally, the experiment shows that compared with the LM-based BA method, the method proposed in this paper has improved time consumption and reconstruction effect. Simultaneously, the EBA algorithm is utilized to apply stereo visual odometry, and experiments demonstrate that our algorithm is better than the previous.