Abstract: In camera trajectory estimation,pose graph is one of the most widely used tools to reduce the cumulative error.How-ever,the scale of the pose graph grows as the cameras move,which would render trajectory estimation impossible in the real-time required applications like AR/VR(augmented reality/virtual reality).To reduce the size of the pose graph optimization,this paper proposes an algorithm that segments the trajectory of cameras incrementally by the change of orientation.The orientation changes significantly where the trajectory is segmented by the proposed algorithm.Efficiency is improved by optimizing the cameras with obvious orientation changes instead of the entire trajectory.In addition,the starting camera and the ending camera of the trajectory segment are utilized as references to estimate different poses for each camera within the trajectory segment,followed by the weighted average method is used on its different poses to solve the final pose.Which not only avoids a large amount of computation of nonlinear optimization,but also reduces the influence of noise to achieve high accuracy.Experiments on EuRoC,TUM,and KITTI datasets show that the proposed scheme reduces the size of the pose graph optimization and ensures the accuracy of trajectory.

Key words: Pose graph, Camera orientation, Trajectory segmentation, Weighted average, Incremental