Exact Point Cloud Downsampling for Fast and Accurate Global Trajectory Optimization

Global trajectory optimization is a crucial step for localization and mappin systems. Because it is unavoidable that trajectory errors accumulate in online odometry estimation that performs real-time optimization using local observations, it is necessary to correct estimation drift by considering the global consistency of the map. Global registration error minimization is one of the most accurate approaches to global trajectory optimization [1]. Unlike the conventional pose graph optimization that minimizes the errors in the pose space [2], global registration error minimization directly minimizes the multi-frame point cloud registration errors over the entire map. This approach avoids the Gaussian approximation of the relative pose constraint and enables accurate trajectory optimization by jointly Figure 1: Dense factor graph for global registration error aligning all frames in the map [3]. However, it is known minimization. The proposed algorithm reduces memory consumption to be computationally expensive compared to pose graph by 99% and processing time by 87% for the optimization, as it requires a re-evaluation of registration optimization of the factor graph. error functions that involve residual computations for many points [4].