Global Consistency Point Cloud Registration: A Multi-Path Bundle Adjustment Approach

We present a novel 3D point cloud registration method based on multi-path registration. The core concept involves utilizing multi-path pairwise registration for pose graph optimization, ensuring globally consistent matching poses. Challenges arise from sensor noise and outliers in overlapping regions, causing errors and incorrect matches during inter-frame registration. These cumulative errors can hinder the alignment of loop closure frames, resulting in a layered effect and a lack of global consistency in the reconstructed model. Single-path registration often relies solely on adjacent frame registration results, leading to the loss of loop closure information and constraints between frames with specific overlapping areas. The optimization method introduced in this paper constructs a graph optimization framework for multiple path registrations. It establishes consistency costs among inter-frame registration, cross-registration, and loop closure registration, tightly integrating the optimization of each frame's pose. Additionally, it assesses the registration confidence based on the similarity of point cloud color intensity information, reducing redundant registration costs to enhance robustness. Experimental results demonstrate that our approach enhances registration accuracy by at least 36.2% compared to frame-by-frame reconstruction method, showcasing its effectiveness and precision.