To address the challenges of significant accumulated errors during large-scale loop closure detection in LiDAR odometry, and the increasing computational cost of inverse operations in conventional iterative SLAM solutions as problem size grows. We propose an incremental smoothing and optimization algorithm based on Bayesian tree reconstruction. The method begins by converting the factor graph into a Bayesian network through an elimination process. Subsequently, a bottom-up approach is used to establish an optimal solution path from parent nodes to leaf nodes. The computation incorporates localized updates to the Bayesian tree, significantly reducing the scale compared to the global reordering required by traditional root mean square incremental smoothing algorithms, thereby improving computational efficiency. This makes the Bayesian tree based incremental smoothing algorithm more suitable for efficient backend optimization in autonomous vehicle localization systems. Validation on the widely used KITTI dataset, as well as data collected from real vehicles, demonstrates the proposed framework’s real-time performance and robust overall effectiveness.