A multi-factor graph optimization-based mobile robot localization algorithm is proposed, utilizing multi-scale point cloud registration, in order to address the issue of unstable performance of point cloud registration affecting the localization accuracy of mobile robots in real-world environments. First, a multi-scale registration model is established, where point and surface element features are associated through attribute filtering in the octree-based feature map, and the pose is iteratively optimized to improve the stability of point cloud registration. Secondly, a sliding window optimization model based on factor graphs is constructed, incorporating multiple constraint factors from LiDAR, IMU, and wheel encoders to optimize the pose and enhance localization accuracy. Experimental results in multiple scenarios show that, in indoor environments, the proposed algorithm reduces the average localization error by 57.79% and 33.71% compared to the LIO-mapping and LIO-SAM algorithms, respectively; in outdoor environments, the average localization error is reduced by 74.00% and 59.37%, respectively, compared to the LIO-mapping and LIO-SAM algorithms. The proposed algorithm significantly improves the localization accuracy and stability of mobile robots.