Abstract: To address the issues of low registration accuracy and long registration time in 3D normal distributions transform (3D−NDT) point cloud registration without a determined initial pose, which fails to meet the real-time localization requirements of mobile robots, an improved 3D−NDT algorithm for three-dimensional point cloud registration was proposed. During the down sampling of the original point cloud, points from the source point cloud were used to replace the calculated centroids, thereby reducing computational complexity while preserving the feature information of the point cloud. A dynamic adjustment of the trust radius for iterative step size was By introduced to improve the accuracy and speed of point cloud registration after down sampling. Additionally, By integrating 3D laser point cloud data with 9-axis inertial measurement unit (IMU) data, the problem of convergence failure or entering local extremum when the pose difference between two sets of point cloud data was too large was solved. Simulation experiments on the algorithm in this paper were conducted using a mobile robot platform built in the laboratory to verify the reliability and accuracy of the real-time positioning of the algorithm. The results show that compared with the traditional 3D−NDT algorithm, the matching accuracy of the 3D−NDT algorithm proposed in this paper is improved by 106% and 108% in outdoor and indoor environments, respectively. The matching success rate increases by 10.9% and 10.7%, respectively, and the average matching time reduces by 51.1% and 47.9%, respectively. The 3D−NDT algorithm proposed in this paper significantly enhances registration accuracy in real-time localization for mobile robots and greatly reduces the time required for single registration, which can meet the needs of real-time positioning of mobile robots.