Abstract: Objectives: Localization is an important module of the LiDAR SLAM system, which provides basic information for perception, control, and planning, further assisting robots to accomplish higher-level tasks. However, LiDAR localization methods still face some problems:the localization accuracy and efficiency cannot meet the requirements of the robot products. In some textureless or large open environments, the lack of features easily leads to dangerous robot kidnappings. Consequently, aiming at the localization problems of mobile robots in large indoor environments, a global localization method based on cylindrical features is proposed. Methods: Firstly, an offline parameterized map is designed, which consists of some map cylinders and a raster map. Because the point cloud map contains a large number of 3D points and complete cylinders, random sample consensus (RANSAC) and geometric models are combined to directly segment the cylindrical points. The raster map is employed to describe the distributions of stable artificial structures. Then, some lightweight binary files are used to offline record the geometric model of cylinders and the feature distribution of the map. Next, based on three unique geometric characteristics of the cylinder (outlier, symmetry, and saliency), a real-time LiDAR point cloud cylinder segmentation method is proposed. Finally, two pose computation strategies are designed. The first is an optimization model based on heuristic search, which searches for the best matching cylinder between the map and real-time point cloud, and calculates the translation and rotation, respectively. The second is an optimization model based on multicylinder constraints, which employs both the topological relation (point-to-point and point-to-line constraints) and geometry attributes to find approximately congruent cylinders, then computes optimal pose. Results: To verify the feasibility of the proposed method, we use a 16-line LiDAR to collect the experimental data in three real-world indoor environments, i.e., lobby, corridor, and hybrid scenarios. The global localization experiment is compared to a similar wall-based localization method, and the loop closure detection is compared to M2DP, ESF, Scan Context, and the wall-based localization. The experimental results show that the proposed method outperforms the baseline methods. The place recognition and localization performance of the proposed method reach the mainstream method level, with a localization success rate of 90% and an error of 0.088m. Some data can reach millimeter localization accuracy, and the fastest speed is within 100ms. Conclusions: The proposed method can effectively realize the global localization and place recognition of the robots in typical open indoor environments. It meets the accuracy and efficiency requirements of autonomous driving for global localization in practical applications. It can be applied to solve the problems of position initialization, re-localization, and loop closure detection.