In order to solve the relocalization problem of humanoid robots under large initial pose uncertainty, long-term drift and kidnapped scenarios in known point-cloud maps, an autonomous relocalization method combining global coarse localization and local fine registration was proposed. A tightly coupled LiDAR–IMU odometry module was used as the basis of local continuous state estimation. After relocalization was triggered, candidate poses were generated in the traversable area of the global map, and global candidate search was carried out according to the current laser observations. Then, local fine registration was performed in high-confidence candidate regions, and a comprehensive evaluation mechanism based on observation matching quality, map structural consistency and short-term stability was introduced to realize robust candidate selection and result verification. A two-dimensional simulation platform for relocalization in known point-cloud maps was established, and comparative experiments between the proposed method and conventional methods, as well as ablation experiments on the evaluation function, were conducted in standard and kidnapped scenarios. The results show that the proposed method achieves a relocalization success rate of 100% in the standard scenario. In the kidnapped scenario, the relocalization success rate reaches 85%, and the position error and heading error are significantly lower than those of the baseline method. The complete evaluation function also outperforms the simplified scoring strategies in success rate, relocation error and pose recovery accuracy, especially under high disturbance and highly repetitive environments. It is concluded that the proposed method can effectively improve the autonomous relocalization ability of humanoid robots in complex indoor environments and provide reliable pose support for subsequent navigation.