Abstract: To solve the large-search-space and low-efficiency problems of traditional global localization algorithms in the large-scale scene, a high-efficient global localization method is proposed for mobile robots on the basis of the feature heat map. Firstly, the point/line information of the environmental grid map is extracted by combining the ray casting and point/line feature detection algorithms. Meanwhile, a robust processing strategy considering random errors is formulated, to construct the feature heat map with robust feature information. In addition, the similar strategy is applied to extracting the point/line feature information from real LiDAR point cloud. Moreover, the extracted feature information is matched with the feature heat map to obtain the initial search space. Finally, the maximum likelihood estimation localization and Monte Carlo localization methods are improved, thus the high-efficient global localization in the initial search space is accomplished. Verification tests are carried out in simulation scenarios and real large-scale workshops. Compared with the traditional maximum likelihood estimation localization method, the computing time is reduced by 90.37% averagely, while the success rate is 5.92 times higher averagely than the Monte Carlo localization method. Results show that the proposed method greatly reduces the localization search space and effectively improves the speed and accuracy of global localization.