Abstract: In complex environments, the geometrical structures aren’t very different and the data features in point cloud may degrade, which easily cause the mismatches of loop closure and the increase of solution errors in back-end map optimization, and restrict the accuracy and reliability of mobile robots in mapping and localization. For this issue, an approach for constructing back-end constraints in LiDAR SLAM（simultaneous localization and mapping） based on an adaptive adjustment mechanism is proposed. Firstly, a dynamic adjustment approach of the search window for loop closure detection is designed based on the matching uncertainty. A shortest distance matrix for submaps is constructed and dynamically updated using an improved Floyd algorithm, enabling real-time quantification of the relative uncertainty between different matching nodes. Based on the quantified model indicators, a dynamic adjustment mechanism is designed to adapt the matching search domain during constraint establishment. Secondly, an adaptive threshold approach is proposed from the perspective of laser point cloud data specificity. By quantifying the point cloud specificity through candidate solution scores, the score threshold in scan matching is dynamically adjusted during the constraint construction and search process. Finally, physical experiments are conducted on a mobile robot platform. Results indicate that the proposed method significantly reduces the establishment of incorrect constraints and effectively enhances the mapping accuracy and localization robustness of robot in complex environments compared with other mainstream graph-based LiDAR SLAM optimization algorithms.