Simultaneous localization and mapping (SLAM) has various application scenarios but is limited due to computational cost. Therefore, a SLAM algorithm (FAST-SAM) based on adaptive features and closed-loop optimization is proposed. The proposed algorithm uses the adaptive feature extraction method Better Feature to ensure the accuracy of the feature extraction at different distances. Then, it uses the ground feature filtering method based on random sample consensus to remove unreliable features and keep the number of features stable. In the scan matching and loop-closure detection modules, we use a matching algorithm combining the normal distribution transformation, nearest point iteration algorithm, and the proposed two-stage loop-closure detection algorithm to output the laser inertial odometry and establish a global point cloud map. The experimental results on the LIO-SAM, KITTI open source datasets, and the measured datasets of Guangxi University show that compared with the mainstream SLAM algorithms, the proposed algorithm improves the accuracy and the computational efficiency of each link by more than 25.6%.