Aiming at the problem that the LIO-SAM algorithm in the SLAM algorithm of LiDAR lacks sufficient localization accuracy in complex environments, this paper proposes an improvement strategy focusing on two key aspects, feature point extraction and back-end point cloud matching. In the back-end matching, in view of its large inter-frame error fluctuation and poor robustness, this paper innovatively proposes an adaptive downsampling method based on pre-matching. The method effectively improves the initial matching accuracy through pre-matching operation, and dynamically adjusts the voxel filtering resolution based on the local density of the point cloud. Thus, the computational efficiency is significantly improved while ensuring the matching accuracy. In the front-end feature extraction, a two-stage feature filtering mechanism combining Early Cutoff and multi-scale voxel spatial covariance analysis is proposed to address the problems of redundant curvature computation, large sorting overhead, and low feature extraction rate of the near point cloud in LIO-SAM. The mechanism mainly focuses on the near point cloud: the redundant points are quickly eliminated by the local geometric change threshold, after which the covariance feature analysis is performed in the multiscale voxel grid. From this, the representative feature points with a balanced spatial distribution and a stable geometric structure are screened out, and the far point cloud is extracted by the original algorithm. Comparison experiments are carried out on the public dataset KITTI by selecting the stable sequence 07. Results show that, while the optimized algorithm provides only slight improvements in X- and Y-axis accuracy, it reduces the average absolute error of Z-axis by 26.44%, the RMSE by 24.43%, and the standard deviation by 30.24%. Furthermore, the algorithm has been deployed on a real-vehicle platform, where its robustness and engineering applicability have been verified.