Abstract: LiDAR SLAM faces challenges in the narrow and confined unstructured environment of underground coal mines, where inaccurate point cloud pose estimation due to few or complex features can result in distortion or even map construction failure. To address the difficulties in LiDAR point cloud feature extraction and registration in this degraded environment, a two-stage method integrating FPFH and ICP algorithms is proposed. Initially, the method constructs kd-tree structures for the source and target point clouds, reduces point cloud numbers through statistical and voxel filtering, extracts point cloud surface normal, and computes fast point feature histogram descriptors for key points. Subsequently, a coarse registration is performed using the sampling consistency initial registration algorithm, followed by fine registration using the ICP algorithm to enhance point cloud registration accuracy and pose estimation precision. Furthermore, enhancements are made to the feature extraction and registration algorithm of the LIO-SAM, along with the optimization algorithm of the back-end loopback factor, to improve key local feature identification and registration capabilities. The addition of the Scan Context global descriptor loop factor enhances loop detection accuracy for consistent global mapping. Experimental testing on the M2DGR public dataset and SLAM experiments in simulated coal mine scenarios demonstrate the effectiveness of the improved algorithm in feature extraction and registration of the point clouds. Compared to the traditional LIO-SAM algorithm, the improved algorithm showcases higher accuracy in pose estimation and point cloud registration, with a 6.52% improvement in average relative position error and an 18.84% reduction in maximum absolute position error. The resulting maps exhibit no obvious distortion and mapping errors are within 1%, allowing for the construction of high-precision consistent global maps in unstructured and degraded environments.