Limited by the requirements of underground explosion protection, the laser radar that can be applied in coal mine is mostly low-power lidar with a few line harnesses, which forms sparse point clouds and lacks detailed description of the environment, resulting in poor mapping performance of robots. A composite residual self-attention network (CRSA-Net) based sampling technique for roadway point cloud is proposed, aiming to achieve the point cloud feature thickening through an end-to-end approach, and to make up for the low accuracy of sensors to a certain extent. Firstly, the dust and water fog noise in the point cloud are removed by outliers, and the point cloud regions with different degradation degrees are segmented by normal vector clustering according to the site characteristics. Secondly, in order to improve the local detail quality of training data, point cloud slices are extracted based on KD tree structure. To solve the problem of weak correlation of data in sparse point cloud, a method of feature calculation and feature dimension extension is proposed to strengthen the constraints on network training results. A cascaded and progressive compound residual self-attention method is adopted to ensure the network to learn the global and local features of the structurally degraded point cloud. Finally, using dense point features, multiple independent MLPs are used for feature extension to output dense point clouds through a fully connected layer based on point features. A dense point cloud dataset was constructed for training and testing by using the data of a lane in Cuncaota Coal Mine of National Energy Group and WHU-LTS open source data. In the up-sampling test, the CD index, EMD index and HD index reached 11.35×10−3, 5.52×10−3 and 112.31×10−3 respectively.