Limited by the underground explosion-proof requirements, the laser beam used in coal mine underground is often low-power with low beam density, resulting in sparse point clouds that lack detailed environmental descriptions. This paper introduces a tunnel point cloud upsampling technique based on the Composite Residual Self-Attention Network (CRSA-Net), aiming to thicken point cloud features in an end-to-end manner to partially compensate for the low accuracy of sensors. Firstly, outliers are removed and the point cloud is divided into regions. Subsequently, a method based on the KD-tree structure is employed to extract Point Cloud Patches for feature extraction modules. To address the low information density issue of low-beam laser tunnel point clouds, a method for computing tunnel point cloud patch fingerprint features is proposed, and the calculated feature values are used to expand the dimensions of the network input data. In the feature extraction module, a cascaded progressive composite residual attention module is introduced to capture long-range features while maintaining permutation invariance of the point cloud data. Finally, using dense point features, multiple independent MLPs are utilized for feature expansion, and dense point clouds are outputted through fully connected layers based on point features. A set of dense point cloud datasets was constructed for training and testing by combining the WHU-LTS open dataset. In the upsampling experiments on collected tunnel point cloud data, the proposed method achieved CD index of 11.35 (e–3), EMD index of 5.52 (e–3), and HD of 112.31 (e–3). The reconstructed point cloud exhibited good fit and uniformity with the real surface, validating the feasibility of using self-attention networks to address the problem of sparse feature upsampling underground.