Abstract: With the development of intelligent mining, 3D laser scanning technology has gradually been applied to coal mine tunnel monitoring. However, due to the complex environmental factors in underground coal mines, the point cloud data obtained through 3D laser scanning usually contains a series of noise points, which adversely affect subsequent point cloud data processing. This study proposes a multi-scale fusion point cloud denoising method based on improved statistical filtering using a joint optimization approach of curvature and statistical filtering. First, a KD-tree is used to construct the topological structure of the raw point cloud in the tunnel to be denoised. For each point in the raw point cloud, a neighborhood is constructed, and the average Euclidean distance between every two points in the neighborhood as well as the curvature of each point are calculated. Then, a curvature threshold is set to differentiate the points, comparing the curvature values of the raw point cloud with large-scale noise thresholds and other noise thresholds to filter out an initial denoised point cloud. Afterward, since the initial denoised point cloud may still contain some scattered points, a radius filter is applied to remove the fewer scattered points, completing the point cloud denoising process. Finally, the proposed point cloud denoising method is experimentally studied in a simulated tunnel environment of gas and coal dust explosion laboratory and arched tunnels. The results show that the proposed multi-scale fusion point cloud denoising method based on an improved statistical filter can effectively remove noise points while retaining the surface points of the tunnel as much as possible. Through experimental comparison with traditional point cloud denoising methods, it was found that the tunnel point cloud denoised by traditional statistical filtering had varying degrees of voids and could no longer maintain the tunnel structure. However, the joint optimization method of curvature and statistical filtering proposed in this paper still retains the complete tunnel structure. In laboratory simulated tunnels, the feature point retention rate can be increased by 4.624%, and in arch simulated tunnels, the feature point retention rate can be increased by 10.27%. Field tests in coal mine roadways have proven that this method can compensate for the shortcomings of traditional point cloud denoising methods and better meet the requirements of coal mine tunnel monitoring.