Abstract: To address the coexistence of multiple types of noise in point cloud data of rectangular coal-mine roadways, as well as the problems of roadway surface voids, boundary loss, and structural fractures that may be caused by traditional radius filtering and statistical filtering methods when dealing with complex internally attached noise, this paper proposes a multi-stage denoising method for rectangular roadway point clouds with mixed noise. The proposed method fully exploits the multi-plane structural prior formed by the roof, floor, and two sidewalls of a rectangular roadway, and constructs a progressive processing framework according to the spatial distribution characteristics and removal difficulty of different noise types. First, based on the spatial distribution characteristics of noise, radius filtering and pass-through filtering are employed to rapidly remove discrete points and external cluster noise. Then, for internal cluster noise that is interwoven with the real roadway surface and difficult to separate using global statistical features, a local denoising strategy based on plane partitioning is proposed. Specifically, the point cloud is divided into equidistant slices using the roadway central axis as a reference. The points in each slice are further partitioned into roof, floor, and two-sidewall regions according to normal-vector components. Within each sub-region, density statistics and normal consistency are combined to achieve differential identification and removal of noise. Finally, a structurally complete denoised rectangular roadway point cloud is obtained through the fusion of slice-wise results and the continuity restoration of overlapping regions. Experiments were conducted in both simulated tunnel and real roadway scenarios, and comparisons were made with radius filtering and statistical filtering methods. The results show that the proposed method performs better in terms of noise removal rate, feature preservation, and boundary integrity. It effectively alleviates the voids and structural fractures caused by traditional filtering methods when removing internal cluster noise. The results verify the reliability and engineering applicability of the proposed method for rectangular roadway point cloud denoising, and indicate that it can provide a stable data basis for high-fidelity three-dimensional reconstruction and intelligent perception of coal-mine roadways.