Abstract: The perception of the environment holds significant research importance for humanoid robots' autonomous navigation and motion planning. It serves as a prerequisite for them to move autonomously and accomplish specific tasks in complex environments. The specialized scenario of staircases has emerged as a formidable challenge in the environment perception process for humanoid robots. Addressing the problem related to the degradation of staircase plane features due to staircase obstacle interference, leading to inaccuracies in staircase parameter acquisition and resulting in complications such as missteps and wrestling, this study combines region growing and plane fitting methods to first identify and remove point cloud obstacles on the stairs, and then performs three-dimensional parameter estimation based on the cleared stairs. Initially, depth cameras are used to capture point clouds of the stair environment and undergo downsampling. Then, the KD-Tree algorithm establishes the topological structure of point cloud data, utilizing the principle of minimizing the sum of projections of neighboring points to accurately extract the horizontal plane of stairs. Subsequently, the region growing algorithm determines stair obstacle clustering, directly removing individually clustered obstacles based on clustering results, and eliminating non-individually clustered obstacles based on plane construction and analysis of point numbers within the plane. Finally, experiments on obstacle removal are conducted using multiple sets of stair data containing obstacles. The results indicate that this method can accurately recognize and eliminate various types of stair obstacles, with an average elimination accuracy of 92.43%. Additionally, the experiments elimination that the presence of obstacles primarily affects the humanoid robot's acquisition of stair height and depth information, with height errors reaching 30% and an overall average relative error of 16%. However, after obstacle elimination, the average relative error of stair parameter perception is approximately 7%. In general, the proposed algorithm improves the accuracy of stair parameter estimation for robots and effectively enhances the perception capabilities of humanoid robots in complex stair environments.