Obstacle recognition and elimination method for humanoid robots based on point cloud processing

Understanding environmental perception is crucial for the autonomous navigation and motion planning of humanoid robots, especially in complex environments. Staircases pose a significant challenge as obstacles on them can disrupt planar features, leading to inaccurate parameter acquisition and potential missteps or falls. This study employs a methodology that integrates region growing and plane construction techniques. Initially, a depth camera captures the point clouds. Improved voxel filtering and straight pass filtering are applied to effectively eliminate noise, reduce data volume, and improve algorithm processing speed. The KD-Tree algorithm is then used to establish point cloud topology. By minimizing the sum of projections of neighboring points, the algorithm estimates normal vectors and accurately extracts staircase levels based on plane normal vector constraints. The region-growing clustering algorithm with adaptive parameters recognizes stair obstacles by defining cluster boundaries using statistical properties and principles. Individually clustered obstacles are then eliminated by assessing the region’s minimum points, whereas non-individually clustered obstacles are identified based on the maximum number of points in the region. Subsequently, the plane is constructed, and obstacles are eliminated by analyzing point mutations within the plane. In this study, obstacle elimination experiments were conducted using data from various obstacle-impaired staircases of inaccessible types. The data and experimental results were recorded and analyzed. Additionally, experiments were conducted to estimate staircase parameters with and without obstacle rejection. The elimination experiments demonstrate that the average correct rate for removing individually clustered obstacle point clouds is 92.13%, whereas non-individually clustered obstacles are removed with a 92.72% accuracy, leading to an overall elimination accuracy of 92.43%. These findings indicate the effectiveness of the proposed method in precisely identifying and eliminating various obstacles in staircase environments. In stair parameter estimation experiments, obstacles significantly hinder the humanoid robot’s ability to accurately measure step height and depth. The experimental results demonstrate that the maximum height error in stair parameter estimation when obstacles were present reached 30.55%, with the overall average relative error being 16%. However, once obstacles were removed, the errors in three-dimensional height measurements decreased to 8.53%, and the overall perception error dropped to approximately 7%. The average relative error in height is reduced to approximately 25% of that when obstacles are present, whereas the overall perception error decreases to about 50% of the error observed with obstacles. These findings highlight the profound impact obstacles have on stair perception and demonstrate that removing them substantially enhances the accuracy of stair parameter estimation.