An Onboard Framework for Staircases Modeling Based on Point Clouds

TL;DR

This work tackles onboard detection and physical modeling of staircases from point clouds to support legged robot mobility in varied environments. It proposes an end-to-end framework that combines PointNet++-based instance segmentation with data augmentations, a curvature suppression cross-entropy loss $L_{csce}$, and pose-based measurement correction to produce traversable regions and their geometric parameters in real time. A dedicated staircase dataset (~800 depth images across 11 groups) is introduced, with comprehensive metrics (DE, HE, FP/FN, CPE, NE) and ablation studies demonstrating improved depth, height accuracy, and generalization. The approach shows strong potential for robust stair navigation in diverse lighting and viewpoints, enabling practical deployment on onboard robotics platforms.

Abstract

The detection of traversable regions on staircases and the physical modeling constitutes pivotal aspects of the mobility of legged robots. This paper presents an onboard framework tailored to the detection of traversable regions and the modeling of physical attributes of staircases by point cloud data. To mitigate the influence of illumination variations and the overfitting due to the dataset diversity, a series of data augmentations are introduced to enhance the training of the fundamental network. A curvature suppression cross-entropy(CSCE) loss is proposed to reduce the ambiguity of prediction on the boundary between traversable and non-traversable regions. Moreover, a measurement correction based on the pose estimation of stairs is introduced to calibrate the output of raw modeling that is influenced by tilted perspectives. Lastly, we collect a dataset pertaining to staircases and introduce new evaluation criteria. Through a series of rigorous experiments conducted on this dataset, we substantiate the superior accuracy and generalization capabilities of our proposed method. Codes, models, and datasets will be available at https://github.com/szturobotics/Stair-detection-and-modeling-project.

Figures (6)

• Figure 1: Staircase environment, staircase detection and modeling. The image illustrates the traversable regions predicted by our framework, with arrows indicating estimated normals. Ellipses positioned beneath the arrows represent center point estimations. The abbreviation 'SH' means step height, while the 'SD' refers to the step depth. The raw output of modeling is refined through measurement correction to derive stair parameters. Our method demonstrates robustness and accuracy across diverse lighting conditions and viewpoints, offering a reliable solution for robotic stair navigation in real-time.
• Figure 2: The training process of the predictor and the complete framework for staircase detection and modeling. Squares in the Data augmentation block represent points, red means steps and blue means non-steps. X marks represent the operation of removing. Arrows between blocks represent the flow of information, while the characters above are their dimensions and attributes (S denotes Step, K denotes points belonging to a specific S).
• Figure 3: The phenomenons of point cloud undulations and overly aggressive stair treads predictions. Red represents FP and FN, blue represents TN and green represents TP in the right sub-image.
• Figure 4: Demonstration of measurement correction: In the illustration, the red points represent the central points after correction, while the blue points represent the central points before correction. The rotation matrix "R" is computed in real time based on the current normal vector to achieve the rotation correction.
• Figure 5: The Visualized results of the impact of the CSCELoss and measurement correction. In sub-image (a), the points in red represent FP and FN after the prediction of baseline. The points in blue mean risers and green mean treads. In sub-image (b), the FP and FN on the risers nearly disappear. There we colored the treads in red and green, alternately. In sub-images (c) and (d), the points in orange represent before the correction, remaining points in cornflower blue represent after the correction.