基于激光雷达的多层楼梯检测和建模方法

Multi-storey staircase detection and modelingmethod based on LiDAR

移动机器人在多层建筑进行自主导航时,需要具备楼梯检测和建模能力。然而,目前楼梯检测和建模主要面向单层直楼梯,对具有回转结构的多层楼梯适用性差,且场景中存在的动态点云会影响建模精度。针对以上问题,提出了一种基于激光雷达的多层楼梯检测和建模方法。首先,构建了轻量化的2.5D点云滤波器,高效滤除动态和噪声点云;其次,提出了一种基于主成分分析的台阶分类方法,实现了对单个及具有回转结构的复杂多层楼梯的鲁棒台阶检测;同时,设计了一种基于区域分割的台阶尺寸估计方法,提高了楼梯模型的构建精度;此外,提出了一种基于多阈值判断的楼梯模型追踪方法,实现了实时楼梯检测和快速动态建模;最后,分别使用Tang数据集和四足机器人在室内外多重场景下采集的点云数据进行试验验证。结果表明,提出的算法在室内外场景下,能够对单层及具有回转结构的多层楼梯进行动态检测和建模,其中台阶尺寸的估计精度在1 cm以内,计算量降低了51.23%。

When mobile robots navigate autonomously in multi-storey buildings,they need to possess the capability of staircase detection and modeling.However,the current staircase detection and modeling methods are mainly for single-storey straight staircase,which have limited applicability to multi-storey staircase with rotational structures.Moreover,the presence of dynamic point clouds in the scene affects the modeling accuracy.A multi-storey staircase detection and modeling method based on LiDAR is proposed.Firstly,a lightweight 2.5D point cloud filter is constructed to efficiently remove dynamic and noisy point clouds.Secondly,a step classification method based on principal component analysis is introduced to achieve robust step detection for single and complex multi-storey staircase with rotational structures.Additionally,a region-based segmentation approach is designed for estimating step dimensions,aiming at improving the accuracy of staircase model construction.Furthermore,a staircase model tracking method based on multiple threshold judgments is proposed to enable real-time staircase detection and fast dynamic modeling.Finally,experiments are conducted to validate the proposed method,using the Tang dataset and point cloud data collected from a quadruped robot in various indoor and outdoor multiple scenarios.The results show that the proposed method is capable of dynamically detecting and modeling both single-storey and multi-storey staircase with rotational structures in indoor and outdoor scenarios.The estimated accuracy of step dimensions is within 1 cm,and there is a 51.23%reduction in computational complexity.