【作者】 袁夏；

【导师】 赵春霞；

【作者基本信息】 南京理工大学 ， 计算机应用技术， 2010， 博士

【摘要】 地面移动智能机器人是一种可以在室内外环境中连续自主运动,集环境感知与理解、动态决策、动态路径规划、行为控制与执行等诸多功能于一体的高度自动化智能装置,智能机器人技术已成为各国高科技领域的一个研究热点。环境理解是实现机器人自主移动的关键技术之一,激光雷达是一种主动探测传感器,由于其具有能够直接得到三维空间信息、几乎不受光照条件影响等特点,因而成为一种重要的环境感知传感器。针对基于激光雷达的地面智能机器人环境理解关键技术,本文从底层数据融合、点云聚类、帧匹配技术、环境特征提取和可通行区域检测等方面展开研究工作。研究中使用的激光雷达传感器包括单线雷达、多线雷达、面阵雷达以及线扫描雷达等多种系列,涉及到了当前国内外地面机器人平台上配备的主要激光雷达类型。本论文的主要研究成果如下：针对点云聚类问题,提出一种基于密度变化和空间分布不同的点云聚类算法,将基于信息理论的聚类优化方法融入基于密度的聚类算法,使用局部压缩编码值作为评判点云聚类结果的依据。该算法可以自适应计算近邻半径,在一定程度上可以区分密度相似但是空间分布明显不同的点云区域。针对可通行区域检测问题,提出一种基于激光雷达数据的可通行区域提取方法。算法首先使用模糊预测模型结合可通行区域特征,在单条激光雷达扫描线数据中提取初始可通行区域,然后利用帧间或扫描线间的时空关联,优化提取结果,提高可通行区域检测正确率。针对激光雷达与摄像机联合标定问题,设计一种箭头形标定板,使用基于点特征的方法进行激光雷达和摄像机的联合标定。针对高密度彩色点云数据,使用结合几何特征和颜色特征的复合特征向量训练分类器,进行地形分类,得到了比单独使用几何特征更高的分类正确率。针对激光雷达数据帧匹配问题,分别提出适用于单线激光雷达数据的基于点-线匹配的帧匹配算法和适用于面阵激光雷达数据的基于点-面匹配的帧匹配算法。算法在激光雷达数据中分别提取线、面等几何特征,根据广义距离关联两帧数据间的几何特征,利用线、面特征匹配和点特征匹配分别估计旋转参数和平移参数,减少迭代计算。更多还原

【Abstract】 Ground moving intelligent robot is a kind of equipment which can move automatically both in indoor and outdoor environment. It integrates technique of environment apperceiving and understanding, dynamic decision-making, dynamic path planning, action control and implementation, etc. Research of ground intelligent robot is an active area of high technology for lots of countries. Environment understanding is very important for a robot to navigate itself automatically. Lidar is a kind of active range finder. It is a kind of primary sensors in robotics as illumination has no effect to it.This dissertation focuses on key technologies of environment understanding of intelligent robot based on lidar. The research area of this dissertation including low-level data fusion, point cloud clustering, lidar scan-matching, environment feature extraction and traversable area detection. Several types of lidar that usually equipped on robots are used in this dissertation’s study, including single row range finder, muti-line lidar,3D scan lidar and PMD lidar.The mainly studying results of this dissertation are as follows:This dissertation proposes a point cloud clustering algorithm which based both on density and spatial distribution. The algorithm combines robust information-theoretic clustering method with DBS CAN algorithm. It uses the value of local volume after compressing to judge the clustering result. The algorithm computes radius of a point’s neighbor adaptively. It can differentiate points which have similar density but different spatial distribution.A traversable area detection algorithm based on lidar data is proposed. It employs a fuzzy cluster algorithm combined with traversable features to find traversable area in a single scan line, and then the algorithm considers space-time association between scan frames or scan lines to refine the extraction results of traversable area.An arrow shape registration board is designed to register a lidar and a camera to get colored point cloud.The dissertation uses a multi-feature vector to classify dense colored point cloud collected by a 3D scan lidar with a camera. The multi-feature vector contains both geometrical features and color feature. The algorithm trains a terrain classifier by using this multi-feature vector and gets better terrain classifying results than method using only geometrical feature. The dissertation studies scan-matching algorithms. Point-line and point-plane based scan-matching algorithms are proposed to match scans of a single row ranger finder or a 3D lidar. The algorithm finds line or plane feature in lidar data and associates them according to their general-distance. The rotations and translations are estimated respectively by associating line or plane feature and find matched points to decrease iterative computing.更多还原