毕业生知识库

随着科技的发展进步，移动机器人逐渐进入人类生活的各个方面。自主导航作为移动机器人代替人类执行各项服务的基础，是移动机器人实现智能化的关键技术，具有重要的理论研究意义和广泛的应用价值。本文面向大范围高可靠性导航问题，研究环境拓扑模型的构建和以此为基础的移动机器人自主导航，以环境感知和导航技能传授为切入点，从机器人自主运动、道路与障碍感知、人工示教跟随和拓扑地图学习等方面展开系统的研究。本文主要的工作和贡献有：

(1) 首先，为了提升机器人自主运动所需信息的完备性，研究基于视觉的环境空间结构感知方法。面结构相交、物体轮廓、明暗分界等环境元素可以形成易被视觉感知的曲线，是反映空间结构信息的重要载体。针对现有方法曲线提取不完整，提取形态单一等问题，借鉴卡尔曼滤波思想提出了一种基于图像结构特征的搜索模型。所设计的图像结构特征提取方法具有良好的尺度适应性，可以在较大的空间跨度上反映图像信息的关联性。搜索模型将图像空间中的曲线提取问题转化为特征空间中的状态向量转移问题。通过不断地预测、搜索和迭代更新模型参数，优化预测结果，实现较为完整和准确地曲线提取。实验表明，该方法能够较好地提取图像中不同形态、不同长短的曲线。

(2) 基于所提出的曲线提取算法，研究环境结构感知方法，设计了一种通过曲线分析确定通路和障碍的方法。针对道路曲线提取任务，结合格式塔认知原理进行曲线分析，设计了道路曲线筛选规则，能够有效地获取图像道路曲线。通过冗余图像匹配实现了相邻帧的曲线跟踪，提高了道路曲线提取的实时性。针对障碍感知任务，借鉴道路曲线提取中相关的格式塔认知规则，筛选道路区域中的潜在障碍曲线，并提出了一种基于规则的曲线链搜索方法，通过判断曲线链的封闭趋势感知潜在障碍物。利用本文中的方法可以有效地感知环境中的可行通路，指导机器人实现自主运动。

(3) 对于具有可靠自主运动功能的机器人，为了实现基本环境拓扑模型的学习，需要示教人员指导。领航者带领机器人沿着环境中某些具有代表性的路线行走是一种较为直观的导航技能传授方式。因此，以长时间跟随导航为目标，提出了一种基于视觉的移动机器人跟随方法。该方法包括视觉跟踪，目标重找回和机器人控制三部分。在视觉跟踪部分，根据目标轮廓带生成滤波模板近似实现目标区域的不规则采样，提高了跟踪的鲁棒性。在重找回部分中，提出了一种结合离线训练和在线学习方法的目标重找回方法，提高了重找回的成功率，并实现跟踪区域的精确定位。在机器人控制部分，采用基于视觉反馈的PI控制方法实现视觉控制。基于移动机器人平台的室内外跟随导航实验表明本文中方法的有效性。

(4) 为了从人工演示的可通行路径中学习环境拓扑，提出了一种基于技能传授的环境拓扑学习方法。该方法包含拓扑节点学习和拓扑关系构建两部分。针对视频序列图像，提出了一种环境拓扑节点检测方法，并利用词袋模型实现了拓扑节点的重识别。使用该拓扑节点学习方法可以检测、记忆并识别环境中的有效拓扑节点。在跟随导航过程通过学习环境拓扑节点获取节点连通链，建立基本环境拓扑地图。在基本拓扑地图的基础上，通过拓扑地图学习可以构建更为完整的环境地图，为此提出了一种自主探索式拓扑地图生长方法。该方法通过路径复现完善基本拓扑地图，然后借鉴Q学习思想，使机器人通过自主试探，寻找环境中的潜在拓扑节点，实现拓扑图的生长。

(5) 为了全面验证本文方法在机器人自主运动、拓扑地图学习和导航中的有效性，在总面积超过两千平米的实验场景中系统性地开展了综合验证实验。首先，机器人根据跟随导航演示，学习建立基本拓扑地图，然后通过复现路径消除由于领航者遮挡造成的环境信息缺失。接下来，在基本拓扑地图的基础上，通过自主探索式方法实现拓扑地图生长。基于所构建的环境拓扑地图，机器人的导航功能获得了充分地验证。

With the progress of civilization and the development of science and technology, mobile robots are more and more widely used in various fields. Navigation, as one of the key technologies for mobile robots, has important theoretical research significance and broad application prospects. This paper studies the construction of environment topology model and robot autonomous navigation. Starting from the environment perception and navigation skill imparting, this paper systematically studies robot autonomous movement, obstacle and road perception, leader following and topology map learning.The main contributions of this paper are summarized as follows:

(1) To realize autonomous movement, the environment structure needs to be perceived. As common geometric element in environment, curve segment can describe environment structure information effectively. Aiming to solve the problems of incomplete curve extraction and monotonous shape extraction in existing methods, a search model based on structural features is proposed and applied to curve extraction. To enhance the spatial correlation of image features, a structural feature extractor with scale adaptability is proposed. Using Kalman filter as a reference, a search model is designed, which can effectively correlate spatial structural features. The model proposed in this paper transforms the curve detection problem in image space into a state vector transfer problem in feature space. By constantly predicting and searching curve points, the model parameters are gradually updated, the prediction process is constantly optimized, and curve segments are detected more completely and accurately. Experiments show the effectiveness of this method.

(2) Roads and obstacles are effective environment structure information for robots autonomous movement. Therefore, a roads and obstacles perception method based on curve analysis is proposed. According gestalt cognitive principle, the road curve screening rules are designed, which can effectively obtain the road curves in images. By redundant image matching, curve tracking between adjacent frames is realized and the real-time performance of road curve extraction is improved. For the obstacle perception, the potential obstacle curves in the road area are screened referring to relevant gestalt cognitive rules, and a rule-based curve chain search method is proposed to perceive potential obstacles by judging the closed trend of the curve chain. The method in this paper can effectively perceive passable areas in the environment and guide the robot to realize autonomous movement.

(3) Target following is an effective way to impart navigation skills. A vision-based mobile robot following method is proposed. The method includes three parts: visual tracking, target re-detection and robot control. In the visual tracking part, irregular sampling is realized approximately using the filtering template generated the target contour band. In the re-detection part, a target re-detection method combining offline training and online learning is proposed, which improves the success rate of re-detection and realizes accurate re-location. In the robot control part, PI control method based on visual feedback is adopted to realize visual control to reduce the relative posture change between the target and the robot as much as possible. The experiments of indoor and outdoor tracking navigation show the effectiveness of this method.

(4) With the basic topological map, a more complete environment map can be built by topological map learning. This method consists topology node learning and topology relation building. An environment topology node re-detection method for video sequence images is proposed to realize the recognition of topology nodes by bag of words model. Based on the topology node learning method, effective topology nodes in the environment can be detected, memorized and recognized. In order to complete the basic topological map, an autonomous exploration method for topological map growth is proposed. This method improves the basic topology map through path repetition, and then uses the idea of Q learning for reference to find potential topology nodes and realizes the growth of topology map. Experiments show the effectiveness of this method.

(5) To verify the effectiveness of the proposed method in robot movement, environmental perception and navigation, comprehensive verification experiments are carried out in an experimental environment with a total area of more than two thousand square meters. Firstly, the basic topology map is esteblished through skill imparting, and then the lack of environmental information is eliminated through path replaying. Then, on the basis of the basic topology map, the topology map growth is realized through the self-exploration method. Based on the constructed environment topology map, the robot navigation is achieved reliably and accurately.