【作者】 刘斐；

【导师】 郑志强；

【作者基本信息】 国防科学技术大学 ， 控制科学与工程， 2007， 博士

【摘要】 机器人足球是目前机器人控制、机器视觉、人工智能、多机器人协调等领域研究的一个热点,也是一个难点。中型组机器人足球比赛(Middle-Size League)是机器人足球世界杯(RoboCup)中的一个重要项目,为彩色全向视觉的相关技术研究提供了一个标准测试环境。本文针对彩色全向视觉系统在中型组机器人应用中的颜色分类、特征提取和传感器模型建立等关键技术进行了研究,研究结果得到了较好的实验验证。本文首先根据中型组足球机器人系统的需要,分析、设计并实现了一套彩色全向视觉系统。该系统的主要特点是使用了组合等比例反射镜面和数字式彩色摄像机,可有效获得比赛环境的彩色数字全景图像。论文分析了该系统的设计原则,介绍了该系统各主要组成部分的特点、主要参数的测量方法,确定了彩色全向视觉系统的距离测量有效范围。其次,针对提高彩色全景图像的颜色分类处理能力,本文提出了一种基于线性分类器的混合颜色空间查找表颜色分类方法。该方法主要解决已有的颜色查找表分类方法的区分能力受颜色空间选择、阈值确定等因素影响而难以区分近似颜色的问题,将模式识别中的线性分类器思想应用于颜色查找表映射关系的建立,并通过同时使用HSI空间与YUV空间的方法提高查找表对近似颜色的区分能力。通过实验证明,基于线性分类器的混合空间查找表颜色分类方法具有查找表建立原则简单、效果直观的特点,对比赛环境中的近似颜色有较强的区分能力,查找表建立快捷、可靠,能够满足机器人足球比赛中彩色全景图像分类的实时性要求。再次,研究了机器人足球比赛环境中,彩色全景图像的特征检测方法。本文分析了光照条件在空间和时间上的不均匀性对机器人足球比赛场地特征检测的影响,通过检测颜色过渡,对场地上的白色标示线实现了与颜色分类结果无关的可靠检测;对蓝/黄色球门与场地的接触点实现了在颜色分类结果较差情况下的可靠检测。其主要思想是利用不同颜色在颜色空间中分布的位置相对关系,结合扫描线检测方法对亮度(白色标示线)或色调(蓝/黄色球门)变化进行判别,根据场地颜色结构化信息,判断变化的性质。论文中分别对两种场地特征给出了判别条件和判别算法,通过实验验证了算法的有效性和鲁棒性。基于上述颜色分类和特征检测的结果,本文结合粒子滤波定位方法,研究了机器人足球比赛环境下的彩色全向视觉传感器的建模方法。将彩色全向视觉系统作为粒子滤波定位方法的环境观测器,将特定的颜色过渡点到机器人的距离作为观测信息,建立了彩色全向视觉传感器模型。对模型的形式进行了分析,给出了模型的数学描述和主要参数计算方法。通过将依靠传感器模型得到的机器人位置、朝向信息与真实位姿进行比较,分析了影响模型精确度以及模型使用效果的主要因素。实验结果表明,本文建立的彩色全向视觉传感器模型能够克服观测信息部分失效的影响,具有一定的鲁棒性,通过使用双查找表方法,还可以进一步提高粒子点置信度的计算速度,从而增强模型的实用性。更多还原

【Abstract】 The robot soccer, as a hard nut to crack, is a research focus in the field of the robot control, machine vision, artificial intelligence, and multi-robots systems. Middle-Size League, as a crucial competition in RoboCup, provides a standard testing environment for the color omni-directional vision techniques. The dissertation made an investigation on the color classification, the feature extraction, and the sensor model establishment in the application of the color omni-directional vision system in the field of the Middle-Size RoboCup competition.Firstly, the dissertation designed and realized a color catadioptric omni-directional vision system according to the need of Middle-Size League in RoboCup. The system features in the use of combined distortionless omni-mirror and the digital color camera in order to obtain digital color panoramic image of the competition environment effectively. Accordingly, the dissertation expounded the principles, the characteristics of the main components and the measurement of the main parameters, and decided the effective range of distance measurement of the system.Secondly, aiming at enhancing the ability of color classification for the color panoramic image, the dissertation proposed a combined color space CLUT (Color Look-Up Table) color classification method based on the linear classifiers. The method solved the problem that the current CLUT method is hard to distinguish the similar colors due to the influence of inaccurate choices of color space and threshold. The improved CLUT method applied the linear classifier in pattern recognition to the establishment of CLUT mapping relationship. Meanwhile, HSI and YUV color spaces were synchronously employed to increase the similar colors classification ability. The experiments indicated that the combined color space classification method, based on the linear classifiers, is convenient to establish CLUT and to take effect. Additionally, the method satisfied the real-time requirement of the color panoramic image classification in RoboCup with an effective way to distinguish similar colors.Thirdly, the dissertation investigated the feature extraction method of the color panoramic image in the RoboCup competition. It analyzed the effects of the inconsistent light condition in area and time on the feature extraction of the RoboCup competition field. Furthermore, by way of the color transition detection, it demonstrated that the white lines on the field can be detected reliably without color classification, and the touch points between the blue or yellow gate and the field can also be detected reliably with the unsatisfied color classification results. The principle of the two kinds of detection was to make use of the relative distribution of various colors in the color space, to decide the change of luminance (for the white lines) and the hue (for the blue or yellow gate) with the scan-line detection method, and to estimate the character of the change according to the color structuralized information of the field. The dissertation, for the two field features respectively, proposed the principles and the arithmetic which were proved effective and robust in the experiments.Finally, based on the results of the color classification and the feature extraction, the dissertation explored the way to establish the color omni-vision sensor model in the RoboCup competition with the particle filter localization method. The color omni-vision sensor, as the environment sensor of particle filter localization method, detected the chromatic transitions of interest as the observation information. Moreover, the dissertation analyzed the color omni-vision sensor model, indicated the mathematics description and the calculation method of the main parameters. In addition, the dissertation clarified the factors to affect the precision of the sensor model by comparing the robot localization information from the sensor model and the real position and orientation. The experiment results indicated that the color omni-vision sensor model, with great robustness, can overcome the invalidation of partial observed information. In particular, the employment of double look-up tables increased the speed of particle belief calculation and the practicability of the model.更多还原