【作者】 安秋；

【导师】 姬长英；

【作者基本信息】 南京农业大学 ， 农业机械化工程， 2008， 博士

【摘要】 农业机器人及其视觉导航系统的研究得到了越来越多的关注,成为探索在农业机械装备中应用智能控制等高新技术的重要研究方向。农业机器人工作在复杂的开放式非结构的农田环境,如何从中提取出导航信息是视觉导航的难点。光照变化是使导航信息提取算法变得不稳定的主要因素,阴影的存在更是严重干扰了图像的处理与分析。因此,必须探索去除光照影响的算法,提高导航系统的鲁棒性。人的视觉系统具有颜色恒常性,如何将人的这种颜色恒常性引入到机器视觉中,是解决机器视觉中光照问题的一种思路。本文以实现农业机器人田间自主导航为目的,首次将颜色恒常性理论用于解决机器视觉导航中的光照问题,并自行设计了农业机器人移动平台,构建了单目视觉导航系统,进行了作物行跟踪和地头转向试验。主要研究成果及结论包括：(1)从图像的物理成象过程微观地探讨光照对图像的影响,引进了颜色恒常性理论从根本上解决光照问题,并把此方法总结为两种思路：一种是转化到标准光照下的图像,另一种是获得光照不变特征量。(2)研究测试了几种将非标准光照下的图像校正到标准光照下的图像的颜色恒常性算法。这些算法根据图像的统计信息,全局地或局部地估计出光照色彩,利用对角模型进行颜色校正,从而获得标准图像。提出了基于色彩学习的彩色图像分割方法,通过对原始图像和各种算法获取的标准图像的分割结果进行统计发现,光照变化影响图像的分割质量,而转换到标准光照下的标准图像的分割效果有明显提高。最终,根据算法耗时统计和分割结果统计两项指标进行算法比较认为,灰度世界假设和三种局部光照估计算法分割效果相当,但灰度世界假设在效率上更高,比较适合于实时导航的应用要求。因此,选取灰度世界假设算法来获取标准图像,以解决农业机器人视觉导航中的光照变化问题。(3)分析了阴影形成的实质,阐述了光照无关图的原理,在此基础上导出了一种基于光照无关图的阴影去除方法。该方法根据光学成像原理通过对图像进行变换来得到一幅与光照无关的灰度图,以达到去除阴影的目的。同时针对该方法需要事先测定摄像机的光照无关角的不便之处,引入了基于最小熵的摄像机光照无关角求取方法。实验表明,采用基于最小熵的光照无关角求取方法得到的光照无关角相差不大,完全可用于获取去除阴影后的光照无关图。针对光照无关图的特点,采用增强最大类间方差法自动求取阈值并进行分割处理,获得了理想的效果。(4)在建立农田导航作业环境模型的基础上,进行了沿作物行行走作业的导航规划,分析了导航信息提取的主要步骤,提出了优化的霍夫变换用于提取单作物行的导航路径,提出了结合最小二乘的线扫描算法提取双作物行的导航路径,提出了水平投影算法用于地头的检测判断。通过对视觉处理算法测试分析,两种导航路径提取算法和基于水平投影的地头检测算法设计合理可靠,整个视觉导航信息提取算法的总花费时间完全可满足视觉导航的实时性要求。(5)引入基于CAN总线的分布式控制思想,自行设计了四轮驱动、四轮转向的农业机器人移动平台,构建了单目视觉导航系统,并开发了基于多线程的视觉导航系统软件。在农业机器人移动平台上验证了视觉处理算法的正确性和可靠性,通过作物行跟踪试验,验证了整个视觉导航系统的可行性和可靠性。最后,提出了基于光电编码器的航位推测算法以实现地头转向功能,模拟测试结果表明,基于光电编码器的航位推测算法可满足农业机器人地头转向的工作要求。更多还原

【Abstract】 Researches on agricultural robot and their vision navigation systems are attracting more and more attentions and becoming an important direction of exploring application of high technology (intelligent control) in agricultural machines and equipments. Agricultural environment is complicated, open and unstructured. How to extract visual information of navigation is a difficulty of vision navigation. Illumination variation makes extracting algorithms of guidance parameters unstable and existence of shadows furthermore interferences image processing and analysis. So some new algorithms of removing illumination influence should be explored to improve stability of vision navigation system. Human visual system has property of color constancy. How to introduce human color constancy into machine vision is a train of thought of solving illumination problem.In the thesis, to realize autonomous navigation of agricultural robot in the field, illumination influence problem in vision navigation system was studied firstly with color constancy theories, a moblie robot platform for agriculture was designed, vision navigation system was constructed, and trials of crop line guidance and headland turn were carried through. Main research contents and conclusions include:(1) Light microscopic influence on images was investigated through image fomation process. Theory of color constancy was introduced to solve illumination influence problem, and its current status was reviewed, two color constancy ways of solving illumination influence problem were concluded. One way is to take the conversion of images from in the non-canonical illumination to in the canonical illumination. Another way is to obtain illumination invariants.(2) Some color constancy algorithms, which to take the conversion of images from in the non-canonical illumination to canonical illumination, were tested. These algorithms estimate illumination color globally and locally according to statistical information of image data, and then obtain color-corrected canonical images by von Kries model. A color learn based color image processing algorithm was proposed, Statistical results of image segmentation experiments to original images and canonical images showed, illumination variation influenced image segmentation quality, but canonical images under canonical illumination could improve segmentation results. Finally, according to time cost and segmentation effects, gray world algorithm and three local illumination estimation algorithm is equivalent in segmentation results, but gray world algorithm is high in efficiency and fit for real-time vision navigation, so gray world algorithm to obtain canonical image is chosen in later vision navigation system.(3) Shadow formation process was analyzed. Theory of illumination invariant image was described. Based on this, a method of shadow removal was derived. To obtain the illumination invariant image, a camera-dependent illumination invariant angle should be measured. Aiming at solving this problem inconvenient, a method of minimal entropy was introduced. Results showed that method of shadow removal based on illumination invariant image could remove shadows efficiently. An enhanced OSTU threshold method was proposed to segment the illumination invariant image, and perfect effects were obtained.(4) Based on founded model of farmland navigation working environment, navigation process of agricultural robot in the field was planed and main steps of navigation parameters extraction were analysised. An optimized Hough transform algorithm was proposed to extract guidance parameters for only single crop row. A line-scan algorithm with least squares theory was proposed to extract guidance parameters for double crop rows. A horizontal projection algorithm was used to detect headland. Tests of these image processing algorithms showed, all the algorithms were reasonable and reliable and total time cost of vision algorithms was enough for vision navigation.(5) A CAN-based distributed control technology was introduced. A 4WD and 4WS robot platform was designed. A monocular vision navigation system was constructed. Multi-thread based vision navigation system software was developed. Validity and reliability of former vision algorithms were verified in the mobile robot platform. Results from cropline tracking tests showed that outdoor vision navigation system was feasible. In the end, an optical-encoder-based Dead Reckoning algorithm was proposed to fulfill headland turn function, and simulated tests showed the algorithm could suffice for headland turn.更多还原