【作者】 唐晶磊；

【导师】 何东健；

【作者基本信息】 西北农林科技大学 ， 农业电气化与自动化， 2010， 博士

【摘要】 基于机器视觉的喷药机器人杂草识别与导航参数获取方法研究,对定点变量投放化学药剂以降低对生态环境的污染,对农业机械自主或辅助定位导航以实现精确农田作业,均具有重要的现实意义。广泛应用智能农业机械,以技术代替资源,实施精准农业,是本世纪农业发展的必然趋势。本文在充分总结国内外先进研究成果的基础上,对喷药机器人杂草识别与定位导航视觉系统的研发进行了基础研究。针对喷药机器人沿导航路径定点变量喷洒的应用目的,以图像处理技术为基础,建立喷药机器人杂草识别与定位导航视觉系统,以快速准确的识别杂草和获取精确的导航参数。本文的主要研究内容及结论如下：(1)以福田欧豹4040四轮拖拉机为平台,搭建了喷药机器人杂草识别与定位导航视觉系统。利用C#面向对象语言为开发工具,在Windows xp环境下开发系统软件,以此为基础,实现了杂草识别图像的采集和定位导航图像的采集。(2)针对杂草识别和定位导航的应用目的,结合实时性要求,在图像基本处理方法上进行研究,探索研究植物和土壤背景的分割方法,为形成针对杂草识别与定位导航视觉系统的专用图像处理方法打下基础。(3)将杂草冠层的形态学特征应用于杂草识别,结合杂草叶片的形状特征,基于4-10-4的BP神经网络实现了麦田伴生杂草的识别,平均正确识别率为88.7%,最高正确识别率可达93.1%。采用最大投票机制构造决策二叉树,提取并筛选玉米幼苗及其伴生杂草的形状和纹理特征,基于SVM的决策二叉树实现了单子叶和双子叶植物的类间分类和类内分类,平均正确识别率为91.4%,最高正确识别率为95.5%。实验结果表明,在识别率和识别精度上,SVM分类器都要优于BP网络分类器。(4)避开计算量巨大、实时性差的双目视觉立体匹配问题,以左摄像机为基准,提出了一种根据左、右图像对中的导航路径信息,获取实际导航路径数据的方法。将实测数据投影到左图像上,依据左图像上的路径信息进行导航。(5)形态学处理点播作物区域后,结合作物区域的形心点,用改进的Hough变换法实现了点播作物导航路径的识别。针对传统Hough变换不能检测条播作物行宽质心的问题,提出了兼顾虚点检测和透视原理的双Hough变换算法(DHT),实现了条播作物导航路径的识别。(6)针对作物行特征不明显的情况,将原图按行间距划分为3个子图像并进行合并,然后在具有明显行特征信息的合并图像上识别导航路径。该方法在作物行特征不明显的情况下,依然能够精确识别导航路径。提出获取精确导航路径的水平扫描法,通过选取合适的d值和Bezier拟合曲线实现对水平扫描法的改进,既平滑了导航路径,又提高了算法的实时性。(7)设置ROI窗口递归跟踪导航路径并提出状态曲率的概念,将反馈和预视原理引入喷药机器人杂草识别与定位导航视觉系统。将反馈导航信息、当前导航信息与预视导航信息相结合,设计了串行的导航路径状态BP网络和导航参数BP网络,基于闭环控制策略实现了导航参数的获取。实验结果表明,横向偏差的输出值与实测值的最大偏差为-7cm,均值为-0.6cm,均方差值为3.2cm。偏转角的输出值与实测值的最大偏差为-3°,均值为-0.45°,均方差值为1.2°,横向偏差和偏转角偏差较小,获取的导航参数具有较高的精度。更多还原

【Abstract】 The machine-vision-based research on weed detection and navigation parameter acquisition of pesticide spraying robot has a great significance both on fixed-point variable spraying of chemicals aiming to reduce the environmental pollution of chemical and on vision navigation of agricultural machinery to achieve the goal of precision agriculture. It is an inevitable trend of agricultural development in this century that intelligence agricultural machinery will be widely applied, resource will be replaced by technology, and precision agriculture will take effect.Basic research has been made on weed detection and location and navigation vision system of pesticide spraying robot on the basis of adequate summary about the advanced achievements at home and aboard. Aiming to realize the application of fixed-point variable spraying along navigation path, this paper build up the weed detection and location and navigation vision system of pesticide spraying robot based on the image processing technology, to achieve rapid and exact weed detection and accurate navigation parameter acquisition. The main contents and conclusions of the research are as follows:(1) Build up the weed detection and location and navigation vision system of pesticide spraying robot on the platform of tractor Foton European Leopard 4040. Develop system software under the Windows XP environment using C# objected-oriented language to collect the weed detection images and location and navigation images acquisition.(2) Combining the real-time demand, the trials on the basic image processing methods and the explorations on segmentation methods of plants and soil background, which aim at the application of weed detection and location and navigation, form the basis of dedicated special image processing methods of the system.(3) The accomplishment of the detection of wheat associated weed based on the 4-10-4 BP neural network takes the advantage of applying the morphological features of weed canopy and combining the shape features of weed blade. Average correct detection rate is 88.7% and the highest correct detection rate is 93.1%. Construct the decision binary tree by adopting the largest voting mechanism to extract and screen the shape and texture features of maize seeding and its associated weeds. The SVM-based decision binary tree achieved the inter-class and intra-class classification between monocotyledons and dicotyledons. Average correct detection rate is 91.4% and the highest correct detection rate is 95.5%. The experimental results indicate that the SVM classifier is superior to BP network classifier on detection rate and detection accuracy.(4) Present an acquisition method of actual navigation path data according to the navigation path information in the left and right images, which based on the camera on the left side, in order to avoid the problems of vast computation and the shortage in real-time of binocular stereo matching. Actually, it projects the measured data to the left image, and then implements navigation in terms of the path information on the left image.(5) After morphology operation on the area of dibbling crops, according to the centroid of the area of crops, achieve the navigation path recognition of dibbling crops using the improved Hough transform. DHT(double-Hough transform), which gives attention to both imaginary point detection and perspective, has been present to settle the issues that traditional Hough Transform fails to detect the centroid of drilling crop row, and achieves the navigation path recognition of drilling crops.(6) Divide the original image into three sub-images and then combine them by line spacing when the crop row features are unobvious, furthermore identify the navigation path by the combined images with obvious row features. Under the condition that crops possess no obvious row features, the above method can still accurately identify the navigation path. Present the horizontal scanning method to get the precise navigation path. Improve the horizontal scanning method by selecting the appropriated value of d and the Bezier curve fitting, which smooth the navigation path and improve the real-time simultaneously.(7) Set ROI window to recursively track navigation path and present the concept of state curvature, and introduce the principles of feedback and previewing to pesticide spraying robot guidance control. By combining the feedback navigation information with the current navigation information and the previewing navigation information, design the serial navigation path state BP network and guidance control parameters BP network, which achieves the pesticide spraying robot navigation parameters acquisition based on closed-loop control strategy. Experimental results show the biggest deviation between practical with network output value in lateral deviation is -7cm and mean value is -0.6cm and mean square deviation is 3.2cm. The biggest deviation between practical with network output value in steering angle is -3°and mean value is -0.45°and mean square deviation is 1.2°.There is a little deviation in lateral deviation and in steering angle, and high precision in navigation parameter acquisiton.更多还原