【作者】 纪超；

【导师】 李伟；

【作者基本信息】 中国农业大学 ， 农业机械化工程， 2014， 博士

【摘要】 果蔬采摘机器人作业于温室非结构环境中,在多变的光线条件下,避让植株茎叶准确定位目标,完成果实无损伤抓取与切割。果蔬采摘信息获取是采摘机器人研究的难点与关键,根据果实与背景颜色差异,采摘对象可分为显著色差系果蔬与近色系果蔬,本文分别选取草莓和小型西瓜作为两类果蔬的典型代表,研究了采摘对象视觉信息获取方法,构建了果蔬采摘机器人系统。应用机器视觉、图像处理、光谱分析等技术,探索了果实识别、空间匹配及三维坐标定位方法,研制了采摘机器人模块化系统,并进行了温室采摘作业试验。主要研究内容和成果如下：(1)研究了基于颜色和形态特征的草莓采摘信息获取方法。根据果实与背景在4种常用色彩空间下的颜色分布特征,选择色差最为显著的R、G通道进行图像分割,通过轮廓补偿法补全果萼遮挡区域,获得完整果实区域图像。利用图像蒙版滤除近色系背景干扰,提取非全熟果实的青色区域,完成草莓成熟度判定。根据采摘机器人末端执行器结构与草莓果梗空间位姿特征,设置采摘高度线与采摘点疑似矩形兴趣区,提取了采摘点图像坐标。对草莓果实采摘信息获取方法进行试验,结果表明草莓识别成功率为94.2%,采摘点定位准确率为93.0%。(2)研究了基于近红外图像的小型西瓜采摘信息获取方法。通过比较立体种植模式下小型西瓜果实、茎、叶片的光谱反射特性差异,选择在850nm附近波段下采集西瓜近红外灰度图像。根据果实与背景灰度分布特征,利用改进的Otsu算法完成了图像分割。采用“米”字型匹配模板识别获得“浓缩版”西瓜果实区域,有效降低阈值分割后果实粘连与小面积干扰影响。根据果实与果梗的空间姿态及相对位置特征,采用分块定位法获得切割点图像坐标。对小型西瓜采摘信息获取方法进行试验,结果表明：不同光照条件下,小型西瓜平均识别成功率为86%,采摘点与切割点定位准确率分别为93.0%和88.4%,为近色系果蔬采摘信息获取提供了一种技术思路。(3)研究了基于双层约束的采摘点空间匹配策略,探索了基于双目立体视觉的采摘点三维信息获取方法。根据草莓形态特征,利用果实区域与采摘点唯一对应关系,提出了基于全局特征与关系特征的草莓区域匹配方法,完成采摘点初次遴选；利用基于极线约束的采摘点空间匹配方法,完成匹配对象二次遴选。搭建了交叉式双目立体视觉硬件系统,完成了摄像机内、外部参数及手眼相对位置参数标定,建立了图像坐标系、摄像机坐标系及机械臂坐标系间的相互转换关系,获得了草莓采摘点三维坐标计算流程。(4)搭建了采摘机器人模块化样机系统,该机器人主要由双目立体视觉系统、机械手系统、中央控制器、导航行走平台、能源系统及其他附件组成。通过归一化色差2r-g-b分割图像并获取了导航线偏移信息,对四自由度关节型机械臂进行运动学逆解分析,获得目标位姿下各关节旋转角度参数。规划了机器人采摘动作流程,对采摘机器人进行了温室作业试验,机器人采摘成功率为86%,单次采摘循环耗时28s,机器人各功能模块运行良好,能够较好的适应温室作业环境。更多还原

【Abstract】 The fruit and vegetable harvesting robot worked in unstructured environment of greenhouse. Under the condition of complex light, the robot should position the target accurately and avoid touching the stem and leaf. And then the fruit would be grasped and cut nondestractively. Information acquisition of fruit was the difficulty and key of harvesting robot. According to the corlor difference between target and background, the study object could be classified into similar-color fruit and different-corlor fruit. Two representative fruits——strawberry and mini-watermelon were selected as study objects, and the havesting information acquisition of fruit was studied, and a fruit havesting robot system was developed. The techniques of machine vision, image processing, spectral analysis were applied to solve fruit recognition, spatial matching and three-dimensional coordinate location. At last, a modular prototype of havesting robot system was developed, and a trail of the robot was done. The main research contents and conclusions were as follows:(1) The information acqusition method of havesting strawberry based on color and morphology was studied. Firstly, according to the color distribution characteristics of fruit and background in four common color spaces, color components (R and G) were selected to image segmentation. Secondly, the region occluded by fruit calyx was inpainted through contour compensation, and the whole fruit region was recognized. Thirdly, the near-color background interference was filtered by image mask processing, and green region of immature fruit was extracted to judge the strawberry immaturity level. Then, according to the structure of robot end-effector and the spatial pose characteristics of strawberry stem, the picking line and rectangle region of interest were set to extract the image coordinate of picking point. At last, the experiment resut showed that the correct recognition rate of fruit was94.2%, as well as the rates of the picking point was93.0%.(2) The information acqusition method of havesting mini-watermelon based on near infrared image was studied. Firstly, the spectral characterisics of mini-watermelon’s fruit, stem and leaf were compared, and optimal wavebands near850nm were selected to capture near-infrared gray image. Secondly, According to the characteristics of fruit and backgroud’s gray pixel distribution, an improved Otsu algorithm was developed to segment image. Thirdly, a matching template likes as circle was proposed to detect fruit region concentrated, and the region adhesion and small area interference were reduced effectively. Then, according to the characteristics of spatial pose and relative position of the fruit and stem, the image coordinate of cutting point was located by "block-location method". Finally, a trail was done to test the algorithm of acqusiting mini-watermelon’s harvesting information, and the result show that the average correct recognition rate of fruit was86%under different illumation condition, as well as the rates of the picking point and the cutting point were93.0%and88.4%respectively. Meanwhile it provides a new technical idea for acqusiting the infromation of harvesting similar-color fruit.(3) The spatial matching strategy of picking point based on double layers constraints was studied and the three-dimensional coordinate location of picking point was explored. Firstly, according to strawberry’s morphology features and the corresponding relation between fruit region and picking point, the picking points were selected preliminary by a region matching method based on global features and relationship features. Secondly, the picking points were determinad by epipolar geometry constraint. Thirdly, the hardware system of binocular stereovision was developed, and then the intrinsic parameters and external parameters of camera and the hand-eye parameters were calibrated. After coordinate transformation model of the image, camera and manipulator was developed, the flow of locating picking point’s three-dimensional coordinate was obtained.(4) A modular prototype of havesting robot system was developed. The robot was constructed of binocular stereovision system, manipulator system, central controller, self-guided moving platform, battery system, and other appendix. Firstly, the image was segmented by normalized color difference (2r-g-b), and then the guide line was extracted. Secondly, the inverse kinematics of4-freedom joint type manipulator was analysised, and then the rotate angle of every joint at target position was obtained. Thirdly, the flow of the robot picking operation was planned, and a trail to test the robot performance was done, and the result show that the success rate of picking fruit was86%, and the execution time of a harvesting cycle was28s. Every fuctional modular of the robot run effectively and well adept to the working environment in greenhouse.更多还原