【作者】 张春龙；

【导师】 李伟；

【作者基本信息】 中国农业大学 ， 机械制造及其自动化， 2014， 博士

【摘要】 机器视觉技术是农业机器人获取作业目标信息的重要手段,可实现自然环境下目标的识别、定位与跟踪,为伺服控制末端执行器完成作业任务提供决策依据。本文针对农业机器人视觉系统面对的两类典型对比性环境——农田和果园,结合农田蔬菜株间锄草,和果园产量监控及预测任务,对目标信息获取方法及视觉伺服策略展开研究。两种环境作业难点和区别在于：田间蔬菜图像以二维田地等为背景,目标与背景颜色区别较大,但难点在于株间锄草对实时性要求较高；果园果树图像难点在于以三维空间物体为干扰背景,绿色苹果与枝叶等背景颜色相近且相互遮挡,但产量监控和预测可离线分析,对实时性要求不高。本文分别设计了视觉伺服系统用于控制株间锄刀避苗锄草,和主动视觉伺服的最佳取景方位搜索策略用于果园产量监控及预测,以降低枝叶遮挡的影响。主要研究内容如下：(1)研究采用G-R>Tr且G-B>Tb因子对田间苗草图像进行背景分割,设计了基于二维直方图的类区域标记法,将对图像的区域标记搜索转换为对图像二维直方图的区域标记搜索,降低了搜索目标数和搜索时间。采用图像二维直方图区域组合特征为判断依据,实现了快速准确的蔬菜苗株间接识别定位,避开了具体识别区分作物和杂草的复杂过程。试验表明,样机系统中目标识别定位算法耗时小于16ms,平均正确识别率为97.34%；(2)针对大间隔种植的蔬菜等作物,设计了在左右对应作物行之间寻找作物包围盒重叠区域的方法,提取了相对准确的蔬菜行导航离散点,进而通过Hough变换对离散点进行拟合,实现了导航路径及导航数据的获取；(3)株间锄草机器人作业时以蔬菜苗株为参照目标,视场内苗株相对于月牙形株间锄刀的位置时刻变化,针对该运动目标的跟踪问题,研究了株间锄刀的视觉伺服控制策略,包括锄刀转速转角跟踪控制原理、视觉盲区补偿计算等；设计了视觉伺服与人机交互系统软件,实现了目标的识别定位与跟踪、机器人作业状态监控、操作指令输入及辅助操作信息的反馈等功能；(4)研究了绿色苹果在近色枝叶等背景中的识别计数方法。利用带有环形辅助光源的相机系统采集果树夜间图像以避免自然光照的影响,设计了以归一化的g分量和H、S颜色分量为特征参数的支持向量机(SVM)分类器,结合基于超绿特征(2G-R-B)的阈值分类器,实现了绿色苹果有效识别；针对果实粘连问题,对粘连区域进行欧氏距离变换,并采用分水岭算法进行分割。对64幅果树夜间图像进行试验表明,该方法识别计数的平均正确率89.3%。(5)针对视觉图像中果实遮挡问题,研究了相机取景方位对果树空间的可见、遮挡区域与果实探测能力之间的关系。通过重建果树三维密集点云并获取相机内外部参数,基于小孔成像模型,采用像素点反向投射方法构建果树点云空间遮挡地图；对单棵果树可见的图像采集点,采取不同的图像采集顺序,通过分析各方法对果树点体素可见能力、遮挡探测能力和树上苹果探测能力发现,可见能力优先法采集图像,能够更快探测到趋近于真实值的苹果数量；(6)设计了具有圆柱面相机运动空间的主动视觉系统,研究了基于粒子群优化算法(PSO)的用于主动视觉伺服的最佳取景方位搜索策略,结合相机点可见能力优先的最佳取景方位判断依据,仿真实现了果树点云空间的主动视觉最佳取景方位搜索策略。更多还原

【Abstract】 As one of the most important method to realize information acquisition for agricultural robot, the machine vision technology could help recognize, locate and track targets to finish the servo control of the end-effector. For agricultural robot the machine vision system offen needs to work in two typical environments, the field environment and the orchard environment. Compare the different purpose of intra-row weeding robot for transplanted vegetables and the orchard yield estimation system based on machine vision, we could find that, the weeding robot needs to locate single vegetables quickly in images with a large color difference between the plants and2D field background. While the orchard yield estimation system could do images anylysis off line, but as the image background contains more3D objects and the color of green apple targets and background is so similar that the situation is more complex. A vision servo control system was studied to realize intra-row weeding without damaging crops, as well as an optimal camera pose search method to reduce the occlusion for the orchard yield estimation based on active vison servo technolog. Following is the main research contents:1. With a background segmentation way of G-R>Tr&G-B>Tb, a region-labeling-kind method based on2D histogram was presented to recognize and locate individual crops. It transformed the image searching to the histogram searching, which could reduce the number of individual areas and the searching time. The individual crops could be recognized indirectly by comparing the local region features of2D histogram. The experiment results showed that the algorithm time cost was16ms, with a correct recognition rate of97.34%on individual crops.2. As the crop space is large, a method of tracking navigation points between corresponding crop rows was studied to fit the navigation line based on Hough transformation. It showed a higher accuracy than the traditional navigation points tracking method.3. A vison servo control method of the intra-row weeding robot was presented to track the relative moving crops in the visual-field, which could realize the angle and rotating speed control of the crescent weeding hoe as well as the tracking of crops in the blind area of vison system. A vison servo control and human-computer interaction system was also designed to realize the vison servo control, the working stadus monitoring, the commond inputting and operation assisting information displaying.4. An image capturing system composed of two color cameras and an active flashing light was used to capture apple tree images at night. And a green apple recognition method was proposed. A hybrid classifier including an SVM method based on the advantage of H, S and normalized g and a Super-G method (2G-R-B) was developed to segement apple areas. To seperate the connected apple regions, the Euclidean distance transformation and a watershed method was used. The analysis of experimental results regarding64images showed that the average rate of correct recognition is89.30%.5. To reduce the occlusion area of the apple tree, the relationship between the visibility, the occlusion exploring ability and apple detection ability of different camera poses was studied. With the3D tree point cloud and the camera parameters from3D reconstruction of the image sequences, the occlusion map was generated based on the pinhole camera model. According to the anylysis of the different order of the existing camera pose sequence, the best visibility decision method could detect more apples after taking the same number of pictures.6. An active vision system was designed with the motion range on a semi-cylinder surface. According to the best visibility decision method, a searching strategy based on the Partical Swarm Opimization was presented to figure out the opitimal camera poses. The simulation experiments showed that this method could help search the opitimal camera poses.更多还原