【作者】 姚立健；

【导师】 丁为民；

【作者基本信息】 南京农业大学 ， 农业机械化工程， 2008， 博士

【摘要】 茄子作为人们生活中常见的蔬菜,在我国大江南北都有广泛的种植。本文以苏皖一带常见的自然生长的长紫茄为研究对象,以实现成熟茄子自动识别、定位和机械臂避障路径规划为研究目的。完成的主要工作有:1根据自然生长茄子图像的灰度和颜色特点,在RGB中,对6种色差组合进行分析,发现目标和背景的R-B、G-B和G-R差别较大;在HSI空间中,H通道的目标和背景有明显差别。因此,在阈值分割中,采用基于遗传算法的自动选取阈值法(Otsu法)对图像进行分割;在去除残留噪音的处理中,采用标注的方法对二值图像的各连通区域进行面积统计。保留最大面积的区域,从而使分割效果大大改善。利用多参数来衡量分割效果,使评价尽可能地做到客观、合理。2尽管单一阈值法对茄子图像分割能取得一定效果,但依然存在较大的残留噪音。在分析茄子图像色差和色调的基础上,选取R-B、G-B和H作为自组织特征映射网络(SOFM)的输入特征向量,利用该网络自组织学习的特征进行聚类;采用信噪比、面积比、分割时间和傅立叶边界描述子等指标来评价分割精度;并据此确定了SOFM网络的输入特征向量的个数、输出神经元个数、训练步数、拓扑函数、距离函数、学习函数等参数。使用傅立叶描述子的能量谱值来评价分割边缘的相似程度,克服了其他边缘描述子依赖边缘点的个数和起始位置等限制。实验证明,基于SOFM神经网络图像分割评价优于单一阈值分割,适合复杂背景的彩色图像分割。3提出了一种基于改进型广义Hough变换的空间有部分遮挡的茄子目标识别方法。用广义柱近似描述样本茄子形状;在描述茄子空间位姿的6个参数中,选择两个位置参数和一个旋转参数等3个主要参数;通过坐标转换获得了不同位姿的立体目标在平面上的投影边缘。将缩放、旋转等运算提前于参数表制作阶段进行,采用了“形状相似度”的方法初选缩放索引和旋转索引,有效降低了搜索时间,提高了搜索精度;建立茄子外形的4维参数索引表;适当提高梯度索引步长,可避免参考点在累加器中的排布过于分散,便于从潜在参考点中找出最终参考点坐标;采用改进型的广义Hough变换计算茄子目标的潜在位置,并通过比较各旋转角度下的“面积比”,筛选出目标实际的位姿。实验表明:改进型广义Hough变换对空间不同位姿、部分遮挡情况下目标的识别具有良好的效果。4根据针孔透视模型,采用了一种平面标定法,对摄像机模型进行分析和标定。左右摄像机标定过程中,利用最小二乘法标定了摄像机的内外参数;结合实验室已有的实验条件,选用平行式的双目立体视觉;根据自然环境中茄子生长的特征,提出了一种基线选择的方法,并在此基础上,通过实验确定两摄像机的合理基线距离2a应大于132mm,合理的测量深度范围z应在[700,1200]mm之间;根据左右摄像机内外参数,对摄像机的位姿进行调节,使之满足Marr约束4;通过对内外参数的求取和基线的选择,建立了茄子收获机器人的双目立体系统;选择目标的形心作为特征匹配点。在此条件下,其测量相对误差可控制在2%以内,能够满足在农业环境里机械手作业要求。5提出了一种机械臂在三维空间的避障方法。将空间障碍物等效为可以用数学建模的轴截面为圆或矩形的圆柱扇环,将三维空间的路径规划简化为二维,提高了控制的实时性;将障碍物等效从工作空间转换到C-空间中,使对机器人的控制直接作用于关节,避免了使用雅克比矩阵的逆阵进行复杂的坐标转换;将C-空间映射到图像矩阵中,通过对图像进行适当的处理,规避了在使用A*算法寻优时可能出现的失败。实验表明,该避障路径规划方法计算量小,实时性好,适合自然生长状态下的茄子自动收获。更多还原

【Abstract】 Eggplant, as a common vegetable in people’s life, is widely planted in all parts of China. In the study of this paper, research object is natural growth long-purple eggplant which is familiar in Jiangsu and Anhui province. Automatic recognition and location of the ripe eggplant and obstacle avoidance planning for manipulator were the research purpose in this paper. The work mainly has been accomplished as followed:1. According to the gray-level and color characteristics of natural growth eggplant image, analysis on six kinds of color-difference combination was implemented in RGB space. It was found that there was great difference between target and background among R-B, G-B and G-R. In HSI space, target and background had great difference in H channel. According to the above analysis, the auto threshold-adaptive method of Otsu operation was specially used for threshold segmentation based on genetic algorithm. And upon the elimination of the residue noise, labeling method for statistical calculation was introduced for the connected regions of the binary image. The maximum areas were preserved to improve segmentation effects. In addition, multi-indices were applied to assess the effect of segmentation, in order that to the great extent the assessment was subjective and reasonable.2. Although the effect had been achieved on eggplant image segmentation by using single threshold method, it also leaved more residue noises. The purpose of this article was to segment eggplant from its complex background. R-B, G-B and H were selected as the input-vectors of the SOFM network based on analyzing the color-difference and hue characteristics of eggplant image. The input-vectors were classified by the characteristics of self-organizing of this network. In order to make the segmentation results objective and reasonable, signal-noise ratio, area ratio, segmentation time and Fourier-Descriptor were adopted to evaluate the segmentation precision. Many SOFM network parameters were determined such as the number of input feature vector and output neurons, training steps, topology function, distance function and learning function, etc. The energy spectrum of Fourier descriptors was already adopted to evaluate the similar degree of segmentation edge. This overcame other descriptor limitations which depended on the number of edge points and initial position. The experiment demonstrated that SOFM network was better than the single-threshold segmentation and more suitable for the color image segmentation with complex background.3. One method has been introduced for recognizing the partially occluded eggplant based on improved generalized Hough transform in three-dimensional space. The shape of eggplant was approximately described by using generalized-cylinders from the view of biology. In order to describe the pose of eggplant, three main parameters such as two location parameters and a rotation parameter were selected from six factors. Different edges of projection were obtained through coordinate transformation. Scaling and rotation operation was ahead of parameter index table establishment. Shape similarity degree method was adopted to primarily select scaling index and rotation index. It could effectively reduce the searching time and improve the searching precision. 4-D parameter index tables were established to describe the shape of eggplant. Properly increasing step size of gradient index could avoid across dispersion of array in accumulator. And it was also easy to search the final reference point coordinates from potential points. Improved Generalized Hough Transform was used to count the potential pose of object and the real pose was screened by comparing the area-ratio of different rotation angles. The experiment demonstrated that it was feasible and effective to recognize different pose and partially occluded object by using Improved Generalized Hough Transform.4. The model of camera calibration was discussed and the calibration results of the left and right cameras were provided based on calibration method. The intrinsic and extrinsic parameters were obtained by the least square method with the pinhole model. The parallel model of stereo vision was selected according to the existing devices of lab. Based on the characteristics of natural growth eggplant, one effective method for selecting baseline length was introduced. The appropriate baseline length between cameras was more than 132mm and the feasible measuring distance range was from 700mm to 1200mm. According to the intrinsic and extrinsic parameters of left and right cameras, we made adjustment on the intrinsic and extrinsic parameters of cameras to fit Marr constraints, and thus the binocular stereovision system was established. We selected centroid of the target as match features. Under this condition, the measure relative error could be controlled within 2% which was satisfied with the requirements of the manipulator in agricultural environment.5. An avoiding obstacle method has been introduced for manipulator in 3D space. Obstacles were made equivalent to cylindrical-rings mathematical model which axial section was circle or rectangular and a three-dimensional planning path problem was reduced to two-dimensional one, which greatly improved real-time control performance. Obstacle transforming from Work-space to Configuration-space could directly control the robot joints so as to avoid complex coordinate translation using Jacobian inverse matrix. Path planning failure under A-star algorithm could be evaded by processing properly image matrix which mapped from C-space. Experimental results showed this algorithm which was small computational amount and good real-time performance was suitable for natural growth eggplant harvest.更多还原