采摘机器人目标识别及定位研究
发布时间：2018-11-23 19:14
【摘要】：在自然环境下橘子目标所处的背景十分复杂,被枝叶遮挡或者果实之间叠加的现象非常普遍,这种环境的复杂性无疑给机器视觉系统的识别带来困难,导致采摘机器人不能有效、准确地识别目标。针对这一问题,本文对复杂环境中橘子目标图像的识别与定位问题进行了仿真与实验研究。主要工作内容如下:对复杂环境中橘子目标轮廓的识别方法进行了研究。介绍了传统的边缘检测算法,并对橘子图像进行了测试,测试效果显示该方法不能有效提取复杂环境中的橘子目标轮廓。将K-means聚类算法与Canny算法融合,采用K-means聚类算法从目标图像中分割目标物区域,结合Canny检测算法检测出目标物区域的轮廓,进而完成目标的识别,橘子图像测试结果验证了该方法的有效性。对重叠橘子目标轮廓分离方法进行了研究。在对腐蚀剥离法及分水岭分割法等方法分离重叠(邻接)目标的原理和特点比较基础上,研究了基于K-means聚类算法分离重叠橘子目标轮廓的方法,该方法对双果邻接、重叠的橘子目标图像进行了测试,测试结果看出重叠目标轮廓分离完整,体现了该方法的有效性。对橘子目标轮廓匹配进行了研究。为描述目标轮廓特征,引入几何不变矩参数作为轮廓的描述子,采用作差法的结果作为两幅图像中轮廓匹配测度值,实验数据表明,几何不变矩参数在橘子目标轮廓特征描述方面具有较好的效果,匹配能力良好。同时引入基于梯度法的Hough变换圆检测方法对类圆形橘子目标轮廓拟合重建,测试效果图显示该方法能够实现果实的有效定位。对基于单目视觉的目标深度进行了计算。移动摄像机采集同一场景下的两幅图像,提取相匹配的特征点,结合摄像机成像原理,计算出空间目标物距离摄像机的深度信息。最后,介绍了实验硬件系统,分别以橘子和大枣为实验对象,完成了复杂环境中目标轮廓的识别与定位实验,验证了本文方法的有效性。
[Abstract]:In the natural environment, the background of orange target is very complex, and it is very common to be occluded by branches or leaves or superimposed between fruits. The complexity of this environment undoubtedly makes it difficult to recognize the machine vision system. As a result, the picking robot can not recognize the target effectively and accurately. In order to solve this problem, the recognition and localization of orange target image in complex environment are studied by simulation and experiment. The main work is as follows: the recognition method of orange target contour in complex environment is studied. This paper introduces the traditional edge detection algorithm and tests the orange image. The test results show that the method can not effectively extract the orange target contour in complex environment. The K-means clustering algorithm and the Canny algorithm are fused, and the K-means clustering algorithm is used to segment the object region from the target image. The contour of the target region is detected by combining the Canny detection algorithm, and the target recognition is accomplished. The results of orange image test show that the proposed method is effective. The separation method of overlapping orange target contour was studied. On the basis of comparing the principle and characteristics of separating overlapping (adjacent) targets by corrosive stripping method and watershed segmentation method, the method of separating overlapping orange target contour based on K-means clustering algorithm is studied. The overlapping orange target images are tested, and the results show that the overlapping targets are separated completely, which shows the effectiveness of the method. The object contour matching of orange was studied. In order to describe the contour feature of the target, the geometric moment invariant parameter is introduced as the descriptor of the contour, and the result of the difference method is used as the contour matching measure value in the two images. The experimental data show that, The geometric moment invariant parameters have good performance in describing the contour feature of orange target, and the matching ability is good. At the same time, the Hough transform circle detection method based on gradient method is introduced to reconstruct the contour of the circular orange target. The test results show that the method can effectively locate the fruit. The target depth based on monocular vision is calculated. Moving camera collects two images in the same scene, extracts matching feature points, and calculates the depth information of the space object distance from the camera in combination with the principle of camera imaging. Finally, the hardware system of the experiment is introduced. Taking orange and jujube as experimental objects, the recognition and localization experiments of target contour in complex environment are carried out, and the validity of this method is verified.
【学位授予单位】：西安理工大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP242
本文编号：2352458
[Abstract]:In the natural environment, the background of orange target is very complex, and it is very common to be occluded by branches or leaves or superimposed between fruits. The complexity of this environment undoubtedly makes it difficult to recognize the machine vision system. As a result, the picking robot can not recognize the target effectively and accurately. In order to solve this problem, the recognition and localization of orange target image in complex environment are studied by simulation and experiment. The main work is as follows: the recognition method of orange target contour in complex environment is studied. This paper introduces the traditional edge detection algorithm and tests the orange image. The test results show that the method can not effectively extract the orange target contour in complex environment. The K-means clustering algorithm and the Canny algorithm are fused, and the K-means clustering algorithm is used to segment the object region from the target image. The contour of the target region is detected by combining the Canny detection algorithm, and the target recognition is accomplished. The results of orange image test show that the proposed method is effective. The separation method of overlapping orange target contour was studied. On the basis of comparing the principle and characteristics of separating overlapping (adjacent) targets by corrosive stripping method and watershed segmentation method, the method of separating overlapping orange target contour based on K-means clustering algorithm is studied. The overlapping orange target images are tested, and the results show that the overlapping targets are separated completely, which shows the effectiveness of the method. The object contour matching of orange was studied. In order to describe the contour feature of the target, the geometric moment invariant parameter is introduced as the descriptor of the contour, and the result of the difference method is used as the contour matching measure value in the two images. The experimental data show that, The geometric moment invariant parameters have good performance in describing the contour feature of orange target, and the matching ability is good. At the same time, the Hough transform circle detection method based on gradient method is introduced to reconstruct the contour of the circular orange target. The test results show that the method can effectively locate the fruit. The target depth based on monocular vision is calculated. Moving camera collects two images in the same scene, extracts matching feature points, and calculates the depth information of the space object distance from the camera in combination with the principle of camera imaging. Finally, the hardware system of the experiment is introduced. Taking orange and jujube as experimental objects, the recognition and localization experiments of target contour in complex environment are carried out, and the validity of this method is verified.
【学位授予单位】：西安理工大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP242
【引证文献】
相关期刊论文 前1条
1 初广丽;张伟;王延杰;丁南南;刘艳滢;;基于机器视觉的水果采摘机器人目标识别方法[J];中国农机化学报;2018年02期
，本文编号：2352458
本文链接：https://www.wllwen.com/kejilunwen/ruanjiangongchenglunwen/2352458.html
