【作者】 张磊；

【导师】 姜军生；

【作者基本信息】 山东科技大学 ， 机械制造及其自动化， 2020， 硕士

【摘要】 重农固本,是安民之基,“三农”问题一直是国家重点关注的问题之一。在持续大力发展工业经济的同时,不断地优化农业产业结构,调整产业政策,促进农业产业进步与发展。随着现代科学技术的发展进步,逐渐向农业产业现代化,智能化,自动化产业新格局迈进。果实的识别与定位是智能采摘的关键。针对青椒果实检测过程中果实与植株背景颜色相近以及温室内环境背景复杂的问题,本文以温室大棚环境下的青椒果实作为研究目标,对采摘机器人的目标检测及定位技术进行研究。本文主要研究如下:首先,研究温室大棚环境下的青椒果实,提出使用Mask-RCNN目标检测算法对青椒果实目标进行的分割识别。研究Mask-RCNN网络结构,并在此基础上优化深度残差网络及特征金字塔网络结构。其次,使用温室环境下采集到的青椒图像,首先使用Labelme对原始图像数据集进行图像标注,生成掩膜(mask)图像,用以计算训练过程中的反向损失以及进行参数的优化。并通过将带有注释的掩膜(mask)图像与掩膜(mask)的预测结果进行比较,从而对用于实例分割的训练模型的性能进行评估。使用ResNet-50作为青椒目标检测模型的骨干网络,设置IOU为0.55,并采用迁移学习方式对标注数据集进行训练,对损失值进行统计,生成训练损失曲线,损失值都能够收敛并稳定在一个范围内。使用150幅测试集图像进行模型性能验证,通过结果统计分析可得:获得平均精准率(AP)为0.944,均值平均精度(mAP)为0.927,检测平均用时为1.2s/张,能够达到采摘机器人对于果实目标实时检测的要求。然后,研究双目摄像机标定及三维立体匹配算法。根据摄像机成像模型、双目摄像机成像原理及深度计算原理,确定使用张正友标定法获取双目摄像机的内外参数。使用基于质心的特征点匹配方法,选择合适的匹配约束条件及相似性度量函数进行三维立体匹配。最后,使用ZED Stereo Camera双目立体摄像机、NVIDIA JETSON NANO嵌入式开发板、显示器、三脚架等组成双目立体视觉系统,并对系统进行标定及立体校正,然后获取果实质心坐标并进行立体匹配,最后进行空间测距实验,实验表明,双目摄像机实际测量距离与实际距离误差范围在0～20mm以内,该误差范围内基本上能够满足青椒果实采摘工作的精度要求。更多还原

【Abstract】 Emphasizing agriculture and solidifying the foundation is the foundation of Anmin.The issue of "three rural issues" has always been one of the key concerns of the state.While continuing to vigorously develop the industrial economy,constantly optimize the agricultural industrial structure,adjust industrial policies,and promote the progress and development of the agricultural industry.With the development and progress of modern science and technology,it is gradually moving towards the modernization,intelligence and automation of the agricultural industry.Fruit identification and positioning is the key to intelligent picking.Aiming at the problems that the background color of the green pepper and the plant are similar during the detection of green pepper fruit and the complex environmental background in the greenhouse,this paper takes the green pepper fruit in the greenhouse environment as the research goal and studies the target detection and positioning technology of the picking robot.The main research of this article is as follows:Firstly,the green pepper fruits in the greenhouse environment were studied,and the Mask-RCNN target detection algorithm was proposed to segment and identify the green pepper fruits.Study the Mask-RCNN network structure,and optimize the deep residual network and feature pyramid network structure on this basis.Secondly,using the green pepper images collected in the greenhouse environment,first use Labelme to label the original image data set to generate a mask image to calculate the reverse loss during the training process and optimize the parameters.And by comparing the annotated mask image with the prediction result of the mask,the performance of the training model used for instance segmentation is evaluated.Use ResNet-50 as the backbone network of the green pepper target detection model,set the IOU to 0.55,and use the transfer learning method to train the labeled data set.perform statistics on the loss values,and generate a training loss curve.The loss values can all converge and stabilize in one.Within range.Using 150 test set images for model performance verification,the results can be obtained through statistical analysis:the average accuracy rate(AP)is 0.944,the average average accuracy(mAP)is 0.927,and the average detection time is 1.2s/piece,which can achieve the fruit picking robot Requirements for real-time target detection.Then,study binocular camera calibration and three-dimensional matching algorithm.According to the camera imaging model,binocular camera imaging principle and depth calculation principle,it is determined to use Zhang Zhengyou calibration method to obtain the internal and external parameters of the binocular camera.The centroid-based feature point matching method is used to select appropriate matching constraints and similarity measurement functions for three-dimensional matching.Finally,use the ZED Stereo Camera binocular stereo camera,NVIDIA JETSON NANO embedded development board,display,tripod,etc.to form a binocular stereo vision system,and calibrate and stereo correct the system,and then obtain the fruit centroid coordinates and perform stereo matching,Finally,a space ranging experiment is carried out.The experiment shows that the error range between the actual measurement distance and the actual distance of the binocular camera is within 0～20mm,and the error range can basically meet the accuracy requirements of green pepper fruit picking.更多还原