【作者】 王亮；

【导师】 赵德安；

【作者基本信息】 江苏大学 ， 控制科学与工程， 2019， 硕士

【摘要】 在我国的农业生产中,果蔬采摘一直是果蔬产业的重要组成部分,而由于农业生产技术水平的落后,采摘工作需要耗费大量的人力物力,增加了农业生产的成本,制约了农业现代化的发展。因此,应加快对采摘机器人的研究,提高采摘机器人的采摘效率,推动农业的转型升级,实现高度自动化和精确化的农业发展目标。本文在国家自然科学基金项目(31571571)“基于快速视觉伺服控制的多照度环境苹果高效机器采摘方法研究”的资助下,主要对苹果采摘机器人的目标检测和路径规划算法进行了研究,主要完成的工作有:1、苹果采摘机器人视觉伺服系统的设计。根据实验课题的目标与要求,设计了苹果采摘机器人的视觉平台,并完成了硬件结构的搭建和软件环境的选择。2、苹果识别算法的研究。要想提高苹果采摘机器人的采摘效率,首先需要在图像中快速且准确地识别出苹果。本文选用了卷积神经网络模型SSD作为基础的网络架构,从默认框设计,匹配原则和损失函数详细介绍了该网络框架的基本原理和构造,并且从默认框和基础网络层两方面对基础的SSD网络框架进行了改进。对于改进后的SSD网络框架,采用了苹果数据集进行训练和测试,并且选择了表现最为优越的模型参数。最终改进的SSD网络框架能够准确且快速地在图像中识别出多个目标果实,得到其二维位置坐标。3、深度信息的获取。本文首先比较了三种不同原理的深度摄像头:结构光,双目视觉和TOF飞行时间法。综合本次课题要求和技术参数,选择了TOF相机作为采集深度信息的设备。其次结合之前获得的苹果二维位置坐标确定了苹果的深度位置,然后利用深度图上的信息确定苹果的周围区域是否存在障碍物,并且结合三维重构图和强度图上的特征信息来区分障碍物的种类。最终获得了目标果实及其周围不同种类障碍物的三维位置坐标,为路径规划的环境建模提供了完整的信息。4、采摘路径规划算法的研究。本文首先介绍了搜索理论基础和环境建模方法,其次结合上文获得的位置信息和栅格法将三维空间的采摘规划转化为在二维平面上路径规划的研究,获得了存在多个目标果实和不同种类障碍物的二维采摘地图,然后研究了标准A*算法的基本原理和主要步骤,并且针对标准A*算法存在的规划时间较慢,搜索节点过多等问题,分别从启发函数,列表的数据结构和移动代价值三个方面进行改进,以此使得改进的A*算法加快了规划路径的速度,减少了路径的长度并且更加适合实际的采摘环境。最后利用苹果和障碍物的三维位置信息将在二维采摘地图上规划的采摘路径扩展成实际采摘空间的三维路径点,为接下来的采摘实验提供准确的路径信息。5、采摘实验结果与分析。本文主要从采摘实验中的果实识别时间、路径规划时间、采摘路径长度和采摘总时间,来验证算法的有效性和实时性。通过采摘实验结果表明:改进的SSD模型不仅在识别速度上比基础模型提高了29.5%,而且也保证了较好的识别准确率和稳定性。而改进的A*算法在采摘总路径上比基础的算法提高了26%,在规划时间方面也提高了40%,从而保证了采摘总时间的缩短,提高了采摘机器人的采摘效率。最终实验结果可得改进后的算法抓取一个苹果的时间在11.4s左右,比改进之前提高了30.06%,满足了苹果采摘机器人的采摘要求。更多还原

【Abstract】 In China’s agricultural production,fruit and vegetable harvesting has always been an important part of the fruit and vegetable industry.Because of the backwardness of agricultural production technology,harvesting work needs a lot of manpower and material resources,which increases the cost of agricultural production and restricts the development of agricultural modernization.Therefore,we should speed up the research of picking robots,improve the efficiency of picking robots,promote the transformation and upgrading of agriculture,and achieve highly automated and accurate agricultural development goals.Under the support of the project of National Natural Science Foundation of China(31571571),"Research on the efficient apple picking method in multi-illumination environment based on fast visual servo control",this paper mainly studies the target detection and path planning algorithm of apple picking robot.1.Design of visual servo system for apple picking robot.According to the requirements and objectives of the experiment,the visual platform of apple picking robot was designed,and the hardware structure and software algorithm were built.2.Research on Apple recognition algorithm.In order to improve the efficiency of apple picking robot,it is necessary to recognize apples quickly and accurately in the image.In this paper,the convolutional neural network model SSD is chosen as the basic network architecture.The basic principle and structure of the network framework are introduced in detail from the default frame design,matching principle and loss function.The basic SSD network framework is improved from the default frame and basic network layer.For the improved SSD network framework,a dedicated Apple data set is designed for training and testing,and the best performance model parameters are selected.Finally,the improved SSD network framework can accurately and quickly identify multiple target fruits in the image,and obtain their two-dimensional position coordinates.3.Deep information acquisition.This paper first compares three different principles of depth cameras: structured light,binocular vision and TOF time-of-flight method.Based on the requirements and technical parameters of this project,TOF camera is selected as the equipment to collect depth information.Secondly,the depth of the apple is determined by combining the two-dimensional coordinates of the apple obtained before,and then the information of the depth map is used to determine whether there are obstacles in the surrounding area of the apple,and the types of obstacles are distinguished by combining the feature information of the three-dimensional reconstruction map and the intensity map.Finally,the three-dimensional coordinates of the target fruit and its surrounding obstacles are obtained,which provides complete information for environmental modeling of path planning.4.Research on the algorithm of picking path planning.Firstly,this paper introduces the search theory and environment modeling method.Secondly,combined with the location information and grid method obtained above,the three-dimensional picking planning is transformed into the path planning on the two-dimensional plane,and the two-dimensional picking map with multiple target fruits and different kinds of obstacles is obtained.Then,the basic principle and main steps of the standard A* algorithm are studied.To solve the problems of slow planning time and too many searching nodes in standard A* algorithm,the heuristic function,data structure of list and mobile cost are improved respectively,so that the improved A* algorithm speeds up the planning path,reduces the length of picking path and is more suitable for the actual picking environment.Finally,using the three-dimensional position information of apples and obstacles,the planned picking path on the two-dimensional picking map is extended to the three-dimensional path points in the actual picking space,providing accurate path information for the next picking experiment.5.Picking experiment results and analysis.This paper mainly verifies the effectiveness and real-time of the improved algorithm from the identification of fruit time,path planning time,picking path length and total picking time in the picking experiment.The experimental results show that the improved SSD model not only improves the recognition speed by 30% compared with the basic model,but also ensures better recognition accuracy and stability.The improved A* algorithm improves the total picking path by 26% and the planning time by 40%,which ensures the shortening of the total picking time and improves the picking efficiency of the picking robot.Finally,the experimental results show that the improved algorithm takes about 11.4 seconds to grab an apple,which is 30.06% higher than before,and meets the requirements of apple picking robot.更多还原