关于深度学习结合软性机械手几何模型进行成堆物体抓取位置检测的研究

发布时间：2018-05-19 06:29

本文选题：深度学习 + 几何模型　；参考：《广东工业大学》2017年硕士论文

【摘要】：机器人利用机器视觉进行物体抓取是机器人应用领域的热门研究之一。目的在于靠机器视觉检测出被抓取物体的可靠抓取位置和方向,进而通过运动规划算法控制机械臂完成抓取动作。但是在这一领域暂未有普适的方法,特别是在需要同时考虑可靠性和安全性的成堆物体抓取应用。此论文提出一种利用深度学习网络结合软性机械手几何模型的方法,用于分析处理安装在机器人中部的单个深度摄像头所获取的3维点云数据,实现在未知的成堆物体中检测出可供机械臂进行可靠安全抓取的物体位置和方向。该论文所提出的方法在考虑成堆物体抓取碰撞检测问题的同时,并没有进行像其他论文所常用的图像分割和物体识别的处理技术。特别地,此论文还分析了深度卷积神经网络的时间和空间复杂度,并得出两个结论用于为深度卷积神经网络的设计提供参考。同时,进行多个实验来得出不同层数的神经网络对应的训练时间和精度,不同层数神经网络进行假设可抓取超平面再筛选的结果对比,从而验证前面由理论推出来的两个结论。经上述理论及实验两方面分析,为深度学习模型的设计起到一定的指导性作用。值得一提的是,本论文中深度学习模型训练数据(所有的假设超平面)的标签根据一些标准进行自动标定。也就是说这些用于判断该假设超平面是否符合抓取条件的标准被深度学习网络模型所取代。首先,由于软性机械手相较于硬性机械手更具灵活性,为软性机械手设计合适的几何模型成为避免碰撞的重点,用于在3维点云数据空间中搜索符合可抓取条件的抓取点和抓取方向。该几何模型内部必须包含足够的点云数据,并且外表面不会与其他点云数据有重合的部分,从而分别保证其抓取的可靠性和安全性。利用几何模型搜索出来符合条件的抓取点和抓取方向,合称之为假设可抓取超平面。第二,进一步考虑抓取可靠性,利用深度学习网络Mod-Le Net对搜索出的假设超平面进行分类和排序,以便找出较为可靠的抓取位置和方向。经过Mod-Le Net与支持向量机技术的对比实验,通过Mod-Le Net的筛选过后的可抓取超平面的质量和可靠性要比通过支持向量机的高,而且数量也相对较少,也就是说在运动控制方面会比较节省时间。
[Abstract]:Robot object capture using machine vision is one of the hot research in robot application field. The aim of this paper is to detect the position and direction of the captured object reliably by machine vision, and then to control the robot arm to complete the grab by motion planning algorithm. However, there is no universal method in this field, especially in the application of stacks of objects which need to consider both reliability and safety. In this paper, a method of using depth learning network combined with geometric model of soft manipulator is proposed to analyze and process 3D point cloud data obtained by a single depth camera installed in the middle of the robot. The position and direction of the object which can be reliably and safely grasped by the manipulator can be detected in the unknown stacks of objects. The method proposed in this paper does not deal with image segmentation and object recognition as commonly used in other papers, while considering the problem of collision detection. In particular, the time and space complexity of the deep convolution neural network is analyzed, and two conclusions are drawn to provide a reference for the design of the deep convolution neural network. At the same time, several experiments were carried out to obtain the training time and accuracy of neural networks with different layers. The neural networks with different layers were supposed to be able to grasp hyperplane and then compared the results of screening, thus verifying the two conclusions deduced from the theory. Through the theoretical and experimental analysis above, it plays a guiding role in the design of the deep learning model. It is worth mentioning that the labels of the depth learning model training data (all hypothesized hyperplanes) in this paper are automatically calibrated according to some criteria. In other words, these criteria used to determine whether the hyperplane meets the grasping condition are replaced by the depth learning network model. First of all, because soft manipulators are more flexible than rigid manipulators, designing suitable geometric models for soft manipulators becomes the focus of collision avoidance. It is used in 3D point cloud data space to search for grab points and grab directions that meet the grabability criteria. The geometric model must contain enough point cloud data, and the outer surface will not overlap with other point cloud data, so as to ensure the reliability and security of its capture. The geometry model is used to search the grasping points and directions which meet the conditions, which is called hypothetically grabbing hyperplane. Secondly, the grabbing reliability is further considered, and the hypothetical hyperplane is classified and sorted by using the deep learning network (Mod-Le Net) in order to find out the more reliable grab position and direction. Through the contrast experiment of Mod-Le Net and support vector machine, the quality and reliability of grabable hyperplane after Mod-Le Net screening is higher than that of support vector machine, and the quantity is relatively small. In other words, it saves more time in motion control.
【学位授予单位】：广东工业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP242

【参考文献】