基于深度学习算法的NAO机器人目标识别与跟踪

合集下载
  1. 1、下载文档前请自行甄别文档内容的完整性,平台不提供额外的编辑、内容补充、找答案等附加服务。
  2. 2、"仅部分预览"的文档,不可在线预览部分如存在完整性等问题,可反馈申请退款(可完整预览的文档不适用该条件!)。
  3. 3、如文档侵犯您的权益,请联系客服反馈,我们会尽快为您处理(人工客服工作时间:9:00-18:30)。
III
目录
第 1 章 绪论 .........................................................................................................1 1.1 课题来源 .................................................................................................... 1 1.2 课题研究背景和意义 ................................................................................ 1 1.3 国内外研究现状 ........................................................................................ 3 1.3.1 深度学习目标识别算法国内外研究现状 ......................................... 3 1.3.2 机器人视觉国内外研究现状 ............................................................. 6 1.4 本文论文结构 ............................................................................................ 8
摘要
随着深度学习算法的快速发展,图像目标识别领域取得了重大突破,无论是 对目标种类的判定还是对目标位置的检测深度学习算法都取得了远高于传统机 器学习算法的准确率。机器人技术是工程和科学的跨学科分支,包括机械工程、 信息工程、计算机科学等。机器人技术涉及机器人的设计制造以及用于控制机器 人运动、传感器反馈和信息处理的计算机系统。机器人视觉是机器人技术中的一 个研究热点,涉及如何利用机器人采集环境的视觉信息并获得数字图像或视频 的高层次理解。从工程角度看,它试图自动化人类视觉系统可以完成的任务。
(2)为了实现 NAO 机器人跟踪目标,对 NAO 机器人摄像头测距进行了研 究。尽管 NAO 机器人有上下两个摄像头,但二者视域没有交叉也不能同步运行, 故本课题构建 NAO 机器人单目测距模型测距。NAO 机器人跟踪目标设定为水 平面上,因此建立一种地平面约束测距模型,实验表明 NAO 机器人单目测距取 得了较高准确率。
(3)针对 NAO 机器人硬件平台,使用 Choregraphe 软件建立了一种 C/S 模 式的 NAO 机器人目标识别与跟踪系统,将深度学习目标识别算法,NAO 机器人通过 HTTP 协议从服 务器调用封装好的算法模块。实验结果表明,NAO 机器人目标识别与跟踪取得 较高准确率。
第 2 章 基于深度学习目标识别算法的优化 ...................................................10 2.1 深度学习算法结构 .................................................................................. 10 2.1.1 损失函数定义 ................................................................................... 10 2.1.2 卷积神经网络结构 ........................................................................... 11 2.2 神经网络优化方法 .................................................................................. 15 2.2.1 反向传播算法和梯度下降算法 ....................................................... 15 2.2.2 学习率设置 ....................................................................................... 18 2.2.3 正则化方法 ....................................................................................... 19 2.3 Faster R-CNN 目标识别算法模型研究................................................... 20 2.3.1 卷积神经网络特征提取 ................................................................... 20 2.3.2 区域建议网络 ................................................................................... 21 2.3.3 RoI 池化层 ......................................................................................... 23 2.3.4 分类与回归 ....................................................................................... 23 2.4 基于 Faster R-CNN 目标识别算法模型的改进..................................... 24 2.4.1 改进卷积神经网络特征提取 ........................................................... 24 2.4.2 改进区域建议网络 ........................................................................... 27 2.5 基于 TensorFlow 的深度学习目标识别实验平台搭建......................... 27 2.5.1 TensorFlow 框架概述........................................................................ 27 2.5.2 TensorFlow 深度学习框架特性........................................................ 28 2.5.3 深度学习目标识别实验软硬件配置 ............................................... 29 2.5.4 本地 TensorFlow 实验环境安装流程.............................................. 29 2.5.5 谷歌云平台环境配置 ....................................................................... 30
(1) Studied the basic principle of deep learning algorithm, deeply studied the Faster R-CNN target recognition model, improved the feature extraction network of Faster R-CNN, and designed a 98-layer volume based on the residual network model. Neural network to extract image features. The Faster R-CNN target recognition model is optimized for small target recognition. Build a training platform on Google Cloud, use PASCAL VOC as a training data set, train the improved Faster R-CNN target recognition model, and obtain higher accuracy than the original model.
关键词:深度学习算法,NAO 机器人,识别,跟踪
I
Abstract
With the rapid development of deep learning algorithms, a major breakthrough has been made in the field of image target recognition. Both the determination of target categories and the detection of target locations have achieved much higher accuracy than traditional machine learning algorithms. Robotics is an interdisciplinary branch of engineering and science, including mechanical engineering, information engineering, and computer science. Robotics involves the design and manufacture of robots and computer systems for controlling robot motion, sensor feedback, and information processing. Robot vision is a research hotspot in robotics. It involves how to use robots to collect visual information of the environment and obtain high-level understanding of digital images or videos. From an engineering perspective, it attempts to automate tasks that the human visual system can accomplish.
II
(3) For NAO robot hardware platform, Choregraphe software was used to establish a C/S NAO robot target recognition and tracking system, and the deep learning target recognition algorithm module and NAO robot monocular distance measurement module were implemented on the server. Package, NAO robots call packaged algorithm modules from the server via the HTTP protocol. Experimental results show that NAO robot target recognition and tracking achieve higher accuracy. Key words: Deep learning, NAO robot, Recognition, Tracking
(2) In order to achieve the NAO robot tracking target, the NAO robot camera ranging was studied. Although the NAO robot has two cameras up and down, but the two sights do not intersect and can not run simultaneously. Therefore, this project constructs the NAO robot monocular ranging model ranging. The NAO robot tracking target is set to the horizontal plane. Therefore, a ground plane constraint ranging model is established. Experiments show that the NAO robot achieves a high accuracy in the monocular range finding.
The deep learning algorithm applied to the field of robot vision will greatly improve the interactive ability of the robot's environment and improve the intelligence of the robot. In this paper, the NAO robot is used as a research platform for robot vision technology. The method of target recognition and tracking based on deep learning algorithm is mainly studied. The specific work of this article is as follows:
深度学习算法应用于机器人视觉领域将大大提高机器人的环境的交互能力, 提升机器人的智能程度。本文以 NAO 机器人作为机器人视觉技术的研究平台, 重点研究了基于深度学习算法的目标识别与跟踪方法。本文的具体工作如下:
(1)研究了深度学习算法基本原理,对 Faster R-CNN 目标识别模型进行了 深入研究,针对 Faster R-CNN 的特征提取网络进行了改进,设计了一种基于残 差网络模型的 98 层卷积神经网络提取图像特征。基于 Faster R-CNN 目标识别模 型针对小目标识别进行了优化。在 Google 云上搭建训练平台,以 PASCAL VOC 作为训练数据集,训练改进的 Faster R-CNN 目标识别模型,取得了高于原始模 型的准确率。
相关文档
最新文档