CASIA OpenIR  > 复杂系统管理与控制国家重点实验室  > 先进机器人
面向服务机器人的抓取位置智能检测方法研究
贾群
Subtype硕士
Thesis Advisor曹志强
2019-06
Degree Grantor中国科学院大学
Place of Conferral中国科学院自动化研究所
Degree Name工学硕士
Degree Discipline控制理论与控制工程
Keyword服务机器人,目标检测,目标分割,多任务深度卷积神经网络,抓取位置智能检测
Abstract

具备操作能力的服务机器人已成为机器人领域的研究热点,在航空航天、军事、助老助残以及家庭服务等领域具有广泛的应用前景。对机器人抓取操作任务来说,抓取位置检测具有重要的地位,是保证抓取成功的关键一环。本文主要面向服务机器人的抓取位置智能检测开展研究,主要内容如下:

首先,介绍了操作服务机器人的研究背景与意义,对目标检测、目标分割、抓取位置检测三方面的研究现状进行了综述,并对论文的内容结构和结构安排做了说明。

其次,开展了基于深度学习的目标检测方法与目标分割方法的选取。分析了Faster R-CNNYOLO两种目标检测方法,以检测精度和速度为指标在自建数据集上进行了测试,选择满足实时性要求且精度较高的Faster R-CNN作为后续多任务深度卷积神经网络的研究基础。同时,开展了FCNGCN两种目标分割方法在VOC2012CityspacesCOCO2014数据集上的精度对比实验,选择精度较高的GCN作为后续研究的基本目标分割框架。

第三,针对级联目标检测卷积网络和抓取位置检测卷积网络的抓取位置检测方法处理时间较长的问题,提出了一种基于多任务深度卷积神经网络的抓取位置智能检测方法。所提方法在Faster R-CNN目标检测模型基础上加入两点回归器,并设计多任务损失函数作为所用深度卷积神经网络优化的目标函数,从而实现通过一个深度卷积神经网络同时完成目标识别、定位、抓取位置检测三个任务。所提方法的有效性通过在自建数据集、Kinect拍摄图像以及机械臂上的实验进行了验证。

第四,为了解决全局语义分割方法处理速度较慢的问题,将轻量级网络和注意力机制融入全局卷积语义分割网络中,提出了一种基于改进全局卷积语义分割的抓取位置智能检测方法。所提方法利用轻量级网络替换全局卷积语义分割中的残差网,并添加空间和通道注意力机制对轻量级网络的输出进行自适应调整,进而结合重心法和主成分分析,实现对目标物体上可抓取位置的检测。在COCO2014数据集、Kinect拍摄图像以及机械臂上的实验结果表明所提方法能够有效从环境中分割出目标物体并检测出可抓取位置。

最后,对本文工作进行了总结,并指出了需要进一步开展的研究工作。

Other Abstract

Service robots with manipulation capability has become a hot research in robotics with broad prospects of applications in aerospace, military, aids of the aged and the disabled as well as family services. For the grasping task of service robots, the detection of grasping position plays an important role to guarantee the quality of the grasping. This thesis focuses on the intelligent detection of grasping position for service robots. The main contents are as follows:

Firstly, the research background and its significance of service robots with manipulation capability are introduced. The research developments of object detection, object segmentation and grasping position detection are reviewed. The contents and structure of this thesis are also given.

Secondly, the selection of object detection and semantic segmentation methods based on deep learning is conducted. The object detection methods Faster R-CNN and YOLO are analyzed and tested on a self-built dataset in terms of the detection accuracy and speed. The Faster R-CNN is selected as a basis of the subsequent multi-task CNN with a higher accuracy and real-time feature. Meanwhile, the accuracy comparison between the semantic segmentation methods FCN and GCN are experimentalized on VOC2012, Cityspaces and COCO2014 datasets. GCN with a higher precision is selected as the basic semantic segmentation framework of follow-up research.

Thirdly, for the problem where the solution by cascading object detection CNN and grasping position detection CNN leads to a longer processing time, this paper proposes an intelligent detection approach of grasping position based on a multi-task deep convolutional neural network. The proposed method adds a two-point regression to the Faster R-CNN object detection model with a multi-task loss function as the objective function of the deep convolutional neural network optimization. Therefore, three tasks including object recognition, localization, and grasping position detection are achieved by a single network. The effectiveness of the proposed approach is verified by experiments on the self-built dataset, images taken by Kinect, and the grasping of a manipulator.

Fourthly, in order to solve the problem of slower processing speed for the global convolutional semantic segmentation method, mobilenet and attention mechanism are integrated into the semantic segmentation network. An intelligent detection approach of grasping position based on the improved global convolutional semantic segmentation is then proposed. The resnet in the global convolutional semantic segmentation is replaced by mobilenet, and the spatial and channel attention mechanism is employed to adjust the feature maps from mobilenet. On this basis, the detection of grasping position is achieved by combining the gravity method and principal component analysis. The experiments on the COCO2014 dataset, images taken by Kinect, and the manipulator demonstrate that the proposed approach can segment the object from the environment with an effective graspable position on the object.

Finally, the conclusions are given and future work is addressed.

Pages1-80
Language中文
Document Type学位论文
Identifierhttp://ir.ia.ac.cn/handle/173211/23768
Collection复杂系统管理与控制国家重点实验室_先进机器人
Corresponding Author贾群
Recommended Citation
GB/T 7714
贾群. 面向服务机器人的抓取位置智能检测方法研究[D]. 中国科学院自动化研究所. 中国科学院大学,2019.
Files in This Item:
File Name/Size DocType Version Access License
面向服务机器人的抓取位置智能检测方法研究(7291KB)学位论文 开放获取CC BY-NC-SAApplication Full Text
Related Services
Recommend this item
Bookmark
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[贾群]'s Articles
Baidu academic
Similar articles in Baidu academic
[贾群]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[贾群]'s Articles
Terms of Use
No data!
Social Bookmark/Share
All comments (0)
No comment.
 

Items in the repository are protected by copyright, with all rights reserved, unless otherwise indicated.