计算机工程与应用 ›› 2020, Vol. 56 ›› Issue (9): 175-182.DOI: 10.3778/j.issn.1002-8331.1910-0399

• 模式识别与人工智能 • 上一篇    下一篇

基于Netvlad神经网络的室内机器人全局重定位方法

陈承隆,邱志成,杜启亮,田联房,林斌,李淼   

  1. 1.华南理工大学 机械与汽车工程学院,广州 510640
    2.华南理工大学 自主系统与网络控制教育部重点实验室,广州 510640
    3.华南理工大学 珠海现代产业创新研究院,广东 珠海 519170
    4.广州地铁设计研究院股份有限公司,广州 510010
    5.日立电梯(广州)自动扶梯有限公司,广州 510660
  • 出版日期:2020-05-01 发布日期:2020-04-29

Indoor Robot Global Relocation Method Based on Netvlad Neural Network

CHEN Chenglong, QIU Zhicheng, DU Qiliang, TIAN Lianfang, LIN Bin, LI Miao   

  1. 1.School of Mechanical and Automotive Engineering, South China University of Technology, Guangzhou 510640, China
    2.Key Laboratory of Autonomous Systems and Network Control, South China University of Technology, Guangzhou 510640, China
    3.Zhuhai Institute of Modern Industry Innovation, South China University of Technology, Zhuhai, Guangdong 519170, China
    4.Guangzhou Metro Design & Research Institute Co., Ltd., Guangzhou 510010, China
    5.Hitachi Elevator(Guangzhou) Escalator Co., Ltd., Guangzhou 510660, China
  • Online:2020-05-01 Published:2020-04-29

摘要:

针对已知地图的室内机器人全局重定位、绑架恢复问题,提出一种基于改进的Netvlad卷积神经网络的室内机器人全局重定位方法,通过激光雷达获取的障碍物信息引导机器人到达空旷区域,粗定位阶段,使用栅格地图最短连通域距离作为正样本判据,并对Netvlad引入残差网络,通过图像检索得到机器人的粗略位置及角度信息。使用粗定位阶段得到的位置和角度信息作为自适应蒙特卡罗定位的初始值来估计机器人的精确位姿。实验结果表明,与传统定位方法相比,该方法可以使机器人从绑架问题中快速恢复准确位姿。

关键词: 激光雷达, 自适应蒙特卡洛, 卷积特征, 重定位, 绑架恢复

Abstract:

Aiming at the problem of global relocation and kidnapping recovery of indoor robots with known maps, a global relocation method of indoor robot based on improved Netvlad convolutional neural network is proposed. The obstacle information obtained by laser radar is used to guide the robot to reach the open area. In the positioning phase, the shortest connected domain distance of the raster map is used as the positive sample criterion, and the residual network is introduced into Netvlad, and the rough position and angle information of the robot is obtained through image retrieval. The position and angle information obtained from the coarse positioning phase is used as the initial value of the adaptive Monte Carlo positioning to estimate the precise pose of the robot. Experimental results show that compared with the traditional positioning method, this method can make the robot quickly recover the accurate pose from the abduction problem.

Key words: laser radar, adaptive Monte Carlo, convolution feature, reset, kidnapping recovery