作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程 ›› 2018, Vol. 44 ›› Issue (6): 182-187. doi: 10.19678/j.issn.1000-3428.0047182

• 图形图像处理 • 上一篇    下一篇

基于精简卷积神经网络的快速闭环检测方法

何元烈,陈佳腾,曾碧   

  1. 广东工业大学 计算机学院,广州 510006
  • 收稿日期:2017-05-12 出版日期:2018-06-15 发布日期:2018-06-15
  • 作者简介:何元烈(1976—),男,副教授、博士,主研方向为图形图像处理、计算机视觉;陈佳腾,硕士研究生;曾碧,教授、博士。
  • 基金资助:
    广东省科技计划重大专项项目(2016B010108004);东莞市产学研合作项目(2015509109107)。

Fast Closed Loop Detection Method Based on Simplification Convolutional Neural Network

HE Yuanlie,CHEN Jiateng,ZENG Bi   

  1. School of Computers,Guangdong University of Technology,Guangzhou 510006,China
  • Received:2017-05-12 Online:2018-06-15 Published:2018-06-15

摘要: 基于深度学习的闭环检测方法在复杂光照下能取得较好的检测效果,但存在提取场景特征维度高、难以满足闭环检测实时性的问题。为此,基于精简深度卷积神经网络,提出一种闭环检测方法。结合级联修正线性单元、批规范化和深度残差模块完成网络模型的设计,并利用大型场景识别的数据集(Places365-Standard)完成网络模型的训练,用训练好的网络模型提取场景特征,通过计算场景特征的相似性得到闭环区域。测试结果表明,与基于位置卷积神经网络与自编码的闭环检测方法相比,该方法在保证较高准确率的同时提高了检测速度。

关键词: 闭环检测, 视觉同步定位与地图构建, 卷积神经网络, 深度学习, 场景识别, 特征提取

Abstract: Loop closure detection methods based on deep learning can achieve good detection performance in complicate illumination environment.But the dimension of extracted scene feature is too high to achieve the real-time detection requirement for closed loop.For that,propose a loop closure detection method based on fast and lightweight convolutional neural network.Concatenated rectified linear unit,batch normalization and deep residual module are combined to design the fast and lightweight network model which is trained with the Places365-Standard data set.The scene feature is extracted by trained neural network model and the regions of closed loop are obtained by measuring the similarities of them.Results from the test show that compared with loop closure detection algorithms based on PlaceCNN and Autoencoder,the proposed method not only achieves the high correct rate but also increases the implementation speed for detection.

Key words: closed loop detection, visual Simultaneous Localization and Mapping Construction(vSLAM), convolutional neural network, deep learning, scene recognition, feature extraction

中图分类号: