基于非监督深度学习的闭环检测方法  

Loop Closure Detection Method Based on Unsupervised Deep Learning

作　　者：汪丹 石朝侠[1] 王燕清 WANG Dan;SHI Chao-xia;WANG Yan-qing(School of Computer Science and Engineering,Nanjing University of Science and Technology,Nanjing 210094,China;School of Information Engineering,Nanjing Xiaozhuang University,Nanjing 210038,China)

机构地区：[1]南京理工大学计算机科学与工程学院,南京210094 [2]南京晓庄学院信息工程学院,南京210038

出　　处：《计算机科学》2020年第10期228-232,共5页Computer Science

基　　金：国家自然科学基金项目(61371040)。

摘　　要：闭环检测是同时定位与建图(Simultaneous localization and mapping,SLAM)的重要组成部分,能够有效减小SLAM系统中的累积误差,并且如果在定位与建图过程中跟踪丢失,还可以利用闭环检测进行重定位。与传统的手动设计的特征(hand-crafted feature)相比,从神经网络中学习到的图像特征具有更好的环境不变性和语义识别能力。考虑到基于陆标(landmark)的卷积特征能够克服整个图像特征对视点变化敏感的缺陷,文中提出了一种新的闭环检测算法。其首先通过卷积神经网络的卷积层直接识别出图像的感兴趣区域生成陆标,然后对图像中识别出的每个陆标提取卷积特征,生成图像的最终表示以检测闭环。为了验证算法的有效性,在典型的数据集上进行了对比实验,结果表明所提算法具有优异的性能,且即使是在极端的视点和外观变化的情况下仍然具有高鲁棒性。Loop closure detection is one of the most critical parts for simultaneous localization and mapping(SLAM)systems.It can reduce the accumulative error in SLAM system.If the tracking is lost during localization and mapping,it can also use the loop closure detection for relocation.Image features learned from neural networks have better environmental invariance and semantic recognition capabilities compared to traditional hand-crafted features.Considering that the landmark-based convolution features can overcome the defect that the whole image features are sensitive to viewpoint changes,this paper proposes a new loop closure detection algorithm.Firstly,it directly identifies the saliency region of the image through the convolutional layer of the convolutional neural network to generate a landmark.And then,it extracts the ConvNet features from the landmarks to generate the final image representations.In order to verify the effectiveness of the algorithm,some comparative experiments were performed on some typical datasets.The rusults show that the proposed algorithm has superior performance,and has highly robust even in drastic viewpoints and appearance changes.

关 键 词：同时定位与建图 闭环检测 人为设计特征 卷积特征 深度学习 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]