基于卷积神经网络的视觉闭环检测研究  

Loop Closure Detection for Visual SLAM Using Convolutional Neural Networks

作　　者：杨孟军 苏成悦[1] 陈静[1] 张洁鑫 Yang Meng-jun;Su Cheng-yue;Chen Jing;Zhang Jie-xin(School of Physics and Optoeletronic Engineering,Guangdong University of Technology,Guangzhou 510006,China)

机构地区：[1]广东工业大学物理与光电工程学院,广东广州510006

出　　处：《广东工业大学学报》2018年第5期31-37,共7页Journal of Guangdong University of Technology

基　　金：国家自然科学基金青年科学基金资助项目(61305069);广东省信息产业发展专项现代信息服务业项目(2150510)

摘　　要：闭环检测是视觉SLAM中很重要的一部分,成功地检测出闭环能减小定位算法所产生的累积里程漂移.鉴于深度卷积神经网络在分类问题上的优越表现,本文首次将应用于图像分类的vgg16-places365卷积神经网络模型应用于视觉SLAM闭环检测中,将配准数据输入训练好的该卷积神经网络,其各个隐藏层的输出对应于图像特征表示.然后通过实验比较选用匹配精度较高的中间层完成场景特征提取,通过计算场景特征的相似性得到闭环区域.最后在闭环检测数据集上进行实验测试.测试结果表明,相比于传统的闭环检测方法,vgg16-places365卷积神经网络模型在相同召回率条件下准确率要高约3%;对于特征提取时间,在CPU上要快约5~10倍,而在GPU上更是比传统人工设计特征的闭环检测快近100倍.The detection of loop closure is a very important part of visual slam.Successful detection of loop closure can reduce the accumulated mileage drift generated by positioning algorithms.In view of the superior performance of deep convolutional neural networks in classification,the network of VGG16-Places 365 is used,which is widely used in image classification to the area of loop closure detection for the first time.The registration data are input into a trained convolutional neural network,and the output of each hidden layer corresponds to the image feature representation.Then,experiments are implemented to get an intermediate layer with higher matching accuracy,which is used to complete scene feature extraction,and then the loop closure region is obtained by calculating the similarity of the scene feature;finally,experimental tests are performed on loop closure detection dataset.Test results show that the accuracy rate of the VGG16-Places 365 convolutional neural network model is about 3%higher than the traditional ways under the same recall rate;and the the feature extraction time is about 5 to 10 times faster on the CPU and 100 times on the GPU.

关 键 词：视觉SLAM 闭环检测 卷积神经网络 特征提取 相似度 

分 类 号：TP242[自动化与计算机技术—检测技术与自动化装置]