摘要
机器人在大尺度场景下开展同时定位与建图(SLAM)任务时,其闭环检测环节会出现较严重的错匹配或者漏匹配问题,因此,采用残差网络(ResNet)对图像序列进行特征提取,并提出一种新的闭环检测算法。通过预训练的ResNet提取输入图像的全局特征,并对该帧图像及之前具有一定长度的图像序列的特征按照下采样的方式进行拼接,将结果作为当前帧图像的特征,保证图像特征的丰富性与准确性。同时,设计一种双层查询的方法以获得最相似的图像帧,并对最相似图像进行一致性检验,确保闭环的准确性。在闭环检测主流公开数据集New College和City Centre上,所提算法在100%准确率下的召回率为83%,在99%准确率下的召回率为85%。与传统的词袋方法和VGG16方法相比,所提算法具有显著的提升。
When robots conduct simultaneous localization and mapping(SLAM)tasks in large-scale scenes,there is serious mismatching or missed matching in loop-closure detection.Focused on this problem,this study proposes a new closed-loop detection algorithm based on a residual network(ResNet)to extract features of image sequences.The global features of an input image are extracted using a pretrained ResNet.The features of the frame image and previous image sequenced with a certain length are stitched by the down sampling method,and the results are taken as the features of the current frame image to ensure the richness and accuracy of the image features.Then,a double-layered query method is designed to obtain the most similar image frame,and the consistency of the most similar image is checked to ensure the accuracy of the loop-closure.The proposed algorithm can achieve an 83%recall rate under 100%accuracy and an 85%recall rate under 99%accuracy in the loop-closure detection mainstream public datasets of New College and City Centre,which is significantly improved compared with the traditional bag of words method and VGG16 method.
作者
占浩
朱振才
张永合
郭明
丁国鹏
Zhan Hao;Zhu Zhencai;Zhang Yonghe;Guo Ming;Ding Guopeng(Innovation Academy for Microsatellite,Chinese Academy of Sciences,Shanghai 201203,China;Key Laboratory of Microsatellites,Chinese Academy of Sciences,Shanghai 201203,China;University of Chinese Academy of Sciences,Beijing 100049,China)
出处
《激光与光电子学进展》
CSCD
北大核心
2021年第4期307-315,共9页
Laser & Optoelectronics Progress
基金
中国科学院战略性先导科技专项(XDA15020305)
中国科学院国防科技重点实验室基金(CXJJ-19S012)。
关键词
成像系统
闭环检测
残差网络
同时定位与建图
机器视觉
imaging systems
loop-closure detection
ResNet
simultaneous localization and mapping
machine vision
