融合显著特征和互信息熵的SLAM闭环检测算法  

SLAM Closed Loop Detection Algorithms Fusing Significant Features and Mutual Information Entropy

作　　者：石祥滨[1,2] 耿凯 刘翠微 SHI Xiang-bin;GENG Kai;LIU Cui-wei(School of Information,Liaoning University,Shenyang 110036,China;School of Computer Science and Technology,Shenyang Aerospace University,Shenyang 110136,China)

机构地区：[1]辽宁大学信息学院,沈阳110036 [2]沈阳航空航天大学计算机学院,沈阳110136

出　　处：《小型微型计算机系统》2020年第1期171-176,共6页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(61170185,61602320)资助;辽宁省博士启动基金项目(201601172)资助;辽宁省教育厅科学研究一般项目(L201607)资助

摘　　要：针对移动机器人SLAM闭环检测存在的闭环错误判定问题,提出了一种融合显著特征和互信息熵的闭环检测算法.首先,通过帧间位姿变化和信息熵差异得到关键帧,利用词袋模型逆向索引查询关键帧获取闭环候选帧.然后,引入图像颜色与深度信息计算当前关键帧与闭环候选帧的显著特征生成显著图.最后,将显著图均匀划分四块,计算两匹配关键帧显著图对应块的互信息熵,由每块对应互信息熵生成一个表征两幅图像相似度的四维向量,通过四维向量将满足互信息熵条件的闭环帧进行几何验证,判断是否产生闭环.实验结果表明,该算法能够降低机器人累积误差,生成全局一致的地图,同时满足实时性要求.Aiming at the problem of closed-loop error determination in SLAMClosed-loop Detection of mobile robots,a closed-loop detection algorithm combining salient features and mutual information entropy is proposed.Firstly,the key frames are obtained by changing the posture between frames and the difference of information entropy,and the closed-loop candidate frames are retrieved by using the word bag model to retrieve the key frames.Then,the salient features of current key frames and closed-loop candidate frames are calculated by introducing image color and depth information to generate salient maps.Finally,the saliency map is evenly divided into four blocks,and the mutual information entropy of the corresponding blocks of the saliency map of two matched key frames is calculated.A four-dimensional vector representing the similarity of the two images is generated from the corresponding mutual information entropy of each block.Through the four-dimensional vector,the closed-loop frames satisfying the mutual information entropy condition are geometrically verified to determine whether the closed-loop is generated.The experimental results showthat the algorithm can reduce the cumulative error of the robot,generate globally consistent maps,and meet the real-time requirements.

关 键 词：SLAM 关键帧 闭环检测 词袋模型 显著特征 互信息熵 

分 类 号：TP242[自动化与计算机技术—检测技术与自动化装置]