基于语义位置验证的鲁棒闭环检测算法  

Robust loop closure detection algorithm based on semantic position verification

作　　者：赵一夫 苗津毓 郭方洪 吴祥[1] 董辉[1] 俞立[1] ZHAO Yi-fu;MIAO Jin-yu;GUO Fang-hong;WU Xiang;DONG Hui;YU Li(College of Information Engineering,Zhejiang University of Technology,Hangzhou 310014,China;School of Automation Science and Electrical Engineering,Beihang University,Beijing 100191,China)

机构地区：[1]浙江工业大学信息工程学院,杭州310014 [2]北京航空航天大学自动化科学与电气工程学院,北京100191

出　　处：《控制与决策》2023年第9期2511-2519,共9页Control and Decision

基　　金：浙江省重点研发计划项目(2020C01109);金华市重大(重点)科学技术研究计划项目(2021-1-012)。

摘　　要：针对现有闭环检测算法因视觉干扰而产生的闭环误判问题,提出一种利用场景语义信息进行验证的闭环检测算法.该算法通过视觉词袋库模型检索闭环候选帧,用后验方法剔除算法可能产生的错误闭环.后验方法首先将场景中的语义信息抽象为语义节点;然后提取包含邻域信息的节点特征,并以此匹配图像间的语义节点,降低算法对于动态物体干扰的敏感性.随后,算法基于图像间的匹配语义节点构建相对位置网络,根据网络相似度验证闭环帧,提高闭环检测算法在具有相似纹理的场景中应对感知混淆的鲁棒性.实验结果表明,语义位置验证算法能够显著提升视觉词袋模型的检测准确率.与其他经典算法相比,所提出的闭环检测算法具备更快速、更准确的检测性能.To address the loop closure misjudgment problem due to visual interference,this paper proposes a loop closure detection algorithm that uses semantic information to verify candidates.The algorithm retrieves loop closure candidates using a visual bag-of-words model,and applies a verification method to eliminate the mismatches.The proposed verification method first extracts semantic nodes on the basis of semantic information in the scene,and then calculates the node descriptors containing neighboring information,so that we can accurately match the semantic nodes across images,reducing the sensitivity of the algorithm to dynamic objects.Subsequently,the algorithm constructs a relative position network based on matched semantic nodes from two views,and verifies loop closure candidates based on the similarity of network,which improves the robustness of the algorithm against perceptual aliasing.Experimental results show that the semantic position verification method significantly improves the detection accuracy of the visual bag-of-words model.Compared with other classic algorithms,the overall loop closure detection algorithm has achieved leading performances in both detection accuracy and computational efficiency.

关 键 词：闭环检测 语义信息 特征描述 节点匹配 相对位置网络 

分 类 号：TP242[自动化与计算机技术—检测技术与自动化装置]