基于多域卷积特征决策融合的SLAM回环检测算法

打开文本图片集

关键词：回环检测；视觉SLAM；卷积神经网络；多任务域；决策融合；特征提取；D-S理论图分类号：TN911.6-34；TP391.4 文献标识码：A 文章编号：1004-373X（2025）20-0181-06

SLAM loop closure detection algorithm based on multi-domain convolutional featuredecisionfusion

YUZhipeng'，WANGChengjun1，ZHAO Yanqiu'，REN Donghai² （1.SchoolofArtificial Inteligence，AnhuiUniversityofScienceandTechnology，Huainan232oo1，China; 2.Huaibei Zhongtai Electromechanical EngineeringCo.，Ltd.，Huaibei 235ooo，China）

Abstract：Loopclosure detection，asacrucial componentin simultaneous localizationand mapping （SLAM），caneffectively reducetheaccumulationerorsofposeestimationandyieldaconsistentglobalmap.Withthedevelopmentofdeeplearning，the convolutinalneuralnetworks（CNNs）begintobeappliedfortheloopclosuredetection.However，thefeaturerepresentationof CNNsisnot widelyappliedduetothedomainspecifity.Inorder toenhancetheaplicabilityofconvolutionalfeaturesinloop closuredetection，aSLAMloopclosuredetectionalgorithmbasedonmulti-domainsconvolutionalfeaturedecisionfusionis proposed.The algorithmcanlearnfeaturerepresentations from multi-task domains andcanbeused forthe featureextraction and decision-makingof theloopclosure detection.Thedecision-makingfrommulti-taskdomain featuresarefused basedonD-S theorytoobtaintheresultsofloopclosuredetection.TheexperimentswereconductedontheNew CollgeandCityCentre datasets.Itdemonstrates thattheloopdetectionalgorithmbasedonmulti-domainconvolutional featuredecisionfusioncan improve the accuracy of 2.292% ， 6.88% ，and 16.98% on the New College dataset， respectively，compared with the scene feature decision，thebjectfeaturedecision，andthesemanticfeaturedecision.OntheCityCentredataset，theaccuracyisincreasedby 6.23% ， 10.33% ，and 13.97% ，respectively.The experimental results indicate that the proposed method can enhance the accuracy ofloopclosuredetectionbasedonconvolutionalfeatures，therebyproviding guarantee forthepreciselocalizationand mapping.

Keywords：loopclosure detection；visual SLAM;convolution neural network;multi-task domain;decision-making fusion; feature extraction; D-S theory

0 引言

LocalizationandMapping，SLAM）中最重要的问题之一。（剩余8149字）