Research On Closed-loop Detection Method Of Mobile Robot Visual SLAM

The intelligentization of mobile robot is mainly manifested in the independent exploration and perception of unknown environment,self-stabilization and dynamic planning of unexpected situations,continuous self-adaptation and learning.Among them,it is difficult to explore and perceive the unknown environment independently,which is a hot topic in the research of intelligent mobile robots,and the core of Simultaneous Localization and Mapping(SLAM)is to solve this problem.Due to the inherent advantages of vision sensors,SLAM technology based on vision has attracted wide attention in recent years.Closed-loop detection is a key component of SLAM system,which can judge whether the robot has reached the previous position through data association,provide closed-loop constraints for back-end optimization,reduce the accumulated error of visual odometer,correct pose,ensure the global consistency of map construction,and ensure the robot to complete SLAM related tasks.In this thesis,the drawbacks of traditional closed-loop detection methods are analyzed,and semantic information is introduced.Three main problems of closed-loop detection,namely,keyframe selection of visual scene,scene modeling and closed-loop decision-making model,are studied in depth.Firstly,the key frame selection mechanism based on time domain,space domain and visual content is compared and analyzed.As the key frame selection method based on visual content is a real-time sequence calculation mode,which meets the image processing characteristics of visual SLAM,the Gist feature of the scene was extracted based on the visual content.The images collected by the mobile robot were classified according to the Gist feature of the scene by the local extreme value method,and the key frame was chosen.The scene image is composed of all the feature points and their introductions.In this thesis,the visual scene description methods based on local,global and spiritual features were compared and analyzed.Because semantic features have seasonal invariance,it is helpful to deal with visual angle changes caused by time,so this thesis chooses the method based on semantic features to model the scene.Semantic features can be acquired by target detection and semantics segmentation.Therefore,this thesis only focuses on indoor environment with many typical unstructured methods.Since semantics segmentation is suitable for indoor fine scene processing,it focuses on the analysis of semantics segmentation methods based on indoor RGB-D data in recent years.Finally,RDFNet network is selected for semantics segmentation of key frames.Obtain the semantics information and construct the semantics topology map using the object’s topology information in the scene to realize the visual scene modeling.Finally,the calculation framework of closed-loop decision model based on probability and image sequence is compared and analyzed.According to the constructed semantic topological graph,the scene similarity is calculated by using the relative spatial position of semantic vectors between key frames,and multiple thresholds are set to exclude the key frames with low similarity and relatively independent high similarity,thus realizing the design of closed-loop decision algorithm.Next,the closed-loop detection experiment is carried out in the indoor test site,and compared with the closed-loop detection method based on FAB-MAP and Bo W.The results show that the closed-loop detection method based on semantics simplifies the scene description,saves the storage space of the system,reduces the complexity and running time of the closed-loop,significantly improves the efficiency and accuracy of the closed-loop,and improves the robustness of the system.