Research On Feature Semantic Representation In Visual SLAM

Simultaneous localization and mapping(SLAM)is the key technology for autonomous mobile robots to perceive the environment and complete complex tasks.With the support of the development of computer vision technology,related algorithms that use semantic information on feature-based visual SLAM systems to solve the constraints brought by pure visual information and improve the system's ability to perceive the environment have attracted extensive attention from researchers.Aiming at the problem that the feature-based visual SLAM systems have insufficient cognition of category attributes in semantic information and ignore the influence of position information in semantic information,this paper studies feature semantic representation in visual SLAM.The main work and contributions are described as follows:1.Based on the effect of object velocity attributes on visual SLAM,a method to adaptively express the impact of objects on the visual SLAM system is proposed.This paper designs a measurement method for calculating the contribution of different category features to the SLAM system,and makes statistics on each category on the public data set.Experiments show that the adaptive representation of the impact of feature categories in this paper can effectively improve the accuracy of the positioning results of the classic ORB-SLAM2 system by about 4%on average,and reduce the standard deviation of the positioning results by about 27%on average.2.This paper points out that there is a problem of unstable spatial structure in visual SLAM,and analyzes the relationship between this problem and the edge feature.A method to reduce the influence of unstable spatial structures by expressing the edge attributes of features has been proposed.This paper summarizes the particularity of edge features.On this basis,the thesis expresses the edge attributes of the features through two steps of applying stricter constraints on the edge-features in the matching process and adjusting the information matrix for the edge-features in the optimization process.Experiments show that the representation of the edge attributes of the features in this paper can effectively improve the accuracy of the positioning results of the classic ORB-SLAM2 system by about 7%on average,and reduce the standard deviation of the positioning results by about 25%on average.3.This paper designs a visual SLAM system for unified expression of feature category attribute influence and edge attribute influence,and sets the semantic representation selection and dynamic interference strategy selection according to the scene applied by the visual SLAM system.The designed system was tested on the TUM RGB-D data set and KITTI data set.The experimental results show that the visual SLAM system designed in this paper has excellent positioning accuracy and the ability to resist dynamic target interference.