Research On Semantic Visual SLAM Algorithm Based On Multiple Information Fusion

Semantic visual SLAM(Simultaneous Localization and Mapping)is a core technology for mobile robots to achieve autonomous mobility.At present,semantic visual SLAM algorithms mainly use single features such as points or lines for data association and pose estimation.However,problems such as object occlusions in real environments lead to ambiguous association matching of semantic objects,which in turn reduces the accuracy of localization.In addition,the constructed semantic maps have large errors in object scale,which reduces the quality of the semantic maps and is not conducive to the implementation of subsequent advanced tasks such as human-computer interaction and object-level navigation.Using the above problem as a starting point,a research on semantic visual SLAM algorithm based on the fusion of multiple information such as geometry and object semantics is carried out in this paper to obtain more accurate localization and to construct a semantic map of the environment with accurate object scale by data association through multiple information.The main work of this research paper is as follows:(1)To address the problem that a single geometric data association method cannot well eliminate the ambiguity of association,a data association method based on geometric features and semantic object appearance features is proposed,which effectively improves the accuracy of data association by combining geometric features and object appearance features and considering the influence of multiple information on data association results.The data association method is applied to a semantic visual SLAM system to jointly optimize object and camera poses in a tightly coupled manner,improving the localization accuracy of semantic visual SLAM.(2)To address the problems of large errors in semantic object scale and the non-uniform distribution of 3D point clouds in space,a semantic SLAM object parameter estimation method based on point cloud spatial projection is proposed,using point features,line features and object semantic features to estimate object scale,effectively improving the accuracy of object scale.The object scale method is applied to a semantic visual SLAM system to construct a 3D semantic semi-dense map with accurate object scale.This paper introduces multivariate information to improve the localization accuracy of semantic visual SLAM and constructs a 3D semantic semi-dense map with accurate object scale.At the same time,the semi-dense object-level semantic map constructed in this paper is beneficial for robots to perceive and understand the real environment,which is important for human-robot interaction and object-level navigation.