| With the research and cognition of the mental map of the blind,the complete cognition of the surrounding scenes is closely related to tasks such as the construction of the mental map of the blind and walking planning.Therefore,the use of convenient means to construct a complete three-dimensional scene semantic map to assist the blind to strengthen their cognition of the surrounding scenes is conducive to helping the blind construct a mental map and facilitate their travel.This paper uses visual SLAM technology to start with the key technology of scene semantic information extraction,and chooses the idea of SLAM to assist the construction of semantic maps,and designs the construction of semantic map system.This paper proposes the RGB-D double Fusion structure,which combines RGB information and depth information,suppresses potential noise information in RGB information and depth information,enhances the effective information features of semantic segmentation,and designs the ESA-double Net model to solve the problem of semantic map Semantic extraction problem.And the effectiveness of the structure has been proved through control experiments.Analyzing the accuracy of detailed category segmentation,it is found that the RGB semantic model is suitable for semantic segmentation tasks with complex object spatial structure but regular color information.The RGB-D semantic segmentation model performs well in semantic segmentation tasks with simple spatial structure and uncertain color information.It explains from the side that RGB information and depth information complement each other,but they also conflict with each other.The structure proposed in this article does not solve this conflict well.Experiments also show that depth holes have a great impact on the RGB-D semantic segmentation model.Although this article uses RGB-D double Fusion,RGB information and depth information are consciously complemented.However,the experimental results show that the structure proposed in this paper does not completely solve the problem of depth information noise.In this paper,ESA-double Net is used as the semantic segmentation model,the ORB-SLAM2 algorithm is used to calculate the position and posture of the camera,combined with the depth map information to solve the spatial coordinates of the target,the octree cube is used to express the simplified map,and the ROS framework is used as the system basis.Designed and developed a semantic map construction system.Through the comparative experiments between the RGB semantic segmentation model and the RGB-D semantic segmentation model,the effect of semantically constructing maps using two segmentation methods is demonstrated.By comparing the number of point clouds before and after using the octree,it shows that the octree has a powerful simplification ability,which can greatly simplify the scene point cloud and reduce the amount of calculation. |
