Three-Dimensions Semantic Map Construction Based On Stereo Vision

In recent years,with the improvement of computer hardware level and the development of artificial intelligence technology,intelligent robots have been applied in various scenarios,such as industrial robots and sweeping robots.The perception and understanding of the environment is the basis for intelligent robots to perform tasks such as autonomous exploration,behavioral decision-making,and human-computer interaction.Robots generally capture peripheral environmental information through sensors such as lasers ridar and cameras,analyze them using algorithms,and finally understand the environment.How to make use of the visual data in the image to make the robot anthropomorphically perceive the environment and make decision is a hot research problem in the field of computer vision.To solve this problem,this thesis proposes a three-dimensional semantic map construction method that combines binocular vision SLAM technology and semantic segmentation technology,so that intelligent robots can build three-dimensional semantic maps based on camera pictures,helping them better understand environmental information and perform autonomous work more intelligently.In this thesis,the method of constructing three-dimensional semantic map based on binocular visual information is the research goal.Firstly,the semantic segmentation technology is studied.The traditional semantic segmentation neural network model has some disadvantages in segmentation efficiency and precision.The design method of high-efficiency neural network is analyzed.The network convolution unit is designed by decomposing convolution,residual connection,convolution channel reordering and multi-scale expansion convolution.The multi-scale pyramid cascade structure is used to construct the attention mechanism based decoder to form real-time semantic segmentation neural network model.The experimental analysis proves that the semantic segmentation network has a good performance in parameter quantity,segmentation precision and running speed,and achieves a good balance between the accuracy requirements and speed requirements of the semantic segmentation task.In the research of 3D reconstruction technology,this thesis analyzes the application range of different 3D reconstruction methods in different scenarios,and selects the SLAM algorithm which is suitable for positioning and mapping in the outdoor environment for real-time 3D reconstruction.Using real-time semantic segmentation neural network and depth estimation neural network to provide semantic information and depth information,combined with the pose informationprovided by SLAM to extract superpixels,surface modeling with facet model,and completed the real-time 3D semantic map construction method design.In the experimental part,the 3D semantic map construction method proposed in this thesis is applied to the KITTI dataset and actual campus data,and a 3D semantic map corresponding to the scene is constructed,which proves the efficiency and feasibility of the method.