Research On 3D Semantic Map Building Of Mobile Robot Based On Deep Learning

As a further form of manufacturing after automated manufacturing,intelligent manufacturing has received much attention.Mobile robots are also receiving attention as one of the key factors to achieving intelligent manufacturing.The ability of mobile robot to sense their position in unfamiliar environments and to acquire semantic information in the environment is the frontier of research in this field.Based on the above reasons,this research focuses on the construction of 3D semantic maps for mobile robot based on deep learning,using a depth camera as a vision sensor to realize the perception of its position and the semantic information of its surroundings and to complete the construction of 3D semantic maps of its surroundings.This paper proposes an overall design of a 3D semantic map construction system,which can realize the semantic perception of objects in the environment and construct 3D semantic maps while estimating the position of the mobile robot itself.The main research contents of the 3D semantic map construction system in this paper are as follows.Firstly,the depth camera-based visual SLAM system is systematically described,including the working principle of RGB-D camera,distortion model,and the basic framework of Visual SLAM system.Subsequently,the overall design framework of the 3D semantic map construction system is proposed.Secondly,this paper elaborates on the basic principle of convolutional neural network and designs the target detection and recognition algorithm based on it to achieve the detection of object location and recognition of object types in the surrounding environment.The network model of the target detection and recognition algorithm is designed based on the YOLOv4 model under the framework of CSPDarknet53,and the model parameters are trained with the MS COCO dataset and tested for its detection and recognition effects.Thirdly,a 3D semantic map construction algorithm is designed.In this paper,the ORBSLAM2 system is used to reconstruct the 3D dense point cloud map of the environment in realtime,and then the point cloud segmentation is carried out for the dense point cloud map based on spatial,geometric,and texture features,and the point cloud map is reconstructed and stored in a K-dimensional tree structure.Finally,the experimental environment is built,and the modules are fused to form a complete3 D semantic map construction system.The system is tested in the experimental environment,and it can build a more accurate and understandable 3D semantic map while estimating its pose and sensing the semantic information of the environment,which verifies the feasibility of the 3D semantic map construction system of this topic.