Research On 3D Edge Visual Semantic Map Construction Method Based On RGB-D Camera In Home Environment

Navigation and perception of unfamiliar environment is one of the key bases for home service mobile robots to realize service tasks.As the realization of navigation and environment perception,the visual semantic SLAM system has great application value.In this paper,a visual semantic SLAM map building system based on RGB-D camera is proposed from the application direction of family service robot navigation.Considering the limitations of the robot’s own resources,the edge visual semantic SLAM map building system is obtained by combining edge computing improvement.Firstly,a visual semantic SLAM map construction algorithm based on RGB-D image is proposed.ORB＿SLAM2 algorithm is used in visual SLAM,and the sparse map generated by ORB＿SLAM2 algorithm is transformed into 3D octree map to display robot position.Rep VGG model is used to extract features from images collected by RGB-D camera.PSPNet semantic segmentation model is used to perform semantic segmentation on the feature images so as to realize the semantic recognition of the objects in the environment by mobile robots.Based on the obtained point cloud information such as location and semantics,the improved Bayesian semantic fusion algorithm and 3D octree map construction algorithm are used to construct 3D map to display 3D map.Then,a map construction algorithm for edge visual semantic SLAM is proposed.The visual SLAM algorithm is decoupled and a local navigation map building module is added to the mobile robot,and the semantic segmentation algorithm and map building algorithm are unloaded to the edge device.According to the task unloading algorithm,the network is designed to increase the independent transmission channel and message transmission channel between the robot and the edge device,so as to realize the data transmission between the mobile robot and the edge device.In the mobile robot side,aiming at the map updating problem of the added local map module,a local map updating method is designed by periodically clearing the local map.On the edge device side,resourceconsuming tasks are implemented.Finally,experimental verification and evaluation are carried out.The SUNRGBD data set of indoor scene was used to train the semantic segmentation model and evaluate the performance of semantic classification.The resource consumption of visual semantic SLAM map construction algorithm is evaluated,and it is verified that the algorithm has high resource consumption for robot.Different versions of the proposed map building algorithms for edge visual semantic SLAM are evaluated experimentally,and the final design algorithm is determined by comparing the experimental results.Verify the running effect of the proposed algorithm and verify that the proposed visual semantic SLAM map construction algorithm meets the requirements of 3D semantic map construction.