Research On 3D Scene Perception Technology Of Mobile Robot Based On Convolutional Neural Networks

As technology advances,the demand for service robots is also increasing.Unlike industrial robots,service robots need not only to understand human instructions,but also to perceive the environment from the perspective of human beings in order to better provide services.The research on segmentation recognition(semantic segmentation)of objects still focuses on two-dimensional pictures,but it's not enough for robots in threedimensional environments.This article focuses on the three-dimensional scene perception technology,specifically expressed as the construction,optimization and description methods of indoor three-dimensional semantic scenes.A construction system of three-dimensional semantic scene is established in the project to construct a three-dimensional semantic scene.Firstly,based on convolutional neural networks,a model merging RGB information and depth information is trained on the data set,and the single frame obtained by Kinect is semantically segmented;the ORBSLAM system is used to extract the picture features and estimate the poses.A dense 3D point cloud carrying semantic information was constructeds.Methods of three-dimensional semantic segmentation and optimization are improved and proposed.Based on former work,combining voxel filtering and Bayesian estimation,this paper carries out voxel probability fusion of point clouds,and obtains a better result than just frames added purely.On the other hand,in the project,the threedimensional semantic segmentation method based on deep learning is combined with the LCCP algorithm based on concavity and convexity,so that the semantic segmentation result of the indoor scene is optimized as a whole.Some of the attributes of objects in the scene are described.On the basis of establishing a three-dimensional semantic scene,the project further describes the objects in the scene,such as the position,size,main body color,and the distance to the surrounding objects of the object,and can give the robot a variety of operations(such as navigation,grabbing).Etc.)Provide more information,.The project explores a path planning method based on natural language commands,which is an application of the construction and description methods of indoor threedimensional semantic scenes.Based on the point cloud,the project generates a grid map that can be used for global path planning;understands natural language and associates it with scene information;after identifying the starting point and target area,global path planning can be performed based on the Dijkstra or A* algorithm.