Research On 3D Reconstruction Method Of Semantic SLAM In Dynamic Scene

With the intelligent development of computing devices,machine vision technology has been widely used.If you want to make a smart device perceive and interact with the real world,you must first restore its three-dimensional space scene.Among them,Simultaneous Localization and Mapping(SLAM),as a key algorithm for intelligent devices to achieve autonomous navigation and scene mapping in unknown environments,has attracted increasing attention.At the same time,the current SLAM 3D reconstruction work is mainly focused on the reconstruction of the static environment.If there are dynamic objects in the environment,a serious "ghosting" phenomenon will occur during the point cloud splicing stage,resulting in poor mapping results.This paper studies the RGB-D 3D reconstruction method based on SLAM framework in dynamic environment.In order to eliminate the impact of dynamic objects,we use deep learning algorithms to detect dynamic objects by introducing semantic information and geometric information between the current frame and key frames,so as to obtain more accurate data associations and transformation matrices and repair the static background blocked by dynamic objects.To achieve 3D reconstruction in a dynamic environment.The main work of this article is as follows:(1)The technical background and related theoretical knowledge of SLAM algorithm are introduced in detail,the depth map imaging principle and calculation method of pose transformation of RGB-D camera are analyzed,and then two kinds of BA and pose map optimization commonly used in back-end optimization are analyzed.The nonlinear optimization method is introduced in detail.(2)In-depth discussion of the SLAM 3D reconstruction algorithm ORB-SLAM2 in traditional static scenes,the key modules of visual odometer tracking,back-end optimization,loop detection in its basic framework are studied in detail.For the problem that a large number of feature points need to be extracted and matched during registration,the time complexity of this process is too high,and the feature extraction algorithm of ORB-SLAM2 is improved.GPU parallel computing is introduced to accelerate the extraction of feature points and improve the 3D reconstruction Overall speed.(3)A three-dimensional reconstruction method based on SLAM algorithm in dynamic scene is proposed,in view of the outstanding results of deep learning in the field of object detection and segmentation.Using Mask RCNN,a semantic segmentation algorithm based on convolutional neural networks,to achieve real-time object detection and semantic segmentation of the scene,and combine the pose information of the current frame and key frames to detect dynamic objects in the scene and mark the segmented area,and then repair the dynamic The static background occluded by the target further separates the static and dynamic two-state point clouds during reconstruction to achieve 3D reconstruction in a dynamic environment.(4)The algorithm proposed in this paper is experimentally verified.First,it is verified that the GPU-accelerated ORB-SLAM2 has higher feature extraction and mapping speed;then use the TUM data set to verify the three-dimensional mapping effect of the static SLAM and the dynamic SLAM proposed in this paper in the static and dynamic environments.The dynamic SLAM proposed in this paper can remove moving objects in dynamic scenes and repair occlusion backgrounds,which can effectively eliminate the miscellaneous generated by static SLAM in dynamic scenes.Point cloud,and has better mapping accuracy.