3D Reconstruction System In Dynamic Scene Based On Depth Camera

3D reconstruction technology is a research hotspot in computer graphics,computer vision,virtual reality,artificial intelligence and other fields.The technology of 3D reconstruction based on visual image is to obtain the data image of the scene object through the sensor,analyze the image with computer vision knowledge,and deduce the 3D information of the object in the real environment.Among them,the VSLAM(visual simple localization and mapping)3D reconstruction method based on depth camera is an online and comprehensive reconstruction method.The traditional VSLAM 3D reconstruction system cannot simply complete the reconstruction of the dynamic environment including moving objects,because the existence of moving objects in the field of vision will affect the effective matching between multi frame images,thus affecting the accuracy of reconstruction results.Based on the classic VSLAM framework ORB-SLAM2,combining the optical flow method and the dynamic point detection algorithm of 3D object detection,this paper proposes a 3D reconstruction system of VSLAM for dynamic environment based on depth camera.In order to complete this research,this paper mainly carried out the following work:(1)Starting from the ORB-SLAM2 framework,this paper summarizes the specific steps of the VSLAM 3D reconstruction system based on depth camera,including camera model,vision odometer,tracking and mapping,and loop detection.Design experiments demonstrate the important role of depth information in 3D reconstruction system,and calibrate the internal parameters of Kinect v2.(2)Based on the application of deep learning target detection and semantic segmentation network in SLAM system,the positive effect of Semantic Module on system reconstruction accuracy is demonstrated.This paper focuses on the 3D semantic information acquisition method combining 2D target detection and depth direction information filtering,and realizes the establishment of the information base of the system at the 3D level.(3)Based on the traditional optical flow method,multi view geometry method and deep learning target detection or semantic segmentation algorithm,three methods are proposed to analyze the dynamic objects in the environment.Based on the analysis of the advantages and disadvantages of the traditional algorithm and the deep learning algorithm,this paper puts forward the logic thinking of the comprehensive dynamic point detection algorithm combining the traditional and the deep learning algorithm.(4)The above-mentioned integrated dynamic point detection algorithm is integrated into the 3D reconstruction system,and a dynamic environment oriented VSLAM 3D reconstruction system based on depth camera is proposed.In order to verify the real-time performance and reconstruction effect of the proposed system,a number of experiments are carried out on the open data set,and the system proposed in this paper is analyzed.The existing shortcomings of the system and the direction of further improvement in the future are summarized.