Research And Implementation Of Real-time 3D Reconstruction System Based On Monocular

3D reconstruction reflects the structure and texture information of objects in real world obtained by various measurement methods through establishing digital models,involves many fields such as augmented reality,robot navigation,and medical CT diagnosis.In the early days,3D reconstruction technology used contact measurement to obtain object information.With the development of sensor technology and the research of related theories,the information acquisition method of non-contact measurement with camera and lidar has become the mainstream of 3D reconstruction technology.Since the introduction of depth cameras,especially the mature application of simultaneous localization and mapping algorithm,real-time 3D reconstruction based on depth cameras has met people's needs for real-time interaction.However,due to factors such as power consumption,size,and cost,depth cameras are not suitable for use in devices with limited payloads and power.Therefore,this paper has carried out research on the real-time 3D system based on monocular camera.The main contents of the research are as follows:(1)Aiming at the time-consuming problem of depth map generation,a depth estimation algorithm based on local stereo matching is proposed.Firstly,perform highly parallel down-sampling random initialization and chessboard iterative propagation on the GPU to obtain a coarse low-resolution depth map.After removing the noise depth by matching loss filter and minimum connected region,the depth map is up-sampled and smoothed to obtain the original resolution depth map.The RGBD data set provided by the Technical University of Munich was used to experiment with the proposed algorithm.The algorithm can generate a depth map with a resolution of 640x480 at a speed of 25 frames per second.The generated depth map has high integrity and low relative error.Experimental results show that the proposed depth estimation algorithm can meet the requirements of real-time 3D reconstruction system in terms of speed and quality.(2)Aiming at the point cloud redundancy problem caused by continuous depth map,a TSDF fusion algorithm based on the incremental update of chunks is proposed.The algorithm divides the world coordinate system into two-level tree structure of chunks and voxels to update voxels quickly.It uses hash table access to chunks to expand the range of reconstruction,and uses frustum to filter reduce the update number of chunks and the incremental update algorithm of chunks to improve the update speed of chunks,and remove invalid chunks through spatial carving to reduce memory usage.Experimental results show that compared with Open Chisel library,the proposed tsdf fusion algorithm is faster and has less noise.In the reconstruction system,we can reduce the updated field of view or enlarge the voxels to achieve real-time fusion depth map.(3)Designed and implemented a real-time 3D reconstruction system based on a monocular camera.The system consists of four modules: image publishing,pose calculation,depth estimation and surface reconstruction.The image publishing module corrects the image acquired by the camera and publishes it.The pose calculation module uses the modified ORB?SLAM system calculates the pose of the camera.The depth estimation module uses the depth estimation algorithm proposed in this paper to generate the depth map,and optimizes the depth map by temporal propagation and temporal filtering.The surface reconstruction module uses TSDF fusion algorithm based on the incremental update of chunks and Marching Cubes algorithm to generate the mesh model.The reconstruction system designed in this paper is built on a computer.Aiming at the problem that the moving range of the monocular camera is limited by the data line,a movable laptop is introduced as a slave computer.The slave is only responsible for the image publishing module,and the computer as the host is responsible for the pose calculation,depth estimation and surface reconstruction modules with a large amount of calculation.The built system has been tested in a real environment.The experimental results show that the system can better restore the texture and structure information in the scene and realize real-time interaction.