Real-time 3D Reconstruction Based On Monocular Camera

3D reconstruction is an important research direction of computer vision,which can be widely used in 3D navigation,online shopping,3D printing,computing pho-tography,video games and cultural heritage protection.With the continuous en-hancement of computing ability and the pursuit of real-time interactive experience,real-time three-dimensional reconstruction is becoming a new research hotspot.Real-time 3D reconstruction based on monocular camera not only requires real-time estimation of the camera’s pose,but also real-time estimation and fusion of the depth map of the image.To this end,we must first,accurately estimate the camera’s pose,and then according to local information to quickly generate accu-rate depth map,and finally the use the method which can real-time interaction for fusion processing.In order to achieve the above objectives,this paper has de-veloped a real-time three-dimensional reconstruction system based on monocular camera.The system can generate real-time 3D images of the scene in real time from the real-time video data collected from the monocular camera.The system flow mainly consists of three parts:Firstly,the camera pose is estimated by using the simultaneous positioning and map construction(SLAM)system,extracted the key frame to optimizate the pose,and depth recovery which uses the key frame as the reference frame.The system of depth recovery adopts the plane sweeping algorithm and the depth-based inference method,and use the local optimization to check the result of the depth recovery.Finally,the depth information of the key frame is merged and the scene is reconstructed.In order t.o meet the real-time application requirements,we have the GPU-parallel acceleration of the depth recovery,and use the methods of time-space consistency and connectivity compo-nent analysis to optimize the depth of recovery,remove the area with poor quality of depth recovery,which ensures the quality of the generated 3D model.The system provides a good user interface and interactive operation,real-time visualization of the current reconstruction of the three-dimensional point cloud or grid surface to guide the user to complete the target object or scene of the full three-dimensional scan,and the reconstruction of the three-dimensional model can be saved to the file.The 3D reconstuction results of online videos demonstrate the effectiveness of the proposed system,which can meet the practical application requirements.