Research On Object Model Reconstruction Based On The Depth Camera

In recent years,as the performance of consumer-grade hardware improves,highquality depth sensors are implanted in the mobile devices and the rapid development of artificial intelligence technology,artificial intelligence's products such as the 3D face recognition,face model reconstruction,and AR / VR are gradually being applied to the mobile devices,promoting the popularization of 3D reconstruction technology.The research on the lightweight specific scene model(such as the objects,human bodies,faces)reconstruction algorithms is the current research hotspot in the field of 3D vision.Modeling objects without relying on the high-configuration hardware has great commercial value and can promote the development of the logistics?robotic grasping and mutual entertainment.Therefore,the research of object model reconstruction method based on the depth camera in this paper is of great significance.In this paper,the depth camera Real Sense D435 and Raspberry Pi 4B are used as hardware platforms.Based on the Linux system,a reconstruction algorithm framework for object modeling is designed and built.Taking the static object as the reconstruction object and the realization of fast and efficient model reconstruction as the research goal,a dense object reconstruction model reconstruction scheme based on the sparse frame fusion is proposed.Ensuring the reconstruction efficiency without relying on a highconfiguration graphics card to accelerate,it quickly build the dense models of objects.In this paper,a method based on the combination of depth camera and artificial checkerboard features is used to obtain the 6D motion state information of the camera;the reference coordinate system and reference point are determined by analyzing the camera's motion state;the selecting frame is introduced to eliminate incorrect positions datas to filter out valid frames to be fused.With the goal of improving reconstruction efficiency,before performing fusion and reconstruction,using the Yolo V2 algorithm in the deep learning to perform target detection on the frames to be fused to obtain the position information of the object in the images and combining the depth image to extract effective point cloud information to be fused.Aiming at improving the utilization of video memory,a point cloud fusion and reconstruction module was designed.The point cloud fusion based on the Voxel Hash structure is adopted and the truncation distance equation(TSDF)is used for the fusion method.The DDA algorithm is used to create and update the hash block,and the volume size is restricted by the boundingbox information to reduce the waste of memory.The marching cube method interpolates the meshed model after fusion to generate a triangular patch model with topology.Aiming at improving the accuracy and quality of the generated model,the mesh model refinement module is designed,which mainly includes the background separation,smoothing,poisson reconstruction and etc,further refining the generated mesh model.At the end of this paper,the designed algorithm module is integrated into an objectoriented model reconstruction system.The feasibility and accuracy of the model reconstruction are verified through reconstructing multiple sets of data tests.