Research On 3D Reconstruction Of Real Scene Based On Depth Sensing

Three-dimensional reconstruction is a technology involving multiple disciplines such as computer vision and computer graphics.Its essence is to reproduce the real scenes of the real world through digital means.Today,this technology is widely used in intelligent driving,cultural relics restoration,and Audio-visual entertainment and other fields.The emergence of commercial depth cameras such as Kinect makes low-cost,cost-effective,and easy-to-operate 3D reconstruction possible.However,due to the limitations of the equipment,the depth images collected by Kinect will be affected by noise and have holes.In addition,the error in the registration after the depth image is converted into a point cloud will also affect the accuracy of the model.Therefore,this article will use the Kinect V2 depth camera as the information acquisition device to explore and improve the entire 3D reconstruction process,and propose a Kinect-based real-scene 3D reconstruction method.The main innovations of this article are as follows:Firstly,aiming at the problems of depth image holes and noise,a depth image restoration algorithm combining improved weighted mode filter and cellular automata is proposed.The algorithm uses edge detection operators to divide the depth image into edge regions and non-edge regions,and uses improved weighted mode filters and cellular automata to repair the two regions under the guidance of the color image.The result shows that it can effectively fill in the depth image holes and suppress noise problems,and at the same time the edge structure of the depth image is kept clear.Then,aiming at the problem of point cloud registration,the process of coarse registration and fine registration are improved respectively.In this paper,the curvature feature of the point cloud is used to establish a multi-scale feature descriptor,the point cloud is correspondingly clustered and divided according to the angle and distance difference of the feature descriptor,and RANSAC(Random Sample Consensus)is used for the obtained point cloud with the corresponding relationship.The algorithm obtains the best transformation of the point cloud block to complete the coarse registration work.Finally,the GMMTree(Gaussian Mixture Model Tree)algorithm is used to convert the point cloud into a probabilistic model to complete the point cloud fine registration work.The Gaussian mixture model tree divides the space hierarchically,and uses depth-first search and pruning to reduce the algorithm complexity to logarithmic time,and complete the fine registration efficiently and accurately.Finally,the model generated by the registration is subjected to texture mapping to generate a geometric model with three-dimensional texture.Experimental results show that the algorithm proposed in this paper improves the quality of depth image restoration and point cloud registration.Although the accuracy of Kinect V2 itself is not high,within the accuracy range of the sensor,the basic outline of the 3D reconstruction model is clear,and it can be applied to the 3D reconstruction work of actual scenes in daily life.