Research On 3D Reconstruction Algorithm Based On Kinect V2

Kinect v2 is a device designed by Microsoft to be used in somatosensory gaming devices.It can be used as a terminal for 3D reconstruction data due to its depth camera and high resolution.Microsoft Research's Newcombe R A uses Kinect v1 to propose a set of real-time 3D reconstruction techniques for indoor scenes that are not exposed to light,but due to global cube and memory limitations,large-scale reconstruction is not possible.The SFM algorithm that recovers the structure from the motion requires a large amount of image data.Otherwise,only a sparse point cloud reconstruction model can be obtained,and the details of the object cannot be accurately characterized.In order to reconstruct the unconstrained scene and the high-reduction three-dimensional object model,this paper proposes a 3D reconstruction system based on Kinect v2,which mainly does the following three aspects.Firstly,for the scene reconstruction,the principle of Kinect v2 to obtain the depth image is analyzed,and the source of the depth image noise is explained.Then an algorithm is designed to tailor the point cloud sampling range according to the distribution characteristics of the point cloud noise.The point cloud outliers are removed,and the point cloud holes are filled to improve the sampling quality of the point cloud.Most of the common 3D scene reconstructions use a global cube scheme of KinectFusion,but only a small range of scenes can be reconstructed.A multi-pair point cloud registration algorithm using queues to control the size of the scene is designed to break the limitation of the size of the reconstruction scene.Secondly,for the object reconstruction,we use Kinect v2 to get the point cloud containing the scene where the object is located and remove the outliers.We use the 3D bounding box to separate the specific object point cloud from the scene.The SAC-IA algorithm is used to coarsely register two adjacent point clouds.The two-two registration ICP algorithm is extended to multiple point clouds.A strategy of approaching from the two sides to the middle is proposed to reduce the accumulation.The error increases the degree of cloud reduction of the object.In this way,we could implement a low-cost,accurate 3D reconstruction system for a single object.Thirdly,we propose an algorithm for registering point clouds using two-dimensional image information under trajectory constraints.The color image of the scene and the corresponding depth image are acquired using Kinect v2.The color image and the depth image are calibrate.We take two common poses(translation or rotation)scenes with a common part of the calibrated color image,and use SURF for feature extraction to find the corresponding points of the two maps.We calculate the projection matrix of the two images and transform the corresponding depth image into a three-dimensional point cloud through the camera internal reference matrix.We use the correspondence between the transformation matrix of the point cloud and the projection matrix to obtain the rotation matrix R or the translation vector t on the two-dimensional plane.In this way,we propose a three-dimensional reconstruction algorithm that can be used for complete indoor scenes.The paper studied in the first two parts are point cloud data,and the third part is a two-dimensional image.The design experiments show that the system designed in this paper can reconstruct the scene and object separately.The experimental results show that the scene or object reconstructed by the three-dimensional reconstruction algorithm proposed in this paper has a high degree of reduction,and the equipment cost is low,the specific steps are simple and easy,and have certain practical value.