Three-Dimensional Reconstruction Technique Application Research Based On Kinect Depth Sensor

Three-dimensional reconstruction is a technology which builds a three-dimensional entityof the objective world in the computer by utilizing computer digitization approaches. As aninterdisciplinary and challenging technology, three-dimensional reconstruction has beenwidely used in the computer animation, augmented reality, human-computer interaction andso on. The core issue of three-dimensional reconstruction technology is to capture sceneinformation with different perspectives and to perform effective registration and integration.Actually, it’s to solve the space coordinate transformation relationship in multiple points ofview. When we use the traditional three-dimensional reconstruction technology to obtainimage information, the device is very complex, the cost is very high, and the operation is verycomplicated. Therefore, how to realize fast three-dimensional model reconstruction with lowcost and simple operation will have high research value and wide application prospects. Giventhe above problems, this paper used Kinect depth sensor as an input device forthree-dimensional reconstruction and analyzed the pipeline including data acquisition,pre-processing, multi-perspective scene alignment, point cloud integration, polygon mesh andtexture mapping. This paper proposed a solution of three-dimensional reconstruction based onKinect depth sensor. The main contents of this paper are as follows:(1) First, we captured depth and color data by Kinect and then converted the depthinformation into three-dimensional point cloud with actual model by usingcoordinate transformation. According to the noise characteristics of depth data, weanalyzed and compared several filtering algorithm. This paper used a fast bilateralfiltering to perform denoising operation from two-dimensional image extended tothree-dimensional point cloud data.(2) Secondly, we aligned the acquired depth data from multi-views. Through theoreticalanalysis on iterative closest point algorithm, we proposed a feature-based registrationalgorithm for point cloud model rigid body registration. The registration consists oftwo steps which are coarse registration and fine registration respectively. For theaccumulate error which generated from point cloud data registration, we researchedand figured out a global registration method to reduce the accumulate error bydistributing error to pair-pair registration.(3) Finally, the point cloud reconstruction and visualization were conducted. Afterfinishing the multi-views point cloud registration, we used integration method with surface element to integrate point cloud model and formed a completethree-dimensional point cloud model. In order to realize the visualization, we usedpolygon mesh representation and texture mapping technique to convert the pointcloud model into a full mesh model with real sense color.Experiment results show that the proposed reconstruction algorithm can make everydayobjects digitized by using an ordinary personal computer and Kinect. This system is easy tooperate during reconstruction process and low-cost without high-performance computerconfiguration. Since Kinect itself is unable to capture high-precision data, the reconstructionmodel is not precise enough, but can meet the requirements of inexact computingapplications.