Like-kinect Based3D Human Body Reconstruction And Application

3D human body reconstruction has always been an important research topic and hotspotin computer graphics and computer vision. The reconstructed3D human body model can beapplied to the garment industry, physical training, medical service and many other fields.Although the data is acquired by the traditional3D scanner is high precision, less noise,high cost and complicated operating. The like-Kinect, launched in recent years, like Kinect,PrimeSense, XtionPRO and so on, also has the function of3D scanning. Its advantage is lowprice and simple operation, but the disadvantage is low resolution and high noisy. Since theraw data is high noise and low resolution, it can’t directly reconstruct the3D human body. Soit is meaningful to study a new method of3D human body reconstruction based onlike-Kinect.Against to the feature of like-Kinect, the thesis improves the traditional algorithms andproposes new algorithms to increase the effect of the3D human body model reconstructionbased on like-Kinect. The main contents include the following:(1) Comprehensively analysis of recent relative research of3D human bodyreconstruction based on like-Kinect, respectively analysis of the algorithms applied in the3Dhuman body reconstruction, And illustrate its important research significance and broadapplication prospects.(2) Analyzing the cause of the noisy and holes based on like-Kinect depth imaging andcomputational theory, we format the energy equation and minimize the energy to reduce noiseand fix holes. After fixing holes and denoising, the depth map is converted to point cloud data.And we obtain point cloud match patch by combining with depth information and colorinformation. Then we build a Gaussian mixture model to register the match patch, which uses the maximum expected algorithm to iteratively solved transformation matrix. For solvingLoop Closure Problem[21], we resample the match patch and use the points from match patchto construct a spatial closure curve, then project the resampling points on the curve, re-calculate the global transformation matrix to complete the global registration. After that weuse the directional distance function to merge point clouds, and reconstruct the mesh byPoisson Surface Reconstruction algorithm. Finally we get a3D full human body model.(3) The reconstructed3D human body model is applied to3D virtual fitting system. Thelast part of thesis overall introduces the system function module and describe the key designand implementation method of individual modules in detail.At last, we make the summarizations of our work and give the further research directions.