Research On3D Human Model Registration Technology Based On Kinect

Accurate evaluation of burn degree has important influence on the treatment of burn patients.The main basis of evaluation is the burned area. So, it is necessary to complete the3Dreconstruction of human body to accurately calculate the burned area. Based on this, the goal ofthis research is to build an automatic evaluation system of burn degree based on3D reconstructionof human body. The system has four modules: data acquisition,3D registration, burned regionextraction, and burned area calculation.This article mainly studies the registration of3D point cloud of human body. Because thetraditional3D body scanning method and registration techniques cannot be conveniently appliedto this system, we present a registration algorithm of body point cloud based on RGB images andRange images. At the same time we use the low-cost Kinect as the scanning equipment, whichreduces the requirement for laboratory equipment.At first, we made a thorough research on3D registration algorithm and the key technologiesand processing steps of3D registration are introduced in this paper. According to the presentedwork flow of the registration algorithm, we describe the contents in this article as the order of dataacquisition, coarse registration, fine registration and model optimization.In the stage of data acquisition and pre-processing, we use Kinect to obtain the RGB imagesand Range images from different perspectives and generate point clouds. Then, we introduce twodownsampling methods including the shuffling algorithm, VoxelGrid filter and two denoisingalgorithms including statistical analysis and bilateral filtering.In the stage of coarse registration, we introduce three coarse registration methods based onthe four-point algorithm, point feature histogram descriptor and the RGB images and range images,the paper uses the coarse registration method based on RGB and range images. The algorithmintroduces the SIFT(scale-invariant feature transform) feature descriptor to achieve the dimensionality reduction of coarse registration and improve the efficiency of registration. Then,we use the RANSAC(random sample consensus) algorithm to match pairs of feature points toenhance the accuracy of the coarse registration to provide a good initial position for fineregistration.In the stage of fine registration, we improve the traditional ICP(iterative closest point)algorithm on the basis of the existing registration algorithm. The strategies of improvementinclude the delineation of the search space based on bounding box, downsampling methods,weighted matching, elimination of error matching points, search acceleration based on Kd-treealgorithm.In the stage of model optimization, we adopt the method based on distance threshold toeliminate background information, the method based on template matching to eliminate whiteedges of objects, and optimize the cumulative error of the model.Finally, this article describes the technical route, designs the experimental program of human3D point cloud data registration, and demonstrates the effectiveness and feasibility of thisalgorithm. Experimental results show that the algorithm can accurately complete the registrationof3D point cloud of human body in case that the overlapping area between the front and side ofhuman body is small.