3D Face Reconstruction Based On Binocular Stereo Vision

Generating realistic 3D human face model is a challenging task in computer graphics and computer vision. In recent years with the extensive applicationof 3D face models in the fields of virtual reality, video surveillance, 3D animation and face recognition, 3D face reconstruction raises a great deal of interest and has been a research focus. In this paper, on the basis of a thorough research on camera calibration, stereo matching, three dimentional reconstruction, Delaunay triangulation and mesh subdivision algorithms, a 3D face reconstruction method based on binocular stereo vision theory is proposed. Experimental results show that the method can generate smooth and vivid 3D face model without expensive devices and generic face model. The main concerns are follows:1. After analyzing and investigating three traditional camera calibration methods including linear camera calibration, radial alignment constraint and Zhang's method, a stereo vision system is built. Considering experimental instruments and environment, Zhang's method is adopted to get the exterior and inner parameters of two cameras (JVC TK-C1481BEC). The reprojection error is within 0.4 pixels.2. Stereo matching algorithms based on feature and region are investigated and improved. Captured by calibrated cameras, a pair of stereo images is rectified to align the epipolar lines and compensated for the image distortions. To obtain accurate matching and dense disparity map, a stereo matching algorithm based on region growing combining with feature technique is presented. First edge feature point with reliable disparity is selected as seed point. Then region growing processes along horizontal scan-lines under multi-constraints.3. In order to build high quality 3D face model, three key techniques including three dimentional reconstruction of space point, Delaunay triangulation and triangular subdivision are researched and analyzed. With camera calibration and stereo matching, 3D coordinates of corresponding points are calculated. The scattered face data is divided by Delaunay triangulation. A new triangular subdivision schemes based on sqrt-3 method and triangular Bernstain-Bézier spline is proposed to generate smooth face surface.4. For the purpose of reconstructing 3D face models in real time and proving practicability of these algorithms proposed upon, a hardware and software platform including image capturing, camera calibration, stereo matching and 3D reconstruction are designed and put up. All work lays a good foundation of further research and development of 3D face reconstruction based on binocular stereo vision.