Research And Implementation On Video-based Fast Hair Capture Method

In recent years, digital virtual characters have become an integral part in3D animationsvideo games, virtual reality, augmented reality and some other areas. Hair is an importantfeature of digital characters. Therefore, hair modeling has been a hot research area in relativefields. Hair modeling first started in computer graphics. So far, some sophisticated modelingmethods have done well in hair simulation, with different hairstyles, physical characteristicsand light conditions token into account. However, these methods need to manually set anumber of parameters, and thus requires more interaction. Users also need to be experiencedin modeling. To overcome these disadvantages, researchers began to consider modeling in thefield of computer vision. For pursuing the approximation of real hair, most of these methodsuse complex capture devices, as well as controlled lighting conditions and specialized stereomatching algorithm, which makes these methods not suitable for applications that use simplecapture devices and have demand on computing speed, such as augmented reality.This study aims to design a quick hair modeling method in everyday light conditions,using only two common cameras. The essence of this method is a combination of both stereomatching and orientation field, based on the traditional3D reconstruction method. We use astandard calibration approach to get camera parameters and a stereo matching algorithm toobtain basic information of depth. After some pre-processings such as image calibration, pixelsampling and contour extraction, hair region has been separated from original image. We usesteerable filters to extract the2D orientation field in hair region and make it combined withthe basic depth information to get refined depth information, which will be used to computethe point clouds and3D orientation field. At last, a hair model is reconstructed, amended anddisplayed, considering the time information.Based on the method mentioned above, we basically completed a hair capture system.The system mainly utilizes OpenCV to implement the capture algorithm, from original imagesto the reconstruction of the hair model, and then the model is imported to OpenGL fordisplaying. Experimental results have shown that in current mid-end personal computer withdual cameras, our method can complete hair capture task in0.4s, which is faster than manyother methods.