3D face reconstruction with RGB-D camera: A practical approach

Abstract

The continuous development of the RGB-D camera technology and it reduce price, have helped them to be used in different fields of applications, including research tasks. In this context, this paper presents a 3D face reconstruction methodology based on a RGB image and its associated depth map, acquired by a RGB-D camera Orbec Astra Pro. A preliminary stage based on face detection with the Haar Cascade classifier is performed in order to find the face area (2D) to be analyzed. These 2D pixel points found are then matched with its corresponding depth values in order to generate a 3D RAW data represented as a point cloud with values in the X,Y,Z coordinates. Several filters are then applied to the point cloud in order to eliminate the noise on the 3D image and smoothing the final 3D face representation. Finally, several tests are performed in order to evaluate the performance of the system, related with parameters as: face distance to the camera, filter tuning, point cloud resolution and the effect of the background scene. The results shown that the appropriate calibration of the filters and the face’s distance to the camera affects directly to the final 3D representation.