3D FACIAL POSE AND EXPRESSION ESTIMATING METHOD USING AAM AND ESTIMATED DEPTH INFORMATION

摘要

PURPOSE: A three-dimensional facial pose and expression estimating method using an active appearance model (AAM) and estimated depth information is provided to learn an input image of a front face by using the AAM and combine the AAM with the depth information about a face estimated from input face images, thereby performing face fitting for various facial poses without depending on prior learning. CONSTITUTION: By marking a landmark on an input image, a triangular mesh is formed (S100). A two-dimensional active appearance model (AAM) face model is generated from an AAM parameter defined in a learning process by an AAM in respect to a front face among the input images in which the landmarks are marked (S200). A two-dimensional transform parameter is applied to the generated AAM face model (S300). Estimated three-dimensional face depth information is added to the AAM face model (S400,S500). A three-dimensional transform parameter is applied to the generated three-dimensional face model (S600). The two-dimensional and three-dimensional transform parameters are updated for every input image frames (S700). [Reference numerals] (S100) Triangular mesh is formed by marking landmarks on an input image showing the front and two sides of a face; (S200) Two-dimensional active appearance model (AAM) face model is generated from an AAM parameter defined in a learning process by an AAM in respect to a front face among the input images in which the landmarks are marked; (S300) Two-dimensional transform parameters are applied to the generated AAM face model and is fitted to the input image; (S400) Face depth information is estimated based on the front and two sides of the face; (S500) Three-dimensional face model is created by applying the estimated face depth information as a Z axis to the AAM face model; (S600) Three-dimension transform parameters are applied to the three-dimensional face model and the three-dimensional face model is fitted to the input image; (S700) Two- and three-dimension transform parameters are updated at every image frame of changing poses and looks of the face in a three-dimensional way by repetitively performing the stage 100 and the stage 600; (S800) Directions of the mesh values for those overlapped areas is identified and the weight of the mesh values of the overlapped areas with their directions contradicting each other is set as 0 by replacing the face texture within the triangle with an average texture