Enhanced 3D geometric-model-based face tracking in low resolution with appearance model

3D geometric-model-based face tracking utilizes 3D geometric face model to extract facial motion information from video. It has the advantages of dealing with oblique views of faces and it is robust to changes in people, expressions and lighting. Although most existing 3D model-based methods work well with high resolution faces (e.g. face regions around 200 × 200 pixels), the face image in real environments (e.g. surveillance video) is often in low resolution. In low resolution, the performance of 3D model-based face tracking can be degraded because the underlying low-level image motion estimation is less reliable. In this paper, we present an appearance-based enhancement to 3D geometric-model-based face tracking in low resolution. First, we estimate the motion parameters of the 3D face model. Then we extract the stabilized texture map image of the face model. The appearance variations in the face texture images are caused by the errors in motion estimation or lighting changes. Therefore these appearance variations are used as additional constraints for refining the motion estimations. Experiments show that the proposed enhancement improves the robustness of 3D model-based tracking in low resolution.