The research on real time face localization and tracking in the video has both significant academic importance and wide applications. Based on the background of expression interaction, this thesis focuses on the problems of face localization and tracking. The requirements of the application and the shortages of the existing algorithms are fully considered. Novel algorithms and improvements are proposed. Research results that have application value are obtained. The main contributions of this thesis are as follows: 1. A face localization algorithm based on eye detection and a fast blinking detection algorithm are proposed. A novel clustering algorithm based on distance measures is used to remove the false positive results obtained form an AdaBoost eye detector. A further search for the precise location of the eye center is conducted, then multiple features are extracted to build statistical models. The likelihoods of different features are combined by Dempster-Shafer theory. Experimental results demonstrate that the localization speed is significantly increased, and the blinking detection is precise. 2. A motion estimation algorithm based on Motion History Image is proposed to improve the stability of the face tracker. A golden section optimization method is used to find the motion direction and speed. Comparing with the existing motion estimation method based on Motion History Image, the proposed one is more robust to the variation of the edge orientation. After further corrected, the estimation is combined with a face tracker based on Active Appearance Model according to two methods. The combined methods reduce the probability of tracking failure, and improve the tracking speed. 3. A 3D head tracking algorithm based on particle filter and feature matching is proposed. A 3D morphable head model is used to fit the shape of the target by minimize the distances between the points on the model and the ones in the input image. Face images with different pose can be rendered based on the head model and face texture obtained at the beginning of the tracking. Feature selection and matching are taken based on the rendered image and input image, then the pose variation can be estimated. RANSAC is used to remove the wrong matching. Average normalized cross correlation is used to evaluate the particles. Then MAP is used to get the tracking result. Experimental results show the algorithm is effective to track the head with large pose variation. 4. A 3D head pose an...
修改评论