Face detection, tracking and orientation estimation
Student thesis: Doctoral Thesis
Related Research Unit(s)
|Award date||3 Oct 2005|
Face orientation estimation is an important topic in computer vision and pattern recognition, and in particular, more recently in some emerging, media-related applications. Due to the non-rigid properties of faces, it is computationally expensive and in general difficult to achieve good estimation accuracy and robustness in face orientation estimation. In this thesis I present a new, robust, and efficient face orientation estimation algorithm. It broadly consists of three parts: first, the face is detected by a new featureinvariant face detection method; then the face is tracked robustly using three visual trackers based on extracted facial features; and finally the orientation of face is estimated by fusing the tracking results obtained independently from the three trackers. In this research I have focused on the development of a reliable method for face orientation estimation, based on the considerations of a good trade-off between estimation accuracy and computational efficiency, and of a robust estimation procedure. First, I propose a new, robust ellipse detection algorithm with a real-time performance. The algorithm is based on a heuristic point-selection scheme and fast Hough transform techniques. By computing the convexity of the point in an edge map, it selects pairs of points according to a convexity matching scheme. Then it decomposes the highdimensional ellipse HT parameter space and computes the parameters by the selected point-pairs in three stages. The new ellipse detection algorithm is the first contribution I have made in this thesis. Since face shapes are in general ellipse-like, I extend the techniques in the proposed ellipse detector to developing a face shape detector. Then I proposed a new, efficient algorithm of face detection, which consists of two modules: a face shape detector and a face texture verifier. In the face texture verifier I present an efficient texture description method and apply it to the matching of face texture patterns. The two modules are combined by a convergence strategy that zooms in the potential candidate regions at each stage and saves computational costs significantly. The new face detection algorithm is the second contribution of this thesis. With the proposed ellipse detector and face detector, I develop a new algorithm for face orientation estimation. The algorithm relies on the combination of three individual tracking-based face orientation estimators that are based on the three properties of the face in question respectively: the variation of face regions, the deformation of face texture patterns, and the trajectory of face motion. The combination is achieved by the data fusion technique upon the principles of Sequential Monte Carlo. In tacking the multiple measurement modalities I propose two data fusion strategies: intermediary-based strategy and assembly-based strategy, which extend the conventional Sequential Monte Carlo methods effectively. By integrating the multiple visual cues through data fusion, the algorithm is reliable and able to estimate face orientation efficiently. The contributions in the face orientation estimation algorithm are twofold: the first and the primary contribution is the innovative idea that uses the state-space modelling approach to seeking solutions to the face orientation recovery problem, and to applying data fusion techniques to the integration of multiple cues. The two data fusion strategies for estimating a high dimensional state vector using multiple measurement modalities can significantly reduce the computational burden. Another associated contribution is the three tracking-based face orientation estimators. I show how to model them probabilistically to participate in the data fusion. The three estimators are able to estimate the face orientation efficiently and independently. To illustrate the performance of the proposed algorithms, I have conducted extensive experiments on both benchmarks and randomly collected real-world samples. In the thesis, I present in detail the experimental results and evaluate the performance of the proposed algorithms. I have also carried out comparative experiments with some typical algorithms to investigate the merits and limitations of techniques developed in this study. And finally, I conclude this study and propose some possible directions for future research.
- Human face recognition (Computer science)