Locally tuned nonlinear manifold for person independent head pose estimation

Date of Award


Degree Name

Ph.D. in Electrical Engineering


Department of Electrical and Computer Engineering


Advisor: Vijayan K. Asari


Fine-grain head pose estimation from imagery is an essential operation for many human-centered systems, including pose independent face recognition and human-computer interaction systems. It is only recently that estimation systems have evolved past a coarse level classification of pose and concentrated on fine-grain estimation. In particular, the state of the art of such systems consists of nonlinear manifold embedding techniques, such as Locally Linear Embedding and Isomap, that capture the intrinsic relationship of a pose varying face dataset. The success of these solutions can be attributed to the acknowledgment that image variation corresponding to pose change is nonlinear in nature. Yet, these algorithms are limited by the complexity of embedding functions that describe the relationship and provide no clear method for projecting novel data to the latent space. On the other hand, linear methods and nonlinear approximation techniques permit a simple embedding process, but lack the representational quality to globally describe the nonlinear image variation. In this dissertation, a pose estimation framework that seeks to describe the global nonlinear relationship in terms of localized linear functions is presented. A locally tuned nonlinear manifold is formulated using a two-layer system based on the assumptions that coarse pose estimation can be performed adequately using supervised linear methods, and fine pose estimation can be achieved using linear regressive functions if the scope of the pose manifold is limited. The localized linear approach results in a simplistic model for which probe input can be embedded through a cascade of linear transformations. Additionally, new methods for modeling the localized structures using feature enhanced Canonical Correlation Analysis are developed, where pose variation is regarded as a continuous variable and is represented by a manifold in feature space. The feature enhanced methods are used to identify the modes of correlation between the observed input images and the head pose angle. These techniques exploit oriented filters which serve two key purposes: (a) eliminate noise features, while boosting image elements that are associated with head pose (b) provide multiple dimensions of the input, allowing the correlation analysis process to extract more basis vectors to provide higher accuracy. A pose estimation system is implemented utilizing simple linear subspace methods, phase congruency, and Gabor features. The framework is tested using conventional test strategies and widely accepted pose-varying face databases. The proposed method is first tested using homogeneous datasets, where the training and testing face images are sampled from the same database. The generalization capabilities of the system are then tested using heterogeneous tests, where the training data is taken from a different database than the testing set. The proposed system is shown to perform fine head pose estimation with competitive accuracy when compared with state of the art nonlinear manifold methods. The results show that the system is capable of predicting head orientation in the yaw direction with as little as 3.46 degrees of error. This research is progressing to expand the multi-layer concept to the case of generalized object pose estimation and predicting pose with multiple degrees of freedom.


Image analysis Mathematical models, Head Models, Human face recognition (Computer science)

Rights Statement

Copyright 2011, author