Clustered manifolds for face pose estimation and driver distraction monitoring

Abstract

Face pose estimation methods try to identify the position and orientation newlineof human faces present in an image. These obtained pose information newlinecould be used to analyze vehicle driver behavior. The main objectives of newlinethis thesis are to model the face pose images as non-linear manifold(s), newline newlinemodel the non-linear manifold into locally linear regions, design unsu- newlinepervised locally linear algorithms for pose estimation, and driver dis- newlinetraction analysis using face pose cues. newline newlineData points with small variations between them are assumed to lie newlineclose to each other on a smooth varying manifold in the feature space. newlineSuch data are hard to classify into separate classes. A sequence of face newlinepose images with closely varying pose angles can be considered as such newlinedata. The pose angles when large enough, create images that are largely newlinediffering from each other and thus the sequence of face images can newlinebe assumed to be on or near a non-linear manifold. Isometric Feature newlineMapping (ISOMAP), Locally Linear Embedding (LLE) and Laplacian newline newlineEigenmaps (LE) are some of the standard non-linear manifold embed- newlineding techniques. The main disadvantage of these techniques is the ab- newlinesence of a projection matrix for out of sample data points. Complexity of these methods increase with the size of the data. Hence in this thesis newlineit is proposed to divide a non-linear manifold into locally linear regions newline newlineand linear manifold embedding techniques are then used for pose esti- newlinemation.The classiand#57346;cation of smooth varying face pose manifolds now boils newlinedown to a multiple subspace problem. The smooth manifold is divided newlineinto multiple disjointed, locally linear, separable clusters. The problem newline newlineof identifying which cluster to use is solved by dividing the entire pro- newlinecess into two steps. The and#57346;rst layer projection using the entire smooth newline newlinemanifold, identiand#57346;es a rough region of interest. Second layer projection newlineuses trained cluster(s) from this rough regions of interest to identify the newlinenearest pose.