Beta 1


Title Face Tracking using an Extended Active Appearance Model on Time-of-Flight data
Author Møller, Christian Hjorth
Supervisor Larsen, Rasmus (Image Analysis and Computer Graphics, Department of Informatics and Mathematical Modeling, Technical University of Denmark, DTU, DK-2800 Kgs. Lyngby, Denmark)
Institution Technical University of Denmark, DTU, DK-2800 Kgs. Lyngby, Denmark
Thesis level Master's thesis
Year 2008
Abstract This thesis describes statistical methods for modelling the shape, texture, volume and appearance of human faces using three-dimensional Time-of-Flight(TOF) data. Traditionally this has been done in two-dimensions due to the easy data acquisition and data representation. But with the emerge of cheap TOF-cameras the three-dimensional domain is now more feasible to explore. Appearance Models are statistical models providing an effective mean of separating inter- and intra-class variation. The models integrate shape and intensity information from images to derive a compact class description. In this thesis an extension has been made to the default model to incorporate the depth map of the third dimension. The Active Appearance Model(AAM) is a learning-based deformable model often used for image segmentation. This is an optimization problem of trying to estimate the parameters of an appearance model to best match a given input. In this thesis statistical models are constructed based on the amplitude images and corresponding depth maps of faces generated by a TOF-camera. The overall aim is to examine AAM's on TOF-data in general and whether or not the information in the third dimension influence on the segmentation/tracking result of the AAM. The primary datasets created contains sequences of more than 500 TOF-images of my own face in different expressions. The overall conclusion is that the use of TOF-data results in mixed answers to whether or not to use the depth-maps for tracking. It seems that models using the depth maps improve the segmentation and tracking on high variance datasets. On low variance datasets, where the default shape/texture model is sufficient, the depth-model only further complicates the model by introducing more parameters and the end-result tends to be worse.
Series IMM-M.Sc.-2008-65
Fulltext
Original PDF ep08_65_net.pdf (2.15 MB)
Admin Creation date: 2008-07-17    Update date: 2008-07-17    Source: dtu    ID: 221859    Original MXD