The dynamic structure of scenes and gestures
Octavia Camps, Benjamin R. Fransen · 2006
This thesis investigates the role of dynamics in computer vision. Feature tracking, super-resolution and gesture identification are each improved through the novel use of dynamics proposed in this thesis. For improvement of feature tracking, identified dynamics are utilized to facilitate location and appearance forecasting for rigid object motion in video. To identify the dynamics of rigid objects, a novel method is proposed that utilizes Caratheadory-Fejer (CF) interpolation with dynamics recovered from a factorization algorithm. Identified dynamics are combined with recovered object structure to predict the appearance and location of rigid bodies. Forecasted object properties are then demonstrated to improve tracking of objects in video sequences. The enhancement of super-resolution is made possible by a novel method for the reintroduction of dynamics to non-planar objects recovered from video. Previous methods for super-resolution have been formulated based on a two dimensional scene model. The use of dynamics with recovered non-planar structures generalizes image registration for scenes with arbitrary shapes. Superior super-resolution for non-planar scenes is demonstrated through the combination of dynamics and recovered scene structure. A novel method for gesture identification of dynamic gestures is the final topic proposed in this thesis. A technique called non-rigid segmental hidden Markov models (NRSHM) is presented and demonstrated to be both robust, accurate and tractable for human gesture recognition. Gesture recognition is performed in real time and capable of producing gesture based computer interfaces for human motion recovered from a moving camera.