Nachricht

Am Mittwoch, den 8. April 2015, hält Robin Vogt um 16:30 Uhr im Raum P 1.4.17 einen Vortrag über seine Masterarbeit mit dem Titel:

Efficient Implementation and Evaluation of Variational Optical Flow Based on Texture Constraints

Abstract:

Estimating the apparent motion of objects, surfaces and edges in a visual scene, caused by the relative motion between a camera and the scene, can be done with the help of optical flow. In an environment with constant brightness, the optical flow is usually estimated using the brightness constancy constraint (BCC). This constraint assumes, that the brightness of a pixel does not change if an object or the camera moves. However, once the illumination changes or objects move to another place with a different illumination condition (e. g. into the shadow of a tree), this assumption is no longer valid. A promising, more robust solution is to use a texture constraint. Such a constraint assumes, that the relation be- tween neighborhood pixels (e. g. edges, gradients) stays constant if an object or the camera moves, while the brightness might vary. There exist several different descriptors that can be used to describe a texture, for example the histogram of oriented gradients (HOG), the local directional pattern (LDP) or the census signature. Although these descriptors provide robust and accurate optical flow results, the high computational cost, which limits the application of these ap- proaches, is a big obstacle. Thus, the goal of this work is to optimize the optical flow estimation based on different texture descriptors by using parallel processing programming on the central processing unit (CPU) and the graphics processing unit (GPU). Furthermore, the robustness, accuracy and performance for different environments will be evaluated. As a result, advantages and disadvantages of each descriptor will be highlighted. Moreover, the descriptors will be extended with color information.