Improved techniques for motion estimation and analysis from videos

Estimation of 2-D motion  is required in many video processing and computer vision tasks. The problem is made difficult to solve by the fact that estimation must based on the changes in image brightness which, in general, depend on several other factors than motion. In some cases, motion may not even cause any observable changes.

Pekka Sangi's doctoral thesis considers the problem of local motion estimation from the perspective of confidence or uncertainty analysis. Such analysis methods can complement motion estimates with useful information about their reliability. Specifically, Sangi's work concentrates on the block matching based motion estimation. A particular form of uncertainty analysis is developed, which - in contrast to the existing approach used in this context - takes into account the local image gradient.

Based on this approach, the thesis addresses two general tasks in estimating projected motions of background and foreground objects in the scene: global motion estimation and motion based segmentation. A new framework is also developed for motion based object detection, segmentation, and tracking. It is experimentally shown that exploitation of uncertainty information can improve the performance of motion analysis.

The study of uncertainty analysis and dominant motion estimation has been used as a basis for a software implementation included in an open access computer vision library. The motion estimation software has also been an important tool in the studies related to digital video stabilisation and human-computer interaction.

Pekka Sangi will defend his doctoral thesis in public at the University of Oulu on January 29, 2013 at 12 noon in auditorium OP-sali, Linnanmaa campus.

Last updated: 23.1.2013