CS7322: Computer Vision II: Final Project Proposal



 

Image Stabilization of Video

Gabriel J. Brostow



 

The Idea:

In order to allow image differencing and object tracking vision algorithms to function in real-world situations, video data must be preprocessed to remove the effects of camera/cameraperson movement. In essence, an automated system can be developed which will be able to distinguish the difference between objects in the scene moving, and the camera itself moving. As humans, we do this rather easily.

The Domain and the Scope:

I will process data captured with a hand-held camera in both indoor and outdoor settings. At the least, I will identify translational motion of the camera relative to the objects in the scene, assuming at least half of the scene is actually static. Hopefully, I will be able to also compensate for camera rotations, zooms, and for scenes where the majority of the image is changing due to objects moving.

Approach:

I have already found several research papers which explore the two most common methods for image stabilization: methods based on optical flow, and methods based on feature tracking. Both methods have been used successfully, but both have drawbacks. Optical flow robustness decreases as the number of moving objects increases. Feature tracking can misidentify the desired feature and result in grossly scewed motion compensations. This method is also especially susceptible to errors caused by occlusions.
 

Proposed Time-line:

Week 1:
Finish reading previous work.
Week 2:
           Perform evaluation of existing packages and algorithms to find out previous sucesses and failures in              real systems.
Week 3+:

           Desing and implement my own system in C/C++.

Final: Incorporate into other vision systems as a preprocessing stage.