|
" A Model-based Approach for Combined Tracking and Resolution Enhancement of Faces in Low Resolution Video. "
Annika Kuhl
| Document Type
|
:
|
BL
|
| Record Number
|
:
|
791737
|
| Doc. No
|
:
|
b611775
|
| Main Entry
|
:
|
Annika Kuhl
|
| Title & Author
|
:
|
A Model-based Approach for Combined Tracking and Resolution Enhancement of Faces in Low Resolution Video.\ Annika Kuhl
|
| Publication Statement
|
:
|
INTECH Open Access Publisher, 2009
|
| ISBN
|
:
|
3902613424
|
|
|
:
|
: 9783902613424
|
| Abstract
|
:
|
We proposed a combined tracking and super-resolution algorithm that increases the resolution online during the tracking process. An object-specific 3D mask mesh is used to track non-planar and non-rigid objects. This mask mesh is then subdivided such that every quad is smaller then a pixel when projected into the image. This makes super-resolution possible and in addition improves tracking performances. Our approach varies from traditional super-resolution as the resolution is increased on mask level and only for the object of interest rather than on image level and for the entire scene. We demonstrated our combined geometric and appearance based tracking approach on sequences of different size faces and showed that our approach is able to track faces down to 28x20 pixels in size. The combination of these two tracking algorithms achieves better results than each method alone. The tracking algorithm is further extended to allow for the non-rigidity of objects. We applied this to faces and expressions. Experiments showed that the proposed method for expression tracking reduces the mean tracking error and thus allows for a better alignment of consecutive frames, which is needed to create super-resolution images. The proposed 3D model-aided super-resolution allows for a high increase in resolution; the finer the 3D mesh the higher the possible increase in resolution. Therefore we experimentally estimated the number of frames needed to achieve a certain resolution increase. In practice about 20 to 30 frames are needed to double the resolution. Increasing the resolution further is limited by the number of frames used as well as the tracking error. Large tracking errors and the averaging process across a large number of frames introduces noise that decreases the quality of the super-resolved image. We demonstrated our method on low resolution video of faces that are acquired both in the lab and in a real surveillance situation. We show that our method outperforms the optical flow based method, and performs consistently better for longer tracking durations in video that contain non-planar and non-rigid low-resolution objects. The combined tracking and super-resolution algorithm increases the resolution on mask level and makes interpolation, the first step of the optical flow algorithm, redundant. The resulting super-resolved 3D model is less blurred by achieving the same or a higher resolution increase. This in turn makes deblurring, the last step of the optical flow algorithm, unnecessary. Furthermore the super-resolved 3D model is created online during tracking and improves with every frame, whereas super-resolution optical flow incorporates consecutive and previous frames which prohibits its usage as an online stream processing algorithm.
|
| Subject
|
:
|
Open Access Collection.
|
| Added Entry
|
:
|
Annika Kuhl
|
|
|
:
|
Svetha Venkatesh
|
|
|
:
|
Tele Tan
|
| |