Abstract
Recently, multiple research groups proposed the concept of direct manipulation,
which makes the task of navigation in videos directly related to content of the
video. However, the developed systems are still too slow for a broad application of
the technique. To increase the accessibility of direct manipulation a system has to
fulfill certain requirements:
• Computation times have to be so small that the system can work in situation
with time constraints.
• The system has to accommodate for the user’s greater comprehension of a
scene. This includes camera motion, fore- and background, and occlusions.
Based on the requirements, we developed and implemented the DRAGONEYE system.
The tracker employs point tracking, with points provided by SIFT [Lowe,
2004], and color tracking with the CAMShift algorithm [Comaniciu et al., 2003]. A
model of the tracked object is built and continuously updated to fit the current conditions.
This model allows the detection and recovery from occlusions.
Secondly, the system estimates the effects of camera motion and represents them as
a homography.
Tests of the system showed that the tracking results are comparable to the current
tracking approaches, and the system outperforms the other algorithm in the case
of occlusions. The tests also showed that improvements are still necessary in case
of rotations around the y-axis and non-rigid movement.
