Integrate KCF tracker
NOTE: may be a nice topic for a seminar (not enough for a whole bachelor/master thesis)
After the work done in #184 in the branch tracking-interface
, we decided that we want to include the Kernelized Correlation Filter (KCF) as a new tracking algorithm in PeTrack. (The CSRT-Tracker, which builds upon the KCF tracker and which is available in OpenCV as well, could also be tested) While it is significantly slower than the Lukas-Kanade-Algorithm, it seems to show more accurate results as well.
The KCF is based on the paper High-Speed Tracking with Kernelized Correlation Filters.
Work that still needs to be done for a successful integration:
- Check different parameters and integrate those most important into the GUI
- Especially the abortion threshold should be checked, since the tracking is aborted rather quickly right now
- Manage tracking in non-benchmark scenarios, this includes problems like
- Tracking manually created points, not only automatically recognized points
- Tracking correctly when manually splitting, deleting, ...
- Correctly reset when pressing the "reset"-button
- Tracking correctly, when jumping to a new frame
- Try to reuse existing tracker-objects as much as possible, to benefit from the "learning"-effect of these tracking methods
- Support changing the tracking method