As the tracking is implemented by association of ROIs based on the position in the world, wrong associations can occur due to calibration errors and wrong motionestimation.
We can assume the following two things:
Boats don’t change their appearance over time
Nature (waves etc.) changes its appearance over time
To avoid associations of totally different ROIs ( boat/nature) and to filter out changing ROIs ( nature ) some image featuers can be implemented to check these assumptions. Some computational simple features can be calculated to check this:
image moments ( Hu´s moments)
correlation
???
After calculation of these features we can filter out wrong trackings by checking against a empirical estimated threshold.
Check slide 22 of the 35c3 talk for further information.
As the tracking is implemented by association of ROIs based on the position in the world, wrong associations can occur due to calibration errors and wrong motionestimation. We can assume the following two things:
To avoid associations of totally different ROIs ( boat/nature) and to filter out changing ROIs ( nature ) some image featuers can be implemented to check these assumptions. Some computational simple features can be calculated to check this:
After calculation of these features we can filter out wrong trackings by checking against a empirical estimated threshold. Check slide 22 of the 35c3 talk for further information.