Editing Projects:Improve starling tracking data

== Overview == 

This problem comes about due to the starling data set. In this experiment we have between 10 and 20 starlings ranging freely in the tracking space with very small tracking backpacks on each with their own ID. While the VICON tracking system is excellent, working with wild animals always brings about new problems. Ours stem from a combination of the tag size limitations (we are pushing the minimum tag size with these birds) and the behaivour of the birds themselves (preening and other behaivours can temporarily cover parts of the tag obscuring its ID from the cameras and causing the tracking to fail until the entire tag is visible again). We already have more than 24 hours of tracking data (6DOF pose of 3D objects) from the VICON system and need a way to process the data in order to remove incorrect data and fill in missing data.


== Contact ==

Nora Carlson, ncarlson@orn.mpg.de


== Aims ==

The main aim of this is to take the very large about of data that we get from the VICON system (no video) and process the data in such a way as to remove any gaps, incorrect tracks, and smooth low-level jitter. Additionally, a helpful graphical interface that would allow visualization of the tracks over a sliding time bar with any number of selected individuals would also be very cool.
Problems to solve include:

* Locate and remove/fix:
:- Incorrect object positions
:- Rotational object flipping
:- Temporary incorrect object identity
:- Excessive object ‘shaking’ that results from poor tracking

* Find and replace/fill
:- Gaps in tracks
:- Identity mismatching


== Provided data ==

* Data set 1: 6DOF pose of 3D objects (12 objects) tracked with 100FPS we have at least 25 hours of data
* Data set 2: videos that correspond to tracking objects (we have three 10 minute videos that correspond to parts of data set 1) with a frame rate of 50Hz and the object locations in 3D at 100Hz.
* Data set 3: 3D points that occur when the tracking could identify points but not object identity (correspond to data set 1)
:– I don’t currently have these and am not entirely sure how to acquire them, but Hemal has gotten these points previously on another tracking project so I would need to talk to him to see how easy it would be to get these data after the recording has finished.



== Suggested/tested approaches ==

While simple threshold methods do not work on these data very well so far, it is very easy to tell when any of these things happen when looking at visualizations of the tracks of the ID (odd behaviour of the tag), especially when a few are graphed at once (for ID mismatch) on a sliding time scale. Therefore some type of visual interface for highlighting or tagging the data may also be a useful tool.