Tracker Reconstruction
Overview
- The current tracking (TkrRecon) is based upon a Kalman Filter for performing both tracking finding
and fitting. However, it is believed that the current code has some problems in implementation,
ranging from how the filter and smoother are implemented to the basic strategy of finding gammas.
- There is general agreement that a new strategy should be adopted which would separate the tasks
of track finding, track fitting and vertexing into separate modules. Further, these modules should
have abstract interfaces so that each of the individual pieces are interchangeable, allowing, for
example, to switch between different pattern recognition or fitting algorithms as desired.
- Pattern recognition would be tasked with identifying possible individual tracks within the detector.
As this is written there is some disagreement as to whether this means attempting to identify all hits
on possible candidates, or simply to identify candidate starting points and slopes. At least two
candidate methods are currently thought of, a "Link and Tree" method and another based on a
Hough Transform.
- For track fitting, the Kalman Filter is the desired fitting algorithm. However, the
implementation should take advantage of a separate track extrapolation algorithm to perform the transport of the
error matrix from the current fit point to the next measured hit point. It is intended that this track
extrapolation algorithm will be a GLAST utility suitable for transporting tracks and their error
matrices throughout the detector.
- The final step would be to form candidate vertices from the resulting found tracks. This algorithm
needs to be thought out in more detail in order to understand how to deal with conversions where
the pair do not split for a layer (or two), where only one track is found, etc. NOTE: A strategy is
that "reconstruction" means only finding and fitting tracks, vertexing would fall into the realm of
"analysis."
- Random note... We should assume that the above can/will proceed with some basic calorimetery
which will provide a rough energy and cluster centroid position. However, the pattern recognition
will be tasked to find track candidates without this information and assume it will only be available
at the track fit stage.