The type of tracking one should use will depend on the nature of the targets, how they move, and the project characteristics. There are three stages to tracking: the initial placement in the image, the search, and the sub-pixel marking.
The user has control over a few options such as whether tracking is done in two or three dimensions, whether to use the previous epoch’s position vs a Kalman-filter based prediction for the initial placement, whether a simple spiral contrast search from the initial placement or NCC matching is done, and lastly what type of sub-pixel marking is done once the target is found. There is a lot of flexibility to account for different scenarios.
The initial placement can be done in two or three dimensions. In 2D placement, the 2D location of the mark in the previous epoch is used in the epoch being tracked. In 3D projection, the 3D location of the point is projected onto the photograph being tracked and that 2D location is used as the initial position.
There are two types of 3D tracking '3D (all photos)' and '3D (previously marked)': the (all photos) option will try to track every 3D point on every photo, the (previously marked) option will only track 3D points on photos that the point was marked on in previous epochs.
Motion Type 2 projects must use 3D tracking because there is no relationship between photos between epochs in type 2 projects.
Target Searching
Regardless of using 2D or 3D points, the first search step is to determine a ‘start location’ in this epoch in 2D. To find this ‘start location’ there are two options: a) if the “Track method” is “Search” then the location from the previous epoch is used (2D location or projected 2D location of the 3D point), b) if the “Track method” is “Predict and Search” then a prediction is made based multiple previous epoch locations (2D or 3D).
“Predict and Search” uses a Kalman filter (for background, search the web for 'Kalman filter in tracking'). The predict and search helps when target motion is large (such as larger than the target spacing) but the targets are moving in 'predictable and non-random' ways. Note that after the initial point is predicted, the search then happens, as described below – so the predicted location does not need to be inside the new target location.
Once we have a starting location, there are two forms of search around that starting location to find the target. The spiral search spirals out from the start location looking for the first pixel transition that is bigger than the ‘Search Radius’ (this method works well for high contrast targets with dark and low-noise backgrounds). The NCC (Normalized Cross Correlation) method uses an image processing correlation method that is slower but is also more robust. NCC works well with Template Targets and works well where there is a natural image or noisy background around the targets. NCC searches in a predetermined area around the start location to find the best match.
The Search radius should be at minimum the largest radius of a target in your images in pixels. If the initial location ends up right in the center of the target, the search for an edge needs to be at least the radius of the target (as the search from inside the target is still looking for a strong edge).
One way of describing the tracking is to ask, “When does a target mark get dropped from tracking? Here are the possible ways a target tracking will fail:
• If the search radius is not big enough (from the copied location or the predicted location starting point) to include any edge of this epoch’s target, then the target will be missed.
• If there is a high contrast noise edge found first, the search will find it and only, and then the marking stage will fail.
• Once a high contrast edge is found in the image (if using spiral search), then the region is expanded to include the whole target, and the point is recentered to give a location for the subpixel marker to work (see below). The subpixel marker can then fail if this region expansion fails due to other edge noise on the target or the target is mishappen (too far from being an ellipse for example).
Target Marking
Once the target has been found, as described above, the marking can be done two ways: LSM (Least Squares Matching) and Centroid. See the Automatic Target Marking section for more information.
When a series of sub-pixel marked points are tracked, the Tracker determines which target type was used to initially mark the point. It then tracks these points into the selected epochs using the same template matching algorithm. For example, if epoch 1 had a number of Circular Targets and two types of Template Targets already marked, then when these points are tracked into epoch 2, the matching target type is used (that is after the search, the Circular Targets are sub-pixel marked with the Circular Target Marking Method, and the Template Targets are marked with the appropriate Target Template.