Tracking
Floe tracking in Ice Floe Tracker is based on object shapes and locations. Given a series of segmented images, the goal is to identify shapes that persist across images, subject to constraints on object similarity and maximum displacements. The algorithm has three main components: (1) the main floe_tracker algorithm, which keeps track of the most recent floe in each trajectory (trajectory heads) and the set of candidate matches in the next image, (2) a "floe filter" function which identifies all the possible pairs between a trajectory head and floes in the candidate set, and (3) a "floe matching" algorithm which identifies a set of unique pairs. By designing and customizing the filter functions and matching functions, the floe_tracker function provides a flexible and powerful platform for developing floe tracking workflows.
Preparing data for tracking
The floe_tracker takes three positional arguments: a list of DataFrames, a filter function, and a matching function, as well as optional keyword arguments to specify the minimum and maximum floe sizes (in pixels) and the maximum time step in between floe pairs. Assuming that a set of images segmented_images has already been produced, the regionprops_table function produces a DataFrame where each row corresponds to a floe, and each column is some measurement or attribute of the floe. DataFrames are highly flexible–entries in the dataframe are not limited to words and numbers, but can include vectors and matrices as well. At the very least, the property tables will need to include the floe ID, floe area and an associated observation time. Other columns and measures depend on what will be used in the filter function. The regionprops_table function defaults to calculating area, convex area, centroid, perimeter, major and minor axis, and orientation. We can initialize the property tables using dot notation to broadcast to a list:
props = regionprops_table.(segmented_images)We include helper functions to add unique IDs to each row and to add image observation times. Assuming passtimes is a list of DateTimes of the same length as segmented_images, we run
add_uuids!.(props)
add_passtimes!.(props, passtimes)The default filter functions include calculations based on binary floe shapes (floe masks) and associated $\psi$-s curves. We include helper functions for these as well. For the floe masks, we also need a list with binary images associated with each segmented image.
add_floemasks!.(props, binary_images)
add_ψs!.(props)Floe Filter Functions
Floe filter functions take two argmuments: a DataFrameRow corresponding to the floe to be matched, and a DataFrame with candidate pairs from the current time step. These functions should operate in-place and result in DataFrame subsetting operation. IceFloeTracker.jl includes four main AbstractFloeFilterFunctions. These functions all follow the same procedure:
- Compute comparisons between the floe and each floe in the candidates DataFrame
- Use an threshold test to evaluate the comparisons
- Reduce the candidates dataframe to only those pairs that pass the threshold test
Each of these functions can be called in series, as they only depend on the columns in the input property tables. TheChainedFilterFunctiontakes a list of filter functions and wraps them into a single function call. Using the struct/functor approach, we can initialize each function with parameters, then pass the function and settings into the floe_tracker function. As an example, let's define a filter function which compares the relative error in area against a step function. We initialize the step function first:
sw = StepwiseLinearThresholdFunction(changepoint_area=500, low_value=0.25, high_value=0.125)The function sw will return true if the relative error is less than 0.25 and the area is smaller than 500 pixels. If the area is larger, then a stricter threshold of 0.125 is applied. Next, we make a filter function
rel_err_ff = RelativeErrorThresholdFilter(variable=:area, threshold_function=sw)This function will add a column :relative_error_area to the filter function, which can be used in the Matching Function.