Tracking particles

The “track” function in diff_classifier operates by creating a temporary script files in the TEMP directory and generating a command sent through the shell to ImageJ. Tracking is implemented using the TrackMate ImageJ plugin. User inputs include tracking parameters and spot and track filter cutoffs.

This implementation is fairly constrained from the complete number of parameters that Trackmate offers, but will be expanded in the future. The parameters that are currently implemented are:

  • radius: estimated radius of spots in videos. In general, should be slightly larger than the average particle size in the videos. Note that the GUI interface uses diameter rather than radius.
  • threshold: not currently in use.
  • do_median_filtering: If set to True, filters the image before performing tracking to minimize noise.
  • quality: Lower threshold on spot quality filter. Usually varies anywhere between 1 to 300.
  • x: Upper threshold on x coordinate spot filter.
  • y: Upper threshold on y coordinate spot filter.
  • ylo: Lower threshold on y coordinate spot filter.
  • linking_max_distance: max distance in pixels that a particle can travel between frames.
  • gap_closing_max_distance: max distance in pixels that a particle can travel when it skips a frame.
  • track_displacement: meant to be duration. Minimum number of frames a trajectory must have to be included in the final dataset.

The algorithm is set to only implement the difference of Gaussians detection algorithm and the simple LAP tracker algorithm.

track(target, out_file, template=None, fiji_bin=None, radius=2.5, threshold=5.,
      do_median_filtering=False, quality=30.0, x=511, y=511, ylo=1,
      median_intensity=55000.0, snr=0.0, linking_max_distance=10.0,
      gap_closing_max_distance=10.0, max_frame_gap=3,
      track_displacement=0.0)

Selecting trajectory parameters

One difficulty with particle tracking technology is that there is no easy way to determine the “correct” tracking parameters for different videos. Even videos collected in the same conditions using the same particles can have different optimal tracking parameters. Parameters can also vary from user to user. How does one select optimal parameters, especially where tracking large numbers of videos?

If users choose to use a single set of parameters for all videos, users risk outputting poor tracking results, especially if there is uneven illumination or different particle distributions with the video frame. This works well for relatively homogeneous videos with similar particle sizes and uniform illumination.

I have implemented an intermediate approach in diff_classifier that uses intensity data from input videos to predict quality cutoff parameters. This assumes that particles in all videos have similar radii and are moving at similar velocities, while adjusting the quality threshold filter. The function imagej.regress_sys allows users to create a regression object that predicts quality parameters from an input training dataset. Users must manually track a small number of randomly selected videos from their entire video collection and use the quality values as inputs to regress_sys e.g.

prefix = 'test'
nvideos = 20
all_videos = []
for num in range(nvideos):
    all_videos.append('{}_{}'.format(prefix, num))
yfit = []
training_size = 4

ij.regress_sys('.', all_videos, yfit, training_size, have_output=False)

When have_output is set to False, regress_sys will output a list of randomly selected videos contained in all_videos which the user must manually track. The quality values are loaded into the variable yfit to produce the regression object using scikit-learn:

yfit = [20.4, 17.9, 10.3, 30.4]
regress = ij.regress_sys('.', all_videos, yfit, training_size,
                         have_output=False)

An example of such an implementation can be found here.