Package boofcv.alg.tracker.circulant

Class CirculantTracker<T extends ImageGray<T>>

java.lang.Object
boofcv.alg.tracker.circulant.CirculantTracker<T>
public class CirculantTracker<T extends ImageGray<T>> extends Object

Tracker that uses the theory of Circulant matrices, Discrete Fourier Transform (DCF), and linear classifiers to track a target and learn its changes in appearance [1]. The target is assumed to be rectangular and has fixed size and location. A dense local search is performed around the most recent target location. The search is done quickly using the DCF.

Tracking is performed using texture information. Since only one description of the target is saved, tracks can drift over time. Tracking performance seems to improve if the object has distinctive edges.

CHANGES FROM PAPER:

  • Input image is sampled into a square work region of constant size to improve runtime speed of FFT.
  • Peak of response is found using mean-shift. Provides sub-pixel precision.
  • Pixels outside the image are assigned random values to avoid the tracker from fitting to them. Ideally they wouldn't be processed, but that is complex to implement

[1] Henriques, Joao F., et al. "Exploiting the circulant structure of tracking-by-detection with kernels." Computer Vision–ECCV 2012. Springer Berlin Heidelberg, 2012. 702-715.

  • Field Details

  • Constructor Details

    • CirculantTracker

      public CirculantTracker(double output_sigma_factor, double sigma, double lambda, double interp_factor, double padding, int workRegionSize, double maxPixelValue, InterpolatePixelS<T> interp)
      Configure tracker
      Parameters:
      output_sigma_factor - spatial bandwidth (proportional to target) Try 1.0/16.0
      sigma - Sigma for Gaussian kernel in linear classifier. Try 0.2
      lambda - Try 1e-2
      interp_factor - Try 0.075
      padding - Padding added around the selected target. Try 1
      workRegionSize - Size of work region. Best if power of 2. Try 64
      maxPixelValue - Maximum pixel value. Typically 255

  • Method Details

    • initialize

      public void initialize(T image, int x0, int y0, int regionWidth, int regionHeight)
      Initializes tracking around the specified rectangle region
      Parameters:
      image - Image to start tracking from
      x0 - top-left corner of region
      y0 - top-left corner of region
      regionWidth - region's width
      regionHeight - region's height

    • setTrackLocation

      public void setTrackLocation(int x0, int y0, int regionWidth, int regionHeight)
      Used to change the track's location. If this method is used it is assumed that tracking is active and that the appearance of the target has not changed
      Parameters:
      x0 - top-left corner of region
      y0 - top-left corner of region
      regionWidth - region's width
      regionHeight - region's height

    • initialLearning

      protected void initialLearning(T image)
      Learn the target's appearance.

    • computeCosineWindow

      protected static void computeCosineWindow(GrayF64 cosine)
      Computes the cosine window

    • computeGaussianWeights

      protected void computeGaussianWeights(int width)
      Computes the weights used in the gaussian kernel This isn't actually symmetric for even widths. These weights are used has label in the learning phase. Closer to one the more likely it is the true target. It should be a peak in the image center. If it is not then it will learn an incorrect model.

    • resizeImages

      protected void resizeImages(int workRegionSize)

    • performTracking

      public void performTracking(T image)
      Search for the track in the image and
      Parameters:
      image - Next image in the sequence

    • updateTrackLocation

      protected void updateTrackLocation(T image)
      Find the target inside the current image by searching around its last known location

    • subpixelPeak

      protected void subpixelPeak(int peakX, int peakY)
      Refine the local-peak using a search algorithm for sub-pixel accuracy.

    • performLearning

      public void performLearning(T image)
      Update the alphas and the track's appearance

    • dense_gauss_kernel

      public void dense_gauss_kernel(double sigma, GrayF64 x, GrayF64 y, GrayF64 k)
      Gaussian Kernel with dense sampling. Evaluates a gaussian kernel with bandwidth SIGMA for all displacements between input images X and Y, which must both be MxN. They must also be periodic (ie., pre-processed with a cosine window). The result is an MxN map of responses.
      Parameters:
      sigma - Gaussian kernel bandwidth
      x - Input image
      y - Input image
      k - Output containing Gaussian kernel for each element in target region

    • circshift

      public static void circshift(GrayF64 a, GrayF64 b)

    • imageDotProduct

      public static double imageDotProduct(GrayF64 a)
      Computes the dot product of the image with itself

    • elementMultConjB

      public static void elementMultConjB(InterleavedF64 a, InterleavedF64 b, InterleavedF64 output)
      Element-wise multiplication of 'a' and the complex conjugate of 'b'

    • computeAlphas

      protected static void computeAlphas(InterleavedF64 yf, InterleavedF64 kf, double lambda, InterleavedF64 alphaf)
      new_alphaf = yf ./ (fft2(k) + lambda); %(Eq. 7)

    • gaussianKernel

      protected static void gaussianKernel(double xx, double yy, GrayF64 xy, double sigma, GrayF64 output)
      Computes the output of the Gaussian kernel for each element in the target region k = exp(-1 / sigma^2 * max(0, (xx + yy - 2 * xy) / numel(x)));
      Parameters:
      xx - ||x||^2
      yy - ||y||^2

    • get_subwindow

      protected void get_subwindow(T image, GrayF64 output)
      Copies the target into the output image and applies the cosine window to it.

    • getTargetLocation

      public RectangleLength2D_F32 getTargetLocation()
      The location of the target in the image

    • getTargetTemplate

      public GrayF64 getTargetTemplate()
      Visual appearance of the target

    • getResponse

      public GrayF64 getResponse()