S. Birchfield, Elliptical Head Tracking Using Intensity Gradients and Color Histograms, CVPR 1998. www
This method seeks to improve tracking robustness by combining complementary tracking criteria: edge gradients and interior color histogram. The head is modelled as an ellipse, and the tracker initialized offline with a three-quarter profile view to include both hair and skin colors. The border and the interior of the ellipse are geometric complements. Gradient and color information should also tend to complement each other in feature space. This latter claim is presented intuitively and supported empirically, but lacks mathematical proof.
For each frame, the best head location in (x,y,s) space (s is scale)
is computed as the maximum value for a matching function. The matching
function is the sum of the range-normalized gradient normals
around an elliptical perimeter plus the range-normalized color similarity
between model and search region. Color similarity is measured as
histogram intersection:
where S is the similarity, I the input histogram, M
the model histogram, and N the number of histogram bins. This is
identical to set intersection, with the two histograms as sets.
A constant-velocity assumption plus a fixed size for the search region constrain the size and location of the search space at each iteration.
| Home | Face Tracking |