Original Project Description


Project Goals:

Our group's name is D.S.P. which stands for Digital SnowPrints. The purpose of our project is to track the movement of up to three animals through a snowy field. We will obtain three sets of snowprints from reputable biological sources. Using Matlab's arsenal of builtin functions, we will convert these images into matrices. We will rotate each snowprint in eight different angular orientations. We will then mirror the print to create the complementary paw, hoof, or foot and rotate it in the same eight orientations. All of these images will be stored as patterns for our matched filter. We will use three animals with eight orientations per print. This gives a total of (3 x 8) 24 images.

The various footprint images will be distributed on a white snowy background. They will be arranged in patterns resembling the paths that animals travel in the snow. This will be known as the "detection area." We will write an *.m file that will scan the detection area row by row, looking for snowprints. The detection area will be divided into small regions or windows. For each window, we will run a matched filter to compare its contents with all 24 possible images. If the correlation is low, we will move on. If the correlation is high, we will mark the window as a match for the appropriate animal. The *.m file will draw a dot in the window where the snowprint appears. After the image has been fully scanned, an interpolation algorithm will connect all the points for a given animal and fit a curve to trace the path. These curves will chart the path travelled by each animal.



Possible Innovations:

If there is sufficient time, we intend to research and implement a selective scanning algorithm. It will make an initial pass through the detection area and identify all regions that contain only white space. Any white region will be marked and passed over. Then we will scan the rest of the area and perform the original snowprint matching algorith (described above). This modification will greatly speed up the search because it will minimize the number of times we check a given window against the 24 possible images.

Another avenue of exploration would be to replace the matched filter with a more sophisticated image comparison technique. Numerous method will work as long as they produce accurate correlations between two image matrices. This implementation would require some research into alternatives to the matched filter. Ideally, we would like to find a method that is computationally faster and more accurate than the matched filter from class.