Prof. Lakaemper,

I found that simply passing the image through a median filter
worked very well on the first movie, almost completely
eliminating the noise.

I tried a lot of different things on the second movie, with
varying amounts of success. At first, I had tried opening the
image to eliminate the lines on the roads and any noise caused
by the weather, then doing histogram equalization to increase
the contrast of each frame (in order to separate the cars from
the road), then performing the analysis we did before. That
didn't work well, probably because the jitter from the camera
was ruining the mean.

I tried creating an image stabilizer by recording the position
of the first nonwhite pixel (since there is a white border
around the image that moves only with the camera), but snowy
cars occasionally touched that border, causing an inaccurate
translation. This may be easier to do in the frequency domain,
but we haven't really discussed 2D image manipulation in the
frequency domain yet and I wasn't quite sure how to proceed.
The literature suggested using wavelets.

In the end, I chose to scale the intensity values from .5 to
.7 to the full range and clip the others using imadjust. This
gave the most promising results by far, as it made everything
but the cars, including the road, completely white. Both
simple thresholding (output < .95) and the analysis from the
previous assignment worked well in detecting the cars.

If we wanted to detect the parked cars as well, it seems
possible to use the results of the initial analysis as a
structuring element for a hit-and-miss transformation.

Using imadjust on the first movie actually harmed accuracy
quite a bit, so removing the imadjust from the doFiltering
file is necessary when analyzing that movie.

Thanks,
Michael