PDF version of this entire document
The experiments were not quite systematic or comparative, as such experiments would require the code to mature and they would also consume a lot of time. Over time I created (i.e. trained) about 30 different classifiers, then tested them by observation and subsequently tweaked classifiers in according with all prior observations. Improvements were rapid but gradual. I took a dozen videos for comparative analysis. The aim has been to find the right sensitivity levels (minimum hit rate, maximum false alarm rate, etc.), sample sizes (24x24 pixels would work well as an abstraction of patterns seen in car rears), and sizes for both sets (it would require a lot of manual work to expand, or in other words a time investment).
Training with large data sets is an essential stage because without good detection of moving obstructions (mostly cars) the alerting would be spurious and therefore counter-productive. The framerate attained is still above 5 frames per second, which is actually very decent for this hardware . Idle, standalone video capture remains is only at about 10 FPS on this inexpensive MIPS/RISC board. It's mediocre equipment on a relative scale.
