Revision history [back]

I would go for some sort of object model detection, making it rotation invariant or applying it in a series of different rotation to cope for the fact that the car will not always be straight horizontal or vertical.

Suggestions:

• Cascade classification using a cascade of boosted classifiers
• Creating a HOG features + SVM approach
• ACF/ICF detectors
• Latent SVM / part based modeling (DPM technique)

All of those are either in OpenCV main or contrib repository. Keep in mind that all of these techniques need training data, and in this case lots of it (since the variance in what is a car exactly).

Also, if the camera is fixed and always on the same position, more simpler techniques could be used, but I did not do that not to restrain the problem to much.