Keep in mind that cascade classifiers are not rotation invariant! You should stick to one orientation for your model and then keep rotating the original image until you hit a detection. Also keep in mind that a up facing car, isn't the same as a down facing car. Ratios of gradient regions are completely different!
More positives will definitely be neccesary! Car models are trained using multiple thousand samples. What you did is training a car detector which has to be identical or almost to the cars trained ... to loose those dependencies more data is needed.
Don't use the distortions of createsamples, instead just work with the second option, which is a file containing all detection images and the region of interest. Distortions only create artificial data, which will not be present in real life situations...
