Using OpenCV cascade of classifier with traffic si

I'm trying to detect some traffic signs like that: http://i.stack.imgur.com/yOTZb.jpg

I managed to isolate the traffic sign (which is, for the moment, always round): http://image.noelshack.com/fichiers/2015/06/1422869734-3.jpg

With bad results with SURF for real-time, and tips of a preceding post, I want to use a cascade of a classifier for the detection of the different traffic sign (30,50,70...).

1) Is traffic signs will be nicely recognized to cascade of classifier like faces are? I mean by that, is a traffic sign is susceptible to haar features detection ?

2) Do I need to train one cascade of classifier (one .xml) for each sign? Do I need to put images of 30,70 traffic signs in the false positive of the 50 ( and conversely) ?

3) If I have understood, the positive images have to have the same size, and i need to delete the background ?

Sorry for my english, thanks :)

I'm of the opinion that you need to train some Haar detector per traffic sign's shape (one for triangular warning signs, an other one for circular signs, etc.). As a result of detection you will have some candidates for further processing and should be decided whether a candidate is true positive or not.
If it is true positive: additional classification needs to recognize the type of a known shape. This classification can be an ANN algorithm or SVMs.

Answers to your questions:

It's strongly depends on the positive/negative database, the features used for training (Haar, LBP, HoG), but I think this cascade structure can be useful for your purposes.
Partly answered above. For the negatives: you should use a very different set of images. E.g. landscapes, animals, etc. It's important to collect a large database because most of the negatives will be rejected during the first steps of the training.
You need to use same scale (for positives) during the training and it is recommended to use some global transformation for reducing the effect of different lighting conditions. But you don't need to remove the background, just crop the images along the border of signs.