using the cascadeclassifier for this might be a bad idea, as you'd have to train 1 cascade per gesture, and later have all cascades in memory and run them consecutively.
also, ASL contains 2 gestures, that include movement which will further complicate it.
imho, a contours based approach ( using e.g the hausdorff distance for classification ) might do much better.
the 3.0 ( master ) version of opencv has a new shape module, that might fit your needs