I don't see where it addresses problem of generating large enough number of shapes that are hard enough to recognize with computer vision.

I'm afraid that nice gesture recognition algorithm is not enough to defend against bots programmed to recognize known solutions & replay (slightly randomized) predefined answers.