Any kind of iterative procedure, in particular mixture model fitting, is quite time-consuming, and thus unlikely to be fast enough for real-time processing, unless you have only a handful of kernels to fit (we fit hundreds to a few thousands of spots, which takes several seconds per frame, despite our fast mixture-model implementation).
Why don't you use a non-iterative spot detector, such as H-Dome? Or do you need sub-pixel positions of your features?