Classify objects in real-time without needing to train the classifier

edit

Hi everyone, I'm just brainstorming here, but is it possible to classify objects (i.e. give each object on screen a unique id and consistently associate that object with that id) without needing to setup training data. It's not desirable to use a training set as this "labeling" needs to be done on screen in real-time. The objects in question are similar sized, similar color, and move all throughout the image. I've tried using the nearest neighbor approach to assigning them numbers and I may take into account their sizes now to help differentiate the objects from one another. Any suggestions on what techniques I could try?