I have the following setup for a prediction task: I want to predict entire pictures from previously given pictures. In my case, only 2 pixels in every frame are neither black nor white, they are some moving objects whose movement I want to predict. The 2 pixels are the centers of some square regions of, say, 10m length/ width. One might be green and the other one might be blue. There are socalled no-go-areas where none of both objects can go, and they are depicted by black pixels, whereas every pixel apart from the 2 coloured and the black pixels are areas where the objects can possibly move to and they are depicted by white pixels.
Now my questions: Is it possible to use this as a prediction setup, i.e. use LSTMs and/ or CNNs to predict the future "image"? The image would stay largely the same, because the two coloured pixels would be the only ones moving, the black or white ones remain in the same spot. Can a CNN/ LSTM combination learn that the white areas are accessible whereas the black ones are not, given enough sequences of images, and can it learn the rules by which the coloured pixels move?