The first problem came with the camera, since in order to control it from Android in a RPi3, only camera2 API can be used. This API is so complex and very different to the deprecated (but still present in mobile devices) camera(1) API.
Second, Android Things is in a 'developer preview' state. This means that it 'works' but there are known bugs, performance issues, and things that simply can't yet be done. There is not much information apart from the examples. But I like it, mostly because it can be used with B4A.
Once I had a minimum basic example for the camera2 API debugged and working in my mobile phone, I tested it in my RPI: nothing worked!!! By debugging, I found the first limitation, it is that TextureViews (needed for the preview) can't be used because of an issue with hardware acceleration (I suppose it is a limitation of the current preview state of Android Things).
Also, the camera2 API is not only complex but there are (again because of the Android Things preview state) some limitations on its use: the number of output surfaces, some lags, ...
Once I have been able to overcome these limitations and had access to the preview data, I just mixed it with my OpenCVforB4A library and,,.surprise!! it worked like a charm!!!