The ability of AI to understand video/images seems to be largely underexplored and underexploited. There are a lot of economically valuable applications to having an AI watch the world in real time, even with errors & limitations, and I have seen few products or papers on it.