This is basically what I've spent the last two years working on - my postdoc has been building an open-source drone payload that does real-time animal detection in visible/thermal images.
Drones aren't practical for total counts. These tend to be e.g. an entire safari park. You need multiple planes flying for several hours each day. Huge expense - hundreds of thousands easily. What's more common is to do transect counts where you fly in a sort of raster pattern and then do your statistics to estimate the population that you didn't see. Drones are still massively limited by flight time, typically an hour at most. Maybe a bit more if you have a really good fixed wing. See https://www.nationalgeographic.com/news/2016/01/160104-great...
On the other hand drones can be used as an effective anti-poaching deterrent - for example if you have a rough location. They can also be used for small scale surveys in locations where you can't really afford a plane and/or it's awkward or dangerous to send people. Some applications like flying very close to cliff faces to monitor birds are only possible to do safely with drones.
Ecology has a whole host of problems that make object detection difficult. You have extreme class imbalance when you capture data - 90%+ of images are empty. Animals are frequently camouflaged. Many animals (like the various gazelles in Africa) look almost identical, especially in low resolution aerial imagery. A lot of animals of interest are small. Even for camera traps, where you have very controlled conditions, the standard approach these days is to have a generic "animal detector" and then have humans do the long tail classification.
On the Africa question though - not a large percentage at all. By the way megafauna covers a lot more than it suggests (40kg+, so that includes deer). Most smaller conservancies have a good idea of how many "rare" species they have, because they're extremely conscious of poaching. But water buffalo? Thompsons gazelle? Who knows exactly. And once you go out of controlled areas into the bush proper, it's anyone's guess. There are charities which organise surveys, but it costs so much to do that it's not feasible to do it annually. See that Nat Geo link - that was the first time anyone had surveyed elephants in Angola. In 2016.
But computer vision is absolutely the future. People are coming round to the idea that we need to start capturing full survey imagery during counts. And while those images are still labelled by people today, once we have that data stored we can always run new models over it in the future.