Computer vision for drones is a set of AI models that recognise what a drone's camera sees, such as a person, a vehicle, a crack in a pipe or a fire. It is used for surveillance and on military unmanned systems. A model is a program that has learned from many labelled pictures. For example, it has been shown thousands of pictures of a truck, each marked "truck".
The model reads each video frame and draws a box around every object it knows. It also tells the kind of object and how sure it is, as a percentage. A person sets a limit, for example 80%, and the model reports only the objects above it. Because a drone is far from the ground, the objects are small. Cameras on drones also shake, and the light changes. So the models are trained on pictures taken from the air.
The model often runs on the drone itself, not on a computer on the ground. This is edge AI, a form of edge computing. It is needed because a radio link may be slow or blocked, and a video is far larger than a short list of results. Imagine a forest department drone in Odisha that flies over a hill at night. A thermal imaging camera shows warm shapes. The model marks 12 of them as elephants and 1 as a person, and sends only that short list to the operator.
The results go to autonomy software, which can change the flight. They also go to video analytics on the ground, which counts and tracks objects. In defence, they help counter-drone systems spot hostile drones and guide a loitering munition. A model can be wrong, for example by mistaking a bird for a drone, so for important decisions a person looks at the picture first.