I'm not sure it depends on image recognition as in visual classification (although a lot of that work was DoD funded). Once you identify a target shape/pattern/object, you can use motion tracking to track it backwards in digitized video.