There's an obvious solution here. Just process the video data and extract whatever possibly descriptive video information and associate it to various objects/ideas in its imaging database and if there's a fairly good match with "fire objects", then it's likely a fire. Couldn't get simpler. ;D