Sounds like it's a pretty simple loop of 1. Find bright spots in last second of exposure (which is then added to a sum for the final picture) 2. Compare their

Sounds like it's a pretty simple loop of 1. Find bright spots in last second of exposure (which is then added to a sum for the final picture) 2. Compare their locations to bright spots in the previous second 3. Move the sensor slightly so the bright spots remain "in the same place" on the sensor. Note that cameras with integral anti-vibration do that by moving the sensor, so they already have the hardware for (3). For star tracking there would be a max speed of stars based on Earth's rotation multiplied by focal length you could use to set a reasonable cap on movement. You could also expect similar movement for each second to the seconds before. The camera could have a data structure per star being tracked, capping the number used to say 32 or 64 "stars" for accuracy. Each "Star" would "vote" how much movement has occurred. You'd gain faith in various stars as analysis frames go by if they keep giving valid results, a black-list of "not stars" would also be useful. Some would be stationary, like lights on the ground. Some would be bogus, such as navigation lights on aircraft. Actual stars should expect similar movement and "agree" though if the zoom was low it might change across the image.