I remember a colleague working on such a project about 7 years ago (he put together a prototype in 2 weeks). The reason the gyro makes a huge difference is that

I remember a colleague working on such a project about 7 years ago (he put together a prototype in 2 weeks). The reason the gyro makes a huge difference is that the image is not produced at the same time, the sensor rows are scanned out gradually, and a gyro allows you to assign a correction to each row of pixels instead only to the whole image at the same time. Doing the same through vision is close to impossible. edit: A gyro-based algorithm is quite simple, while computer vision is still not a fully solved problem in general