Real-time weed/crop discrimination through fast direct image registration

This paper presents a computer vision system that is able to discriminate between weed patches and crop rows in real-time, from videos taken directly from a tractor moving through the field. Weed/crop discrimination is highly simplified thanks to video stabilization. We present a simple but effective variant of the inverse compositional algorithm for image alignment, and show that on our videos, our optimized version of the algorithm performs just as well as key-point matching methods, while being up to 2x faster. Once the video stabilized, crop rows remain almost constant through short periods of time, and be detected by a simple image processing. We tested our approach on several videos, taken in different maize fields on different dates, and presenting a variety of weed/crop conditions. Our final approach achieves a mean recognition of 84% on weeds and 91% on crop pixels, improving on our previous work 9% and 29% respectively.