Plant height

But the challenge taken up by the researchers at Gembloux is of a completely different nature: what is the solution when the crops and weeds are mixed? Obviously, in this case the most important stage is that of recognition: how will the robot be capable of distinguishing between the plants it must destroy and those that it must not touch?

The first idea was the most natural as it was the most “human”: differentiating plants by their color. However, this soon proved to be a failure because it is not a sufficiently discriminating characteristic, at least not with the current technology, non-uniform lighting both in terms of intensity and light wavelength interfered with recognition. A second equally natural possibility was to differentiate the plants by their shape and form.  Yet here too, we encountered problems: the variability of the foliage even within the same species and the fact that the leaves often overlap makes it impossible for the recognition systems to distinguish the leaves correctly. Professor Destain explains, “We then took a simple idea as our basis. In principle, the healthy plants grow at the same speed as they were planted on the same day. The weeds have differing rates of growth. Therefore, on a given day, they will be of different sizes!” This is indeed a simple idea yet it is difficult to achieve. This is because continuously measuring the height of plants in real time by a system poses some problems.

In order to succeed, Marie-France Destain’s team decided to use stereoscopic vision. The best-known and most simple system is passive stereoscopy. In this case, two cameras are used which must focus on the same point; the system calculates height from these two points, but given the wide variability of cases this was seen to be insufficient. Therefore, explains Marie France Destain, “we used an active stereoscopy technique: a video projector projects a series of coded luminous black and white fringes similar to a barcode onto the plant.  The scene read by the cameras allows distortion lines and surface information to appear. Evidently we were not satisfied with only one frame, it is necessary to take a large number; each image is therefore made up of several intermediate images. The decoding principle is based on the correlation between signals emitted and signals received.”