A useful tool for agricultural robot developers to have is the ability for the robot to be able to detect crop rows. There have been many methods developed for this purpose, but this repository contains a real-time algorithm that works in a variety of conditions.

Although thas algorithm is far from perfect (it only detects about 50% of the rows, I've put ways to improve below), there are several good things about the algorithm:

1. It took an average of 32.1ms to process each image (max of 70ms) on my laptop (2.5Ghz intel i5, 2Gb ram). This means it can be run in real-time
2. Because of the grayscale transform, which does 2*g-r-b, the green values come out more, and it also has some lighting invarience. This hasn't been tested though.
3. The algorithm iteratively increases the Hough threshold parameter, which means it is more robust to different kinds of crops at different stages
4. Even on curved crop rows, it prioritizes the angle of the rows closer to the camera simply because closer rows have more pixels (perspective). This means that a robot would detect the angle of the rows closer rather than further away.

Should you want to develop this algorithm further, these are the most important aspects I've picked out:

At the moment the algorithm seems to be the most effective at finding precisely three crop rows. Firstly, it should be modified such that any number of rows works roughly evenly, and then some criteria made that means that the number of rows to find does not have to be manually set.

One of the first steps of the skeletonization process is to apply a threshold. In order to get the most effective skeleton, the threshold could be repeated until some criteria achieved that gives the best output for the Hough transform. This would give better outputs for both more sparse crops (image 5 in CRBD) and more dense crops.

Instead of removing faulty rows based purely on angle, they can be removed if it doesn't go to a similar vanishing point to the rest of the rows. Ideally it wouldn't just be a plain threshold if the top point of the line is in the centre - the vanishing point may be to one side if the camera is not looking down the rows or if the rows are curved.

The first step in this and many other algorithms is to find the green values in the image in one way or another. This works for many of the images in the CRBD, but there are three cases that this method has been untested and may not function:

1. The crop is not green
2. The plants are too small
3. The plants are too large and there are no gaps between rows

In these cases, it could be worth investigating other colour spaces for whichever consistently has the most contrast between where the crop rows are and aren't. Such colour spaces that could be compared are YUV and HSV, with the RGB that the proposed method uses. If none of these methods are able to detect rows when there are no gaps, it would also be worth investigating if the robot is able to drive between them at all without damaging the crops. It could be that no detected rows is a beneficial feature.

Many crop rows are not linear due to circular irrigators or various landscapes. In these cases, a simple set of lines may not be enough. There are two ways one could go about this: Either use a Hough transform to find circles treating rows as circles of various radii, or aligning the lines to the bottom part of the crop rows, as that is the important part for a mobile robot.