Clustering Algorithm Based Straight and Curved Crop Row Detection Using Color Based Segmentation

Crop row detection is the first step towards automatic navigation of agricultural robots and a critical step for precision agriculture. Based on crop and weed growth, four scenarios could arise during crop row detection. First, crop rows are clearly visible and no weed growth in between rows. Second, crop rows are clearly visible and low weed growth in between rows. Third, crop rows are clearly visible and high weed growth in between rows. Fourth, crop missing within crop rows and high weed growth in between rows. The proposed crop row detection algorithm has to work reasonably well for all four scenarios. The algorithm has to be robust against different field conditions (different crop growth stage, different crop inter-row spacing, different types of crops etc.) so that it can be deployed to any type of crop field. Moreover, the algorithm has to be robust against different lighting and weather conditions. The primary objective of this algorithm is precise autonomous guidance, so low processing time is an important factor. To cope with all the above mentioned scenarios, an algorithm mainly consists of three linked stages is proposed: (1) color based segmentation for differentiating crop and weed from background, (2) differentiating crop and weed pixels using clustering algorithm and (3) least square straight and curved line fitting over crop pixels. First a region of interest (ROI) is selected which contains 3 or more crop rows. All the row detection steps are applied within this ROI. Selecting this ROI has two advantages. First, green pixel near horizon are congested and hard to separate. As a result, they increase false detection. Selecting this ROI eliminates the need to process those green pixels. Secondly, this ROI is almost one-fourth the size of the original image. So only operating in this section reduces the computational cost and processing time. The proposed color-based segmentation is robust against lighting and noise. Performance of a wide variety of clustering algorithm is tested based processing time and crop vs. weed pixels clustering. Statistical methods and domain knowledge are used to differentiate between correctly and incorrectly fitted lines over crop rows. The proposed algorithm is tested over a wide variety of scenarios and performance is compared against existing strategy of similar kind. Most of the existing methods are good at detecting straight rows (not curved) and applicable in real-time. Some of the existing methods are good at detecting straight and curved rows but not applicable for real-time application. This work will discuss the effects of clustering algorithms on separating weeds from rows. We will also discuss how each step of this algorithm will affect processing time. Limitations of these type of algorithms are mentioned and possible solutions are also discussed.