Segmentation of plants and weeds for a precision crop protection robot using infrared images

Abstract

The work presented is part of a project whose goal is to guide automatically a vehicle along crop rows to apply herbicide or pesticide accurately on weeds or plants. The paper deals with the segmentation of near-infrared images for discriminating plants, weeds, and soil. A restricted problem domain has been chosen, that of transplanted vegetable crops using cauliflower as a model. Furthermore, the natural lighting has been limited to diffuse illumination conditions. The algorithms described have been developed with a real-time implementation in mind. The algorithms use techniques based on global and multithresholding methods, gradient detection, various methods of noise elimination, blob filtering, and mathematical morphology. They can provide information on the size and the position of plants and weed patches as well as information that could be used either for guiding the vehicle or tracking plants and weeds. Successful segmentation has been done on several image sequences. It is shown that it is possible to discriminate plants, weeds, and soil with simple algorithms for different fields of view.