Crop row recognition using computer vision

H. T. Søgaard

Danish Institute of Agricultural Sciences, Department of Agricultural engineering,

Research Centre Bygholm, P.O. Box 536, DK-8700 Horsens, Denmark

HenningT.Sogaard@agrsci.dk

Introduction

In an ongoing project at the Danish Institute of Agricultural Sciences (DJF), cereals are drilled with an interrow spacing of 0.24 m, which is twice the spacing conventionally used in Scandinavia. With an interrow spacing of 0.24 m it is possible to guide an implement, e.g. a fertilising machine or a row weeder, along the rows with an accuracy of about ±10 mm. In principle, a human operator can perform the guidance job, but this would be a very demanding task in the long run. For this reason a project has been started with the aim of automating the guidance task by means of a computer vision controlled system.

Scope

The goal of the project is to develop a system, which is able to guide an agricultural implement with an accuracy of about one centimetre in row crops. The system should be able to work under natural light conditions and normal forward speeds.

Material and method

An experimental field of approx. 100 m in length was sown with barley (interrow spacing of 0.24 m). A colour video camera (three CCD) connected to a computer with a Targa 2000 PCI frame grabber was used to record images in the field. The camera and electronics were mounted on a light hand-operated carrier (Figure 1). Prior to the recording in the field, the camera was calibrated with respect to its perspective parameters and its white balance. Subsequently images from multiple tracks of each approx. 70 m were recorded and saved on CD-ROM. These recordings were performed under different light conditions and growing stages of the crop.

The database of images obtained in that way has been used as a basis for developing a method for determining the crop row position in colour images. The image processing method, which has thus been developed does not include a segmentation step, which is the case for most other methods for plant detection reported in the literature. The segmentation step has been replaced by computation of centres of gravity for row segments in each image. This approach has proven to reduce the computational burden of the image processing software. The estimation of the orientation and the lateral offset of the centre lines of the rows is accomplished by weighted linear regression. An important feature of this method is that the accuracy of the estimated lateral offset can be determined and used as a reliability measure in the control system.

Results and discussion

Preliminary results indicate that the above-described method is able to determine the row position with accuracy about ±1.5 centimetre. As the camera must look a little ahead of the implement to have a free field of view it will be necessary to extend the estimated centre line of the guiding row backwards to the operating area of the implement. This extrapolation step introduces additional uncertainty and suggests that the camera should be positioned in such a way that the extrapolation distance is reduced as much as possible.

To obtain a precise guidance of the implement the speed of the image processing is crucial. The method developed in this project has been programmed in C++ and tested on a 143 MHz UltraSPARC (» 300 MHz Pentium). The tests show that the processing speed is about 10 frames per second which is sufficient for precise implement guidance.

Future work

The ideas developed and tested during this project have been utilised by the Danish company ECO-DAN in their row guidance system, which has been available on the market since spring 2000.

The research interest in DJF now move towards robotic weed identification. Recently a new Danish project involving DJF, Aalborg University and The Royal Veterinary and Agricultural University has been started. The aim is to construct a small autonomous vehicle, which is able to map the occurrence of selected weed species and the relative crop cover of agricultural fields. The identification of weed species should be carried out by machine vision, but it is still an open question which particular features should be focused on to accomplish this task.