1 Goal and Approach
 

The goal of part II of the project is to develop a vision-guided intra-row cultivator, able to perform mechanical weed control within the row of sugar beets plants, thus totally eliminating the need for chemical weed control.  We decided to develop an autonomous mobile robot of about 120 cm long and 70 cm wide as shownabove. The robot should be able to follow a row of plants guided by the row-following vision system which was developed in part I of the project. Another vision-system will then be utilized to recognize the individual sugar beet plants among weed. This vision system controls a weed-removing device mounted on the robot. The task of the weed-removing device is thus is to remove or destroy the weed. Eventually the robot should also be able to recognize the end of the rows, turn, and to find the next row autonomously.

2 Autonomous Mobile Robot

During 1998 we designed, build and successfully tested an autonomous mobile robot for agriculture operations. The robot is driven by its two wheels at the back, each independently driven with a DC-servo motor equipped with encoder, tachometer and brakes. The steering mechanism is an Ackermann-steering controlled with a DC-servo-motor. The electrical power is provided by batteries or by a fuel-driven generator. Two students projects were performed in this area [Becker a] and [Becker b]. The row-following system developed in Part I of the project has been adapted and implemented on the robot. In November 1998 we were able to successfully follow a row of plants at a speed of 0.2 m/s within + 2 cm.

In 1999 we mounted a color camera on the robot to provide the images for the plant recognition system and acquire a considerable amount of images to be able to develop and test the vision algorithms for plant recognition. Secondly, we plan to develop an accurate position estimation system, which allows the robot to exactly estimate the distance traveled along the row. This information will be very valuable for our plant recognition system based on contextual information, as described below. The development of the position estimation system is planned to be a master project in computer systems engineering.

3 Vision Algorithms for Plant Recognition

In 1998 we have investigated two approaches to recognize sugar beet plants among weed. The first one is a more traditional approach, where the vision systems analyses one plant at the time and decides whether this plant is a sugar beet plant or weed. Fig 1 shows an example of how the input image looks like. We collected color images from different fields: a total of 214 of sugar beet plants and 373 of weeds. The pictures were taken with a normal color photo-camera. Different classifier such as Bayes Quadratic, Neural Networks and k-Nearest Neighbor were tested. We selected 19 features, such as color, shape and size, of which all or a subset were evaluated. With three features we obtained a classification rate of 96 %. A master project has been performed in this area carried out by three students [Bondesson]. The succesrate of 96 % looks satisfying at first sight. However, the process of extraction of individual plants out of a scene has been done manually. In a final system, this should also be done by the vision system. This will reduce the succesrate with about 10 to 15 % according to preliminary results.

The second, novel approach, tries to classify a plant using contextual information. Instead of looking at one plant at the time, the systems looks at a certain environment containing several plants. Knowing that the sugar beet plants are sown in rows and with a certain constant distance among them, it is possible to classify the plants based on this information without looking at individual features of a plant as described above.

Fig. 2 Input for the approach based on contextual information. The plants marked with a
white circle are classified as sugar beet plants, as they fit best to the context:
sugarbeet plants lie on a row spaced apart with a certain constant distance.

This is illustrated in fig 2. We obtained a classification rate of 92 % with this method. A master project has been performed in this area carried out by two students [Fredriksson].

In 1999 we will combine these two approaches into one classifier, refine them and optimize them with respect to processing time. We will also acquire images taken from the mobile robot, while the mobile robot is following a row. We also have to verify our results based on more images. Two student projects are planned in 1999 which have the goal to speed-up the algorithms by using the MMX-utilities of today’s Pentium-processors. The approach based on contextual information seems very promising and will certainly lead to new scientific result

List of Student Degree Projects 1998

[Becker a] Bernd Becker, "Development of the Process Hardware and the Device Drivers for an Agricultural Robot", Final Project in Technical Computer Science, Spring 1998, Halmstad University.
Exchange student from Fachhochschule Ulm, Germany.

[Becker b] Bernd Becker, "Development of the Real-Time Software for an Agricultural Robot", Diploma Project in Technical Computer Science, Autumn 1998, Halmstad University. Exchange student from Fachhochschule Ulm, Germany.

[Bondesson] Fredrik Bondesson, Musie Minas and Ulf Winberg, "Classification of Sugarbeet Plants among Weeds Using Machine Vision", Master Project Electrical Engineering, Halmstad University, 1998.

[Fredriksson] Johan Fredriksson and Tony Turujlija, "Classification of Sugarbeet Plants Based on Contextual Information", Master Project in Computer Systems Engineering, Halmstad University, 1998.