Agricultural robots: individual plant recognition.

Tilneac, Mihaela ; Dolga, Valer

1. INTRODUCTION

In this chapter, we give an overview of the state of the art in the field of vision guidance and plant recognition. Many technical solutions used in the field of agricultural robots are not efficient enough. Some of the frequently encountered problems in the field of agricultural robots, are: problems in distinguishing crop from weed (Oberndorfer, 2006); problems caused by variations of environmental conditions (illumination, wind, fog, occluded appearance of the plants, damages caused by insects) (Andersen, 2002); problems caused by the change of seasons; recognition problems caused by intra-row gaps; recognition problems caused by uneven rows; recognition problems by using thermal images, when fruit and neighboring elements have the same temperature.

Plant recognition is one of the most difficult problems that the horticultural robots industry confronts. This is caused by the fact that there is no available system, with enough high computational power to distinguish between crops and weeds (Oberndorfer, 2006). Generally, the problem is to find a set features that may be used in a classification scheme (Andersen, 2002).

The main plant recognition methods are : gaussian distribution models; near infrared images; texture ; shape features; spectral properties of reflection; spatial frequency based properties; dichromatic reflection model; other wavebands; RGB (red, green, blue) colour model; HSV(hue, saturation, value) colour space; YCrCb colour space (Klose et al, 2005); Grey Level Co-occurrence Matrix (GLCM); Fast Fourier Transform (FFT); Scale Invariant Feature Transform (SIFT); binarization; EGRBI transform (Excessive Green, RedBlue, Intensity) (Maksimow et al, 2007); Minimum Distance Function (MDF); clustering; seed mapping.

An efficient individual plant discrimination can be achieved only by simultaneous use of several plant recognition methods. One of the study's objectives is to find sets of methods that, when combined together, have high efficiency. Electronics and software are basic elements in the modern agriculture. Multi-sensor fusion systems are used for individual plant recognition.

2. PLANT RECOGNITION EXPERIMENT

Plant recognition is one of the most difficult problems. Changes in light intensity cause major difficulties in this issue. To solve this problem, we started to develop a plant recognition method that can be efficient in variable lighting conditions. The method will be implemented to an autonomous mobile robot for weed control. Each plant species has a specific green hue and a specific texture, which is different from other plants. Individual plant identification can be achieved by digital image processing, using the RGB values.

The RGB color model is an additive color model in which red, green, and blue light are added together in various ways to reproduce a broad array of colors. The name of the model comes from the initials of the three additive primary colors; red, green and blue. This color space has the disadvantage that the RGB components values are strongly influenced by the variation of light intensity. To solve this problem we proposed the correlation between the RGB values and the photodiode voltage values. Leaves belonging to different species of plants, were analyzed using the experimental stand, represented in Fig. 1. We analyzed the RGB components values, taking into account the following parameters: the photodiode voltage, the distance between photo camera and plant, the distance between light source and plant, the angle between photo camera and leaf, the distance between photodiode and light source. The photos taken with the photo camera were downloaded on the computer in JPEG file format. The JPEG files were imported into MATLAB environment. After this, we extracted the RGB values. The data were processed using Microsoft Excel. In Fig. 2 is represented the electrical schema, where i denotes the alternative current intensity, I denotes the light intensity, and u denotes the photodiode voltage. Using a photodiode, the light intensity is converted into voltage. The voltage is measured with a voltmeter. Light intensity is directly proportional to current intensity. Voltage is directly proportional to the light intensity. The light source used in the experiment is a frosted 100W light bulb. The light intensity can be adjusted using a dimmer switch. A photodiode and a voltmeter are used to measure the light intensity.

[FIGURE 1 OMITTED]

[FIGURE 2 OMITTED]

In Tab. 1 we can observe some of the results obtained from the analysis of tomato, pepper and aubergine leaves. The results were obtained with the following parameters: the distance between photo camera and plant is: horizontal 0 mm, vertical 265 mm; the distance between light source and plant is: horizontal 130 mm, vertical 140 mm; the distance between light source and photodiode is: horizontal 130 mm, vertical 140 mm; the angle between the camera and leaf is 0 grade. The experiment has been performed in the laboratory during the night. Each R,G,B value from Tab. 1 is the arithmetic mean of ten distinct values taken from the same leaf. The graphs in Fig.3 correspond to the values from Tab. 1. The graphs contain three curves. Each curve corresponds to the three species: tomato, pepper and aubergine, respectively. On the same principle we have processed over 100 photos taken at different light intensities, different distances, and different angles. A database of the experimental results was created.

[FIGURE 3 OMITTED]

3. FUTURE RESEARCH

The experiment was done in the laboratory. Next experiments will be done outdoors. Individual plant discrimination will be done using fuzzy logic. Further research will be done by studying of RGB values of a large number of plants belonging to the same species. After multiple measurements, a statistical data processing will be done. The aim is to find out which RGB values space corresponds to which plant species. Also, the influence on RGB values, caused by seasons, growth stages and humidity must be studied. It happens that in the same image, different plant species have a very similar hue. This create problems in plant discrimination. This problem could be solved by usage of optimal parameters. This requires the finding of optimal light intensities, optimal distances, and optimal angles. Each plant reflect the light differently. The reflection model is described in the following equations:

C = [F.sub.b] + [F.sub.s] (1)

[F.sub.b] = [m.sub.b](n,s) [[integral].sub.[gamma]] [f.sub.c] ([gamma])e([gamma])[c.sub.b]([gamma])d[gamma] (2)

[F.sub.s] = [m.sub.s](n,s,v)[[integral].sub.[gamma]] [f.sub.c]([gamma])e([gamma])[c.sub.s]([gamma])d[gamma] (3)

C denotes one of the R, G, B components; [[florin].sub.c] ([gamma]) denotes the spectral sensitivities of R, G, B respectively; e([gamma]) denotes the incident light; [c.sub.b]([gamma]) denote the albedo reflectance; [c.sub.s]([gamma]) denotes the Fresnel reflectance; y denotes the wavelength, n denotes the surface patch normal; s denotes the direction of the illumination source; v denotes the direction of the viewer; [m.sub.b] denotes the geometric dependence on body reflection; [m.sub.s] denotes the geometric dependence on surface reflection (Gevers & Smeulders, 1997).

4. CONCLUSIONS

The graphs (Fig.3) corresponding to tomato, pepper and aubergine respectively, are distinct. It means that individual plant discrimination can be achieved by correlating the RGB values with the light intensities values (in this case, the photodiode voltage).

5. REFERENCES

Andersen H. J. ( 2002). Outdoor Computer Vision and Weed Control, Ph.D. Dissertation, Aalborg University, Denmark, Available from: http://www.cvmt.dk/~hja/publications /all.pdf Accessed: 2008-12-12

Gevers T. & Smeulders A.W.M. (1997). Color Based Object Recognition, Available from: http://www.springerlink.com /content/r90643413m4jx652/ Accessed: 2009-06-12

Klose R.; Meier M.; Linz A. & Ruckelshausen A. (2005). Field robot "optoMAIZER" : Development of a mechatronic system based on sensor fusion, a real time operating system and WLAN, Proceedings of the 3rd Field Robot Event 2005, Wageningen, June 16& 17, 2005, pp.51-65, Available from: http://www.fieldrobot.nl/ Accessed: 2008-12-12

Maksimow T.; Holtta J.; Junkkala J.; Koskela P.; Lamsa E.J.; Posio M.; Oksanen T. & Tiusanen J. (2007). Wheels of Corntune, Proceedings of the 5th Field Robot Event 2007, Wageningen, June 14, 15 & 16 2007, pp.75-87, Available from: http://www.fieldrobot.nl/ Accessed: 2008-12-12

Oberndorfer T. ( 2006). Embedded vision system for intra-row Weeding, Master's Thesis in Computer System Engineering, School of Information Science, Computer and Electrical Engineering Halmstad University, Available from: http://dspace.hh.se/dspace/bitstream/2082/533/1/ 0623%20TO.pdf Accessed: 2008-12-12

Tab. 1. Correlation table between RGB values and photodiode
voltages

photodiode tomato pepper aubergine
 voltage
 [mV] R G B R G B R G B

272 52 63 87 35 50 67 75 91 102
355 42 58 64 28 45 46 64 83 79
386 75 94 84 34 51 50 75 94 84
401 43 60 55 31 48 41 72 93 72
410 86 57 20 69 47 13 129 87 31
417 80 53 19 62 43 14 108 78 26
418 76 51 19 59 41 12 116 79 28