Terrain Classification for Track-driven Agricultural Robots

(1)

Procedia Technology 15 ( 2014 ) 775 – 782

Peer-review under responsibility of the Organizing Committee of SysInt 2014. doi: 10.1016/j.protcy.2014.09.050

ScienceDirect

2nd International Conference on System-Integrated Intelligence: Challenges for Product and

Production Engineering

Terrain classification for track-driven agricultural robots

Khairul Azmi Mahadhir

a

, Shing Chiang Tan

b

, Cheng Yee Low

a,c

, Roman Dumitrescu

d

,

Adam Tan Mohd Amin

a

, Ahmed Jaffar

a,c

a_{Humanoid Robot and Bio-Sensing Center, Faculty of Mechanical Engineering, Universiti Teknologi MARA, 40450 Shah Alam,Malaysia} b_{Faculty of Information Science and Technology, Multimedia University, Malaysia}

c_{Brain and Neuroscience Communities of Research, Universiti Teknologi MARA, 40450 Shah Alam, Malaysia} d_{Fraunhofer Project Group Mechatronic Systems Design, Zukunftsmeile 1, 33102 Paderborn,Germany}

Abstract

A long-term goal of agricultural automation is to deploy intelligentrobots to facilitate labor-intensive tasks such as crop care or selective harvesting with minimum human supervision. To achieve this goal, the agricultural robots must be able to adapt themselves in response to various terrain conditions.The reason is that the terrain characteristics can jeopardize the performance of a robotin carrying out a taskor even causing it being trapped in the field.The aim of this work is to evaluate the effectiveness of using an intelligent algorithm, i.e. support vector machine (SVM) in recognizing various terrain conditions in an agricultural field. For this purpose, asmall tracked-driven mobile robot together witha terrain test bed has been developed. The terrain test bed emulates three types of terrain conditions, i.e. sand, gravel and vegetation.The tracked-driven robot is embedded with a low power MEMS accelerometer for measuring vibration signals resulted from the track-terrain interaction.An experimental study was conducted usinga SVMtrained with three different kernel functions, i.e. linear function, polynomial function and radial basis function (RBF). The results showed that the SVM can recognize different terrain conditions effectively. This work contributes to devising a self-adaptive agricultural robot in coping with changing terrain conditions.

Peer-review under responsibility of the Organizing Committee of SysInt 2014.

Keywords:Support vector machine; track-driven robot;agricultural automation; intelligent technical systems

_{* Corresponding author. Tel.:+603-5543-6276; fax: +603-5543-5160.} E-mail address: [email protected]

(2)

1. Introduction

The traditional agriculture industry is labor intensive. In the last decades, a proliferation number of robotic systems have been developed to assist human workers in agricultural activities, for instance,robot-assisted methods forfertilization, spraying,fruit harvesting and transferring process [1-3]. Recent advances in software have allowed the robots to possess the ability to adapt to their environment [4-5]by learning from the data about the surrounding.One of the approaches is the deployment of machine-learning techniques [6].

In agricultural automation, robots can be equipped with a computer vision system to perform visual navigation [7]. For example, a low-cost robot is equipped with a vision control system to provide a visual navigation for fertilization and spraying artificial pollination [8] in a greenhouse environment. Computer vision systems are also installed on a robot to guide it to travel between the crop rows [9] and to perform automatic recognition on the fruit conditions before harvest [10]or for fruit grading [11]. On the other hand, there is also research on fusing the agricultural robots with machine-learning techniques [12-13]. For example, a harvesting robot [14] is installed with a statistical machine-learning method to recognize the maturity of apples. A computer vision system is integrated with artificial neural networks to perform leave image classification for sun flower crops of which the application can ease weed control [15-16].

In an agricultural field, the terrain conditiona robot is traversing on affects the performance of the robot in carrying out a task.Gravel,for instance, produces high vibration to robots traversing on such surface. In this work, a machine-learning technique based on support vector machine (SVM) is proposed as a learning algorithm to distinguish different terrain conditions in an agricultural field. To evaluate the effectiveness of the algorithm, atrack-driven mobile robot is embedded with a MEMS accelerometer used to measurevibration data which is then analyzed and classified using SVM. Having known the terrain condition, the control of the motor drive can be adapted to produce the thrust required for the mobility of the robot when traversing on changing terrain conditions in the field.

This paper is organized in the following manners: Section 2describes the system architecture of the track-driven agricultural robot; Section 3explains the theoretical background of the support vector machine used for terrain classification; Section 4 presents the experimental results of terrainclassification; and Section 5 concludes the paper.

2. System architecture of a track-driven agricultural robot

The aim of this work is to evaluate the effectiveness of using support vector machine (SVM) in recognizing terrain conditions in an agricultural field,i.e., sand, gravel and soil. For this purpose, a small tracked-driven mobile robot together with a terrain test bed has been developed. The key system elements of the agricultural robot developed for this work is shown in the active structure [17] at the upper part of Figure 1.The behavior–state diagram is shown at the lower part of Figure 1.

As shown in the active structure in Figure 1, the track-driven robot is driven by DC motors, and the robot maneuvers by means of a differential drive system [18]. A small and low power MEMS accelerometer (ADXL335 from Spark Fun Electronics) is integrated into the robot for measuring the vibration produced by the track-terrain interaction. Based on the acquired vibration signal, the SVM performs classification and determines a terrain type,

i.e. sand, soil or gravel. According to the terrain type, a Finite State Machine (FSM) [19] is used to adapt the system

behaviour by triggering a terrain specific system state. The controller then sends command signals to power electronics to adjust the input current and voltage of the motor drive.

As shown in the behavior–state diagram in Figure 1, if a system state for soil is triggered, the DC motors will be controlled produce medium speed and torqueto propel the robot in a normal condition. If a system state for gravel is triggered, the DC motors will be adapted to produce high torque at low speed. If a system state for sand is triggered, the DC motors will be adapted to produce higher toque at lower speed. By recognizing the terrain condition and adapting the system behavior accordingly, the agricultural robot is able to traverse across changing terrain conditions without being trapped in the field.

(3)

Fig.1. Active structure (upper part) and behavior–state diagram (lower part) for a track-driven agricultural robot.

Fig. 2 shows a sample of data collected from three different types of terrain within a constant period of time in a controlled laboratory environment. Before the classification of the terrain condition can be performed, the data from the accelerometer is filtered to remove noise. The resolution used during data collection is around 100Hz. The software components are developed in MATLAB and C programming language.

(4)

Fig. 2.Example of vibration signal measured on different types of terrain, i.e. sand, gravel and soil.

3. Support Vector Machine for terrain classification

SVM is a computing method based on statistical learning and optimization theories [20]. It is chosen for terrain classification because of its robustness in representing the information at the boundary class [21]. During the training process of SVM, it finds a set of hyperplanes to maximize the margin among themselves and the nearest data samples of arbitrary classes so that these hyperplanes are separable for data classification. An example of a linear SVM is illustrated in Figure3. SVM is initially designed to handle data of two classes where they are separated by

(1)

wherex is the data sample, w is the weight vector, and bis bias for constant offsets.

wtx + b = 0

Sand

Gravel

Soil

(5)

Fig. 3. Separating hyperplane in the SVM between two data sets.

In many circumstances, a real-world data is complex. A linear SVM system may be not effective to separate this complex data that are non-linear. A way is to introduce a soft margin approach to handle non-linear problems. Another way to overcome this limitation of the SVM model is to include a linear kernel trick to make non-linear transformation of the data space to improve its recognition ability. In this case, the kernel tricks such as radial basis function, polynomial function and etc [22]can provide mapping from linear to non-linear classification.

SVM adopts two strategies to classify the data samples of multi-classes, i.e., either One-versus-One (OVO) or One-versus-All (OVA). The OVO strategy is firstly introduced in SVM [23] and it is also known as pairwise coupling or round robin. It is actually a basic form of binary classification. Let say n data pairs

n

m

y

D

x

{

_m

,

_m

},

1 ,...,

are available for training, where

x

_m

p is a feature vector indicating the m sample, and

y

_m

{

1 ,

2 ,...

K

}

is the class label ofx . The SVM model that implements OVO will consist of _m

2 /

)

1 (

K

binary SVMs. On the other hand, the OVA strategy is applied to build K SVMs where the i-th SVM is trained with all the data samples of the i-class coded as 1, and the data samples of other classes coded as -1. In this work, the SVM model is built to solve a problem by using an OVA strategy, as follows.

Minimize

_¦

ns j i j i i C ₁ 2 * 5 . 0 ,ȟ w [ w (2) subject to

, ( )

1 i, j i j i j b z w I x t [ sign

z_j i

, ( )

1 i, j i j i j b z w I x d [ sign

z_j zi 0 t i j [

whereC is a predefined parameter being introduced according to a soft margin approach and it controls the trade-off between training accuracy and generalization (note: an example of the effect of C on a linear SVM is illustrated in Figure 4; wjis the weight vectors of SVM trained with data samples from two classes; I(xj) is the kernel function;

i

b is a scalar; [ij is the slack variable that permits i 1 , ,ns constraints to be violated; zj{1,1} is the class

label for the classifier. Given a data sample x, the decision function of the SVM is

x argmaxi K

wi,x bi

Class 1,..., (3)

wtx+ b = 0 wtx+ b = -1

(6)

(a) C=10 (b) C=100

Fig. 4. The effect of soft margin constant C. On the left side (a) C = 10 and at the right side (b) C = 100. The figure shows that the positive and negative samples can be separated by a hyperplane. In the case of (b), when the margin value increases, the hyperplane is closer to the boundary. By selecting an appropriate value of the parameter C, the SMV model can perform with optimum classification results by reducing its training errors. [23]

4. Classification results

In the experiment setup, three types of terrain of agriculture field are used in the classification, i.e. sand, gravel and soil. The data obtained from the accelerometer is measured in three axes, i.e.Xaxis, Yaxis and Zaxis within the range of 2G. The experiment uses ONE-versus-ALL strategy,i.e. the data (gravel) gained from the vertical acceleration, Zaxisis used as the “ONE” and the other data (sand and soil) is combined and used as “ALL”. Then the experiment is repeated for sand as the “ONE” and the others combined as “ALL” for soil and gravel. In the support vector machine, a 10-fold cross-validation is used during the experiment to generate the separating line on the hyperplane,i.e. one part data out of 10 parts from gravel is used as testing and other nine parts are used as training. The experiment then is repeated for sand and soil data. In the experiment, three types of kernel are used in the classification process. As shown in Table 4, the kernels usedinclude the linear function, polynomial function and radial basis function (RBF).

Table 1. Percentage of Classification Rate and Mean for X-Y-Z-axis using different kernel functions.

Type of terrain Gravel Sand Soil

Axis X Y Z X Y Z X Y Z Kernel F u nct ion Linear Test 40% 34% 70% 89.6% 88% 53.5% 48% 60% 64% Mean 14.14 18.37 13.72 7.76 2.88 11.47 19.23 10 20.02 RBF Test 80% 77% 93% 92.4% 83% 92.4% 79.8% 80% 84% Mean 10 9.48 11.41 4.85 5.77 11.81 15.49 7.42 15.85 Polynomial Test 44% 32% 46% 94.8% 86.7% 86% 44% 60% 53% Mean 8.94 20.44 15.06 5.59 5.77 11.43 20.74 10 20.02

For the first approach, the experiment uses only a single data from the vertical acceleration along theZ axis collected from the accelerometer. The experiment shows that the classification rate of gravel in the Zaxis have

(7)

higher classification rate as compared to the other terrains and expected in the results caused by high acceleration rate obtained from the accelerometer for the gravel. From the experiment it is observable that by varying the kernel function,the learning algorithm generatesa different result. For learning process, one set of data is used for testing and the rest for training, and the result is observable in Table 1. The results on the radial basis function (RBF) kernel shows higher consistency in the classification rate compared to the linear function andpolynomial function kernelsin the Z axis.

The second approach also used the ONE-versus-ALL strategy for linear function, radial basis function (RBF) and polynomial functionkernels comparing only the acceleration obtained only from the Yaxisfor all the terrains, and then repeated only for theX axis. The data obtained from the support vector machine when using the Yaxisand

Xaxisis not consistent with the Z axisdue to the data collected from the horizontal and lateral acceleration do not

represent the actual vibration from the terrain surfaces.

The mean values obtained in Table 4 demonstrate the variant of the data from the mean results. The lower deviation specifies the numbers that are closer to the mean results or expected data, and with higher deviation indicates the numbers is scattering out over a huge range of values.

5. Conclusion

Asupport vector machineis embedded on board a track-driven agricultural robot to realize the functionality of terrain classification. The vibration signals as a result of track-terrain interaction have been measured and used as an input for the support vector machine. The effectiveness of three kernel functions, i.e. linear function, polynomial function and radial basis function (RBF) has been compared using a 10-fold cross validation method. The resultshows that the SVM model trained with an RBF kernelcanbetter identify the terrain conditions in comparison to linear function and polynomial function kernels.

Acknowledgements

The authors thank the Ministry of Science, Technology and Innovation Malaysia and the Ministry of Education

Malaysia for funding the research work through ScienceFund [Ref. 100-RMI/SF 16/6/2 (3/2014)] and Exploratory Research Grant Scheme [Ref. 600-RMI/ ERGS 5/3 (13/2013)]. The authors thank the Research Management

Institute of Universiti Teknologi MARA for managing the research funds.

References

[1] Allotta, B., G. Buttazzo, P. Dario and F.A. Quaglia. Force/torque sensor-based technique for robot harvesting of fruits and vegetables. Proceedings. IEEE International Workshop on Intelligent Robots and Systems (IROS) ; 1990 , p. 231-235.

[2] Edan, Y., D. Rogozin, T. Flash and Miles G.E. Robotic Melon Harvesting. IEEE Transactions on Robotics and Automation 2000:831–835. [3] Sistler, F. Robotics and intelligent machines in agriculture. IEEE Journal of Robotics and Automation 1987:3–6.

[4] Gausemeier J, Kahl S, Low CY, Schulz B. Systematic Development of Controllers based on the Principle Solution of Self-Optimizing Systems. In Proceedings of the DESIGN 2008, 10th International Design Conference, Dubrovnik, Croatia, 2008; 1263–1276.

[5] Low CY, Aldemir M, Aziz N, Dumitrescu R , Anacker H , Mellado M. Strategy Planning of Collaborative Humanoid Soccer Robots based on Principle Solution. Prod. Eng. Res. Devel. 2013; 7:23–34.

[6] Mousazadeh H. A Technical Review on Navigation Systems of AgriculturalAutonomous Off-road Vehicles, Journal of Terramechanics 2013:211–232.

[7] Santosh A, Gerie W.A.M. van der Heijden, Frits K. van Evert, Alfred S, Cajo J.F. ter Braak. Laser range finder model for autonomous navigation of a robot in a maize field using a particle filter. Computers and Electronics in Agriculture;100: 41-50.

[8] Belforte G, Deboli R, Gay P, Giglio A.Robot Design for Applications in Intensive Agriculture. 2002 IEEE International Conference on Industrial Technology; 2002, p. 519-523.

[9] Zhao C.-J., Jiang G.-Q, Baseline detection and matching to vision-based navigation of agricultural robot, in International Conference on Wavelet Analysis and Pattern Recognition; 2010,p. 44-48.

[10] JiW, Zhao D, Cheng F, Automatic recognition vision system guided for apple harvesting robot. Computers and Electrical Engineering 2012:1186-1195.

(8)

[11] Roseleena J, Nursuriati J, Ahmed J, Low CY. Assessment of Palm Oil Fresh Fruit Bunches using Photogrammetric Grading System. International Food Research Journal 2011; 18(3).

[12] Mousazadeh H. A technical review on navigation systems of agricultural, Journal of Terramechanics 2013:211–232.

[13] Huang Y, Lan Y, Thomson SJ, Fangc A, Hoffmann WC, Lacey RE. Development of soft computing and applications in agricultural and biological engineering. Computers and Electronics in Agriculture 2010:107-127.

[14] Wang J, Zhao D-A, Ji W, Tu J-J, Zhang Y. Application of support vector machine to apple recognition using in apple harvesting robot, IEEE Int. Conference on Information and Automation; 2009, p. 1110-1115

[15] Young J.R. Neural Network Control for Visual Guidance System of Mobile Robot. Lecture Notes in Computer Science 2007: 685-693. [16] Astrand B, Baerveldt A. J. An Agricultural Mobile Robot with Vision-Based Perception for Mechanical Weed Control. Journal of

Autonomous Robots 2002:21-35.

[17] Gausemeier J, Low CY, Steffen D, Deyter S. Specifying the Principle Solution in Mechatronic Development Enterprises. In: 2nd Annual IEEE Systems Conference, Montreal, Canada; 2008, p. 1–7.

[18] Low CY, Rosdayanti F, Azmi K, Che Zakaria NA. Steering Behavior of a Track-Driven Paintball Robot. Procedia Engineering 2012. [19] Kohavi Z. Switching and Finite Automata Theory. 2nd ED. New York: McGraw-Hill; 1978.

[20] Vapnik V. The Nature of Statistical Learning Theory. London: Springer-Verlag; 1995.

[21] Weiss C, Holger F, Andreas Z. Vibration-based Terrain Classification Using Support Vector Machines.Proceedings of the 2006 IEEE/RSJ International Conference on Intelligent Robots and Systems;2006, p.4429-4434.

[22] Shawe-Taylor J,Cristianini N. Kernel Methods for Pattern Analysis. Cambridge: Cambridge University Press; 2004.

[23] Hsu CW, Lin CJ. A Comparison of Methods for Multiclass Support Vector Machines. IEEE Transactions on Neural Networks 2002:415-425.