Implementation Vehicle Classification On Distributed Traffic Light ...

0 downloads 0 Views 693KB Size Report
Abstract— Distributed Traffic System Control. System is a real-time adaptive traffic light system with traffic condition for minimize the probabilty of traffic ...
ICACSIS 2011

ISBN: 978-979-1421-11-9

Implementation Vehicle Classification On Distributed Traffic Light Control System Neural Network Based Big Zaman1, Wisnu Jatmiko1, Adi Wibowo2, and Elly Matul Imah3 1 Faculty of Computer Science , Universitas Indonesia 2 Faculty of Mathematics and Natural Science, Diponegoro University 3 Mathematics Department, Universitas Negeri Surabaya E-mail: [email protected]

Abstract— Distributed Traffic System Control System is a real-time adaptive traffic light system with traffic condition for minimize the probabilty of traffic congestion. So far, the research of Distributed Traffic Light Control System has been developed with Principle Component Analysis (PCA) as the recognition method to identify vehicle object. The recognizition can be optimized using classification system that can identify an object to more spesific class as large cars like bus and truck, or minicars like van, jeep, and sedan. Classification systems has be implemented with neural network algorithm specifically Backpropagation, Fuzzy Learning Vector Quantization (FLVQ), and Fuzzy Learning Quantization Particle Swarm Optimization (FLVQ-PSO).

T

I. INTRODUCTION

RAFFIC is one of the most important thing of the city because almost all vital activities like capital flows, informations, logistics, and many more will be back to the road. Unfortunately, nowadays, traffic congestion is one of the main problem of a huge city as Jakarta. It can be measured based on data from Samapta Directorate, Police and Central Bureau of Statistics that in 2003 vehicle populations in Jakarta has reached 4.159 million. Most of the population of such vehicles are motorcycles (66.15%) followed by private cars (22.01%), public transport buses (3.57%) and goods vehicles or trucks (8.25%). In addition, based on data from the Indonesian National Police Commission, the number of registered motor vehicles in Jakarta (except for the TNI / police) in June 2009 were as many as 9,993,867 vehicles (9,739,633 private cars and public transportation 254 234), while the number of residents DKI Jakarta in March 2009 as many as 8,513,385 people [1]. By comparing between the number of vehicles to population, it can be concluded that in one family in Jakarta, there are approximately three vehicles. It can be predicted that in 2014, Jakarta will have a total traffic congestion, because growth of vehicle numbers is offset by growth or manufacture of road widening. An appropriate traffic light control system is needed to minimize the possibility of traffic congestion.

Conventional traffic lights that exist today, are still oriented to time, standalone, and not adaptive with road conditions. Above conditions became the main motivation holding of research on Distributed Traffic Light Control System (DTLCS). DTLCS expected to be a form of evolution from conventional traffic lights. This system can perform scheduling and multi-junction network settings based real-time data. DTLCS purposes is to understand the road conditions using the method of Principal Component Analysis (PCA). This method is able to recognize a vehicle from an image captured by a video camera that is directed to the highway and it purposes to monitor the state of the road in real-time. This system will also connect the lights at each intersection using Distributed Constraint Satisfaction Problem (DCSP). This method will make the existing road crossing can communicate each other, share data, and collaborate. This research is part of the whole research of Distributed Traffic Lights Control System. This research aims to improve recognition accuracies and to count the number of cars with vehicle classification process. Type of car information will be useful to conclude road density in number, so conclusions can be drawn precisely. Illustration in Fig.1 will explain clearly that type of vehicle is necessary to be known.

Fig. 1. Illustration type of cars that should be classified

With the vehicle type information, we can guess better forecasting for traffic densities. It can be done with classification process. Classification process will be implemented using three artificial neural network (ANN) algorithms namely Backpropagation, Fuzzy Learning Vector Quantization (FLVQ), Fuzzy Learning Vector Quantization Particle Swarm Optimization (PSO-FLVQ). All of these algorithms will be tested by

107

ICACSIS 2011

ISBN: 978-979-1421-11-9

entering the image of the same car, then evaluated the level of precision. II. RELATED RESEARCH

Distributed Traffic Light Contol System research has been conducted last two years. This research is led by Wisnu Jatmiko in [2]. Abdul Arfan in [3] and Adhitya Novian Raidy in [4] have tried to simulate traffic light by implementing the Kuramoto model. Abdul Arfan have compiled Kuramoto model with Kosuke Sekiyama's research in a two-dimensional simulations using the Java programming language. While Adhitya Novian Raidy have implemented Kuramoto model in a three-dimensional simulation using C + + programming language using the ODE library. Adi Wibowo in [5] also have tried to develop a large system consist of an engine to perform calculations on data provided by previous system which was developed by Syarif A.R in [6] and Hendra Marihot in [7]. Visualization process that describe the components as road traffic, cars, traffic light has also been developed by Michael Wicaksana in [8]. Then car sensors whose task was to calculate how many vehicles in a road, and calculate average speeds of all vehicles passing through these roads have been developed by Eka Manggala Adideswar in [9], and Faris Al Afif in [10]. III. BACKGROUND

A. Principle Componenent Analysis Principal Component Analysis (PCA)is a mathematical procedure that analyzes data statistically in order to extract some variables that have correlation possibilities each other that why its called principal components. In some application, the main component is commonly referred as feature. PCA function to calculate the level of variability in the data was found, and calculate each component of the pass. In addition, PCA purposes also calculated variation rates remaining, and possibly correlated. In theory, PCA is commonly known as the most appropriate computational model to represent an object by extracting the most relevant information contained in the object. Unfortunately, not all data can't be extracted using PCA, one of data type that often is processed using PCA is digital image data. PCA extraction on image introduces us to the next concept which is eigen vector. Eigen vector approach on PCA is an approach where a small group of image characteristics is used to obtain variance between images of certain objects, such as sedans, mini-buses, trucks, buses, and other types of cars. The aim is to find eigen vectors from covariance matrix of the object distributions that spread by training process using images data. Each subsequent image will be represented by linear combination of these eigen vector. Object recognition is done by projecting new image into subspace that spanned by the eigen vector

then categorize vehicles by comparing its position in space objects with a certain position. Suppose for a face, its eigenvalue is represented by the both shape and position of eyes, position of nose, and lips. In vehicle image data, eigenvalues can be represent the location of the windshields, two headlights, which are considered as principal component of a car. B. Artificial Neural Network Artificial neural network (ANN) is a concept that adopt interconnection between neurons in human nerve tissue. ANN was developed to create systems that can perform learning process itself like human being without having any time to change the input condition or the system and have tendency to perform better at the next decision. ANN consists of neurons that are connected each other. Each neuron can be connected with number of neurons, this connection has same characteristics as synapses in human nervous system tissue that connects axon to dendrites. ANN system is built based on the workings of human neural network. Human nervous tissue consists of nerve cells called neurons that are composed of dendrites that receive input from other neurons. Then continue functioning axons of these inputs to other neurons. Learning paradigms in ANN are categorized into two types namely learning by direction (supervised) and learning without direction (unsupervised). C. Backpropagation Backpropagation is one of the neural network classification algorithms. Backpropagation is a learning method that was developed from multilayer perceptron rules. Backpropagation algorithm was introduced by Arthur E. Bryson and Yu-Chi Ho in 1969 [11], later its developed by Werbos, Rumelhart, Hinton, and Williams [12] in 1986.

Fig. 2. A neuron content in Backpropagation

Backpropagation algorithm known as one of the most simple and easy neural network to implement. Backpropagation consists of three layers namely input layer, hidden layer, and output layer. Each layer has each neurons with varying number based on by types, number of features, and other considerations. Each neuron in hidden layer and output layer aims is to calculate based on neuron inputs and weights, this

108

ICACSIS 2011

ISBN: 978-979-1421-11-9

function known as activation function. Calculation results of these functions will be accumulated into the next neuron, to raise the classification of the input values (see Fig.2).

represented by the membership function of reference vectors in cluster layers, while the activation function is a slice of the membership function between input vectors and reference vectors.

D. Fuzzy Learning Vector Quantization Fuzzy Learning Vector Quantization (FLVQ) is one of classification algorithms that are supervised neural network. FLVQ is a modification of LVQ algorithm that was combined with fuzzy concepts in initial inputs. FLVQ have winner take all principle, suposse for any data represented as a vector, the closest vectors to input vectors will be decided as winner. In FLVQ, vector also do shifting position according to the output condition, in example when the vector is winner vectors it will be brought near, otherwise will be kept. As we know before, FLVQ is consist of LVQ dan fuzzy concepts. This combination makes FLVQ has advantages in terms of computing time advantage quickly as LVQ and higher recognition rate ablity than Backpropagation or other probabilistic neural networks. In addition FLVQ also very good in terms of data classification that has high level of similarity. FLVQ network architecture can be described as Fig.3 as follows.

E. Fuzzy Learning Vector Quantization FLVQ-PSO is an improvised form of FLVQ methods previously described. FLVQ-PSO was first developed by Benjamin Kusumoputro and colleagues in [12]. FLVQ concept as a modified form of the LVQ algorithm is still considered not quite powerfull since FLVQ have significant dependences on the initial initialization vector, if not match with the expected reference vectors then classification accuracies will be very low. This is resolved in FLVQ-PSO by forming reference vectors more than one, and were randomly selected as many particles are desired (see Fig.4).

Fig. 3. FLVQ’s Architecture

Fig. 4. Group of reference vectors illustration that was formed from input vectores.

Similar to other neural network algorithms, which is composed of three layers namely input layer (input layer), the output layer (output layer), as well as the cluster layer. Input layers is filled by input neurons correspond to data that will be classified. Output layer are filled by planned output neurons, while cluster layer or hidden layer is a layer that serves to calculate the similarity value of each input vector to reference vectors. The number of neurons in cluster layer is the result of multiplying the number of neurons in input layer and the number of neurons in output layer. These neurons are referred to reference vectors. What distinguishes FLVQ with other neural network algorithms is that FLVQ network weights is

FLVQ-PSO concept is a combination of FLVQ concepts, Similarity Matrix Analysis (MSA), and Particle Swarm Optimization (PSO). FLVQ is used to increase speed in conducting training process, MSA is used to determine fitness values, while the PSO is also adopted to make changes position fuzzy vectors to produce more optimal position. In FLVQ-PSO, during training process, each referece vector position is constantly updated by PSO, until n iteration, reference vector is increasingly closer each other and eventually become convergent. This is one of the excess of FLVQ-PSO method because the addition of PSO concepts, so vectors converges to the

109

ICACSIS 2011

ISBN: 978-979-1421-11-9

solution point.

Fig. 7. Object Extraction Process

IV. VEHICLE CLASSIFICATION ON DISTRIBUTED TRAFFIC LIGHT CONROL SYSTEM

2.

DTLCS gets input of data objects obtained from the vehicle recognition and tracking phase. Tracking results are represented using vehicle images that has been adjusted as shown in Fig.5. Later, image data is classified using neural network algorithms that have been implemented.

Scalling Object Object scaling is a process that equalize object sizes to be identified further as a vehicle or not (see Fig.8). Scaling process is necessary because before entering the stage of recognition, which will be the input image must have the same pixel size. In this case, all result objects is converted into an image object with a size of 50 x 50 pixels.

Fig. 5. Vehicle image illustration provided by PCA training process

Before going to classification process, first image data have to do several steps of data pre-processing. In Fig.6, it can be seen flow of data pre-processing prior to the classification process.

Fig. 8. Scaling Object

3.

Fig. 6. Data-Preprocessing Flow

1. Vehicle Object Extraction Object extraction a process purposes to get object to be recognized as a vehicle or not (see Fig.7). Extraction will provide the output of the image objects that are identified as a vehicle with IplImage data structure which would then become the input of the subsequent recognition process to be performed by the neural network algorithm. This extraction process uses PCA algorithm. At this early stage, PCA training process with number of positive image. Positive image is an image with dominant vehicle objects. In addition positive image, its also required negative image that represent the background of vehicle objects, such as empty road image with other static objects such as trees, houses and traffic lights.

Eigenvalue Convertion Each network can not receive direct input in image form, so it needs to be converted into a series of values that can represent that image. These values known as eigenvalue. Eigenvalue is obtained from the PCA method. This PCA method will convert image in a file to be eigenvalues. In this study, we use an experiment on 120 images that consist of two classes, which is small vehicle class (the class that represents small vehicles like sedans, jeeps, minivans, etc.) and large vehicle classes (classes that represent large-sized vehicle such as buses, trucks, etc). Each class is selected 60 images. Codification of each input will be stored as text files (. txt), with the file name information along with the target class. Examples of text files can be seen in Fig.9.

Fig. 9. Eigenvalue in a file input

110

ICACSIS 2011

ISBN: 978-979-1421-11-9

TABLE II ACCURACY VALUES OF BP, FLVQ, AND FLVQ-PSO (IN PERCENT)

4. Neural Network Classification This is the main process of this research. Classification process performs on input data file that was created earlier from eigenvalue convertion phase. As described before, classification process in this research was conducted using three algorithms namely Backpropagation, FLVQ, and FLVQ-PSO. Experimental tests are carried out by the experimental design as follows: a. Input Data Data input is a list of eigenvalues that represent each vehicle image that would be classified. List of eigenvalues are stored in a file with .txt extention with 120 image data. Cross validation method with 10-fold is used to determine the ratio of training and testing data. There are seven categories of data ratio 20:80, 30:70, 40:60, 50:50, 60:40, 70:30, and 80:20. b. Target Class Target class being tested consists of two classes, small and large vehicles. Small vehicles are represented by value [1 -1], whereas a large vehicles are represented by value [-1 1]. c. Feature Features used to represent an image consisting of some combination of features. We decided to use four features obtained by PCA method because with four features only, it represents about 85% representation of the actual image. d. Learning Rate and Epoch Learning rate values and number of epoch that used to perform a comparison test is the best performance of each algorithms that have been studied previously. The results of experiments uses various values of learning rate and number of epoch from each method. Experiment results can be seen in Table I as follows TABLE I LEARNING RATE (ALPHA) AND EPOCH EXPERIMENT RESULTS Algorithm

Alpha

Epoch

FLVQ

0.02

100

PSO

0.02

100

BP

0.045

45

Ratio

BP

FLVQ

FLVQ-PSO

20-80

97.5

81.69

55.81

30-70

99.05

84.43

61.71

40-60

100

89.58

70.39

50-50

100

95

83.43

60-40

100

94.5

83.92

70-30

100

93.5

84.22

80-20

100

96.25

83.83

Mean

99.51

90.71

74.76

Std Dev

0.953

5.677

12.11

Fig. 10. Comparison results of each neural network algorithms

Based on experiment results, Backpropagation seems outperform than other neural network algorithms. Backpropagation obtains high accuracy in any ratio and maintains accuracy above 97% with average at 99.51%. FLVQ have a pretty good performance, in which the lowest accuracy only 81.69% but has an average accuracy 90.71%. FLVQ-PSO is still not good and far from the expectations which had average values 74.76% with the worst accuracy reach to 55.81%. V. CONCLUSION AND NEXT RESEARCH

Experiment were also carried ten times for each scenario and averaged to obtain a more objective conclusion. Test results for each algorithm is shown in Table II. Comparison charts can also be seen in Fig.9.

Classification cars on DTLCS can be implemented using several neural network algorithms such as Backpropagation, FLVQ, and FLVQ-PSO. Among that three methods, Backpropagation showed the best accuracy performance with average accuracy values 99.51%, followed by FLVQ with 90.71%, and FLVQ-PSO with an average accuracy 74.76%. Several subsequent researhs can be done to support the realization of these vehicle classification system such as follows: 1. Classification method was developed to more target classes with a wide variety of vehicles in traffic conditions. 2. Exploring methods of vehicle tracking, phase

111

ICACSIS 2011

ISBN: 978-979-1421-11-9

before classification process, in order to identify vehicle objects from the video stream into vehicle objects that is ready to be classified. The researchs above are expected to make this product one step further ready to be launched so it can quickly become one of solution to ease traffic congestion in big cities in Indonesia.

[5] [6]

[7]

ACKNOWLEDGEMENT

[8]

This work was supported by Competitive Research Grant 2011 Universitas Indonesia.

[9]

REFERENCES

[10]

[1]

[2]

[3] [4]

[1] Komisi Kepolisian Indonesia. Jumlah Kendaraan Bermotor Juni 2009. 2009. http://www.komisikepolisianindonesia.com/main.php?page=a rtikle&id=1187 (diakses Juni 3, 2011). W. Jatmiko, et. al., “Self-Organizing Urban Traffic Control Architecture With Swarm-Self Organizing Map In Jakarta: Signal Control System And Simulator”. International Journal on Smart. 2010 A. Arfan. “Sistem Pengaturan Lalu Lintas Terdistribusi dengan Menggunakan Nonlinear Coupled Oscillator”.2010. A. N. Raidy. “Pengembangan Simulasi Pengaturan Lampu Lalu Lintas Secara Terdistribusi dengan Menggunakan Open Dynamics Engine”. 2010.

[11] [12] [13]

112

A. Wibowo. “Optimasi Pengaturan Lampu Lalu Lintas Secara Terkoordinasi Berbasiskan Self Organizing Control”. 2011. S. Abdurrahman. “Usulan Pengem-bangan Model Matematika Pengaturan Lampu Lalu Lintas Simpang-N Secara Terdistribusi Berbasiskan Nonlinear Coupled Oscillator dan Sinkronisasi Kuramoto”. 2010. H. Marihot. “Desain dan Implementasi Engine Simulasi Pengendalian Lampu Lalu Lintas secara Terdistribusi Menggunakan Model Kuramoto” . 2010. M. Wicaksana. “Penerapan Model Component Based Software Engineering untuk Implementasi Visualisasi Simulasi Keadaan Lalu Lintas” . 2010. M. E. Adideswar. "Implementasi Pendeteksian dan Penghitungan Mobil dengan Pemrosesan Video untuk Sistem Pengendalian Lampu Lalu Lintas Terdistribusi." 2010. F. A. Afif. “Implementasi Prototipe Sistem Pengaturan Lampu Lalu Lintas Terdistribusi dengan Optimasi Pengenalan dan Penjejakan Kendaraan Berbasis Pemrosesan Video”. 2011. A. E. Bryson and Y. C. Ho. “Applied Optimal Contol : Optimization, Estimation, and Control. Blaisdell Publishing Company”. 1969. D. E. Rumelhart, G. E. Hinton, and R. J. Williams.”Learning representations by back-propagating errors”. 1986.Nature, 323, pp. 533--536. B. Kusumoputro, H. Budiarto, and W. Jatmiko, “Fuzzy-neuro LVQ and its comparison with fuzzy algorithm LVQ in artificial odor discrimination system”, ISA Transactions 41, 2002, pp 395 – 407.