neural network based vehicle classification

1 downloads 0 Views 431KB Size Report
especially in toll gates for effective traffic management. This project presents a NN based approach for vehicle classification. This proposed vehicle classification ...
Proceedings of “National Conference on Information Processing & Remote Computing dated 27 & 28 feb. 2014

ISSN:2320-0790

NEURAL NETWORK BASED VEHICLE CLASSIFICATION R.Suganya II-M.E (Embedded System Technologies ) KSR Institute for Engineering and Technolog y Tiruchengode.

Mr.T.Srihari Associate Professor KSR Institute for Engineering and Technolog y Tiruchengode.

ABSTRACT: Automated vehicle classification based on static images is highly practical and directly relevant for various operations in real world such as traffic related investigations. It is important to reduce the traffic density especially in toll gates for effect ive traffic management. This project presents a NN based approach for vehicle classification. Th is proposed vehicle classification approach extracts various features from vehicle images, normalizes and classifies them into one of the three classes. The resulting system robustly identifies vehicles, by rejecting its background, and then the attributes of the vehicles such as height, width, perimeter and area are extracted by means of feature extraction techniques in image p rocessing. These extracted features will then be used by the Neural Network (NN) to classify the vehicles to one of the known classes. The NN make use of Levenberg-Marquardt(LM ) algorith m wh ich provides faster convergence than any other types of existing algorith ms. A proposed classification system have improved the classification rate to 95.1% and also reduced the misclassification rate up to 4.9% as co mpared to that of existing classifiers. Keywords-Veh icle classificat ion, Feature ext raction, NN, LM I. INTRODUCTION probability of two vehicles fro m two cameras being similar. The authors define the vehicle matching There is a need to develop automatic processing problem as a two-class classification problem, and classification of vehicles in order to face many thereafter apply a weak classificat ion algorith m to real wo rld problems such as toll collection, fare obtain labeled samples for each class. The main collection, break-down roadside services etc. classifier is trained by a unsupervised learning Vehicle classification is important in the algorith m built on Gibbs sampling and Fisher’s computation of the % of vehicle classes that use linear discriminant using the labeled samples. A state-aid streets and highways. The current key limitation of this algorithm is that it can only situation is described by outdated data, and it is perform on images that contain similar vehicle pose human operator to manually count vehicles at a and size across mult iple cameras. specific street. Thus, one uses an automated system to obtain perfect design of asphalts with apparent Julius Stroffek et al.[2] proposed that the highway results in cost and quality. A human being may be toll control requires automated and real-time capable of identify ing the class of a vehicle with a classification of hurtling motor vehicles. In this quick glance at the digital data but accomplishing proposal, it presents a modular software solution to that with a computer is not as straight forward. NN the procedural problem of how to classify vehicles are capable of learning and adapting complex nonon a highway with a tollgate furnished with a laser linear patterns. The most of the NN imp lemented scanner. Software identifies individual vehicles and for this classification purpose make use of Multi passes a set of descriptors to the classification Layer Perceptron(MLP).But here the NN make use procedure itself. The classification algorithm uses of LM backpropogation algorithm. The main the shapes of vehicles, in the form of threeadvantage of this algorithm is its high speed of dimensional reconstructions of scanned vehicles convergence. called wire models, together with a series of incidental feature descriptors. Most of the II. RELATED WORKS computations described here depends on the strength and paleness of the main sensor used, Shan et al. [1] presented an edge-based method for which is a standard industrial laser scanner. . vehicle matching for images fro m non-overlapping For the rear v iew, Dlagnekov and Belongie [3] cameras. This feature- based method computed the develop a vehicle make and model recognition

Proceedings of “National Conference on Information Processing & Remote Computing dated 27 & 28 feb. 2014

system for video surveillance using a database of partial license plate and vehicle visual description data, and they report a recognition accuracy of 89.5%.Visual features are extracted using two feature-based methods(SIFT and shape context matching) and one appearance-based method .The drawbacks of the proposed system are that it is relatively slow, and only the license plate recognition stage is done in real-t ime. Similarly Qin [4] used the basic technique for vehicle classification by analyzing images taken fro m the video. Firstly he set up road background relative to the different serial image. Then by using background division he segmented vehicle region and calculated the moment invariant features. These features were passed as an input to the BP neural networks. The network is div ided into three layers such as Light Vehicle, Heavy Vehicle and Intermediate Vehicle. The neural network helped in classifying the vehicle type more efficiently and effectively. III. EXIS TING S YSTEM Vehicle classification is an efficient technique used for transportation systems such as toll plaza, t raffic surveillance etc. As like in Figure 1 ,two common techniques are used for the classificat ion of vehicles, such as  

Hardware Based Veh icle Classificat ion Software Based Vehicle Classification

Hardware based vehicle classification include different tools such as magnetic, radar, in frared detectors etc. Different toll stations defined a standard for classifying vehicles. This standard is based on different types of vehicles. This classification technique has various drawbacks     

Such detectors are large in size and have difficu lt in maintenance. Installation cost is very high. Its installation causes damages to the road. Costly detectors Limited information.

While software based vehicle classification includes image based classification. Image based vehicle classification have been implemented in order to overcome the problems of hardware based vehicle classification. Generally, there are three phases in a vision-based vehicle classificat ion system are vehicle segmentation, feature extraction and vehicle classification. Image based classificat ion has many

advantages over hardware based classification such as,  No damage to roads.  Provide much mo re informat ion than hardware based detectors. 

Less maintenance.

Figure 1. General Classification Techniques

The existing classification technique mainly concentrates on side view and rear view based classification. The side view have a problem of extracting the parameters from the lengthy vehicle such as truck, and the rear view have a problem of occlusion of images which is behind our target vehicle 3.1 Existing Neural Classification The existing neural based classification scheme makes use of MLP network. The M LP make use of gradient descent back propagation algorithm. The gradient descent algorithm have slower convergence rate than that of proposed two-layer feed-forward network with LM back propagation algorith m. It is better to habit additional hidden layers of just a few units is well than using fewer hidden layers requiring too many units, especially for networks that need to learn a function with discontinuities. So, that it is possible for an M LP netwo rk to handle the discontinuity data. The proposed technique makes use of two layer feed forward network by means of LM algorithm. Multi-layer networks use a mu ltip licity of learn ing techniques, the most widespread being backpropagation technique. Here, the output values are related with the correct answer to compute the value of some predefined error-function. By various techniques, the error is then fed back through the network. Using this informat ion, the algorith m adjusts the weights of every single connection in order to shrink the value of the error function by some slight amount.

Proceedings of “National Conference on Information Processing & Remote Computing dated 27 & 28 feb. 2014

After repeating this process for a sufficiently enormous number of training cycles, the network will typically converge to some state where the error of the calculations is small. In this case, one would say that the network has learned to a certain target function. The subsequent are the types of algorith ms used in NN for existing classificat ion purpose, which leads to the evolvement of Levenberg-Marquardt algorithm,  

Gradient descent back propagation. Conjugate gradient descent propagation technique.

4.1.1.1 Grey Image Conversion and Threshol di ng This conversion converts the RGB images to gray scale by eliminating the hue and saturation informat ion while retain ing the lu minance. Thresholding of the image depends on the pixel value of its background image. The pixel values which are less than the certain pixel value are initially ext racted separately. Thresholded image should undergo some of the morphological operations such as erosion and dilation.

back

Typical problems of the back-propagation algorith m are the speed of convergence and the likelihood of ending up in a local minimu m of the error function. The existing algorith ms took longer time for the network to reach its target value. Thus the problems in existing method lead to the proposed system which overcomes the above discussed drawbacks. IV. PROPOS ED S YS TEM The purpose of this proposed system is to classify the vehicles to different classes with the help of NN Toolbox. The existing systems make use of M LP to implement the classification technique. The main problem in M LP based classifier is its slower rate of convergence. The type of algorithm used in this proposed technique is that LM backpropogation algorith m. The type of network used in this proposed technique is two-layer feed forward network. In prior to the training of the network, the features required for training are extracted by means of certain image processing algorithms.

4.1.1.2

Morphological Operations

Before making an image to get involved in morphological operations, we should init ially open an image in morphological manner. The erosion operation mainly concentrates on smoothening the edges of the images. There may be an availability of some holes in the extracted object, so in order to prevent this fill command is used in image processing command. The following Fig.3 represents, the Vehicle object which undergone different morphological operations. Opening consists of an erosion tracked by a dilation and can be used to eradicate all pixels in regions that are tiny to contain the structuring element. In this case the structuring element is often called a probe, because it is probing the image looking fo r small objects to filter out of the image. Closing consists of a dilation follo wed by erosion and can be used to fill in holes and small gaps. In closing operation there will be an effect of filling in holes and closing gaps.

The parameters to be used are extracted from the vehicles front view. The main reason to take the frontal image of the vehicles to extract the feature is to prevent the occlusion which could happen in the case of considering the rear view of the vehicle. The main parameters to be extracted are Width, height, area and perimeter. The different steps involved in extracting the features from the vehicles were d iscussed in fo llo wing Fig.2. 4.1 STEPS INVOLVED IN EXTRACTING THE FEATURES FROM AN IMAGE 4.1.1 Image Pre-Processing Initially the image has to undergone various steps before being it to get involved in extracting the various parameters. The pre -processing steps involved are grey image conversion, thresholding, erosion and filling of the gaps in the image.

Figure 2. Flowchart For Extracting the Features From an Image

Proceedings of “National Conference on Information Processing & Remote Computing dated 27 & 28 feb. 2014

for the input to reach towards its target value. Here the network converges towards the best at epoch 3.

Figure . 3. Image of Class1 Vehicle Obtained From Morphological Operations Figure 5. Performance Plot of NN

V. TRAINING AND TES TING THE NN The type of NN tool used for training is Neural Fitting tool(nftool) wh ich permanently make use of Levenberg-Marquardt algorithm. Here the network gets converged quickly by means of training the network with different number of hidden neurons. The ext racted feature will be further provided to the NN for classification purpose. There is a lot much availability of NN tools which mainly concentrates on classification purpose. Here the proposed system mainly concentrates on Levenberg-Marquardt algorithm. This lead to faster convergence of the network towards its target value.The steps involved in training and testing are shown in Fig.4

Next ,the fitness curve at Fig.6 determines the fitness of different samples between input and output along with the target value of the specific network wh ich have been involved in training the network.

Figure 6. Fitness Curve of NN

The fitness curve determines the number of intermediate values occur between the target values. Training mu ltip le times will generate different results due to different in itial conditions and sampling. 5.2 TES TING THE NEURAL NETWORK

Figure .4. Training and Testing the Neural Network

5.1 TRAINING THE NEURAL NETWORK

The trained network have been tested now with different number of samp les i.e apart fro m those samples which have been used for training. The Train ing dataset have been verified with the help of Simu link block. The accuracy of the testing set on the system is determined with the help with the graphical plot. The 3 types of vehicles are considered here for the classificat ion purpose. So there are 3 different types of target values have been fixed here. The Simu link b lock for testing the trained network have been shown in Fig.7.

There is a possibility to choose the number of hidden neurons in the network for its faster convergence in minimu m time period. Each time the network is tested with its performance and Regression Plot to determine its efficiency. Here the performance plot at Fig.5 is validated by means of determin ing number of epochs required

Figure 7. Simulink Block for Testing the Network

Proceedings of “National Conference on Information Processing & Remote Computing dated 27 & 28 feb. 2014

The 3 types of vehicles are considered here for the classification purpose are shown in Table 1.Such that there are 3 different types of target values have been fixed here . Table 1. Vehicle Type vs. Class Type

VEHICLE TYPE Car

CLASS TYPE 1

Bus

2

Lorry

3

Figure 9. Output of Class1 Vehicle.

The graph in Fig.9 represents the output of the class1 vehicle, whose target value gets fixed as 1. .

The regression plot in Fig.8 shows the number of samples that lies on the target plane. Maximu m no of samples on the line shows that the maximu m convergence of the output towards the target value.

Figure 10. Output of Class2 Vehicle.

The graph in Fig.10 represents the output of the class2 vehicle, whose target value gets fixed as 2. The follo wing graph represents the output of the class3 vehicle, whose target value gets fixed as 3. Figure 8 Regression Plot of NN

The above values gets trained several number of times in NN for its effective convergence. Then after this the network gets tested with some new data sets. The convergence have been shown for training ,testing and validation datasets. VI

REPRES ENTATION OF DIFFERENT CLASSES OF VEHICLES

The LM algorithm improves the efficiency of the network by means of its faster convergence. Faster Convergence leads the network to reach its target value in reduced number of epochs. The Regression plot of the network give the details of convergence of the network output towards the target. The NN have classified the extracted parameters in to 3 different classes such as class1,class2 and class3. The following graph represents the output of the class1 vehicle, whose target value gets fixed as 1.

Figure 11. Output of Class3 Vehicle

The graph in Fig.11 represents the output of the class3 vehicle, whose target value gets fixed as 3. VII

COMPARIS ION WITH EXIS TING S YSTEM

The proposed LM based NN seems to be imp roved Classification rate and reduced misclassified rate as compared to that of existing classifiers. The efficiency of the proposed classifier have been compared with different classifiers such as Bayesian, SVM (Support Vector Machine), and MLP based NN in Table.2.The following table shows the comparative analysis of different existing classifiers.

Proceedings of “National Conference on Information Processing & Remote Computing dated 27 & 28 feb. 2014

Table 2. Analysis of Different Classifier

place with the help of LM algorith m. Performance of the trained system has been analyzed by means of testing the network with different set of parameters. The NN has found to be efficient in classifying the vehicles than any other types of classifiers with efficiency rate of 95.1%. Furthermore the classificat ion of vehicles can be also done for more types of vehicles as different number of classes using different types of algorith ms and also through hybrid systems such as fuzzy along with Neural Net works. REFERENCES

The Tables 3 and 4 shows the list of parameters extracted fro m the vehicle frontal view for training and testing the LM algorith m based Neural Net work. Table 3. Sample Parameters for Training the Network

Table 4. Sample Parameters for Testing the Network

VIII CONCLUS ION The system for vehicle classificat ion have been proposed with the help of image processing and neural networks. The primary object ive of this proposed system is to classify the vehicles in to different types of classes in tollgate applications. Here an image processing technique is added along with the neural network in order to classify the vehicles. The image processing technique does the job of ext racting the parameters fro m the vehicles. Then these ext racted parameters are provided to the neural network for train ing. The training takes

[1]Y. Shan, H. Sawhney, and R. Kumar, “Unsupervised learning of d iscriminative edge measures for vehicle matching between non overlapping cameras,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 30, no. 4, pp. 700– 711, Apr. 2008. [2] Julius Stroffek, Eduard Kuriscak, and Petr Marsalek, ”Real Time Classification of Motor Vehicles,” IEEE vehicular technology magazine December 2010. [3] L. Dlagnekov and S. Belongie, “Recognizing cars model” Univ. California, San Diego, CA, Tech. Rep. CS2005-083, 2005. [4] Zhong Qin,” Method of Vehicle Classificat ion Based on Video”, Proceedings of the 2008 IEEE/ASM E International Conference on Advanced Intelligent Mechatronics July 2 - 5, 2008, Xi'an, Ch ina. [5] Wei Wang ,Yu long Shang,Jinzh i Guo,Zh iwei Qian” Real-t ime Vehicle Classification Based on Eigenface”, 2011 IEEE. [6]Dubuisson Jolly M. P(1996)’Vehicle segmentation and classification using deformable templates’, Pattern Analysis and Machine Intelligence, IEEE Transactions on, vol. 18, pp. 293-308. [7] Fab io Del Frate, Fabio Pacifici, Giovanni Schiavon, Chiara So limin i (2007),’Use Of Neural Networks For Automat ic Classification Fro m HighResolution Images’, IEEE Transactions On Geoscience And Remote Sensing, Vo l. 45, No. 4. [8] Mehran Kafai ,Student Member, Bir Bhanu(2012)’ Dynamic Bayesian Networks for Vehicle Classification in Video’, IEEE Transactions On Industrial Informat ics, Vol. 8, No.1.. [9]Yong-Kul Ki, Doo-Kwon Baik(2006)’VehicleClassification Algorith m for Single-Loop Detectors Using Neural Netwo rks’, IEEE Transactions On Vehicular Technology, Vol. 55, No.6. [10] Jun-Wei Hsieh, Shih-’Automatic Traffic Surveillance System forVeh icle Tracking and Classification

Suggest Documents