STEP TOWARDS INTELLIGENT TRANSPORTATION SYSTEM WITH VEHICLE CLASSIFICATION AND RECOGNITION USING SPEEDED-UP ROBUST FEATURES

Janak, T. et al: Step towards intelligent ...... Archives for Technical Sciences 2023, 28(1), 39-56
Technical Institute Bijeljina, Archives for Technical Sciences. Year XV – N0
28. 39
Original scientific article
http://dx.doi.org/10.59456/afts.2023.1528.039J
STEP TOWARDS INTELLIGENT TRANSPORTATION
SYSTEM WITH VEHICLE CLASSIFICATION AND
RECOGNITION USING SPEEDED-UP ROBUST FEATURES
Janak Trivedi1
, Mandalapu Sarada Devi2
, Brijesh Solanki3
1
Faculty of Electronics & Communication Engineering Department, Government Engineering College
Bhavnagar, Affiliated to Gujarat Technological University, A.I.C.T.E., India,
e-mail; trivedi_janak2611@yahoo.com
2
Faculty of Electronics & Communication Engineering Department, Ahmedabad Institute of
Technology, Affiliated to Gujarat Technological University, A.I.C.T.E., India.
3
Faculty of Electrical Engineering Department, Government Engineering College Bhavnagar, Affiliated
to Gujarat Technological University, A.I.C.T.E., India.
SUMMARY
Vehicle classification is a crucial task owing to vehicles' diverse and intricate features, such as
edges, colors, shadows, corners, and textures. The accurate classification of vehicles enables
their detection and identification on roads and facilitates the development of an electronic toll-
collection system for smart cities. Furthermore, vehicle classification is useful for traffic signal
control strategy. However, achieving accurate vehicle classification poses significant
challenges due to the limited processing time for real-time applications, image resolution,
illumination variations in the video, and other interferences. This study proposes a method for
automated automobile detection, recognition, and classification using statistics derived from
approximately 11,000 images. We employ SURF-based detection and different classifiers to
categorize vehicles into three groups.
The Traffic Management System (TMS) is crucial for studying mobility and smart cities. Our
study achieves a high automobile classification rate of 91% with the medium Gaussian
Support Vector Machine (SVM) classifier. The paper's main objective is to analyze five object
classifiers for vehicle recognition: Decision Tree, Discriminant Analysis, SVM, K-Nearest
Neighbor Classifier (KNN), and Ensemble Classifier. In the discussion section, we present the
limitations of our work and provide insights into future research directions.
Keywords: Classifier, feature extraction, Support Vector Machine (SVM), Speeded Up Robust Features
(SURF), k-mean.
INTRODUCTION
With the development of hardware and the reduction in manufacturing costs, there has been a
significant increase in surveillance devices, and high-resolution video cameras have been increasingly
employed in such systems. Consequently, many video sources generate large volumes of information
that require analysis and interpretation, but this amount of information needs to be reduced for human
operators to handle. Thus, researchers utilize technology such as the Intelligent Transportation System
(ITS). Automatic extraction of vehicles from surveillance videos using artificial intelligence is a

28. 40
prominent research topic. Vehicle classification using traditional methods poses two main problems:
the need for human supervision to classify vehicles and the availability of variations in different
vehicle models. Computer vision-based traffic management is more intelligent, faster, and more
efficient than human operators. Artificial Intelligence (AI) plays a significant role in ITS using
computer vision. One of the most exciting aspects of AI is its potential to revolutionize the computer
industry and any industry that affects our lives. Deep learning and machine learning are part of AI, and
they offer a new direction with high accuracy for an automatic vehicle recognition system in
computer-vision techniques. Deep learning works similarly to the human brain in the proposed system,
and its layer-by-layer structure extracts useful information for the given dataset.
Recent progress in artificial intelligence (AI) has significantly impacted the research related to vehicle
detection, identification, and classification. Vehicle classification has applications in multiple
domains, including recognizing emergency vehicles in automated toll-collection systems. Janak et al.
have presented a comprehensive review of the utilization of computer vision-based Intelligent Traffic
Control systems in Smart Cities [1]. Vehicle classification can be implemented through online, offline,
and real-time applications. Automated vehicle classification can be specifically advantageous in toll
plazas, where it can rapidly identify the vehicle type and automatically collect the toll charge.
Furthermore, vehicle classification and recognition can assist in developing an adaptive traffic control
system and identifying emergency vehicles. These advancements in vehicle classification and
recognition using AI have the potential to pave the way for the development of more efficient and
intelligent transportation systems, thus making our cities smarter and safer.
The deep learning algorithm represents a recent trend in object detection and recognition in artificial
intelligence systems utilizing computer vision. In deep learning, tiny feature extraction plays a crucial
role in identifying the type or class of a vehicle from selected images. Feature extraction is crucial in
acquiring helpful information about the detected features. Feature matching is essential in applications
such as motion tracking, image retrieval, object recognition, object classification, and human activity
recognition. The limitation of the changing pixel values due to light intensity can be overcome by
selecting the most appropriate features. These features include edges, corners, blobs, gradient values,
magnitude, and phase.
Various classifiers are used to analyze images selected from various web search tools. Feature
extraction, a general term for creating combinations of variables to solve classification problems while
adequately representing the information, is a critical step. Many machine learning practitioners believe
that properly optimized feature extraction is key to successful model construction. Features are pieces
of information obtained from an image. Feature extraction is achieved through various image
characteristics, such as color, shape, corners, position, and dominant edges. These features are critical
for recognition, matching, and detection. The number and size of features play an important role in
different real-time applications. Feature detectors such as SHIFT and SURF are popularly used. In our
research, we classified vehicles into bikes, cars, and trucks, using k-means clustering algorithms and
SURF on a generated dataset of around 11k images.
Various methods and approaches for vehicle detection, recognition, and classification contribute to
versatile traffic light control systems at intersections. Vehicle recognition helps determine the number
and type of vehicles on specific roads. This system is also useful in identifying emergency vehicles,
such as ambulances and traffic management systems (TMS), which can handle difficult situations in
real time. The application of deep learning and artificial intelligence (AI) is essential in developing
Intelligent Transportation Systems (ITS) in smart cities. These techniques are valuable and trending in
the field of ITS research.
Vehicle classification helps in traffic signal control strategy. For example, high-emissions vehicles
stopped less frequently, so recognizing these vehicles and clearing the traffic jams beneficial in terms
of air quality. Vehicle classification helps optimize infrastructures and increase the return on toll gates
by traffic flows. Vehicle classification can classify (A) According to purpose: (1) Passenger
vehicles: like car, bus, taxi, etc. (2) Goods vehicles: like truck, tempo, container, etc. (3) Special
purpose vehicles: like ambulance, fire brigade, etc. (B) According to load carrying capacity: (1)

28. 41
Light Motor Vehicles (LMV): like cars, jeep, etc. (2) Medium Motor Vehicles (MMV): like tempos,
pick up vans, etc. (3) Heavy-duty Motor Vehicles (HMV): tractor, truck, container, etc. (C)
According to number of wheels: (1) Two wheelers: like motor-cycle, scooters, etc. (2) Three
wheelers: like auto rickshaw (3) Four wheelers: like cars, jeep, mini vans, etc. (4) Six wheelers: like
truck, bus, etc. (D) According to fuel used: (1) Petrol vehicles: like motor-cycle, cars, etc. (2) Diesel
vehicles: like cars, trucks, buses, etc. (3) Gasoline vehicles: like LPG vehicles, CNG vehicles, etc. (4)
Electric vehicles: like Yo bikes, cars, etc. (5) Hybrid vehicles: like vans, cars, etc. (E) According to
drive of vehicles: (1) Front wheel drive vehicles: like cars (2) Rear wheel drive vehicles: like truck,
buses, etc. (3) Four-wheel drive vehicles: like military vehicles. This article has classified bike, car,
and truck-vehicle accordingly to their purpose category using computer vision technique.
This study investigates using an alternative classifier for a dataset of approximately 11,000 images
randomly collected from various web search tools. A comprehensive literature review of various
methods is presented in Section 2, followed by a detailed explanation of different feature detection
techniques in Section 3, including Scale Invariant Feature Transform (SIFT) and Speeded-Up Robust
Feature (SURF) detection. Section 4 provides the details of five primary classifiers and 22 sub-
classifiers. The vehicle recognition process's workflow is presented in Section 5, utilizing SURF and
k-means algorithms, as depicted in Figure (1). Section 6 presents the results regarding feature
selection, performance analysis, and verification and testing on random images, as shown in Figures
(4-6). The study concludes with a discussion of limitations, a summary of the findings, and future
research directions.
LITERATURE SURVEY
Recently, computer vision techniques using deep learning and artificial intelligence have been
extensively employed for developing Intelligent Transportation Systems (ITS). These systems require
automatic detection, recognition, and classification of vehicles to enhance the Traffic Management
System (TMS) efficiency in a smart city. Francesco et al. [2] have presented the European Process of
smart city development using three interconnected processes, which include smart city planning
theories, smart city development, and smart city rules and policy. For effective smart city planning,
smart TMS is necessary.
The TMS comprises real-time vehicle information, accident detection, vehicle speed measurements,
parking information, bus-train route information for users and commuters, and an automatic toll
collection system. Antonello et al. [3] have discussed the integration of conventional vehicle reviews
(which provide little information) with extensive knowledge from Information and Communication
Technology (ICT) in constructing Transport System Models (TSMs).
The effective implementation of Intelligent Transportation Systems (ITS) in a smart city requires
automated vehicle detection, recognition, and classification. The classification of vehicles is crucial
for a wide range of applications, including vehicle speed detection, autonomous driving, intelligent
parking systems, and toll collection. In the literature, various classification approaches have been
proposed, such as the nearest neighbor classification approach for multispectral images as
demonstrated in Hardin et al. [4], and the use of KNN classification by Ghosh Anil [5]. The creation
of feature-based vector trees and the AdaBoost method-based classification with random forest and
linear estimation of regression trees are discussed in Breiman [6] and Loh [7]. The decision tree-based
application of classification from tree graphs is also explained by Paensuwan et al. [8]. These
approaches have been shown to classify different types of vehicles effectively and can be applied to
real-time ITS systems in smart cities.
Various vehicle classification methodologies have been proposed in the literature, including length-
based or shape-based, information-based, vision-based, movement-based, and distinctive feature
extraction-based approaches, as explained by Wen et al. [9]. Kim et al. [10] demonstrated on-street
vehicle detection using the Pi-Histogram of Oriented Gradients (HOG) strategy with support vector

28. 42
machines (SVMs). Stocker et al. [11] utilized a supervised learning approach with multilayer
perceptron (MLP) feed-forward artificial neural networks (ANNs) for vehicle classification and
identification. Arinaldi et al. [12] proposed SVM and Faster Region Convolutional Neural Networks
(RCNN) based on vehicle identification and classification using MIT traffic data. Atiya [13] discussed
simple pattern recognition-based algorithms, support vector classifiers, and support vector regression
with different kernel principles.
In Intelligent Transportation Systems (ITS), automatic vehicle detection, recognition, and
classification are essential for developing smart cities. The classification of vehicles is necessary for
various purposes, such as vehicle speed detection, autonomous driving vehicles, intelligent parking
systems, and toll collection.
Sarikan et al. [14] implemented K-Nearest Neighbors (KNN) and decision tree-based vehicle
classification. Gorges et al. [15] clarified the utilization of street profile estimation calculation for
grouping two-wheeled vehicles. Machine learning classifiers such as KNN, decision trees, and support
vector machines (SVMs) have been employed to develop automatic vehicle classification systems.
However, the accuracy rate of SVM and Artificial Neural Network (ANN)-based classification
approaches is low, as discussed by Bautista et al. [16]. Convolutional Neural Network (CNN) has
shown a higher accuracy rate for vehicle recognition than SVM and ANN.
In recent years, Bag of Features (BOF) has been used to classify Magnetic resonance images. Deepika
et al. [17] utilized SVM for BOF-based classification and achieved 93% accuracy. Similarly, Pranata
et al. [18] used BOF-based classification with SURF for detecting fractures in medical images. Sykora
et al. [19] utilized SIFT and SURF for gesture recognition and SVM for classification. Anca and Ioan
[20] discussed palm print characterization and recognition using SURF. Anzid et al. [21] employed
SURF and different types of SVM (linear, non-linear, and multiclass) for multimodal image
classification. These works demonstrate the potential of machine learning methods and feature
extraction techniques for image classification and recognition.
FEATURE DETECTION
Feature detection is a crucial step in many computer vision applications, which involves extracting
relevant features from an image for subsequent processing. One common method for feature detection
is corner detection, which identifies points of interest in an image by detecting changes in the X and Y
gradient values. Corners are useful features because they are invariant to changes in illumination or
viewpoint. The Harris corner detector is a widely used corner detection algorithm that identifies
corners based on changes in gradient in all directions. It operates in three cases: flat regions where
there is no change in gradient, edges where there is a change of gradient value in one direction and
corners where there is a change in gradient value in all directions. SIFT and SURF are popular feature
detection methods that utilize the Hessian matrix to calculate and identify features.
SIFT Feature calculation
The SIFT (Scale-Invariant Feature Transform) feature calculation process involves selecting a 16 x
16-pixel neighborhood and dividing it into 4 x 4 block neighborhoods. Each neighborhood is rotated
and aligned with the previously calculated orientation, and eight orientations are calculated at a 4 x 4
bin array which results in 128-dimensional features which are used for object detection and
recognition applications. The unit magnitude can be used to normalize the illumination problem. The
SIFT features have several properties, including finding feature points, repeatability of key points,
scale-rotation invariance, and robustness to viewpoint changes, as explained by David in [22].
SURF feature calculation
In feature extraction, three important properties are repeatability, matching speed, and descriptiveness.
The number and size of features are also critical factors. Feature extraction consists of two steps:
feature detection and feature extraction. The SIFT method has been widely used due to its high

28. 43
descriptiveness but comes at a high computational cost. In contrast, the SURF method improves upon
the scale-invariant feature detector by using the Hessian matrix to detect the gradient magnitude and
orientation to select points of interest. The Hessian matrix is calculated in Equation (1), and the sum of
the Haar wavelet can also be used to improve orientation performance. Bay et al. [23] discussed the
use of SURF and its advantages over the Difference of Gaussian (DoG) method.
H =
(1)
For multivariate function, the Hessian matrix contains all the second partial derivate of function D.
Dxx represents twice the partial derivative for x, Dyy represents twice the partial derivative for y, Dxy
represents the partial derivative first with x, then for y Dyx represents the partial derivative first for y
than with x.
The machine learning approach enables the automatic classification of objects without requiring
manual labeling or feature extraction. Matlab is a powerful tool for classifying different objects in a
given dataset. The accuracy, prediction, and output results are then validated. In machine learning,
data can be divided into numerical and categorical classes. Numerical data are represented by numbers
(floating point or integer), while discrete label groups represent categorical data. Using Matlab, five
classifiers are tested, further subdivided into twenty-two classifiers.
DIFFERENT CLASSIFIERS
Decision Tree
A decision tree is a hierarchical structure that resembles a flowchart used for classification and
prediction. Each node in the tree represents a test for a specific attribute, and each branch represents
the result of that test. The terminal nodes of the tree, also known as leaf nodes, contain the class labels.
Decision trees are one of the most widely used and powerful tools for classification and prediction.
They can handle both categorical and numerical data, and different decision trees can be used based on
the number of splits made during the classification process.
(1) Fine Trees: Many leaves for acceptable discrimination.
(2) Medium Trees: Moderate number of leaves for finer discrimination between classes.
(3) Coarse Trees: some leaves to roughly distinguish between classes [24].
Discriminant analysis
Discriminant analysis is a statistical method used to classify objects into two or more groups based on
their measured characteristics or features. When the predicted variable is binary, discriminant analysis
can be viewed as a regression analysis. Suppose we have a set of objects that belong to two categories,
and we measure several quantities that help determine the category of the object, which we assign the
values 0 and 1. The discriminant analysis classifier is used to classify numerical data only. Different
types of discriminant analysis can be used to create linear or nonlinear boundaries between classes.
(1) Linear Discriminant: This creates linear boundaries between classes.
(2) Quadratic Discriminant: This creates nonlinear class boundaries [24].
SVM
The SVM algorithm seeks to find the one that maximizes the margin between the two classes, i.e., the
distance between the hyperplane and the nearest data points from each class. The data points closest to
the hyperplane are called support vectors, which play a critical role in determining the hyperplane's
location. The SVM algorithm can also handle nonlinearly separable data by transforming the data into
a higher-dimensional space using kernel functions. SVMs are popular machine learning classifiers due
to their ability to handle high-dimensional data and generalize well to new, unseen data.

28. 44
Figure 1 projection of vector x, vector w perpendicular to the hyperplane,
and decision boundary c [25]
c
• =
x w
c
• 
x w
c
• 
x w (2)
The inner product, which is the projection of one vector onto another, is utilized in the support vector
machine (SVM) algorithm. The dot product of the x vector and the w vector is taken, and if the
resulting value is greater than a threshold 'c', the point is classified as belonging to the positive
samples category. Conversely, if the dot product is less than 'c', the point is classified as belonging to
the negative samples category. If the dot product is equal to 'c', the point is on the decision boundary
between the two categories. The SVM classifier is commonly used for both categorical and numerical
classification tasks.
Different types of SVM depend on their kernel selections. A kernel function is a method used to take
data as input and transform it into the required format for processing data.
(1) Linear SVM: This makes a linear separation between two classes. Data is classified with the help
of hyperplane and straight-line.
(2) Quadratic SVM: This makes a non-linear separation between classes. It cannot be easily
separated from a linear line. The quadratic kernel is used to perform this task.
(3) Cubic SVM: This makes a non-linear separation between classes. Use cubic kernels to make non-
separable data into separable data.
(4) Fine Gaussian SVM: Gaussian kernels can separate nonlinearly separable data by mapping the
input vectors into Hilbert space. Makes finely detailed distinctions between classes, with kernel scale
set to sqrt(P)/4.
(5) Medium Gaussian SVM: Medium distinctions, with kernel scale set to sqrt(P).
(6) Coarse Gaussian SVM: Makes coarse distinctions between classes, with kernel scale set
to sqrt(P)*4, where P is the number of predictors [24].
Nearest Neighbor Classifier (KNN)
The k-nearest neighbor (KNN) algorithm, also known as ANN or k-NN, is a nonparametric supervised
learning classifier that leverages the concept of proximity to make predictions or classifications about
the clustering of single data points. KNN classifiers can be used to classify both categorical and
numerical data types and typically utilize either Hamming distance calculation for categorical data or
Euclidean distance calculation for numerical data. The specific type of KNN algorithm employed
depends on factors such as the distinctions between classes and the number of nearest neighbors set.
(1) Fine KNN: Finely detailed distinctions between classes. The number of neighbours is set to 1.
(2) Medium KNN: Medium distinctions between classes. The number of neighbours is set to 10.
(3) Coarse KNN: Coarse distinctions between classes. The number of neighbours is set to 100.
(4) Cosine KNN: Medium distinctions between classes, using a Cosine distance metric. The number
of neighbours is set to 10.

28. 45
(5) Cubic KNN: Medium distinctions between classes, using a cubic distance metric. The number of
neighbours is set to 10.
(6) Weighted KNN: Medium distinctions between classes, using a distance weight. The number of
neighbours is set to 10 [24].
Ensemble Classifier
Ensemble learning is a technique used to generate a diverse set of base classifiers, from which new
classifiers are derived that outperform any individual classifier. The base classifiers may differ in
algorithms, hyperparameters, representations, or training sets. Ensemble methods aim to reduce bias
and variance in the classification process. The Ensemble Classifier is mainly used to classify
categorical data, except for Subspace Discriminant, which is used for numerical data. Depending on
their specific characteristics, there are different types of ensemble methods in KNN.
(1) Boosted Trees: The ensemble method is AdaBoost, with Decision Tree learners.
(2) Bagged Trees: The ensemble method is a Random Forest Bag with Decision Tree learners.
(3) Subspace Discriminant: The ensemble method is Subspace, with Discriminant learners.
(4) Subspace KNN: The ensemble method is Subspace, with Nearest Neighbor learners. This kind of
operation is suitable for many predictors.
(5) RUSBoosted Trees: The ensemble method is RUSBoost, with Decision Tree learners. This is
good for skewed data (with many more observations of 1 class) [24].
WORKFLOW
5.1.1 Detailed discussion of the dataset creation
The first step is to collect images of bikes, cars, and trucks from various sources, resulting in a dataset
of approximately 11,000 images, with 3.5 thousand bike images, 5.8 thousand car images, and 1.8
thousand truck images. These images were resized to a fixed resolution of 640 (width) x 480 (height)
to ensure consistency, although this may impact the accuracy of the results. The dataset was split into
training and testing sets, with 70% of images used for training and 30% used for testing from each
vehicle category.
Operation on input images
We use the bag of visual words approach with k-means clustering for feature extraction, which
involves iteratively clustering the feature descriptors extracted from representative images of each
vehicle category into k mutually exclusive clusters. The resulting clusters are compact and separated
by similar properties, and the center of each cluster represents a visual word or function. Feature
descriptors can be extracted based on a feature detector or defined raster.The Speeded Up Robust
Features (SURF) detector is widely used to extract features invariant to scale and rotation. These
features are extracted from the representative images of each category using keypoint extraction,
creating feature vectors, and clustering of features through the bag of visual words approach using k-
means clustering. Based on these feature descriptors, a histogram of the frequency of visual words is
then created for each image. These histograms are then used to train an image category classifier.
After the classifiers are trained, they must evaluate their performance using appropriate metrics. The
confusion matrix is a commonly used tool to evaluate the performance of classification models. The
confusion matrix summarizes the number of true positive (TP), false positive (FP), true negative (TN),
and false negative (FN) predictions made by the model. From this, we can calculate various
parameters such as True Positive Rates (TPR), False Negative Rates (FNR), Positive Predictive
Values (PPV), and False Discovery Rates (FDR). These metrics can help assess the classifier's
accuracy, precision, and recall. The workflow for evaluating the classifiers is typically shown in a
figure, such as a Figure (2), which provides an overview of the entire process from training to
evaluation.

28. 46
Figure 2 Workflow
RESULTS
In this study, the training dataset was used to validate various classifiers for categorizing vehicles into
one of three categories: bike, car, and truck. A dataset of size 4800 x 451 was used for this purpose. A
total of 22 different classifiers were evaluated using metrics such as True Positive Rates (TPR), False
Negative Rates (FNR), Positive Predictive Values (PPV), and False Discovery Rates (FDR) with the
help of a confusion matrix. The results for all the decision tree classifiers are shown in Figure 3(a).
To elaborate, Figure 3(a) demonstrates the accuracy of decision tree classifiers, where the fine tree
classifier outperforms the medium and coarse trees. Additionally, Figure 3(b) shows support vector
machine classifiers' accuracy and prediction speed, where a maximum accuracy of approximately
91% is achieved. The accuracy of a classifier is determined by its ability to accurately classify true
positives (Tp), true negatives (Tn), false positives (Fp), and false negatives, which are taken into
account during the calculation of classification accuracy.
Figure 3 (a) Confusion matrix for decision tree classifiers
The number of accurate guesses divided by the number of predictions is the categorization accuracy
as per equation (2) [17].
Accuracy = (Tp + Tn) / ((Tp + Tn + Fp + Fn) (2)

28. 47
In this study, we present our findings, categorized into three distinct areas: feature selection,
performance analysis and comparison, and verification and testing on random images. Feature
selection was conducted to identify the most relevant and informative features contributing
significantly to the classification task.
Performance analysis and comparison were performed to evaluate the effectiveness of the proposed
method against state-of-the-art approaches. Lastly, we conducted verification and testing on a set of
random images to validate the generalizability and robustness of the proposed method. Our results
provide valuable insights into the efficacy of the proposed approach and demonstrate its potential for
real-world applications.
Figure 3 (b) Confusion matrix for SVM classifiers
Feature selection
This study investigates the effect of different feature selection methods on vehicle classification. We
selected different images from each class of vehicles and extracted the number of features for each
category. Specifically, we tested the k-means clustering method using around 250 words to classify
bikes, cars, and trucks. We used the Speeded-Up Robust Features (SURF) algorithm to extract
features, which were represented using the bag of visual words approach. Key points and descriptors
were extracted from the images to identify features invariant to changes in image size, orientation, and
compression. The number of features extracted from each vehicle category is presented in Annexure-I.
We compared the performance of the k-means clustering method using 250, 350, and 450 visual words
for the same dataset of bikes, cars, and trucks. The results of our experiments, including the number of
iterations and required time for each method, are summarized in Table 1.
Performance analysis and comparison
Multiple classifiers, including Support Vector Machine (SVM), K-Nearest Neighbors (K-NN),
Decision Trees, Discriminant Analysis, and Ensemble classifiers using MATLAB, were evaluated for
their performance in vehicle classification. We used the features extracted from the images of bikes,
cars, and trucks, which were classified into different categories. The accuracy of each classifier was
measured using cross-validation on the dataset.

28. 48
The results of our experiments are summarized in Table 2, which shows that the medium Gaussian
SVM classifier achieved the highest accuracy of nearly 91% for vehicle identification. These findings
demonstrate the effectiveness of the proposed approach for vehicle classification and suggest that a
medium Gaussian SVM classifier may be a suitable choice for similar applications.
Table 1. Number of iterations and the computational time for different images
Annexure II presents the results of vehicle classification for bikes (B), cars (C), and trucks (T) in terms
of true positive rate (TPR), false positive rate (FPR), positive predictive value (PPV), and false
discovery rate (FDR) as percentages. We used different decision tree classifiers to identify the three
types of vehicles, and the confusion matrix in the first row was used to assess the classification
performance.
The diagonal elements in the confusion matrix indicate the correct identification of the selected
vehicles. These results provide valuable insights into the effectiveness of the proposed approach for
vehicle classification and can be used to optimize the system's performance in real-world applications
Sr .No. Number of Test
Images
Bag of Visual words Number of
Iterations
Time Required (S)
1 100 250 25 84.755746
350 17 76.971196
450 20 80.409061
2 200 250 26 178.666412
350 17 157.776405
450 28 196.360222
3 300 250 22 256.148293
350 18 237.015765
450 30 341.758059
4 400 250 18 365.215308
350 25 359.875082
450 30 393.088333
5 500 250 28 523.370598
350 29 497.012188
450 18 442.437712
6 600 250 27 573.092221
350 24 621.244559
450 37 697.177003
7 700 250 75 1164.067380
350 50 908.966623
450 35 855.142737
8 800 250 22 859.710340
350 20 843.597179
450 19 826.786462
9 900 250 38 1083.519846
350 22 943.515072
450 19 1016.507187
10 1000 250 18 1024.145876
350 27 1075.037355
450 23 1130.071606

28. 49
Table 2. Accuracy of different classifiers.
Testing on random images
We assessed the performance of different classifiers for predicting the type of vehicle in a randomly
selected image. Specifically, we used the medium Gaussian SVM classifier to test the system on a set
of random images, and the results are presented in Figures (4) and Figure (5). These figures
demonstrate the effectiveness of the proposed approach for vehicle classification and illustrate the
classification results for a range of images. The results suggest that the medium Gaussian SVM
classifier may be a suitable choice for similar applications, and provide important insights into the
system's performance under realistic conditions.
Figure 4. True detection of bike and car
Sr .
No. Classifier Name accuracy )%( prediction speed (Obs/Sec)
1
Decision Trees
Fine Trees 68.6 2600
2 Medium Trees 66.3 2500
3 Coarse Tress 60.7 3100
4
Discriminant
Analysis
Linear Discriminant 88.1 2400
5
Quadratic
Discriminant 80.4 2000
6
Support Vector
Machines
(SVM)
Linear SVM 89 2100
7 Quadratic SVM 90 790
8 Cubic SVM 90.3 660
9 Fine Gaussian SVM 43.7 250
10
Medium Gaussian
SVM 90.8 530
11 Coarse Gaussian SVM 88.5 550
12
K-Nearest
Neighbor
Classifier
(KNN)
Fine KNN 77.7 220
13 Medium KNN 80.9 260
14 Coarse KNN 81.1 260
15 Cosine KNN 81.7 240
16 Cubic KNN 79.8 10
17 Weighted KNN 82.3 270
18
Ensemble
Classifier
Boosted Trees 78.1 1300
19 Bagged Trees 83.3 1100
20
Subspace
Discriminant 88.4 630
21 Subspace KNN 82.6 20
22 RUSBoosted Trees 65.6 1500

28. 50
Figure 5. True & false detection of the truck
DISCUSSION
The images of bikes, cars, and trucks were classified and trained using Speeded Up Robust Features
(SURF) and k-means clustering algorithms, with different visual feature word sizes of 250, 350, and
450 in MATLAB (as shown in Annexure-I). The most significant features were selected from each
vehicle class, and the total number of features was calculated using MATLAB. While some features
remained unchanged regardless of the visual feature word size selection, others varied with the change
in visual feature word size. This study compared five object classifiers for the three vehicle categories:
bike, car, and truck, using the identified features. The results demonstrate the effectiveness of the
proposed approach for vehicle classification and provide valuable insights into the system's
performance. These findings can be used to optimize the system for real-world applications, and the
approach can be extended to other types of objects beyond vehicles.
After identifying the features of the objects, we applied twenty-two classifiers to classify different
types of vehicles. These classifiers comprised sub-classifiers from five leading classification
algorithms: Decision Tree (three sub-classifiers), Discriminant Analysis (two sub-classifiers), Support
Vector Machines (six sub-classifiers), K-Nearest Neighbors (six sub-classifiers), and Ensemble
Classifiers (five sub-classifiers). The vehicle dataset was used for training and classification into three
categories: bikes, cars, and trucks. The medium Gaussian SVM classifier method of SVM
demonstrated superior accuracy compared to the other twenty-one classifiers. Therefore, we
recommend using the medium Gaussian SVM classifier for accurate vehicle recognition.
Variations in image resolution can impact the accuracy of inbuilt architecture and network results. To
mitigate this issue, we created a custom dataset and evaluated it using a Convolutional Neural
Network (CNN), as described by Trivedi et al. [26]. However, our current machine learning approach
provided better accuracy for vehicle classification of bikes, cars, and trucks than the CNN approach,
particularly for smaller datasets. Our machine learning approach yielded high accuracy, with the bike
classification achieving a 99% prediction accuracy and the highest confidence values. In contrast, the
truck classification demonstrated a lower highest confidence value than bikes and cars, with a
confidence value of 85%.
Next, we applied various classifiers with our machine-learning approach to the same vehicle dataset
used in this study. Prior work by Janak et al. [27] has discussed vehicle counting using morphology in
different real-time traffic videos. However, the method was less accurate for congested traffic,
resulting in a false count of on-road vehicles. To overcome this limitation, we focused on classifying
three passenger vehicles using a machine-learning approach for static images only.

28. 51
The present study has certain limitations. First, the results may vary for different iterations of trained
datasets with varying environmental conditions. Additionally, all the images were resized to achieve
better evaluation, which may require additional computation to implement these classifiers in real time
directly. It may be necessary to increase the number of iterations and the size of the image dataset to
increase the vehicle recognition process. It is worth noting that the Math Works website classifies the
speed measurements as fast, medium, and slow for 0.01 s, 1 s, and 100 s, respectively, while memory
usage is defined as small, medium, and large for data sizes of 1 MB, 4 MB, and 100 MB, respectively.
The classifiers' memory usage and prediction speed differ based on the selected algorithm. For the
ensemble classifier, the speed and memory usage can range from medium to fast for SVM, high to
low, or vice versa. For tree classifiers, memory usage is small, and speed is fast. The discriminant
analysis classifier can have small or large memory usage, and the speed can be fast or slow depending
on the selected method. KNN has medium memory usage and speed. It is worth noting that the
prepared dataset differs from deep convolutional architecture networks, such as AlexNet, VGGNet,
and ResNet. This study relies on hand-crafted features like SIFT and SURF, whereas CNN can learn
features from data (images) and derive scores from the output that may impact the generalizability of
the study's findings.
CONCLUSION AND FUTURE SCOPE
In this study, the accuracy and prediction speed of twenty-two different classifiers were evaluated
using a machine learning approach for vehicle recognition of three categories - bikes, cars, and trucks.
The medium Gaussian SVM classifier was more accurate than the other twenty-one classifiers. The
study also compared the performance of different classifiers on image datasets of varying sizes and
iteration times, as presented in Table 1 and 2. Additionally, the study compared the machine-learning
approach with the deep-learning approach, as discussed in earlier work using CNN [26], for the same
image dataset. The accuracy of the deep-learning approach was measured using a confidence value,
while the machine-learning approach measured accuracy using TPR, FPR, PPV, and FDR values. The
twenty-two classifiers were used to identify bikes, cars, and trucks.
The study found that SVM and KNN classifiers have True Positive Rates of almost 90%, while the
Positive Predictive Values in SVM classifiers are above 93%. The False Positive Rates are almost
below 10% in SVM classifiers, and False Discovery Rates are below 8% in SVM classifiers. Among
the classifiers, the medium Gaussian SVM classifier provides the best accuracy of around 91%.
Furthermore, the medium Gaussian SVM classifier was verified and tested on random images, as
shown in Figures (4) - (5), which present the predicted and actual objects.
In summary, this study shows that the machine learning approach is effective for vehicle classification,
especially when working with smaller amounts of data. The results of this approach outperform those
of a deep learning approach, specifically a CNN, for the same dataset. However, the machine learning
approach requires structured data and label information, whereas CNN can work with unstructured
data and learn from it. In addition, the machine learning approach requires more computational time
when working with larger amounts of data, while CNN requires more data to achieve accurate results.
Therefore, the approach choice should be based on the specific data and the problem at hand.
The present study demonstrates the potential of machine learning approaches for accurately classifying
different types of vehicles from static images. The future scope of this work is to extend the proposed
classifier module for real-time applications that classify various objects. In real-time scenarios,
classifying different objects requires the simultaneous extraction of feature information and image
localization. The proposed work can be further extended to enable the classification of different
objects with real-time scenarios using available surveillance systems. For this purpose, a more
extensive dataset of various environmental conditions and sizes can be prepared to train the proposed
models for accurate and robust object recognition in real time, which can lead to the development of
intelligent surveillance systems capable of real-time object recognition with applications in various
fields, including security, transportation, and healthcare.

28. 52
In addition to the limitations mentioned, it should also be noted that the proposed work uses hand-
crafted features like SIFT and SURF, which may not be as robust as the features learned by deep
convolutional neural networks. Future work could investigate the use of deep learning methods to
extract features directly from the image data.
The offered procedure in the study is a machine learning-based vehicle recognition and classification
approach, which has several advantages and limitations. One of the significant advantages is that it can
achieve high accuracy in vehicle recognition with smaller amounts of data compared to deep learning
methods like CNN. Additionally, it provides a more structured approach to data processing, requiring
labeled information to classify objects accurately. However, it also has limitations, including a longer
computational time when working with larger amounts of data and using hand-crafted features that
may not be as robust as deep learning-based features.
The suggested approach has practical applicability in various traffic flow situations, including
continuous traffic flow, parking lots, toll facilities, interchanges and intersections, and identifying
congestion on urban roads. In a continuous traffic flow situation, the proposed method can monitor
traffic and recognize different types of vehicles, such as cars, trucks, and bikes, for better traffic
management. In parking lots, the proposed method can help identify different vehicles and prevent
unauthorized parking while in toll facilities, it can be integrated into the automatic toll collection
system for better efficiency and accuracy.
The presented solution can be used for traffic control and monitoring at interchanges and intersections
to prevent accidents and traffic jams. Additionally, it can identify urban road congestion and adjust
traffic signals and routes accordingly for better traffic management. Overall, the proposed method can
contribute significantly to developing intelligent surveillance systems capable of real-time object
recognition, with practical applications in various fields, including transportation and security.
Furthermore, while the proposed work focuses on vehicle classification, it may generalize poorly to
other object categories. Thus, future work could explore the applicability of this approach to other
object recognition tasks. Additionally, the proposed work assumes that images are captured from a
stationary camera, which may not be the case in real-world scenarios. Future work could investigate
this approach's use in moving camera scenarios.
Finally, the proposed work has potential practical applications in the field of transportation,
particularly in developing more efficient and accurate toll-collection systems. Further research could
explore integrating this approach into such systems to improve their functionality and reduce traffic
congestion.
(Received February 2023, accepted February 2023)
REFERENCES
[1] Trivedi, J., Devi, M.S. and Dave, D. (2017), ” Review Paper on Intelligent Traffic Control system using
Computer Vision for Smart City”, Int. Jour. of Sci. & Engi. Res., Volume 8, Issue 6, pp.14-17.
[2] Russo, F., Rindone, C., Panuccio Panuccio P. (2016)," European plans for the smart city: from theories
and rules to logistics test case, European Planning Studies," 24:9, 1709-1726,
DOI:10.1080/09654313.2016.1182120
[3] Croce, A., Musolino, G., Rindone, C., and Vitetta A. (2019), “Transport System Models and Big Data:
Zoning and Graph Building with Traditional Surveys, FCD and GIS”, Int. J. Geo-Inf., 8(4), 187,
DOI:10.3390/IJGI8040187
[4] Hardin, P. and N. Thomson, C. (1992), "Fast Nearest neighbor Classification Methods for Multispectral
Imagery", Professional Geographer 44 (2): 191–202, DOI:10.1111/J.0033-0124.1992.00191.X
[5] Ghosh, A. K. (2007), "On Nearest Neighbor Classification Using Adaptive Choice of K" Journal of
Computational and Graphical Statistics 16 (2): 482–502, DOI:10.1198/106186007X208380
[6] Breiman, L. (2001), “Random Forests”, Machine Learning 45, 5–32,
[7] DOI: https://doi.org/10.1023/A:1010933404324
[8] Loh, W. (2002), "Regression Trees with Unbiased Variable Selection and Interaction Detection",
Statistica Sinica 12, 361-386.

28. 53
[9] Paensuwan, N., Yokoyama, A., C. Verma, S., and Nakachi Y. (2012) "Application of Decision Tree
Classification to the Probabilistic TTC Evaluation" Journal of International Council on Electrical
Engineering 1(3):323–30, DOI:10.5370/JICEE.2011.1.3.323
[10] Wen, X., Shao, L., Xue, Y., Fang, W. (2015), "A Rapid Learning Algorithm for Vehicle Classification"
Information Sciences 295: 395–406, DOI:10.1016/j.ins.2014.10.040
[11] Kim, J. Jeonghyun Baek, J. and Kim, E. (2015), "A Novel On-Road Vehicle Detection Method Using
ΠHOG" IEEE Transactions on Intelligent Transportation Systems 16 (6): 3414–29,
DOI:10.1109/TITS.2015.2465296
[12] Stocker, M., Silvonen, P., Rönkkö, M., and Kolehmainen, M. (2016), "Detection and Classification of
Vehicles by Measurement of Road-Pavement Vibration and by Means of Supervised Machine Learning"
Journal of Intelligent Transportation Systems: Technology, Planning, and Operations 20 (2): 125–37,
DOI:10.1080/15472450.2015.1004063
[13] Arinaldi, A., Arya Pradana, J., and Arventa Gurusinga, A. (2018), “Detection and classification of
vehicles for traffic video analytics”, Procedia Computer Science 144:259-68,
DOI:10.1016/j.procs.2018.10.527
[14] Atiya, A. (2005), “Learning with Kernels: Support Vector Machines, Regularization, Optimization, and
Beyond”, IEEE Transactions on Neural Networks, 16 (3), DOI:10.1109/tnn.2005.848998
[15] Sarikan, S., Ozbayoglu, A., and Zilci, O. (2017), "Automated Vehicle Classification with Image
Processing and Computational Intelligence" Procedia Computer Science 114:515–22,
DOI:10.1016/J.PROCS.2017.09.022
[16] Gorges, C., Öztürk, K., Liebich R. (2018), "Road Classification for Two-Wheeled Vehicles" Vehicle
System Dynamics 56 (8): 1289–1314, DOI:10.1080/00423114.2017.1413197
[17] Bautista, M. C., Dy, A. C., Manalac, I. M., Orbe, A. R. and Cordel, M. (2016), "Convolutional neural
network for vehicle detection in low-resolution traffic videos," IEEE Region 10 Symposium, pp. 277-281,
DOI:10.1109/TENCONSPRING.2016.7519418
[18] Bansal, D., Khanna, K., Chhikara, R., Kumar Dua, R. and Malhotra, R. (2020), “Classification of
Magnetic Resonance Images using Bag of Features for Detecting Dementia”, Procedia Computer Science
167, pp. 131–137, DOI:10.1016/j.procs.2020.03.190
[19] Pranata, Y. D., Wang, K.C., Wang, J.C., Idram, I., Lai, J.Y., Liu, J.W., and Hsieh, I.H. (2019), “Deep
learning and SURF for automated classiﬁcation and detection of calcaneus fractures in CT images,”
Computer Methods and Programs in Biomedicine 171, pp.27–37, DOI:
https://doi.org/10.1016/j.cmpb.2019.02.006
[20] Sykora, P., Kamencay, P., and Hudec, R. (2014), “Comparison of SIFT and SURF Methods for Use on
Hand Gesture Recognition Based on Depth Map”, AASRI Conference on Circuits and Signal Processing
(CSP 2014), pp- 19 – 24, DOI:10.1016/j.aasri.2014.09.005
[21] Ignat, A. and Pavaloi, I. (2021), “Keypoint Selection Algorithm for Palmprint Recognition with SURF”,
25th Inter. Conf. on Knowledge-Based and Intelligent Information & Engineering Systems, Procedia
Computer Science 192, 270–280, DOI: 10.1016/j.procs.2021.08.028
[22] Anzid, H., Le Goic, G., Mammass, D., Bekkari, A. and Mansouri, A. (2019), “Multimodal Images
Classiﬁcation using Dense SURF, Spectral Information and Support Vector Machine”, Second
International Conference on Intelligent Computing in Data Sciences, Procedia Computer Science 148, pp-
107–115, DOI:10.1016/j.procs.2019.01.014
[23] LoweDavid, G. (2004), “Distinctive Image Features from Scale-Invariant Keypoints”, International
Journal of Computer Vision 60 (2): 91–110.
[24] Bay, H., Tuytelaars, T., and Gool L. (2006), “SURF: Speeded Up Robust Features.” Computer Vision,
ECCV 404–17, DOI:10.1007/11744023_32
[25] https://in.mathworks.com/help/stats/choose-a-classifier.html.
[26] https://www.analyticsvidhya.com/blog/2021/10/support-vector-machinessvm-a-complete-guide-for-
beginners/
[27] Trivedi, J., Devi, M.S. and Dave, D. (2021), “Vehicle classification using the convolution neural network
approach,” Scientific Journal of Silesian University of Technology. Series Transport, 112, 201-209.
ISSN: 0209-3324. DOI: https://doi.org/10.20858/sjsutst.2021.112.7.16.
[28] Trivedi, J., Devi, M.S. and Dave, D. (2018), “Vehicle Counting Module Design in Small Scale for Traffic
Management in Smart City”, IEEE, 3rd Int. Conf. for Convergence in Tech. (I2CT), DOI:
10.1109/I2CT.2018.8529506.

28. 54
ATTACHMENTS
Annexure-I Number of features for different vehicles.
Sr .
No.
Number of
test images
Image class /Extracted Number features
Bikes
A
Cars
B
Trucks
C
Total Number of
Features
1 100 95783 67389 77560 161733
2 200 201081 140479 157553 337149
3 300 294684 204318 232975 490362
4 400 399111 273086 305896 655407
5 500 506155 344331 388061 826395
6 600 603543 420279 465669 1008669
7 700 696414 489869 549068 1175685
8 800 789421 552722 637420 1326534
9 900 886088 622349 724289 1493637
10 1000 986129 688459 826715 1652301

28. 55
Annexure-II Vehicle classification using different classifiers.
Decision Trees
Fine Trees Medium Trees Coarse Trees
B C T
TPR
(
)%
FPR
)%( B C T
TPR
)%(
FPR
)%( B C T
TPR
)%(
FPR
)%(
B 1170 296 134 73 27 1153 336 111 72 28 966 520 114 60 40
C 231 1013 356 63 37 241 1023 336 64 36 222 940 438 59 41
T 114 376 1110 69 31 98 498 1004 63 37 114 479 1007 63 37
PPV )%( 77 60 69 77 55 69 74 48 65
FDR )%( 23 40 31 23 45 31 26 52 35
Discriminant Analysis
Linear Discriminant Quadratic Discriminant
B C T
TPR
)%(
FPR
)%( B C T
TPR
)%(
FPR
)%(
B 1466 109 25 92 8 1340 242 18 84 16
C 78 1353 169 85 15 95 1450 55 91 9
T 20 168 1412 88 12 65 466 1069 67 33
PPV )%( 94 83 88 89 67 94
FDR )%( 6 17 12 11 33 6
Support Vector Machines (SVM)
Linear SVM Quadratic SVM Cubic SVM
B C T
TPR
)%(
FPR
)%( B C T
TPR
)%(
FPR
)%( B C T
TPR
)%(
FPR
)%(
B 1471 109 20 92 8 1489 96 15 93 7 1496 85 19 94 6
C 77 1376 147 86 14 75 1382 143 86 14 75 1386 139 87 13
T 22 151 1427 89 11 15 135 1450 91 9 14 136 1450 91 9
PPV )%( 94 84 90 94 86 90 94 86 90
FDR )%( 6 16 10 6 14 10 6 14 10
Fine Gaussian SVM Medium Gaussian SVM Coarse Gaussian SVM
B C T
TPR
)%(
FPR
)%( B C T
TPR
)%(
FPR
)%( B C T
TPR
)%(
FPR
)%(
B 209 1391 0 13 87 1488 95 17 93 7 1467 114 19 92 8
C 0 1600 0 100 0 57 1434 109 90 10 79 1376 145 86 14
T 0 1313 287 18 82 17 147 1436 90 10 33 163 1404 88 12
PPV )%( 100 37 100 95 86 92 93 83 90
FDR )%( 0 63 0 5 14 8 7 17 10
Nearest Neighbor Classifier (NN)
Fine KNN Medium KNN Coarse KNN
B C T
TPR
(
)%
FPR
)%( B C T
TPR
)%(
FPR
)%( B C T
TPR
)%(
FPR
)%(
B 1431 105 64 89 11 1480 80 40 93 7 1457 106 37 91 9
C 212 986 402 62 38 200 1123 277 70 30 182 1218 200 76 24
T 76 210 1314 82 18 87 232 1281 80 20 37 285 1218 76 24
PPV )%( 83 76 74 84 78 80 84 76 84
FDR)%( 17 24 26 16 22 20 16 24 16
Cosine KNN Cubic KNN Weighted KNN
B C T
TPR
)%(
FPR
)%( B C T
TPR
)%(
FPR
)%( B C T
TPR
)%(
FPR
)%(
B 1517 47 36 95 5 1501 56 43 94 6 1483 75 42 93 7
C 252 1054 294 66 34 282 1001 317 63 37 179 1108 313 69 31
T 111 137 1352 85 15 118 156 1326 83 17 70 173 1357 85 15

28. 56
PPV )%( 81 85 80 79 83 79 86 82 79
FDR )%( 19 15 20 21 17 21 14 18 21
Ensemble Classifier
Boosted Trees Bagged Trees Subspace Discriminant
B C T
TPR
)%(
FPR
)%( B C T
TPR
)%(
FPR
)%( B C T
TPR
)%(
FPR
)%(
B 1340 203 57 84 16 1456 125 19 91 9 1478 100 22 92 8
C 178 1176 246 74 26 176 1232 192 77 23 81 1364 155 85 15
T 74 291 1235 77 23 74 217 1309 82 18 22 179 1399 87 13
PPV )%( 84 70 80 85 78 86 93 83 89
FDR )%( 16 30 20 15 22 14 7 17 11
Subspace KNN RUSBoosted Trees
B C T
TPR
)%(
FPR
)%( B C T
TPR
)%(
FPR
)%(
B 1516 49 35 95 5 1147 351 102 72 28
C 221 1052 327 66 34 277 967 336 62 38
T 75 130 1395 87 13 102 484 1014 63 37
PPV )%( 84 85 79 75 54 70
FDR )%( 16 15 21 25 46 30

STEP TOWARDS INTELLIGENT TRANSPORTATION SYSTEM WITH VEHICLE CLASSIFICATION AND RECOGNITION USING SPEEDED-UP ROBUST FEATURES

More Related Content

Similar to STEP TOWARDS INTELLIGENT TRANSPORTATION SYSTEM WITH VEHICLE CLASSIFICATION AND RECOGNITION USING SPEEDED-UP ROBUST FEATURES (20)

More from JANAK TRIVEDI (8)

Recently uploaded (20)

STEP TOWARDS INTELLIGENT TRANSPORTATION SYSTEM WITH VEHICLE CLASSIFICATION AND RECOGNITION USING SPEEDED-UP ROBUST FEATURES