International Journal of Reconfigurable and Embedded Systems (IJRES)

Vol. 14, No. 1, March 2025, pp. 89~99

ISSN: 2089-4864, DOI: 10.11591/ijres.v14.i1.pp89-99  89

Comparative analysis of feature descriptors and classifiers for

real-time object detection

Vikas J. Nandeshwar1, Sarvadnya Bhatlawande1, Anjali Solanke2, Harsh Sathe1, Shivanand Satao1,
Safalya Satpute1, Atharv Saste1
1
Department of Engineering Sciences and Humanities (DESH), Vishwakarma Institute of Technology, Pune, India
2
Department of Electronics and Telecommunication Engineering, Marathwada Mitra Mandal’s College of Engineering, Pune, India

Article Info ABSTRACT

Article history: Detecting objects within complex environments, such as urban settings,
holds significant importance across various applications, including driver
Received Apr 9, 2024 assistance systems, traffic monitoring, and obstacle detection systems.
Revised Oct 4, 2024 Particularly crucial for these applications is the accurate differentiation
Accepted Oct 18, 2024 between cars and roads. This study introduces a novel approach that
leverages traditional feature descriptors and classifiers for real-time object
detection. It conducts an exhaustive comparative analysis of feature
Keywords: descriptors and classifiers to identify the most effective model for real-time
object detection. Handcrafted features of images are extracted using
Classifiers algorithms such as scale invariant feature transform (SIFT), oriented fast and
Comparative analysis brief (ORB), fast retina key-point (FREAK), and local binary pattern (LBP).
Computer vision Seven classifiers are employed, including support vector machine (SVM),
Driver assistance system K-nearest neighbors (KNN), random forest (RF), decision tree (DT), logistic
Feature descriptors regression (LR), Naive Bayes, and extreme gradient boosting (XGBoost).
Obstacle detection systems The performance of the 28 generated combinations of feature descriptors and
Supervised machine learning classifiers is evaluated based on the parameters of accuracy, precision, F1
score, and recall. The model utilizing LBP and XGBoost achieves the
highest accuracy, reaching 83.59%. The system architecture comprises a
camera, a high-speed computing unit, a display, and an audio subsystem,
with the algorithm implemented on a Raspberry Pi 4B (8 GB).
This is an open access article under the CC BY-SA license.

Corresponding Author:
Anjali Solanke
Department of Electronics and Telecommunication Engineering
Marathwada Mitra Mandal’s College of Engineering
Pune, Maharashtra, India
Email: [email protected]

1. INTRODUCTION
The task of object detection holds pivotal significance within the field of computer vision, finding
extensive utility across diverse sectors. As per the findings of the U.S. National Highway Traffic Safety
Administration, a substantial majority, exceeding 88% of traffic incidents arise due to errors in judgment or
delayed reactions on the part of drivers [1]. As a solution, an intelligent driver assistance system for road
safety must be developed which would alert drivers in dangerous situations [2]. Technological advancements
in embedded systems and vehicle electronics are being driven by the goal of enhancing vehicle safety and
energy efficiency [3]. The emergence of real time technologies like driver assistance systems and obstacle
avoidance systems has highlighted the need for precise and instantaneous object detection methodologies.
Notably, in these application domains, vehicles and roadways emerge as central entities, having profound
influence over transportation networks and safety frameworks.

Journal homepage: http://ijres.iaescore.com

90  ISSN: 2089-4864

This study is dedicated to devising an object detection framework ported on an embedded system for
instantaneous segmentation of cars and roads. The method employs hand crafted features extracted from
traditional feature descriptors and statistical machine learning models. These models are assessed based on
the parameters of accuracy, precision, F1 score, and recall. Through combination of traditional feature
descriptors with classifiers, our objective is to enhance the effectiveness of object detection in real-time
environments while maintaining the rapid response times and low computational power both critical for real-
time applications.

2. LITERATURE SURVEY
In recent years, deep learning algorithms like region-based convolutional neural networks (R-CNN),
you only look once (YOLO), and single shot multibox detector (SSD) have gained widespread adoption
across various domains for object detection tasks [4]. Convolutional neural networks (CNNs) are widely
recognized as one of the most representative and influential models in the domain of deep learning [5]. A
method enhancing faster R-CNN for obstacle detection and recognition using U-V disparity maps and an
improved network structure involving context aware modules trained on datasets from KITTI and CCD
cameras is proposed in [6]. Dairi et al. [7] introduced a stereo vision method for obstacle detection in urban
environments. They utilized a deep stacked auto-encoders (DSA) model and an unsupervised K-nearest
neighbors (KNN) algorithm to detect obstacles reliably. A YOLO-v5-based object detection model capable
of detecting objects of various sizes, including large, small, and tiny objects utilizing a multi-scale
mechanism to adaptively determine the optimal scales for vehicle detection within a scene was proposed by
Carrasco et al. [8]. Zhou et al. [9] proposed a road detection and tracking algorithm that accurately delineates
the road region using a graph-cut-based approach. It employs a fast homography-based tracking technique,
utilizing features from scale invariant feature transform (SIFT) and speeded up robust features (SURF)
algorithms alongside the fast feature detector. A system designed to identify drivable and non-drivable roads,
with the goal of mitigating traffic accidents was introduced which employed a combination of SIFT and
binary robust invariant scalable key-points (BRISK) algorithms to extract features. These extracted features
were subsequently input into classifiers. The system attained a 70.9% accuracy rate when utilizing the
support vector machine (SVM) with a radial basis function (RBF) kernel in [10]. An algorithm utilizing
principal component analysis-SIFT (PCA-SIFT), exploiting scale-invariant key points to ensure rotation and
scale invariance was introduced where the PCA method, based on variable covariance matrices, efficiently
handles, compresses, and extracts feature vectors was proposed in [11].
Retraining a YOLOv3 model for vehicle detection using a dataset of 400 unmanned aerial vehicle
(UAV) images on a graphics processing unit (GPU), while demonstrating the computational demands of deep
learning algorithms like YOLO, the model exhibited occasional failure to detect objects and generated false
positives in some instances [12]. This article presents a comparative study of three detectors histogram of
oriented gradient (HOGs), local binary pattern (LBP), and Haar features, for the detection of animals. These
features were individually trained on a dataset using the AdaBoost algorithm. LBP-AdaBoost/HOG-SVM
demonstrated effective performance under daytime conditions. However, limitations were observed during
the nighttime operation [13]. An ensemble learning-based decision model for predicting vehicle lane-
changing behavior employing random forest (RF), demonstrated higher precision in its predictions [14]. An
embedded system for traffic surveillance that utilizes the NVIDIA Jetson TX1, integrating a deep detector
named MF faster R-CNN which can identify various classes of traffic objects, including pedestrians, cars,
buses, and motorbikes, concurrently was proposed in [15]. Zhang [16] proposed that employing the Canny
edge detection algorithm can yield high accuracy as it identifies and removes unstable edge points. A
solution to integrate multiple low-cost sensors, including infrared and ultrasonic technologies, offering
improved reliability and reduced mathematical complexity was introduced in [17]. A lossy compression
framework for input images was proposed to reduce memory traffic and power consumption, balancing
accuracy and compression performance [18]. Conversion of image-based depth maps to light detection and
ranging (pseudo-LiDAR) representations for object detection was suggested. This approach achieved a 74%
detection accuracy within a 30 m range [19]. A method exploiting K means and PCA for large-scale data sets
was presented that leads to both computational time and memory benefits. It utilized randomized
preconditioning transformations to achieve accurate data sparsification, reducing variance in estimates [20].
A system utilizing ATMega328p microcontroller, ultrasonic sensor for object detection, MQ3 sensor for
alcohol detection and Image processing to detect driver’s drowsiness is proposed in [21]. A configuration
including Raspberry Pi board-based control system, a monocular camera and an earphone for recognizing the
position of an object was exploited in [22]-[24]. SIFT feature descriptor for feature extraction and classifiers
such as KNN and RF are leveraged in [25]. A security system integrating facial features from FaceNet and
Mediapipe for face detection was observed to achieve an accuracy of 80.5% [26].

Int J Reconfigurable & Embedded Syst, Vol. 14, No. 1, March 2025: 89-99
Int J Reconfigurable & Embedded Syst ISSN: 2089-4864  91

3. METHOD
This research paper presents a real time system designed for the classification and detection of
vehicles on roadways, comprising a camera and a fast-computing processor-based system, as illustrated in
Figure 1. The system captures the data from its environment, which is then provided as an input to the
fast-computing processor. A specialized obstacle detection algorithm, implemented on a Raspberry Pi 4B
(8 GB) board, is designed to understand and interpret road surroundings. This algorithm segments vehicles
and roadways, deduces priority information, and provides details about the drivable road. The system
translates vehicle presence into textual and auditory feedback, which is then relayed to the driver through a
compact audiovisual system integrated into the car's dashboard, ensuring efficient communication.

Figure 1. System for detection of cars and roads

3.1. Details of the dataset and pre-processing

The dataset utilized in this study comprises a total of 5,400 images, encompassing 2,700 positive
images depicting cars on roads and an equal number of negative images depicting empty roadways. The
images in the dataset were curated by the authors. The authors employed a 50-megapixel camera on an
android mobile phone to capture vehicle photographs in landscape orientation. The composition of images in
the dataset is presented in Table 1. Each image within the dataset underwent a standard resizing process,
resulting in dimensions of 280×430 pixels, and was subsequently converted to grayscale. The edge detection
algorithms of Prewitt and Canny were applied uniformly across the entire dataset. These techniques were
employed to identify and extract edges from the images while minimizing noise and extraneous elements,
thereby enhancing the overall visual quality of the images.

Table 1. Details of the dataset

Image type Train images Test images
Positive images 2160 540
Negative images 2160 540
Total images 4320 1080

3.2. Feature extraction

The traditional feature descriptors SIFT, ORB, FREAK, and LBP were employed to extract key-
points and hand-crafted features for each image 𝑥𝑖 in the dataset (where i ranges from 0 to 𝑁-1 where, N
represents the total number of images, i.e., 5400). SIFT was opted for due to its well-established robustness
and resilience to affine transformations and noise change factors [27]. ORB was selected for its combination
of the FAST detector's speed with the BRIEF descriptor's durability, offering rotation invariance and real-
time performance [28]. Fast retina key-point (FREAK) was selected owing to its accelerated performance
compared to SIFT and SURF, leveraging insights from human visual subregion structure to obtain
information [29]. Notably, SIFT key-points were provided to the FREAK descriptor in this study. LBP was
preferred due to its computational efficiency, ease of implementation, and generation of a concise feature
vector. Additionally, LBP finds widespread application in texture analysis and facial recognition tasks [30]. 𝐹
represents the set of feature descriptors {SIFT, ORB, FREAK, and LBP} where the application of a feature
descriptor 𝐹𝑗 on 𝑥𝑖 image generates a feature vector 𝐹𝑗 (𝑥𝑖 ) representing the image's characteristics. The
feature extraction process is denoted by (1):

Comparative analysis of feature descriptors and classifiers for real-time … (Vikas J. Nandeshwar)
92  ISSN: 2089-4864

𝑉𝑙 = ⋃𝑁−1
𝑖=0 𝐹𝑗 (𝑥𝑖 ) (1)

where 𝑉𝑙 denotes the feature vector obtained by appending the feature vectors of 𝑥𝑖 images. The feature
vectors generated by different feature descriptors (𝐹𝑗 ) are as follows: SIFT produces a vector with dimensions
(3,419,678 rows×128 columns), ORB generates one with (3,314,807 rows×32 columns), FREAK results in
(2,731,168 rows×64 columns), and LBP creates a vector of (5,400 rows×26 columns).

3.3. Feature transformation

The resulting feature vectors obtained from each descriptor tend to be high-dimensional, presenting
challenges such as heightened computational complexity and vulnerability to overfitting. To mitigate these
challenges, a two-step novel approach for feature transformation, encompassing K-means clustering followed
by PCA was adopted. K-means clustering was employed, which is an unsupervised machine learning
technique, to group similar data points into a predetermined number of clusters (K), determined using the
elbow method [31], resulting in K=14. Following the application of the K-means clustering algorithm, the
initial high dimensional feature vectors obtained from feature descriptors were transformed into a reduced
dimensionality of (5,400×14). PCA is further utilized to acquire uncorrelated features and reduce the
dimensionality of the feature vector while maintaining the essential structure and variability of the original
data. Two scenarios are investigated: one focused on preserving 90% of the information and the other
targeting 95% within the transformed feature vectors. The determination of the retained information
percentage is performed by analyzing the cumulative explained variance ratio and subsequently transforming
the feature vector through the selection of principal components (𝑝). The overall process of dimension
reduction is shown in Figure 2.

Figure 2. Workflow for feature transformation process

This dimension reduction can be denoted by (2):

𝛹(𝑉𝑙 ) = (𝑉𝑙 ∗ 𝐶[: , 0: 𝑝]) (2)

where, 𝛹(𝑉𝑙 ) represents the final dimensionally reduced feature vector and 𝐶 denotes the principal
components matrix. After obtaining the final feature vector, labels are appended to the last column to prepare
the data for subsequent training and testing of classifiers. Specifically, a label of 0 represents the class "Car
on Road," while a label of 1 signifies “Not Car on Road”. The final feature vector obtained after the
application of K-Means and PCA is presented in Table 2.

Table 2. Final feature vector obtained after applying K-Means and PCA
Case 1: (90% of information retained) Case 2: (95% of information retained)
Sr No.
Feature descriptor Feature vector Feature descriptor Feature vector
1. SIFT 5,400×5 SIFT 5,400×9
2. ORB 5,400×9 ORB 5,400×12
3. FREAK 5,400×6 FREAK 5,400×10
4. LBP 5,400×19 LBP 5,400×23

Int J Reconfigurable & Embedded Syst, Vol. 14, No. 1, March 2025: 89-99
Int J Reconfigurable & Embedded Syst ISSN: 2089-4864  93

3.4. Object detection and classification

The feature vector obtained post PCA was divided into training and testing the data with a ratio of
80:20. The training set, comprising 80% of the data, was utilized to train the classifiers, while the remaining
20% served as the testing set for evaluation. The classifiers employed in this study include: i) decision tree
(DT) with a maximum depth set to 13; ii) RF with the number of estimators set to 100; iii) RBF kernel-SVM;
iv) KNN with a parameter for the number of neighbors set to 20; v) logistic regression (LR); vi) extreme
gradient boosting (XGBoost); and vii) Gaussian Naive Bayes (GNB). These classifiers were aimed at
detecting obstacles categorized into two classes: "Car on Road" (label 0) and "Not Car on Road" (label 1). G
represents the set of classifiers where G={DT, RF, RBF-SVM, KNN, LR, XGBoost, and GNB}. The
classification process can be represented by (3):

𝑐𝑖𝑘 = 𝐺𝑘 (𝐹𝑗 (𝑥𝑖 ))) (3)

where 𝑐𝑖𝑘 denotes the predicted class label for the i-th image by the k-th classifier from the set G
and 𝐺𝑘 (𝐹𝑗 (𝑥𝑖 )) represents the model of the j-th feature descriptor from F and the k-th classifier from the set
G. Considering feature descriptors (SIFT, ORB, FREAK, and LBP) and classifiers (SVM, DT, RF, LR,
KNN, XGBoost, and GNB), the total number of models (|𝑀|) generated can be expressed using (4):

| 𝑀| = ( |𝐹| 𝐶𝑓 ∗ |𝐺|𝐶 𝑔 ) (4)

|𝐹|! |𝐺|!
| 𝑀| = ∗ (5)
𝑓 ! ∗ (|𝐹|−𝑓)! 𝑔 ! ∗ (|𝐺|−𝑔)!

where, |𝐹| and |𝐺| denote the total number of feature descriptors and classifiers in sets F and G, respectively.
Conversely, 𝑓 and 𝑔 denote the number of feature descriptors and classifiers selected from sets F and G. The
values of, 𝑓 and 𝑔 always remains one as the total number of models (|𝑀|) produced are evaluated on a set of
parameters (𝑃𝑀), which include accuracy, precision, F1, and recall scores. Figure 3 showcases how the |𝑀|
i.e., 28 combinations of feature descriptors and classifiers are generated. In (5) provides a detailed
mathematical foundation for (4), explaining how the total number of models ∣M∣ is computed by considering
all possible combinations of feature descriptors and classifiers from sets F and G.

Figure 3. Illustrates the generation of 28 combinations of feature descriptors and classifiers

.
The performance of classifiers such as KNN, RF, and DT was found to exhibit minimal
improvements when their respective hyperparameters were optimized to enhance accuracy. This
hyperparameter tuning process encompassed experimenting with diverse parameter settings, including
modifying K-neighbor values for KNN, optimizing the number and depths of trees for RF, and fine-tuning
the maximum depths of DT. However, the resultant changes in accuracy were surprisingly slight, generally
fluctuating within a narrow range of 1% to 1.4%. The hyperparameter values which provided the optimal
results were utilized in the respective models.

Comparative analysis of feature descriptors and classifiers for real-time … (Vikas J. Nandeshwar)
94  ISSN: 2089-4864

3.5. Comparative analysis

To determine the most reliable model for real-time object detection, each of the | 𝑀| combinations,
undergoes meticulous evaluation, which relies on an exhaustive examination of every combination's
performance across critical parameters of accuracy, F1 score, recall score, and precision. Through this
comparative analysis, insights into the effectiveness and robustness of each model are obtained, allowing for
the identification of the optimal feature descriptor-classifier combination suited for real-time object detection.
The novelty of this system lies in its utilization of a comparative analysis method aimed at identifying the
model (𝑀(𝐹𝑗 , 𝐺𝑘 )) of the j-th feature descriptor and k-th classifier from the sets F and G, with the paramount
gradings across all parameters. If the scores for the selected model decline, this comparative analysis is
performed again, which returns the model with superior scores than the previous model. This iterative
process can be described by (6):

𝑀𝑚𝑎𝑥 (𝐹𝑗 , 𝐺𝑘 ) = arg max ( ∑ 𝑤𝑃𝑀 𝑃(𝐹𝑗 , 𝐺𝑘 , 𝑃𝑀)) (6)

where, 𝑃𝑀 = {𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦, 𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛, 𝐹1, 𝑎𝑛𝑑 𝑅𝑒𝑐𝑎𝑙𝑙 𝑠𝑐𝑜𝑟𝑒𝑠 } and 𝑃(𝐹𝑗 , 𝐺𝑘 , 𝑃𝑀) defines the performance
scores of the combination of feature descriptor 𝐹𝑗 and classifier 𝐺𝑘 with respect to the parameters in the set
𝑃𝑀. Additionally, 𝑤𝑃𝑀 denotes the weight assigned to each parameter in the set 𝑃𝑀. The gradings of model
𝑀𝑚𝑎𝑥 across all parameters are observed and if there is a decline in these scores, (5) is utilized again for
ascertaining the model 𝑀𝑚𝑎𝑥 which has superior results across all metrics mentioned in set 𝑃𝑀. The
functioning of this framework ported on an embedded system is executed as shown in Algorithm 1.

Algorithm 1. Feature extraction and feature transformation

Input: Image dataset of N Images
Output: Classification of Car and Road
1. Data 𝐹𝑗 =[ ]; // data frame for features extracted by feature descriptors |𝐹|
2. For image in dataset do
3. Preprocessing
4. Extract SIFT, ORB, FREAK, and LBP features of 𝑥𝑖 images
5. 𝐷𝑎𝑡𝑎𝐹𝑗 =append features //for 𝐹𝑗 𝑡ℎ feature descriptor
6. 𝑉𝑙 =Data 𝐹𝑗 . Append (𝐹𝑗 (𝑥𝑖 ))
7. End For
8. For each image feature 𝐹𝑗 (𝑥𝑖 ) in 𝑉𝑙 :
9. 𝐾(𝑉𝑙 )=Pre-trained K-Means [K=14] ( 𝑉𝑙 )
10. //Dimension of 𝑉𝑙 is (5400 x 14)
11. Nm(𝑉𝑙 ) =Normalize (𝑉𝑙 );
12. Data 𝐹𝑗 .append(N)
13. End For
14. 𝛹(𝑉𝑙 ) = 𝑃𝐶𝐴(𝑉𝑙 ) [ with principal components (𝑝)]
15. //Dimension of 𝛹(𝑉𝑙 ) is (N x 𝑝 )
16. Train classifiers from 𝐺 on 0.8N images
17. 𝑐𝑖𝑘 = 𝐺𝑘 (𝐹𝑗 (𝑥𝑖 ))) //𝑐𝑖𝑘 represents the prediction
18. (𝐹𝑗 , 𝐺𝑘 ) = arg max ( ∑ 𝑤𝑃𝑀 𝑃(𝐹𝑗 , 𝐺𝑘 , 𝑃𝑀))
19. For each parameter in 𝑃𝑀:
20. if 𝑃𝑀𝑚𝑎𝑥 (𝐹𝑗 , 𝐺𝑘 , 𝑃𝑀) < 𝑃𝑀𝑐𝑢𝑟𝑟𝑒𝑛𝑡 (𝐹𝑗 , 𝐺𝑘 , 𝑃𝑀)
21. Set 𝑀𝑚𝑎𝑥 = 𝑀𝑐𝑢𝑟𝑟𝑒𝑛𝑡
22. Return 𝑀𝑚𝑎𝑥
23. End For

4. RESULTS AND DISCUSSION

The results of the proposed methodology are extensively discussed herein. The implementation was
conducted using Jupyter Notebook, optimized for GPU utilization to enhance computational efficiency. The
entire model was then ported on a Raspberry Pi 4B (8 GB) board. Following the training of classifiers on
80% of the dataset images, testing was conducted on the remaining 20%. The evaluation of each model
among the 28 combinations was performed based on key metrics such as accuracy, precision, F1 score, and
recall. Noteworthy is the model utilizing the combination of LBP and XGBoost, which demonstrated the
highest accuracy, achieving 81.48% when retaining 90% of information post PCA (case 1), and 83.5% when
retaining 95% of information (case 2).

Int J Reconfigurable & Embedded Syst, Vol. 14, No. 1, March 2025: 89-99
Int J Reconfigurable & Embedded Syst ISSN: 2089-4864  95

In case 1, with 90% data retention post-PCA, SIFT with RF achieved a testing accuracy of 73.8%,
with class 0 (car) precision, F1, and recall at 75.03%, 74.0%, 72.05%, and class 1 (road) scores at 73.9%,
76.1%, and 78.02%. ORB with RF attained a testing accuracy of 78.98%, with class 0 scores of 79.2%,
78.6%, and 79.0%, and class 1 scores of 78.0%, 78.1%, and 78.6%. FREAK exploited with KNN achieved a
testing accuracy of 75.09%, with class 0 scores at 76.4%, 75.0%, and 75.6%, and class 1 scores at 74.8%,
76.1%, and 76.2%. LBP with XGBoost achieved the highest accuracy at 81.48%, with class 0 scores of
82.1%, 81.1%, and 80.2%, and class 1 scores at 81.0%, 82.3%, and 82.5%. In case 2, with 95% data
retention, scores across all parameters improved. Among 28 models, SIFT and FREAK with RF and SVM
outperformed other SIFT-based models, while ORB with SVM and LR yielded higher scores compared to
other ORB models. LBP, combined with RF and XGBoost, delivered the best overall results. Case 2 results
are presented in Table 3.

Table 3. Illustrates the results of the models when 95% of the information is retained (case 2)
Feature descriptor+classifier Testing accuracy (%) Label Precision (%) F1 (%) Recall (%)
SIFT+RF 78.05 0 79.2 77.4 76.3
1 77.3 79.1 81.6
SIFT+SVM 77.22 0 81.0 76.8 71.4
1 74.0 79.3 84.2
80.46 0 82.0 79.1 76.4
ORB+SVM
1 78.0 81.9 84.8
79.53 0 79.4 79.7 80.0
ORB+LR
1 79.1 79.2 78.6
76.57 0 77.2 76.5 76.0
FREAK+RF
1 76.8 77.8 77.0
75.5 0 78.4 74.5 71.2
FREAK+SVM
1 73.6 77.0 80.0
LBP+RF 82.77 0 81.0 83.7 86.3
1 85.1 82.2 80.1
LBP+XGBoost 83.5 0 82.1 83.5 85.0
1 85.0 84.0 81.4

In both case 1 and 2, it was observed that LBP features, in combination with ensemble classifiers,
achieved superior scores for both class labels (0 and 1). Notably, LBP outperformed other classifiers by
having the highest testing accuracy ranging from 76.14% to 83.57%, thus indicating its ability to capture
essential feature values for real-time obstacle detection. In Case 1, a fusion of LBP and XGBoost resulted in
an F1 score of 85% for car detection (class 0), demonstrating balance between precision and recall. SIFT
performed well for class 1 (“Not Car on Road”), but it was less accurate in car detection (class 0) compared
to LBP and ORB, suggesting that SIFT’s features may be less efficient at spotting cars in diverse
backgrounds. Yet, SIFT performed well when categorizing roads, attesting to its capacity for capturing
nuances in ‘Not Car on Road’ elements. ORB was also able to achieve an accuracy of 80.46% with SVM.
On the other hand, FREAK underperformed across all these measures consistently implying that it fails to
capture intricate patterns in the dataset. Notable shifts were observed among the most accurate models of
each feature descriptor when there was a change in amount of information retained post PCA. Specifically,
concerning SIFT, there was a notable decline in the results `when only 90% of the information was retained.
For SIFT, KNN relatively demonstrated superior performance when a lesser amount of information was
retained. As for ORB, RF exhibited better performance compared to RBF-SVM in case 1. Interestingly,
ORB exhibited resilience to changes in the retained information by producing slightly lower scores across
the metrics. Contrarily, for FREAK, KNN, and LR emerged as FREAK’s most accurate models, replacing its
combination with RF and SVM in case 2. On the other hand, LBP maintained its effectiveness with
ensemble-based classifiers such as XGBoost and RF by delivering comparatively superior results across all
the parameters. It was observed that the performance of different models of feature descriptors and
classifiers was influenced by the percentage of information retained post PCA in dimensionality reduction.
Notably, all models encountered relatively more challenges in predicting cars (class 0) compared to road
(class 1), perhaps owing to the complexity involved in distinguishing cars from various distinct and diverse
environmental factors.
Balancing computational efficiency and accuracy is crucial for selecting models suitable for real-
time applications. By assessing the time complexities of feature descriptors and classifiers, including their
training and prediction phases, we can estimate the model's overall time complexity. SIFT emerges as the
most computationally intensive, while ORB and FREAK are observed to be significantly more efficient [32].
LBP stands out for its computational efficiency and strong performance in human detection tasks [33]. After
analyzing the time complexity across the 28 models incorporating feature descriptors and classifiers, it was
Comparative analysis of feature descriptors and classifiers for real-time … (Vikas J. Nandeshwar)
96  ISSN: 2089-4864

observed that combinations implicating LBP achieved lower time complexities than other feature
descriptors. The pairing of LBP with the GNB classifier resulted in the least time complexity of [O(n*d)],
where 'n' signifies the samples in the training dataset and 'd' denotes the dimensionality or the number of
features in each sample. However, despite this efficiency, the achieved accuracy stood at 76.8%, which was
relatively lower compared to other LBP-based models. Furthermore, the KNN model exhibited a time
complexity of [O(n*d)], similar to GNB but with negligible training time complexity. However, it exhibited
a lower accuracy rate of 76.1%, positioning it as one of the least accurate among the LBP-based models. The
model exhibiting the worst-case time complexity ranging from [O(d*n^2) to O(d*n^3)] was observed when
SIFT was coupled with the RBF-SVM classifier. Conversely, when LBP was combined with ensemble-
based classifiers, superior performance across all parameters was observed. However, this came at the cost
of increased time complexity. With RF, the time complexity reached [O(K*d*nlog(n))], and with XGBoost,
the time complexity rose to [O(K*d*|x| log n)], where K is the number of trees, d is the height of the trees, n
represents the total number of data points used to train the model, and |x| is the number of non-missing
entries in the training data [34]. These complexities notably surpass those achieved with KNN, GNB, and
LR. The LBP-based model demonstrated a worst-case time complexity of [O (d*n^2)], where n is the
number of data points in the training set and d is the number of features in the data when coupled with the
SVM classifier. Contrarily, models employing the GNB classifier achieved the lowest time complexity,
albeit with inferior scores across all parameters. Despite KNN and LR classifiers performing well in certain
scenarios, their overall performance was surpassed by ensemble-based classifiers and SVM. This
observation indicated the trade-off between accuracy and time complexity delivered by the model in this
particular case. The receiver operating characteristic (ROC) curve and area under the ROC curve (AUC) for
the XGBoost classifier are illustrated in Figure 4. Notably, the AUC for both the "Car on Road" class with
label 0 and the "Not Car on Road" class with label 1 was observed to be 91%.

Figure 4. ROC curve

5. CONCLUSION
This study proposed a solution aimed at detecting and classifying cars and roads using handcrafted
features extracted from traditional feature descriptors and employing seven classifiers. A total of 28 models
were created and assessed based on accuracy, precision, F1, and recall scores, alongside considerations of
time complexity. Among these models, the one leveraging LBP and XGBoost attained a testing accuracy of
83.57% and outperformed others across all parameters. The adoption of classical machine learning
techniques rendered the system computationally efficient. Furthermore, a novel dimension reduction
technique integrating K Means and PCA was employed. The implemented solution was deployed on a
Raspberry Pi 4B (8 GB) board, demonstrating satisfactory performance in real-world scenarios. While the
system effectively detected cars and roads, it encountered misclassification under low illumination

Int J Reconfigurable & Embedded Syst, Vol. 14, No. 1, March 2025: 89-99
Int J Reconfigurable & Embedded Syst ISSN: 2089-4864  97

conditions. To address this limitation, the authors intend to explore a deep learning-based approach capable
of handling multiple images.

REFERENCES
[1] H. Li, J. Li, Z. Su, X. Wang, and J. Luo, "Research on active obstacle avoidance control strategy for intelligent vehicle based on
active safety collaborative control," in IEEE Access, vol. 8, pp. 183736-183748, 2020, doi: 10.1109/ACCESS.2020.3029042.
[2] H. Bilal, B. Yin, J. Khan, L. Wang, J. Zhang, and A. Kumar, "Real-time lane detection and tracking for advanced driver assistance
systems," 2019 Chinese Control Conference (CCC), Guangzhou, China, 2019, pp. 6772-6777, doi: 10.23919/ChiCC.2019.
8866334.
[3] Y. Torres-Berru and P. Torres-Carrion, "Development of machine learning model for mobile advanced driver assistance (ADA),"
2019 International Conference on Information Systems and Software Technologies (ICI2ST), Quito, Ecuador, 2019, pp. 162-167,
doi: 10.1109/ICI2ST.2019.00030.
[4] S. V. Mahadevkar et al., "A review on machine learning styles in computer vision—techniques and future directions," in IEEE
Access, vol. 10, pp. 107293-107329, 2022, doi: 10.1109/ACCESS.2022.3209825.
[5] Z. -Q. Zhao, P. Zheng, S. -T. Xu, and X. Wu, "Object detection with deep learning: a review," in IEEE Transactions on Neural
Networks and Learning Systems, vol. 30, no. 11, pp. 3212-3232, Nov. 2019, doi: 10.1109/TNNLS.2018.2876865.
[6] J. Leng, Y. Liu, D. Du, T. Zhang, and P. Quan, "Robust obstacle detection and recognition for driver assistance systems," in IEEE
Transactions on Intelligent Transportation Systems, vol. 21, no. 4, pp. 1560-1571, Apr. 2020, doi: 10.1109/TITS.2019.2909275.
[7] A. Dairi, F. Harrou, Y. Sun, and M. Senouci, "Obstacle detection for intelligent transportation systems using deep stacked
autoencoder and k -nearest neighbor scheme," in IEEE Sensors Journal, vol. 18, no. 12, pp. 5122-5132, Jun. 15, 2018, doi:
10.1109/JSEN.2018.2831082.
[8] D. P. Carrasco, H. A. Rashwan, M. Á. García, and D. Puig, "T-YOLO: tiny vehicle detection based on YOLO and multi-scale
convolutional neural networks," in IEEE Access, vol. 11, pp. 22430-22440, 2023 doi: 10.1109/ACCESS.2021.3137638.
[9] H. Zhou, H. Kong, L. Wei, D. Creighton, and S. Nahavandi, "Efficient road detection and tracking for unmanned aerial vehicle,"
in IEEE Transactions on Intelligent Transportation Systems, vol. 16, no. 1, pp. 297-309, Feb. 2015, doi:
10.1109/TITS.2014.2331353.
[10] S. Bhatlawande, A. Patil, S. Shilaskar, K. Patil, and V. Pingat, "A monocular vision based system for detection of drivable road,"
2023 First International Conference on Cyber Physical Systems, Power Electronics and Electric Vehicles (ICPEEV), Hyderabad,
India, 2023 pp. 1-5, doi: 10.1109/ICPEEV58650.2023.10391905.
[11] J. Dai-Hong, D. Lei, L. Dan, and Z. San-You, "Moving-object tracking algorithm based on PCA-SIFT and optimization for
underground coal mines," in IEEE Access, vol. 7, pp. 35556-35563, 2019, doi: 10.1109/ACCESS.2019.2899362.
[12] Q. Wu and Y. Zhou, "Real-time object detection based on unmanned aerial vehicle," 2019 IEEE 8th Data Driven Control and
Learning Systems Conference (DDCLS), Dali, China, 2019, pp. 574-579, doi: 10.1109/DDCLS.2019.8908984.
[13] A. Mammeri, D. Zhou, and A. Boukerche, "Animal-vehicle collision mitigation system for automated vehicles," in IEEE
Transactions on Systems, Man, and Cybernetics: Systems, vol. 46, no. 9, pp. 1287-1299, Sep. 2016 doi:
10.1109/TSMC.2015.2497235.
[14] X. Gu, J. Yu, Y. Han, M. Han, and L. Wei, "Vehicle lane change decision model based on random forest," 2019 IEEE
International Conference on Power, Intelligent Computing and Systems (ICPICS), Shenyang, China, 2019, pp. 115-120, doi:
10.1109/ICPICS47731.2019.8942520.
[15] A. Mhalla, T. Chateau, S. Gazzah, and N. E. B. Amara, "An embedded computer-vision system for multi-object detection in
traffic surveillance," in IEEE Transactions on Intelligent Transportation Systems, vol. 20, no. 11, pp. 4006-4018, Nov. 2019, doi:
10.1109/TITS.2018.2876614.
[16] W. Zhang, "Combination of SIFT and canny edge detection for registration between SAR and optical images," in IEEE
Geoscience and Remote Sensing Letters, vol. 19, pp. 1-5, 2022, doi: 10.1109/LGRS.2020.3043025.
[17] N. Gageik, P. Benz, and S. Montenegro, "Obstacle detection and collision avoidance for a UAV with complementary low-cost
sensors," in IEEE Access, vol. 3, pp. 599-609, 2015, doi: 10.1109/ACCESS.2015.2432455.
[18] L. Guo, D. Zhou, J. Zhou, S. Kimura, and S. Goto, "Lossy compression for embedded computer vision systems," in IEEE Access,
vol. 6, pp. 39385-39397, 2018, doi: 10.1109/ACCESS.2018.2852809.
[19] Y. Wang, W. -L. Chao, D. Garg, B. Hariharan, M. Campbell, and K. Q. Weinberger, "Pseudo-LiDAR from visual depth
estimation: bridging the Gap in 3D object detection for autonomous driving," 2019 IEEE/CVF Conference on Computer Vision
and Pattern Recognition (CVPR), Long Beach, CA, USA, 2019, pp. 8445-8453.
[20] F. Pourkamali-Anaraki and S. Becker, "Preconditioned data sparsification for big data with applications to PCA and K-Means," in
IEEE Transactions on Information Theory, vol. 63, no. 5, pp. 2954-2974, May 2017, doi: 10.1109/TIT.2017.2672725.
[21] S. Shilaskar, V. Patil, S. Pawar, A. Poke, R. Jambhulkar, and S. Bhatlawande, "Driver safety system using microcontroller and
image processing," 2023 International Conference on Intelligent Data Communication Technologies and Internet of Things
(IDCIoT), Bengaluru, India, 2023, pp. 340-344, doi: 10.1109/IDCIoT56793.2023.10053512.
[22] S. Bhatlawande et al., "AI based handheld electronic travel aid for visually impaired people," 2022 IEEE 7th International
conference for Convergence in Technology (I2CT), Mumbai, India, 2022, pp. 1-5, doi: 10.1109/I2CT54291.2022.9823962.
[23] K. M. Masal, S. Bhatlawande, and S. D. Shingade, "Development of a visual to audio and tactile substitution system for mobility
and orientation of visually impaired people: a review," Multimedia Tools and Applications, vol. 83, pp. 20387–20427, 2024, doi:
10.1007/s11042-023-16355-0.
[24] J. Madake, S. Bhatlawande, S. Purandare, S. Shilaskar, and Y. Nikhare, "Dense video captioning using BiLSTM encoder," 2022
3rd International Conference for Emerging Technology (INCET), Belgaum, India, 2022, pp. 1-6, doi:
10.1109/INCET54531.2022.9824569.
[25] S. Bhatlawande, D. Khapre, M. Kinge, and T. Khairnar, "Vision based assistive system for fall detection," 2022 2nd International
Conference on Intelligent Technologies (CONIT), Hubli, India, 2022, pp. 1-7, doi: 10.1109/CONIT55038.2022.9847697.
[26] S. Bhatlawande, S. Shilaskar, T. Gadad, S. Ghulaxe, and R. Gaikwad, "Smart home security monitoring system based on face
recognition and android application," 2023 International Conference on Intelligent Data Communication Technologies and
Internet of Things (IDCIoT), Bengaluru, India, 2023, pp. 222-227, doi: 10.1109/IDCIoT56793.2023.10053558
[27] T. Yawen and G. Jinxu, "Research on vehicle detection technology based on SIFT feature," 2018 8th International Conference on
Electronics Information and Emergency Communication (ICIEC), Beijing, China, 2018, pp. 274-278, doi:
10.1109/ICEIEC.2018.8473575.

Comparative analysis of feature descriptors and classifiers for real-time … (Vikas J. Nandeshwar)
98  ISSN: 2089-4864

[28] C. Yao, H. Zhang, J. Zhu, D. Fan, Y. Fang, and L. Tang, "ORB feature matching algorithm based on multi-scale feature
description fusion and feature point mapping error correction," in IEEE Access, vol. 11, pp. 63808-63820, 2023, doi:
10.1109/ACCESS.2023.3288594.
[29] Y. Li, "A novel fast retina keypoint extraction algorithm for multispectral images using geometric Algebra," in IEEE Access, vol.
7, pp. 167895-167903, 2019, doi: 10.1109/ACCESS.2019.2954081.
[30] Y. Ding, Q. Zhao, B. Li and X. Yuan, "Facial expression recognition from image sequence based on LBP and Taylor expansion,"
in IEEE Access, vol. 5, pp. 19409-19419, 2017, doi: 10.1109/ACCESS.2017.2737821.
[31] F. Liu and Y. Deng, "Determine the number of unknown targets in the open world based on Elbow method," in IEEE
Transactions on Fuzzy Systems, vol. 29, no. 5, pp. 986-995, May 2021, doi: 10.1109/TFUZZ.2020.2966182.
[32] A. Canclini, M. Cesana, A. Redondi, M. Tagliasacchi, J. Ascenso, and R. Cilla, "Evaluation of low-complexity visual feature
detectors and descriptors," 2013 18th International Conference on Digital Signal Processing (DSP), Fira, Greece, 2013, pp. 1-7,
doi: 10.1109/ICDSP.2013.6622757.
[33] J. Xu, Q. Wu, J. Zhang, and Z. Tang, "Fast and accurate human detection using a cascade of boosted MS-LBP features," in IEEE
Signal Processing Letters, vol. 19, no. 10, pp. 676-679, Oct. 2012, doi: 10.1109/LSP.2012.2210870.
[34] T. Chen and C. Guestrin, "XGBoost: A scalable tree boosting system," KDD '16: Proceedings of the 22nd ACM SIGKDD
International Conference on Knowledge Discovery and Data Mining, 2016, pp. 785-794, doi: 10.1145/2939672.2939785.

BIOGRAPHIES OF AUTHORS

Vikas J. Nandeshwar Assistant Professor at Vishwakarma Institute of

Technology, Pune, specializes in sensors and applications, robotics, and mechatronics. His
expertise includes data science, machine learning, deep learning, and natural language
processing (NLP). His work is marked by contributions to conferences and journals,
emphasizing his commitment to advancing technology in engineering disciplines. He can be
contacted at email: [email protected].

Sarvadnya Bhatlawande is a first-year engineering student at Vishwakarma

Institute of Technology, Pune, with a deep passion for machine learning and computer vision.
He actively explores advanced techniques in artificial intelligence and machine learning,
aiming to expand his expertise in these dynamic fields. Despite being in the early stages of his
academic journey, Bhatlawande exhibits a strong commitment to machine learning and
artificial intelligence, driven by a desire to contribute significantly to their advancement. For
further correspondence. He can be contacted at email: [email protected].

Anjali Solanke is currently working as an Associate Professor in the Department

of Electronics and Telecommunication Engineering, Marathwada Mitra Mandal's College of
Engineering, Pune, India. She is having 23 years of teaching experience. Her area of research
is medical image processing and machine learning. She can be contacted at email:
[email protected].

Harsh Sathe is a first-year B.Tech. student majoring in information technology

(IT). He eagerly anticipates his inaugural work opportunity. He contributes a burgeoning
perspective and a proactive approach to learning and research. They aspire to make meaningful
contributions to the field of machine learning through collaborative efforts. He can be
contacted at email: [email protected].

Int J Reconfigurable & Embedded Syst, Vol. 14, No. 1, March 2025: 89-99
Int J Reconfigurable & Embedded Syst ISSN: 2089-4864  99

Shivanand Satao is currently a first-year engineering student at Vishwakarma

Institute of Technology, Pune. He is committed to continuous learning, participating in
workshops and seminars that enhance their understanding of computer vision. His involvement
underscores a proactive approach in acquiring essential knowledge and skills for future
research endeavors. He can be contacted at email: [email protected].

Safalya Satpute is currently a first-year engineering student at Vishwakarma

Institute of Technology, Pune. Embarking on his academic journey, they are eager to explore
the intricacies of machine learning and robotics. He actively engages with foundational
concepts in artificial intelligence and machine learning and exhibits enthusiasm towards
contributing to research initiatives. He can be contacted at email: [email protected].

Atharv Saste possesses proficiency in programming and automation. His primary

interests encompass artificial intelligence (AI), machine learning (ML), and robotics, fields in
which he aspires to make meaningful contributions. He can be contacted at email:
[email protected].

Comparative analysis of feature descriptors and classifiers for real-time … (Vikas J. Nandeshwar)