Implementing an Efficient Speed Bump Detection System using

Adaptive Threshold Gaussian over Support Vector Machine for

Improved Detection

R. Priyanka

and W. Deva Priya

†

Department of Electronics and Communication Engineering, Saveetha School of Engineering,

Saveetha Institute of Medical and Technical Sciences, Saveetha University, Chennai, Tamil Nadu, 602105, India

Keywords: Adaptive Threshold, Support Vector Machine, Gaussian Thresholding, Intelligent Vehicle System, Roads,

Novel Speed Bump Detection.

Abstract: This research endeavours to identify speed bumps from provided images using Adaptive Thresholding for

enhanced detection. A total of 120 samples were divided equally into two groups. The first group, comprising

60 samples, underwent testing using the Support Vector Machine, while the second group was tested with the

Adaptive Threshold-Gaussian. Each group underwent 10 iterations. The dataset, comprising 6000 images

sourced from Kaggle.com, allocated 4800 images for training and the remaining for testing. With a G power

roughly at 80%, the Gaussian Adaptive Threshold yielded an accuracy of 85.60%, surpassing the Support

Vector Machine's 81.40%. A significance value of 0.002 (p<0.05) indicates that the results between the two

groups are statistically significant. The Gaussian Adaptive Threshold, therefore, stands out for its superior

accuracy.

1 INTRODUCTION

Adaptive thresholding is a type of image processing

method extensively employed in digital image

processing. In adaptive thresholding, image

segmentation is accomplished by setting the threshold

value equal to the sum of neighbouring values, where

weights are a Gaussian window (KuKuXia 2018).

Speed bumps are deemed a crucial component of the

road traffic control system. They are designed and

placed on roads to diminish vehicle speed and

enhance neighbourhood safety (Kosakowska 2022).

The project's most promising application is to offer

safer navigation for drivers, mitigating accidents

caused by overlooked speed bumps (Arunpriyan

2020). Moreover, human-caused abnormalities can

diminish when drivers are warned of approaching

speed bumps (Dewangan and Sahu 2021) (Palanivelu

et al. 2022).

A compilation of approximately 200 articles from

platforms such as Google Scholar, IEEE Xplore, and

Springer has been amassed over the past five years.

These articles propose distinct approaches to a

Research Scholar

†

Research Guide, Corresponding Author

specific issue. A notable strategy emphasised in the

literature involves a multivariate genetic algorithm, as

illustrated by Celaya-Padilla (2018). This algorithm

capitalises on data from IoT devices fitted with

accelerometers, GPS, and gyro sensors. Another

significant paper by Dewangan and Sahu (2021) uses

Raspberry Pi to detect speed bumps within an

Intelligent Vehicle System, achieving commendable

accuracy and precision (G. R et al. 2014).

In the quest to enhance the Advanced Driver

Assistance System, the identification of speed bumps

on well-marked roads is tackled by Devapriya, Babu,

and Srihari (2016) using Gaussian filtering, median

filtering, and connected component analysis.

Intriguingly, a more straightforward method for speed

bump detection and recognition using basic image

processing methods, which simultaneously triggers

driver alerts, is presented by Devapriya, Babu, and

Srihari (2015). Notably, an engaging paper that

utilises image processing techniques for speed bump

recognition, combined with gyro and sensor inputs, is

credited to Celaya-Padilla (2018). The research gap

for novel speed bump detection lies in the

inaccuracies when detecting a speed bump from an

Priyanka, R. and Priya, W.

Implementing an Efﬁcient Speed Bump Detection System Using Adaptive Threshold Gaussian over Support Vector Machine for Improved Detection.

DOI: 10.5220/0012603100003739

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 1st International Conference on Artiﬁcial Intelligence for Internet of Things: Accelerating Innovation in Industry and Consumer Electronics (AI4IoT 2023), pages 527-532

ISBN: 978-989-758-661-3

527

image taken from a long distance, and occasionally,

there is false detection in the images. Existing

methods seem to possess lesser accuracy in

recognising speed bumps and require a significant

amount of time to alert the drivers. The primary aim

of this research is to achieve better accuracy by

comparing two algorithms: Adaptive Gaussian

Thresholding and Support Vector Machine for speed

bump detection.

2 MATERIALS AND METHODS

The research for this project took place within the

Compiler Design Laboratory at the Saveetha School

of Engineering, part of the Saveetha Institute of

Medical and Technical Sciences. This study

comprises two distinct research groups: Adaptive

Gaussian Thresholding and Support Vector Machine,

with differentiation based on their ability to detect

speed bumps in images. Of the 120 samples used, 60

samples each are assigned to these groups, as detailed

by Arunpriyan (2020). Both groups undergo 10

iterative processes, ensuring reliable results. The

dataset, sourced from Kaggle.com, contains 6000

images; 4800 for training and the rest for testing. The

sample size was influenced by previous research,

particularly that of Devapriya, Babu, and Srihari

(2016). The study employs a 95% confidence interval

and maintains a pretest power of 80% for robust

statistical outcomes.

The research framework demands 4GB of RAM

for prompt program processing. An Intel(R) CPU @

1.10GHz or its superior version is recommended. The

study used Windows 11 as its operating system. A

storage space of 30GB is essential to house the

collected dataset images, store the code, and

accommodate necessary plugins. The Jupyter

Notebook is utilised to operate the framework and test

the program on images featuring speed bumps.

Adaptive Threshold- Gaussian

Adaptive thresholding, a facet of Image Binarisation,

encompasses two primary methods: Adaptive

Threshold Mean and Adaptive Threshold Gaussian.

The following table 1 consists of accuracies of a

sample size of 10 for both the Adaptive Threshold

Gaussian (ATG) algorithm and the Support Vector

Machine (SVM) algorithm.

Table 1.

S.No ATG SVM

1 83 84

2 84 81

3 89 83

4 87 80

5 85 79

6 87 85

7 84 78

8 89 80

9 80 83

10 88 81

The table below presents comprehensive statistics for

two distinct groups, each comprising a sample size of

N=10. The mean percentage accuracy achieved by the

Adaptive Gaussian Thresholding method is

documented at 85.60%, while the accuracy

percentage attributed to the Support Vector Machine

algorithm registers at 81.40%.

Table 2.

Groups

Mean

Std.

Deviation

Std. Error

Rate

Accuracy

ATG 10 85.602.91357 0.92135

SVM 10 81.40 2.270580.71802

This technique separates the desired object from

the background based on the varying pixel intensities

throughout the image. Instead of a manually specified

threshold or fixed constraints, adaptive thresholding

automatically determines the threshold value in

relation to the image pixels, converting them into

grayscale or a binary format. This method enables an

automatic selection of the threshold value to

differentiate the main subject from its backdrop,

especially beneficial in scenarios with fluctuating

lighting, colour, or contrast in the image.

AI4IoT 2023 - First International Conference on Artiﬁcial Intelligence for Internet of things (AI4IOT): Accelerating Innovation in Industry

and Consumer Electronics

528

Figure 1: (a) The output of Adaptive Gaussian Threshold and (b) The output of Support Vector Machine.

Here's the Adaptive Threshold-Gaussian algorithm

steps, rephrased in British English:

Step 1: Images are imported from the dataset for

preprocessing, ensuring they're appropriately

formatted for computational processing.

Step 2: During preprocessing, each image is resized

to a standard dimension of 200x350, which trims

unnecessary parts of the images.

Step 3: Once resized, RGB colour images are

transformed into grayscale using the luminosity

method. This colour conversion simplifies the

computational demands.

Step 4: The Adaptive Threshold Gaussian algorithm

is then applied to the grayscale images, turning them

into binary images.

Step 5: A state-of-the-art speed bump detection

system is then utilised, employing the Haar Cascade

classifier, which has been trained on various speed

bump images.

Step 6: Successfully identified speed bumps are

emphasised with bounding boxes, signalling their

detection.

Support Vector Machine

The Support Vector Machine (SVM) is pivotal in the

suite of supervised learning algorithms, being

especially prominent in both classification and

regression tasks. Specifically, when diving into object

recognition, which includes detecting speed bumps,

SVM's expertise falls under the classification sphere.

At its core, SVM focuses on pinpointing and

earmarking the key vectors and boundary points in

data distributions. These elements are critical in

setting the hyperplane that segregates distinct classes

in the data realm. As the name suggests, support

vectors are these integral data points underpinning the

SVM approach.

Steps for the Support Vector Machine algorithm:

Step 1: Employing deep learning techniques such as

HOG, the dataset images are ingested and key

features extracted.

Step 2: Resizing images enhances detection precision.

Subsequently, the dataset is partitioned into training

and testing subsets.

Step 3: With the data prepped, an SVM classifier is

formulated using sklearn, and it's trained using the

provided dataset.

Step 4: Object localisation is undertaken in two

phases: initially via sliding windows and then through

heatmapping. The sliding window outlines the

potential detection area with a bounding box.

Step 5: Bounding boxes in static regions signal false

detections, whilst those in recognised regions denote

true detections.

Step 6: As mentioned, object localisation follows a

dual process: firstly through sliding windows, and

then supplemented by a heatmap to reaffirm bounding

box demarcations.

Step 7: Depending on detection outcomes, a

confusion matrix is developed. This matrix

subsequently facilitates the calculation of system

accuracy.

Statistical Analysis

For the statistical analysis of the results, IBM SPSS

version 29 was utilised. The mean accuracies were

compared using an independent sample t-test, with the

confidence level set at 95% and the standard deviation

error fixed at +/-2SD (Elliott and Woodward 2020).

The research's independent variables encompass

accuracy and the quantity of input images, whilst there

isn't a dependent variable identified. The research

utilised a sample size of (N=10), and the observed

significance value stands at 0.002.

Implementing an Efﬁcient Speed Bump Detection System Using Adaptive Threshold Gaussian over Support Vector Machine for Improved

Detection

529

3 RESULTS

Both the Adaptive Threshold Gaussian and Support

Vector Machine algorithms were executed multiple

times using the Jupyter notebook platform. The

results showed that Group1, employing the Adaptive

Threshold Gaussian algorithm, achieved an accuracy

of 85.60%. In contrast, Group 2, utilising the Support

Vector Machine algorithm, registered an accuracy of

81.40%. This indicates that the Adaptive Threshold

Gaussian algorithm outperforms the Support Vector

Machine algorithm in terms of accuracy.

Table 1 enumerates the accuracy values derived

from both the Adaptive Threshold Gaussian and

Support Vector Machine algorithms. Table 2

delineates the mean accuracy values and standard

deviation calculations for both algorithms.

Meanwhile, Table 3 lists samples from the

independent t-tests comparing the two algorithms. A

graphical representation showcasing these values for

both algorithms is provided.

Figure 1 showcases the outputs of both the

Support Vector Machine and the Adaptive Gaussian

Thresholding. While the Support Vector Machine

detects the entire background along with the speed

bump, the Adaptive Gaussian Thresholding zeroes in

on a specific section of the speed bump.

Figure 2 visually contrasts the mean accuracies of

the Adaptive Threshold Gaussian and Support Vector

Machine algorithms. The concluding data infers that

the Support Vector Machine is less effective than the

Adaptive Threshold Gaussian algorithm, with the

mean accuracy detection positioned at +/-2SD.

4 DISCUSSION

From the research results presented, it is

unequivocally clear that the Adaptive Gaussian

threshold algorithm outperforms the Support Vector

Machine algorithm, thus validating the hypothesis.

The accuracy percentage gleaned for the Adaptive

Gaussian threshold algorithm stands at 85%, whereas

the Support Vector Machine algorithm achieved an

accuracy of 81%. The Adaptive threshold algorithm

finds precedent in various related research pursuits,

such as face detection, motion detection, and content

recognition (Devapriya 2015). While the Support

Vector Machine is renowned for its robustness as a

classification and regression algorithm across myriad

application fields (Cervantes 2020), the Adaptive

Gaussian Threshold emerges as particularly suited for

speed bump detection. This is attributable to its

proficiency in converting RGB images into grayscale,

enhancing the accuracy of detection (Celaya-Padilli

2018). Though sophisticated hardware like the

NVIDIA GPU and Stereolabs ZED Stereo camera can

facilitate the recognition and detection of speed

bumps (Varma 2018), they invariably demand

specific prerequisites and considerable investments.

The approach of stereo vision, applied to local binary

pattern images, presents a unique technique for

detecting delineated speed bumps (Ballinas-

Hernández, Olmos-Pineda, and Olvera-López 2022).

However, the accuracy of such detection can

sometimes leave room for improvement. Various

machine learning algorithms, including Naive Bayes,

Multi-Layer Perceptron, and Random Forest, have

Table 3: Independent Sample t-Test for Accuracy Comparison with 95% Confidence Interval and Equal Variance Assumption.

Levene’s test for

equality of

variances

T-test for equality of means

Sig.

Sig. (2-

tailed)

Mean

difference

Std.

Error

difference

95% confidence

Interval of the

difference

Lower Upper

Accuracy

Equal

Variance

assume

0.815 0.037 3.596 18 0.002 0.053 4.20 1.74 6.65

Equal

variance

Not

assumed

3.596 16.9 0.002 0.053 4.20 1.73 6.66

AI4IoT 2023 - First International Conference on Artiﬁcial Intelligence for Internet of things (AI4IOT): Accelerating Innovation in Industry

and Consumer Electronics

530

Figure 2: Comparison of Mean Accuracies between Existing and Proposed Algorithms. This bar chart illustrates a comparison

of mean accuracies, with the Y-axis representing accuracy values and the X-axis denoting the proposed and existing

algorithms. The mean accuracy for the Adaptive Gaussian Thresholding is recorded at 85%, while the Support Vector

Machine algorithm achieves an accuracy of 81%.

been trialled for detecting speed bumps using datasets

sourced from GoPro cameras (Marques 2021).

Nonetheless, the research indicates that deep learning

algorithms consistently deliver the most accurate and

superior performance.

This research is not without its limitations. In

scenarios where images are captured under deficient

lighting conditions, the speed bump detection

deteriorates, leading to compromised accuracy. The

research's overarching objective is multifaceted: it

aims to detect speed bumps from various perspectives

and discern unmarked speed bumps. However,

challenges arise when an unmarked speed bump is

used as input, resulting in a substantial drop in

detection accuracy. One key feature that sets a speed

bump apart from a crosswalk is its elevation. Yet, in

certain images, this elevation isn't distinctly outlined,

making detection arduous. Looking ahead, the future

of novel speed bump detection holds promise in real-

time video detection of both marked and unmarked

speed bumps. As techniques evolve, advanced DNN

methodologies could provide an avenue for even

more precise speed bump detection on roads.

5 CONCLUSION

The conducted research clearly delineated the

comparative efficiencies of the Adaptive Gaussian

Threshold algorithm and the Support Vector Machine

algorithm in the realm of speed bump detection. With

the former securing an accuracy of 85.60% and the

latter managing 81.40%, the Adaptive Gaussian

Threshold algorithm unequivocally outshone its

counterpart. This outcome was not merely a casual

observation; it held statistical weight, given the

achieved significance value of 0.002. This value,

being less than the conventional threshold of 0.05,

reinforced that the performance discrepancy between

the two algorithms was not due to random chance but

was indeed statistically significant. In essence, for

those seeking to deploy an algorithm for detecting

speed bumps with optimal accuracy, the Adaptive

Gaussian Threshold algorithm emerges as the more

promising choice over the Support Vector Machine

algorithm, as corroborated by the empirical evidence

presented in this study.

REFERENCES

Arunpriyan, J., V. V. Sajith Variyar, K. P. Soman, and S.

Adarsh. 2020. “Real-Time Speed Bump Detection

Using Image Segmentation for Autonomous Vehicles.”

In Advances in Intelligent Systems and Computing,

308–15. Advances in Intelligent Systems and

Computing. Cham: Springer International Publishing.

Ballinas-Hernández, Ana Luisa, Ivan Olmos-Pineda, and

José Arturo Olvera-López. 2022. “Marked and

Unmarked Speed Bump Detection for Autonomous

Vehicles Using Stereo Vision.” Journal of Intelligent &

Fuzzy Systems 42 (5): 4685–97.

Celaya-Padilla, Jose M., Carlos E. Galván-Tejada, F. E.

López-Monteagudo, O. Alonso-González, Arturo

Moreno-Báez, Antonio Martínez-Torteya, Jorge I.

Galván-Tejada, Jose G. Arceo-Olague, Huizilopoztli

Luna-García, and Hamurabi Gamboa-Rosales. 2018.

“Speed Bump Detection Using Accelerometric

Features, Gaussian Thresholding: A Genetic Algorithm

Implementing an Efﬁcient Speed Bump Detection System Using Adaptive Threshold Gaussian over Support Vector Machine for Improved

Detection

531

Approach.” Sensors 18 (2).

https://doi.org/10.3390/s18020443.

Cervantes, Jair, Farid Garcia-Lamont, Lisbeth Rodríguez-

Mazahua, and Asdrubal Lopez. 2020. “A

Comprehensive Survey on Support Vector Machine

Classification: Applications, Challenges and Trends.”

Neurocomputing 408 (September): 189–215.

Devapriya, W., C. Nelson Kennedy Babu, and T. Srihari.

2015. “Advance Driver Assistance System (ADAS) -

Speed Bump Detection.” In 2015 IEEE International

Conference on Computational Intelligence and

Computing Research (ICCIC), 1–6.

———. 2016. “Real Time Speed Bump Detection Using

Gaussian Filtering and Connected Component

Approach.” In 2016 World Conference on Futuristic

Trends in Research and Innovation for Social Welfare

(Startup Conclave), 1–5.

Dewangan, Deepak Kumar, and Satya Prakash Sahu. 2021.

“Deep Learning-Based Speed Bump Detection Model

for Intelligent Vehicle System Using Raspberry Pi.”

IEEE Sensors Journal 21 (3): 3570–78.

Elliott, Alan C., and Wayne A. Woodward. 2020. “Quick

Guide to IBM® SPSS®: Statistical Analysis With Step-

by-Step Examples.”

https://doi.org/10.4135/9781071909638.

G. Ramkumar and M. Manikandan, "Uncompressed digital

video watermarking using stationary wavelet

transform," 2014 IEEE International Conference on

Advanced Communications, Control and Computing

Technologies, Ramanathapuram, India, 2014, pp.

Kosakowska, Katarzyna. 2022. “Evaluation of the Impact

of Speed Bumps on the Safety of Residents - Selected

Aspects.” Transportation Research Procedia 60

(January): 418–23.

Kishore Kumar, M. Aeri, A. Grover, J. Agarwal, P. Kumar,

and T. Raghu, “Secured supply chain management

system for fisheries through IoT,” Meas. Sensors, vol.

25, no. August 2022, p. 100632, 2023, doi:

10.1016/j.measen.2022.100632.

KuKuXia. . 13_Adaptive_thresholding.py at Master ·

KuKuXia/OpenCV_for_Beginners. Github. Accessed

December 27, 2022.

https://github.com/KuKuXia/OpenCV_for_Beginners.

Marques, Johny, Raulcezar Alves, Henrique C. Oliveira,

Marco MendonÇa, and Jefferson R. Souza. 2021. “An

Evaluation of Machine Learning Methods for Speed-

Bump Detection on a GoPro Dataset.” Anais Da

Academia Brasileira de Ciencias 93 (1): e20190734.

Varma, V. S. K. P., S. Adarsh, K. I. Ramachandran, and

Binoy B. Nair. 2018. “Real Time Detection of Speed

Hump/Bump and Distance Estimation with Deep

Learning Using GPU and ZED Stereo Camera.”

Procedia Computer Science 143 (January): 988–97.

Marques, Johny, Raulcezar Alves, Henrique C. Oliveira,

Marco MendonÇa, and Jefferson R. Souza. 2021. “An

Evaluation of Machine Learning Methods for Speed-

Bump Detection on a GoPro Dataset in driver assistance

system.” Anais Da Academia Brasileira de Ciencias 93

(1): e20190734.

Palanivelu, J., Thanigaivel, S., Vickram, S., Dey, N.,

Mihaylova, D., & Desseva, I. (2022). Probiotics in

functional foods: survival assessment and approaches

for improved viability. Applied Sciences, 12(1), 455.

Rokhim, Abdul, Hanendyo Putro, and Dani F. Brilianti.

2022. “Automatic Speed Bump (ASeB) to Reduce the

Rate of Traffic Accident in image processing” RSF

Conference Series: Engineering and Technology 2 (2):

79–87.

Varma, V. S. K. P., S. Adarsh, K. I. Ramachandran, and

Binoy B. Nair. 2018. “Real Time Detection of Speed

Hump/Bump and Distance Estimation with Deep

Learning Using GPU and ZED Stereo Camera.”

Procedia Computer Science 143 (January): 988–97.

V. P. Parandhaman, "An Automated Efficient and Robust

Scheme in Payment Protocol Using the Internet of

Things," 2023 Eighth International Conference on

Science Technology Engineering and Mathematics

(ICONSTEM), Chennai, India, 2023, pp. 1-5, doi:

10.1109/ICONSTEM56934.2023.10142797.

Zheng, Zengwei, Mingxuan Zhou, Yuanyi Chen, Meimei

Huo, and Dan Chen. 2019. “Enabling Real-Time Roads

Anomaly Detection via Mobile Edge Computing.”

International Journal of Distributed Sensor Networks

15 (11): 1550147719891319.

AI4IoT 2023 - First International Conference on Artiﬁcial Intelligence for Internet of things (AI4IOT): Accelerating Innovation in Industry

and Consumer Electronics

532