Neural Computing and Applications (2022) 34:19015–19033

https://doi.org/10.1007/s00521-022-07436-6 (0123456789().,-volV)(0123456789().
,- volV)

ORIGINAL ARTICLE

Road anomaly detection using a dynamic sliding window technique

Noureddine Chibani1 • Faouzi Sebbak1 • Walid Cherifi1 • Khadidja Belmessous1

Received: 20 January 2022 / Accepted: 11 May 2022 / Published online: 29 June 2022
 The Author(s), under exclusive licence to Springer-Verlag London Ltd., part of Springer Nature 2022

Abstract
The need to use roads is vital. In reality, smooth asphalt roads help people in their daily lives by saving time, avoiding
traffic, and preserving the means of transportation. Recently, road anomaly detection using smartphone sensors such as
accelerometers, gyroscopes, and GPS has become an important topic in the field of Intelligent Transportation Systems
(ITS). In this context, many solutions have been proposed using Static Sliding Window (SSW), which is based on fixed
window length. However, in the real world, the window length of the anomaly changes according to the speed value and
the anomaly width, which is considered as a major drawback of SSW. In this paper, we propose a new technique called
Dynamic Sliding Window (DSW), which aims to improve the quality of road anomaly detection by preprocessing the
accelerometer signal. The proposed technique is applied to the same dataset and under the same conditions as the SSW. To
cover all scenarios, thirty different virtual roads and several types of anomalies (speed bumps, metal bumps, and potholes)
were used as training and test data. The resulting outputs of the DSW and SSW have been used by seven heuristic
algorithms proposed by previous researchers and seven classifiers based on twelve feature detectors. The obtained results
using the proposed DSW have been compared to those obtained using the SSW to demonstrate the efficiency of the former.
Indeed, based on the comparison, the proposed DSW has proven its potential to outperform all previous road anomaly
detection methods.

Keywords Dynamic sliding window  Machine learning  Mobile sensing  Road anomalies detection  Static sliding
window

1 Introduction problem as they represent the eighth leading cause of death

in the world [1]. Among existing factors that contribute to
According to the World Health Organization (WHO), the road traffic crashes, such as excessive speed, reckless
number of people who died each year due to road traffic driving, alcohol, and traffic violations, some important
crashes has already reached 1.35 million in 2016, and factors are directly related to poor road quality, including
between 20 and 50 million people suffered from non-fatal potholes, damaged roads, and road cracks. Indeed, there are
injuries, including total or partial disability [1]. In fact, two main reasons that cause anomalies on the roads and
road traffic crashes have become a major public health lead to road accidents. The first is the climatic variation, as
pointed out by Qiao et al. [2], while the second is related to
the poor quality of road construction, which creates speed
& Faouzi Sebbak bumps, road joins, manholes, etc.
[email protected] To save lives and prevent injuries around the world, the
Noureddine Chibani present work aims to monitor road surface conditions, draw
[email protected] people’s attention to the location of the anomalies, and
Walid Cherifi continuously inform the authorities about newly detected
[email protected] anomalies. Recently, three main approaches have been
Khadidja Belmessous used for road surface monitoring as introduced by Shahram
[email protected] et al. [3]:
1
Ecole Militaire Polytechnique, PO Box 17, Bordj El Bahri,
Algiers, Algeria

123
19016 Neural Computing and Applications (2022) 34:19015–19033

• A 3D laser scanning approach [4, 5], based on reflected 1. We propose a Dynamic Sliding Window (DSW) that
laser pulses to generate accurate surface samples. These splits the collected signal data into different lengths
samples are then compared to base samples to detect depending on the speed of the vehicle;
anomalies in the road surface; 2. The proposed technique is incorporated into different
• A Vision approach [6–9], based on image processing road anomaly detection methods and has been evalu-
analysis of already captured images by a camera/video ated using the Pothole lab dataset [29] as well as the
system, which are then processed to automatically Carsim dataset [30];
detect road anomalies; In more detail, this work aims to overcome the SSW limits,
• A Vibration approach [10–18], based on sensing by targeting the problem of road anomaly detection using a
vehicles’ vibrations captured by motion sensors like vibration approach with a DSW. It also evaluates the
accelerometers and gyroscopes. In reality, a vehicle has impact of the use of a DSW instead of an SSW on that
different vibration behaviors according to the type of field. To this end, we used two datasets: one developed by
road anomaly. It vibrates differently when passing over Carlos et al. [29], which was generated from a web plat-
a pothole than over a smooth road surface. form, called Pothole lab,1 and another generated from
3D laser scanning and vision approaches are very costly, Carsim [27, 30], which included 5379 anomalies.
especially when targeting large-scale road networks. In order to evaluate the impact of using DSW, we have
However, the vibration-based sensing approach has been compared the obtained results using both DSW and SSW
used in various fields [19], especially in the road surface with seven classifiers, namely Support Vector Machine
monitoring domain, using mobile sensors and smartphones (SVM), Decision Trees (DT), Gradient Boosting (GB),
that seem to offer a low cost. Recently, with the rapid Random Forrest (RF), Naı̈ve Bayes (NB), Neural Networks
development of motion embedded sensors (cameras, (NN), and k-Nearest Neighbor (KNN). Besides, we have
accelerometers, gyroscopes, and GPS), smartphone-based also compared the obtained results from seven heuristic
sensing has become an interesting alternative to the two road anomaly detection algorithms, namely G-ZERO,
previous, costly approaches. As reported in [20], the cur- Nericell, Pothole Patrol (P2 ), STDEV(Z), Swarm,
rent number of mobile phone users is around 7.26 billion, Z-THRESH, and Z-DIFF. The obtained results show that
representing 91.54% of the world’s population. Hence, in the proposed DSW technique overcomes the limitations of
the current work, we target the problem of road anomaly the SSW and provides better performance under the same
detection using a smartphone-based sensing vibration experimental conditions. The remainder of the paper is
approach. In fact, the majority of existing vibration organized as follows: Sect. 2 discusses related work. The
approaches are based on filters, which are used during the proposed technique is discussed in Sect. 4. Problem
data collection phase. The real-time vibration sensor data Statement is discussed in Sect. 3. Section 5 presents the
are then continuously processed in order to avoid some experimental results. Finally, concluding remarks are pro-
issues associated with big data overload [21]. Besides, vided in Sect. 6.
these approaches are based on the processing of the
accelerometers’ readings and detecting, from a time-series
of readings, where an anomaly has been encountered. 2 Related work
These time-series are generally obtained using the sliding
window concept, which is regarded as a critical step pre- Road anomaly detection using smartphone sensors, such as
ceding the preprocessing stage. The latter is employed in a accelerometers, gyroscopes, and GPS, is becoming an
variety of domains prior to the classification phase [22–28]. interesting area of research for the scientific community.
The sliding window technique aims to frame the time- Among the existing approaches for road surface monitor-
series data. Additionally, existing sliding window approa- ing, we focus in this paper on a vibration-based approach.
ches are based on the Static Sliding Window (SSW), which Generally, vibration-based approaches are based on the
means that the window length is fixed over time, and each use of three sensors: an accelerometer to measure the
time-series data occur at exactly the same duration. How- vibration in the three axes (X, Y, and Z), a gyroscope to
ever, using an SSW may result in low-quality road anomaly measure angular velocity in the three axes (X, Y, and Z),
detection since anomalies do not have the same size. In and a GPS to get the location of an object based on the
addition, the speed at which an anomaly is detected influ- latitude and longitude coordinates. Therefore, before using
ences the detection process. this sensor’s data, a preparation phase is needed to extract
To overcome these SSW limits, the current work useful features. Actually, most road anomaly detection
addresses the issue of finding sliding window length. The
contributions of our research are twofold:
1
http://www.accelerometer.xyz/pothole_lab.

123
Neural Computing and Applications (2022) 34:19015–19033 19017

approaches use the SSW to extract these features length of 100 without overlap. Besides, a GPS was used to
[15, 16, 18, 29, 31–36]. Indeed, the SSW technique consists detect road anomalies’ location. The obtained results report
of splitting the raw data stream derived from sensors into a true positive rate (TPR) increasing from 73 to 92%. In
individual segments of a fixed window length with possible more detail, the Z-THRESH, Z-DIFF, STDEV(Z), and
overlap or not between adjacent windows. So far, road G-ZERO algorithms with 0.4 g, 0.2 g, 0.2 g, and 0.8 g as
anomaly detection has been carried out using three main optimal threshold values reach 78%, 92%, 81%, and 73%
approaches: Threshold-based approach, Machine Learning of true positives, respectively. From the above results, the
(ML), and Dynamic Time Warping, as discussed by Z-DIFF seems to be the most important parameter for road
Shahram et al. [3]. anomaly detection.
Eriksson et al. [31] developed the so-called P2 system
2.1 Threshold-based approach for detecting road anomalies which is based on three-axis
acceleration sensors, along with a GPS device (1 Hz) to tag
This approach is based on predefined thresholds that are the anomaly’s location. The authors used the Z-axis to
fixed according to statistical values calculated with a measure vibration anomalies while the X-axis is used to
heuristic detector (e.g., variance, standard deviations), measure speed. Peak acceleration in the Z-axis, as well as
using the input data captured by sensors. X-axis acceleration, was used as thresholds to identify road
In the field of road anomaly detection, several threshold- anomalies. Besides, the authors used an SSW with a fixed
based approaches have been proposed, including the stud- length and overlap, with a sampling rate of 380 Hz, and
ies carried out by Mednis et al. [33], Eriksson et al. [31], various vehicle models to enhance their dataset. Threshold
and Mohan et al. [35]. values were determined by a training P2 algorithm using a
Mednis et al. [33] have proposed a real-time pothole hand-labeled dataset as well as a loosely labeled dataset.
detection system based on processing accelerometer data. The reported results show that P2 algorithm has a very low
For this, the authors proposed real-time heuristics for event false-positive rate with over 90% of true-positive detected
detection and focused on potholes as road anomalies. They potholes.
proposed a simple event detection algorithm called Mohan et al. [35] proposed the Nericell system in order
Z-THRESH which thresholds the acceleration amplitude at to monitor the road and the traffic conditions. The proposed
the Z-axis (Table 1). According to this algorithm, potholes system is based on the use of smartphone sensors (ac-
are identified by values exceeding specific thresholds. celerometer, microphone, GSM radio, and GPS) to detect
Additionally, the authors proposed Z-DIFF algorithm, road bumps and potholes, brake, and honk. To this end, the
which detects rapid changes in the Z-axis acceleration data authors proposed a speed-aware heuristic bump detector:
by calculating the difference between two consecutive (1) the Z-PEAK heuristic for high speeds (  25 kmph) as
Z values. proposed in [31], (2) the new Z-sus bump detector (sus-
Also, the authors implemented the algorithm for a tained dip in Z-axis) for low speeds (\ 25 kmph). Both
standard deviation of the Z-axis STDEV(Z). Finally, the heuristic bump detectors were tested on ‘‘bumpy roads’’
authors provided a G-ZERO algorithm, which aims to (44 bumps or potholes) as well as on ‘‘mixed roads’’
detect the temporary free fall of the vehicle according to (bumps or potholes with smooth highways). For training,
data values of the three axes (values close to 0 g). Based on optimal thresholds of 0.8 g and 1.45 g were chosen for
a tri-axial accelerometer, data were acquired and collected Z-PEAK and Z-sus, respectively.
over a 4.4 km track using four different smartphones (with
sampling rates of 26, 47, 52, and 98 Hz) using an SSW

Table 1 The real-time heuristics proposed by Mednis et al. [33]

Algorithms Description

Z- According to this algorithm, anomalies are identified by acceleration amplitude (Z-axis values) exceeding specific thresholds
THRESH
Z-DIFF This algorithm detects rapid changes in the Z-axis acceleration data by calculating the difference between two consecutive
Z values. When the obtained value exceeds the predefined thresholds, an anomaly exists
STDEV(Z) This algorithm is based on computing the standard deviation for each window and comparing the resulting value to predefined
thresholds to make a decision
G-ZERO This algorithm is designed to detect events when the accelerometer’s three axes values are close to zero, indicating the presence of
an anomaly

123
19018 Neural Computing and Applications (2022) 34:19015–19033

In addition, the Nericell system alternates between the the tri-axis accelerometer and the GPS of the smartphone
two heuristic detectors depending on the value of the fixed to the vehicle’s windshield. In fact, in order to cover
speed. In reality, at low speeds, the system selects Z-sus as all possible cases of potential anomalies, experiments were
the best detector, while at high speeds, the system switches conducted using five types of vehicles and 100.3 km of
to Z-PEAK. Therefore, the authors used the SSW with two roads. After preprocessing the collected data, the collected
different lengths to follow the two heuristic-based detec- data were divided, using the SSW technique, into two
tors. However, the obtained results showed that the false- different window lengths of 256 and 170 samples using an
negative rate is still high (20–30%) due to the difficulty in accelerometer frequency of 96 Hz and 47 Hz, respectively,
establishing the ground truth as noted by the authors. while the overlap length was 66% of the window length.
SVM was used to correctly identify three distinct classes of
2.2 Machine learning approach anomalies. RoADS was based on the Radial Basis Function
Kernel (RFB) and enhanced by the combined feature set
The ML approach has been successfully applied in large- from time-domain (TD) as well as the Stationary Wavelet
scale developments in several fields, including road Transformation (SWT). The proposed system performed
anomaly detection. In fact, road surface monitoring was with an accuracy rate of 91.1%.
addressed in many ML-based studies using mobile devices
and sensor technologies [15, 16, 29, 32, 34, 36]. The main 2.3 Dynamic time warping (DTW) approach
goal of the proposed solutions was to come up with the best
ML model able to detect road anomalies. The DTW approach has been used to measure the simi-
Carlos et al. [29] selected SVM as the best classifier larity between any two time series. It is frequently used in
based on the radial basis function kernel (RFB). This latter the field of speech recognition [37], but it is rarely used in
was enriched by twelve features, among which were vari- the field of road anomaly detection [27]. The DTW uses the
ance, standard deviation, and Z-DIFF. The authors pro- training data, which consists of time series values captured
posed a heuristic detector named ‘‘Swarm’’ that was by the accelerometer sensor of several anomalies (i.e.,
inspired by the majority vote approach and is based on six potholes, bumps, metal pumps, etc.), in order to store a
threshold-based heuristics. Besides, the authors used the variety of anomaly window models. Next, the current test
SSW with a fixed-length window of 30 samples and a half data window will be compared to the stored anomaly
overlapping window. They compared the results to eight window models by determining the degree of similarity
heuristic detectors that had been used by other researchers. between them.
The proposed method outperformed the existing ones with The Quick Filter Based Dynamic Time Warping (QFB-
an average F1 score of 0.78. DTW), based on the tri-axial accelerometer readings, was
In addition, Perttunen et al. [15] developed a surface proposed by Zengwei et al. [27] to detect three kinds of
condition monitoring system that detects road surface anomalies (potholes, speed bumps, and metal bumps). The
anomalies using smartphone sensors (GPS and authors used the low and the high thresholds in the vertical
accelerometer) fixed on the windshields of several vehicles. axis of the accelerometer (ZTL = 8.3, ZTH = 11) to
The data were collected from a drive of about 40 minutes detect the length of the candidate window. Then, they used
over 25 km (including several categories of anomalies) at an RF classifier as a filter to select only the anomaly’s
frequency values of 38 Hz for the GPS and 1 Hz for the window in order to reduce the time needed for classifica-
accelerometer. After that, the collected data were filtered tion. Finally, the authors used KNN and DTW classifiers to
by using the Kalman filter and a heuristic filter to reduce identify the kinds of anomalies for each candidate window.
noise and remove unnecessary data. Then, it was framed Three datasets were used in the experiment to make their
using a sliding window with multiple frame lengths from method more robust and realistic. The obtained results
0.5 to 2 s with an overlap of less than half of the window. outperformed the other methods previously mentioned in
Since speed bumps and potholes have lower frequency their paper.
components, the authors applied the Fast Fourier Trans- The drawback of the DTW approach is that it needs to
form (FFT) to the accelerometer signal. Besides, SVM has reference all anomaly models for each different condition
been used to classify the preprocessed data, including 95 (i.e., various vehicles, anomaly forms, vehicle’s speed,
features. The proposed system obtained an accuracy of etc.), which requires extensive processing resources and is
80%. often unreliable.
Also, SVM was used in the road pavement anomaly To sum up, the major drawback of all of the previous
Detection System (RoADS), proposed by Seraj et al. [16] studies related to the road anomaly detection field is the use
to detect three types of anomalies (severe anomalies, mild of the SSW technique to split the collected signals from the
anomalies, and span). The proposed system was based on vibration sensors (more details will be provided in the

123
Neural Computing and Applications (2022) 34:19015–19033 19019

following section). In order to overcome the limitations of eliminating errors and improving the performance of the
SSW, a new windowing technique known as Dynamic anomaly detectors.
Sliding Window (DSW) has been proposed that splits the For instance, at low speed, the SSW technique generally
collected signal data into different lengths depending on divides the anomaly’s signal into several windows, as
the speed of the vehicle. The DSW has been incorporated shown in Fig. 1b. In this case, the entire anomaly cannot be
into different road anomaly detection methods and evalu- covered by the SSW because its size is less than the
ated using two real-world datasets. anomaly’s length. In this case, a longer time is needed to
pass over the anomalies when the vehicle is at low speed.
Consequently, the anomaly may not be captured by the
3 Problem statement detectors. Additionally, at high speeds, the time required to
cross the anomaly is very short. Therefore, when applying
The vibration approach has been widely used by the SSW technique to the collected signal, the selected
researchers [10, 15–18, 29, 31, 33, 35] in the field of road window may include the anomaly’s signal as well as other
anomaly detection in order to collect vibration data after smooth signals, as shown in Fig. 1a. In this situation, the
passing over road anomalies. This approach is based on window length is larger than the anomaly’s length, and
vibration sensors such as accelerometers and gyroscopes thus, the anomaly may not be detected.
being attached to various locations throughout the vehicle. Finally, to improve the performance of the anomaly
The collected data are then preprocessed by splitting them detectors, the selected window length of the signal should
into a countable number of windows. To do so, each of the be as the same as the anomaly’s length. Consequently, in
previous researchers used the SSW (i.e., fixed window this paper, we propose a DSW technique that overcomes
length) to prepare the collected data for subsequent the limitations of the SSW by selecting sliding windows
detection steps. with a variable lengths according to the vehicle speed.
Unfortunately, the SSW can have a negative impact on In fact, the idea of using DSW has been used in many
the performance of the road anomaly detection techniques. areas, such as the work of Laguna et al. [22], Qin et al.
Indeed, the SSW lacks consideration in the detection of [23], and Noor et al. [38] used the DSW technique in the
road anomalies for two reasons. On one hand, the lengths activity recognition domain. Smithy et al. [25] used the
of the anomalies are different in the real world, and a DSW technique in the health care domain, while Yu et al.
longer anomaly may require longer windows to cover [39] used it to deal with multi-label streaming feature
exactly the entire anomaly. On the other hand, in the pre- selection. Bruno et al. [40] used the DSW in network traffic
vious works, the authors varied the range of the window prediction, while Xin et al. [41] used the DSW in reputa-
size from a minimum to a maximum size in order to find tion commercial systems (eBay, Amazon, etc.). Addition-
the optimal size; therefore, the outcomes vary accordingly. ally, Guoqing et al. [26] suggested, in their future work, to
As a consequence, each work proposed its own optimal combine the DSW with malware classification.
dataset-dependent window size, as shown in Table 2. In previous studies, authors that used dynamic sliding
However, it is worthwhile to note the fact that there is a relied on certain characteristics or parameters to define the
huge difference between the anomaly’s window length at length of the window to be adapted during the classifica-
high speed and low speed, which is considered as an tion. In the current work, we define the speed as the win-
important drawback of the SSW. Indeed, the time required dow length’s key parameter. Actually, the main purpose of
to pass the anomaly is proportional to the speed at which our approach is to fit the window length with high accuracy
the vehicle passes over it and thus the size of the window. to the anomaly’s data, as shown in Fig. 2 (windows 6 and
Therefore, selecting sliding windows with a variable 10).
lengths according to the vehicle speed is crucial to

Table 2 Length and overlap

Authors Window length (number of samples) Overlapping length (%)
window parameters used in the
grid search process for static Carlos et al. 30 50%
windowing
Mednis et al. 100 No overlapping
Mohan et al. 310 No overlapping
Perttunen et al. 38 No overlapping
Silva et al. 125 60%

123
19020 Neural Computing and Applications (2022) 34:19015–19033

(a) (b)
Fig. 1 SSW with low and high speeds

Fig. 2 DSW usage with real data

4 The proposed approach acceleration signal can exactly fit the shapes of the
anomalies, and as a consequence, their detection rate can
To overcome the main limitation of the SSW technique, we be significantly improved. The whole process of detecting
propose in this paper the DSW technique, which is a speed- anomalies using the proposed DSW technique is illustrated
dependent vibration technique. The key idea of DSW is to in Fig. 3. The proposed approach is based on three main
dynamically select the window length and the overlapping steps: the preprocessing step, the processing step, and the
length according to the vehicle speed, as presented in evaluation step.
Figs. 2 and 4. By doing that, the windows of the

123
Neural Computing and Applications (2022) 34:19015–19033 19021

4.1 Preprocessing step In Eq. (1), D represents the distance between the front
wheel and the rear wheel of the car (see Fig. 5), which
The aim of this step is to compute the length of the sliding varies from 2 to 6 m, while Si represents the vehicle’s
window and the overlap length according to the vehicle’s speed, which is time-dependent (Table 3). DTi is the period
speed. This can be done by using Eqs. (1, 2, and 3): that the vehicle takes to pass over the anomaly, it can be
D calculated as the difference between the start time tk and
DTi ¼ ð1Þ the end time tkþn as detailed in Fig. 5. According to
Si
Eq. (2), for each anomaly, the window’s length is refer-
Lwi ¼ DTi  F ð2Þ enced as Lwi . The length value is calculated according to
Lwi the speed Si . F displays the accelerometer sample rates (in
Loi ¼ ð3Þ Hz), which is dataset dependent. Therefore, the main
X
problem is to find the dynamic window length, which must

123
19022 Neural Computing and Applications (2022) 34:19015–19033

Fig. 3 Main steps of the proposed DSW approach

Table 3 The used notations and

Notation Description
their definitions
Lwi The window’s length
Loi The overlap’s length
Si The vehicle’s speed
D The distance between the front wheel and the rear wheel of the car (see Fig. 5)
DTi The period to pass over the anomaly
F The accelerometer sample rates (in HZ)
X The optimal number chosen to achieve the best overlap length

100
Length (sample)

80

Overlapping Length Window Length

60

40

20
Fig. 5 Illustration of the car encountering an anomaly
0
5 10 15 20 25 30 35 40 45 50 55 60 65 70
Speed (km/h)
Table 4 Vehicle types used for pothole laboratory experiment col-
Fig. 4 Variation in window length and overlap length according to the lection [29]
speed, using Eqs. 2 and 3
Vehicle Wheelbase (m)

be changed according to DTi , as shown in Figs. 2 and 4. In VW Jetta 2002 2.512

the current work, and in order to compute the dynamic Chevrolet pickup S10 1991 2.75
window length (Lwi ), we have fixed the frequency F at Chevrolet Chevy 2009 3.022
50 Hz (as done in [29]), and the distance D at 2.784 m, Nissan Frontier 2001 2.948
which is the average of all vehicle wheelbases presented in Toyota Camry 1999 2.672
Table 4. Nissan Altima 2003 2.8
The overlap also changes as a function of dynamic
window length and speed, as can be seen in Fig. 4. This
overlapping window is given by Eq. 3, which determines
how many samples we take from the previous window. To the length of every window (Lwi ) by X, which varies from 2
find the optimal overlap length for each window, we divide to 10. For example, when X = 2, the overlap takes 50% of

123
Neural Computing and Applications (2022) 34:19015–19033 19023

0.8 in the case of DSW, where the window length fluctuates

Max of F1 score = 0.835
F1 score average

with speed, the DTW technique is inapplicable. Further-

0.7
DSW SVM more, this approach is extensively employed in speech
0.6
recognition. In the case of the ML approach, we have used
0.5 seven classifiers, namely DT, GB, RF, NB, NN, KNN, and
0.4 SVM.
2 3 4 5 6 7 8 9 10
X To accomplish this, it is necessary to extract the features
of each window from the training data set using a feature
Fig. 6 The optimal overlap value based on the variation in X extraction algorithm that detects abnormal changes in the
accelerometer signal in order to train the ML model, which
the previous window, and when X = 3, the overlap takes will be used later to classify the real data. On the other
33% of the previous window. Note that, starting with hand, in the case of the heuristic detectors, we have
X = 1 is not possible, because that gives Lwi ¼ Loi which implemented seven existing heuristics, namely Z-
causes an endless circle in the windowing. The DSW THRESH, Z-DIFF, STDEV(Z), G-ZERO, Swarm, P2 , and
algorithm remains fixed in the same position. Furthermore, Nericell, based on their own optimal threshold (see
we stopped at X = 10 because the short window size in our Table 5) to detect road surface anomalies. These heuristics
data set is 10 (Lwi ¼ 10). In this case, if we divide Lwi by are applied directly to the test data set.
X = 11 the result of Loi \1, which means there is no
overlap. In order to select the optimal value of X, we tested 4.3 Evaluation step
its values with the training data set presented in [29]. The
results are plotted in Fig. 6. As can be seen, the best F1 To evaluate the performance of the ML classifiers, as well
score was obtained with X = 4. as the heuristic detectors, we apply three popular evalua-
The details of the DSW technique are described in tion metrics, namely precision, sensitivity, and F1 score,
algorithm 1, which uses the accelerometer signal and the for the purpose of selecting the best classifier for the field
vehicle’s speed as inputs and generates as output a set of of road surface anomaly detection.
accelerometer signals of different windows size, depending
on the vehicle’s speed.
5 Experimental results and discussion
4.2 Processing step
In this section, we present the experimental settings,
This step concerns both the ML and the threshold-based including the dataset, the evaluation metrics, and experi-
approaches. Our comparison excludes the DTW-based mental comparison. The experiments were conducted on
technique. Due to the fact that the two time-series, the the Pothole-Lab dataset using seven ML classifiers and
model window and the test data window, must all have the
same length in order to establish their similarity, as a result,

Table 5 Thresholds values for features detector

Feature Threshold Definition

Mean No threshold Average Z-axis value of every window

STDEV(Z) No threshold Standard deviation
Var No threshold Standard of variation
CV No threshold Coefficient of variation
Z-DIFF No threshold Difference between the maximum and the minimum Z axis vlue for each window
SC No threshold Sum of Confidence Score
CS-Mean g*0.3 Confidence score of the mean
CS-STDEV(Z) g*0.15 Confidence score of the STDEV(Z)
CS-CV g*0.015 Confidence score of the CV
CS-Z-DIFF g*0.2 Confidence score of the DIFF (max, min)
CS-Var g*0.15 Confidence score of the variance
CS-SC 3 Confidence score of SC

123
19024 Neural Computing and Applications (2022) 34:19015–19033

seven heuristic detectors. The environment setting for the percentage of accuracy and is defined as the harmonic
experiments is as follows: mean of precision and recall, as shown in Eq. (6).
• Operation system: Windows 10 Professional 64-bit with TP
sensitivity ¼ ð4Þ
16 GB of DDR4 RAM and an SSD hard drive; TP þ FN
• Processor: Intel Core i7-8550U @ 1.99 GHz. TP
precision ¼ ð5Þ
TP þ FP
5.1 Dataset 2TP
F1 score ¼ ð6Þ
2TP þ FP þ FN
In order to test our approach, we used two datasets. The
first one, presented by Carlos et al. [29], is generated by the 5.3 Experimental comparison
Pothole Lab,2 which is considered a freely available
benchmark web platform, using triaxial acceleration mea- In order to evaluate the classification efficiency in the
surements. This web platform generates virtual road windowing stage, we compared the results of the proposed
anomalies based on the physical anomaly’s data. The DSW technique with the classical SSW techniques using
dataset was generated using six different kinds of vehicles. two different approaches of road anomaly detection. The
It includes 187 anomalies, among them, 81 were asphalt first approach is the ML approach, which is widely used in
speed bumps, 50 were metal bumps, and 56 were potholes. the road anomaly detection domain, while the second one
The dataset contains thirty roads, as shown in Fig. 7. is the Threshold-based approach. However, an efficient
The data set is divided into training data and testing data, comparison between classifiers needs to use the same
with two sets of roads: homogeneous (with one type of datasets, under the same conditions. Regarding the SSW,
anomaly) and heterogeneous (with different types of we set the window length to 30 samples and the overlap
anomalies). In order to cover the maximum number of length to 15 samples as presented in Carlos et al. [29] work,
possible cases, the number of anomalies for each type of where they fixed the optimal length at 30 after varying the
anomaly (speed bumps, potholes, metal bumps), as shown window length from 10 to 100 samples.
in Fig. 8, ranges from 5 to 100. All the anomaly patterns
data were detected on real roads using different vehicles as 5.3.1 Machine learning approach
shown in Table 4 and also using different types of
anomalies as shown in Table 6. The experiments were Concerning the ML approach, Table 8 shows the parame-
conducted using a 50 Hz sampling frequency rate, adding a ters selected for each classifier. The present work is based
few seconds of smooth road data after and before each on the implementation provided by the Python-based
anomaly in order to create the distance between two suc- platform namely Scikit-Learn.3 In fact, seven classifiers are
cessive anomalies. Various speed profiles were used with used in this paper to detect road surface anomalies namely
standard roads in residential streets and high-speed roads. DT, GB, RF, NB, NN, KNN, and SVM.
The second dataset was generated using the Carsim Also, the same features used in [29] were re-used in this
platform, based on the idea of Andrew et al. [30] and work, along with some statistical features inspired by
Zengwei et al. [27]. This platform has been used to simu- Mednis et al. [33] study, such as the Mean, the Standard
late cars driving over three types of anomalies: potholes, Deviation (Stdev), the Variance (Var), the coefficient of
metal bumps, and speed bumps. The resulting database variation (CV), and the difference between the Max and the
includes 1518 speed bumps, 1485 metal bumps, and 2376 Min (Z-DIFF). Also, as shown in Table 5, five features are
potholes (see Table 7). used to calculate the confidence score of the previous
features. These confidence scores are obtained after com-
5.2 Evaluation metrics paring the previous features with their fixed threshold. The
remaining two features are the sum of the confidence score
In addition to precision and F1 score, sensitivity has been and its corresponding confidence value.
also used in this work to measure the percentage of the
True Positive Rate (TPR), as shown in Eq. (4). The pre- 5.3.2 Threshold-based approach
cision is used to measure the percentage of the Positive
Predicted Value (PPV), as shown in Eq. (5). Besides, the In this second approach, seven heuristic algorithms were
F1 score, also called the F-measure, is used to calculate the used. Among them, Z-THRESH uses fixed threshold values

2 3
https://www.accelerometer.xyz/datasets. http://scikit-learn.org/stable.

123
Neural Computing and Applications (2022) 34:19015–19033 19025

50
Training dataset
40

30
Potholes Speed bumps Metal bumps
Number of Anomalies

20

10

0
0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30

50
Road number

40 Test dataset
30

20

10

0
0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30
Road number

Fig. 7 Anomalies included in the training and test data of dataset 1

(a) (b) (c)

Fig. 8 Example of vertical acceleration signal for a speed bumps, b metal bumps, c Pothole, generated by Pothole Lab

Table 6 Anomalies types included in the pothole lab web platform to analyze all raw vertical acceleration samples (in the Z-
axis). If a sample is outside of the defined margin, that is
Category Samples
considered as the presence of a peak caused by an anomaly
Asphalt speed bumps 81 in the road. G-ZERO algorithm aims to detect the situation
Pothole 56 where three accelerometer axes values are close to 0.
Metal bump 50 Z-DIFF algorithm detects the rapid changes in the
Plane road 50 Z-axis acceleration data by calculating the difference
Total 237 between two consecutive Z-axis values, and the
STDEV(Z) algorithm targets calculating the Standard
Deviation of Z-axis. In addition, we used the Nericell
system, which includes two heuristic detectors: a Z-SUS
detector (sustained dip in Z-axis) and a Z-PEAK detector
originally presented by Mednis et al. [33].
Table 7 Anomalies types included in the second dataset The P2 heuristic proposed by Eriksson et al. [31] uses
Potholes Speed bumps Metal bumps
high-pass and low-pass filters to eliminate the low-pass
data and analyze the high-pass data by the Z-PEAK
The number of anomalies 2376 1518 1485 detector following the same logic as Z-THRESH to detect
the location of the pothole. Swarm presented by Carlos
et al. [29] is based on a majority vote of six popular

123
19026 Neural Computing and Applications (2022) 34:19015–19033

Table 8 Features and

Classifier Parameters
parameters used in the ML
algorithms KNN n_neighbors = 9, weights = ’distance’, leaf_size = 20, algorithm = ‘‘kd_tree
DT min_samples_split = 36
RF n_estimators = 100, criterion = ’gini’
SVM C = 10, kernel = ’rbf, tol = 0.001
GB n_estimators = 25
NN solver = ’lbfgs’, alpha = 1e-5, hidden_layer_sizes = (14,5), random_state = 1
NB No parameters

threshold-based heuristics, to detect road anomalies. All of SSW technique on 28 roads. However, it suffers on road
those heuristics algorithms were used with both the pro- 20, with an F1 score rate of 0.592, which is also due to
posed DSW and the SSW techniques. training data that contains only the metal bumps as
anomalies.
5.4 Results On the other hand, the RF classifier combined with the
DSW technique performs better, with an F1 score average of
The results of our experiments using two datasets (see 0.832 against 0.663 obtained by the SSW technique. For
Fig. 7 and Table 7) are presented in Tables 9, 10, and 11. instance, RF performs better for 28 roads. The same behavior
We compared the classical SSW technique with the pro- is observed when using the NN classifier. The obtained
posed DSW technique, using the same seven ML classi- results for the DSW technique exceed those obtained using
fiers, seven heuristic detectors, and DTW approach as the SSW technique on 25 roads, as can be seen in Fig. 9e.
explained in the previous section. In the same way, the NB classifier combined with DSW
gives better results on 29 roads, and the F1 score average
5.4.1 Machine learning approach evaluation has improved from 0.637 to 0.761. Finally, with the DT,
which also performed better with the proposed DSW
Figures 9 and 10 illustrate the results obtained by the ML technique, the F1 score was enhanced from 0.532 to 0.726,
approach on the first dataset. First, we re-implemented as shown in Table 9. Also, it outperforms the classical
Carlos et al. [29] work, which is based on the SVM clas- SSW technique on 27 roads.
sifier, using the DSW technique instead of the SSW tech- Among existing cases, the NN classifier in Road 19
nique. As a result, and based on the SVM classifier, the gives a value of 0.902 F1 score for SSW, in contrast to
proposed DSW performs better on five roads (1, 7, 11, 12, 0.894 for DSW. Also, the SVM classifier in road 4 obtains
and 22) among commonly used ML classifiers. an F1 score of 0.907 compared to 0.789 obtained using the
In addition, for 20 roads, the proposed DSW technique proposed DSW approach. Despite that, we can conclude
with SVM classifier performs better than using SSW. Also, that the proposed DSW technique outperforms the classical
the F1 score average rate using DSW is higher than the SSW technique in road anomaly detection using all ML
average obtained using SSW. According to Fig. 9a, the F1 classifiers. The obtained results prove the efficiency of
score varies between 0.746 and 1 on 27 roads using the adopting the proposed DSW technique instead of the
proposed DSW approach with SVM classifier. However, classical SSW technique on time-series readings in the road
SVM did not perform well on roads 13, 15, and 16 due to anomaly detection field.
the training data set, which did not contain enough data to Also, to improve the robustness of our comparison, we
train the SVM classifier. applied fivefold cross-validation to the previously descri-
Among the ML classifiers that perform well in this bed ML classifiers on the second dataset generated with
paper, the KNN classifier gives the best results even before Carsim, which included 5379 anomalies. As illustrated in
using the DSW technique, with an F1 score average of Tables 10 and 11, the DSW consistently outperforms the
0.862 when used with the DSW technique against 0.718 SSW. The lack of sufficient training data explains the large
when combined with the SSW technique. The KNN clas- discrepancy between the first and second datasets’ results.
sifier, along with the proposed DSW technique, is In the first dataset, the training data were smaller than the
stable over 30 roads, as can be seen in Fig. 9f, despite the test data, which resulted in the insufficient training for the
insufficient size of training data on some roads such as 11, implemented methods.
12, and 13. The GB classifier is ranked in the third position On the other hand, the second dataset yielded 80% of the
when using the DSW approach. It achieves an F1 score training data. That is why the outcome is enhanced, espe-
average of 0.833 and outperforms itself when using the cially with the machine learning approaches. According to

123
Neural Computing and Applications (2022) 34:19015–19033 19027

Table 9 F1 score results of seven classifiers and seven heuristic algorithms tested in 30 roads
Road DT GB RF NB NN kNN SVM
SSW DSW SSW DSW SSW DSW SSW DSW SSW DSW SSW DSW SSW DSW

1 0.780 0.727 0.741 0.845 0.734 0.829 0.673 0.785 0.780 0.833 0.747 0.868 0.821 0.892
2 0.468 0.690 0.686 0.832 0.505 0.895 0.633 0.772 0.593 0.726 0.638 0.829 0.833 0.780
3 0.598 0.776 0.859 0.882 0.751 0.906 0.778 0.876 0.886 0.867 0.865 0.949 0.891 0.925
4 0.460 0.772 0.786 0.800 0.590 0.731 0.671 0.818 0.800 0.891 0.755 0.844 0.907 0.789
5 0.495 0.839 0.563 0.852 0.589 0.852 0.479 0.548 0.493 0.821 0.708 0.847 0.603 0.778
6 0.536 0.735 0.771 0.847 0.646 0.878 0.619 0.676 0.680 0.805 0.720 0.827 0.676 0.784
7 0.531 0.789 0.893 0.947 0.738 0.941 0.769 0.868 0.859 0.863 0.807 0.932 0.872 0.958
8 0.314 0.837 0.683 0.859 0.527 0.757 0.609 0.793 0.644 0.883 0.676 0.841 0.837 0.882
9 0.554 0.598 0.728 0.785 0.569 0.832 0.561 0.745 0.671 0.882 0.598 0.810 0.755 0.875
10 0.488 0.782 0.708 0.915 0.626 0.823 0.644 0.795 0.713 0.850 0.659 0.872 0.690 0.911
11 0.900 0.900 0.889 0.857 0.947 0.889 0.900 0.824 0.947 0.889 0.761 0.889 0.947 0.947
12 0.769 0.800 0.869 0.818 0.666 0.952 0.952 1.000 0.909 0.952 0.909 0.947 1.000 1.000
13 0.562 0.182 0.588 0.909 0.636 0.778 0.500 0.571 0.428 0.842 0.375 0.833 0.625 0.588
14 0.444 0.800 0.571 0.857 0.529 0.842 0.563 0.727 0.769 0.842 0.818 0.900 0.888 0.800
15 0.327 0.727 0.545 0.640 0.529 0.750 0.529 0.667 0.700 0.842 0.800 0.800 0.750 0.615
16 0.512 0.762 0.692 0.727 0.762 0.667 0.640 0.667 0.667 0.600 0.736 0.778 0.777 0.696
17 0.517 0.828 0.600 0.828 0.718 0.824 0.636 0.750 0.718 0.824 0.615 0.867 0.774 0.788
18 0.403 0.746 0.757 0.857 0.649 0.906 0.533 0.750 0.774 0.870 0.724 0.880 0.920 0.840
19 0.623 0.602 0.772 0.815 0.786 0.920 0.602 0.800 0.902 0.894 0.745 0.941 0.862 0.894
20 0.427 0.438 0.512 0.592 0.517 0.807 0.460 0.612 0.542 0.706 0.578 0.808 0.750 0.755
21 0.434 0.852 0.857 0.873 0.811 0.903 0.615 0.807 0.786 0.830 0.757 0.933 0.827 0.857
22 0.559 0.737 0.793 0.833 0.683 0.730 0.848 0.877 0.857 0.784 0.865 0.877 0.800 0.933
23 0.574 0.718 0.684 0.769 0.698 0.844 0.600 0.800 0.716 0.759 0.704 0.857 0.823 0.833
24 0.443 0.746 0.754 0.863 0.623 0.845 0.503 0.752 0.689 0.828 0.713 0.907 0.853 0.845
25 0.583 0.598 0.755 0.840 0.614 0.809 0.644 0.773 0.748 0.756 0.723 0.864 0.860 0.810
26 0.484 0.676 0.638 0.750 0.676 0.677 0.605 0.754 0.622 0.857 0.745 0.746 0.826 0.769
27 0.500 0.824 0.807 0.917 0.745 0.821 0.657 0.732 0.760 0.762 0.656 0.920 0.717 0.894
28 0.596 0.812 0.778 0.947 0.745 0.920 0.684 0.851 0.818 0.857 0.656 0.884 0.829 0.933
29 0.638 0.795 0.704 0.880 0.679 0.795 0.647 0.747 0.780 0.806 0.744 0.780 0.757 0.822
30 0.436 0.685 0.769 0.842 0.608 0.847 0.562 0.682 0.682 0.735 0.673 0.835 0.656 0.821
Avg 0.532 0.726 0.725 0.833 0.663 0.832 0.637 0.761 0.731 0.822 0.718 0.862 0.804 0.834
Road P2 STDEV(Z) Swarm Z-DIFF G-ZERO Z-THRESH Nericell

SSW DSW SSW DSW SSW DSW SSW DSW SSW DSW SSW DSW SSW DSW

1 0.723 0.775 0.861 0.842 0.794 0.835 0.631 0.727 0.717 0.827 0.736 0.819 0.690 0.732
2 0.666 0.772 0.865 0.848 0.813 0.851 0.644 0.729 0.784 0.833 0.666 0.827 0.685 0.756
3 0.769 0.784 0.895 0.866 0.863 0.848 0.456 0.348 0.884 0.857 0.770 0.784 0.701 0.733
4 0.770 0.808 0.937 0.891 0.901 0.845 0.580 0.416 0.888 0.863 0.763 0.804 0.765 0.783
5 0.432 0.566 0.653 0.746 0.625 0.667 0.545 0.712 0.675 0.833 0.428 0.607 0.482 0.571
6 0.675 0.667 0.764 0.868 0.666 0.800 0.755 0.786 0.847 0.846 0.690 0.701 0.703 0.657
7 0.814 0.900 0.914 0.905 0.866 0.880 0.616 0.564 0.838 0.848 0.803 0.891 0.848 0.882
8 0.699 0.782 0.837 0.841 0.805 0.820 0.585 0.489 0.783 0.861 0.694 0.783 0.675 0.732
9 0.660 0.722 0.777 0.845 0.767 0.750 0.576 0.575 0.807 0.828 0.677 0.755 0.641 0.711
10 0.692 0.761 0.843 0.882 0.795 0.832 0.585 0.555 0.831 0.806 0.699 0.798 0.677 0.734
11 1.000 1.000 0.947 0.750 0.888 0.824 0.588 0.533 0.947 0.889 1.000 1.000 0.824 0.824
12 0.947 0.947 1.000 1.000 1.000 1.000 0.842 0.889 0.952 1.000 1.000 1.000 0.947 0.947
13 0.421 0.429 0.625 0.818 0.666 0.842 0.666 0.750 0.900 0.842 0.454 0.533 0.353 0.462

123
19028 Neural Computing and Applications (2022) 34:19015–19033

Table 9 (continued)
Road P2 STDEV(Z) Swarm Z-DIFF G-ZERO Z-THRESH Nericell

SSW DSW SSW DSW SSW DSW SSW DSW SSW DSW SSW DSW SSW DSW

14 0.640 0.667 0.842 0.889 0.888 0.800 0.516 0.640 0.571 0.889 0.740 0.741 0.696 0.667
15 0.727 0.667 0.705 0.750 0.705 0.778 0.636 0.526 0.800 0.889 0.727 0.609 0.696 0.556
16 0.608 0.556 0.750 0.842 0.750 0.842 0.454 0.333 0.75 0.857 0.666 0.667 0.571 0.588
17 0.666 0.733 0.888 0.839 0.814 0.788 0.465 0.593 0.764 0.813 0.666 0.667 0.611 0.667
18 0.666 0.772 0.840 0.852 0.851 0.800 0.379 0.364 0.654 0.766 0.657 0.780 0.645 0.755
19 0.727 0.863 0.938 0.906 0.893 0.842 0.619 0.769 0.714 0.851 0.640 0.793 0.755 0.833
20 0.571 0.542 0.682 0.851 0.700 0.745 0.575 0.607 0.758 0.840 0.606 0.593 0.571 0.522
21 0.710 0.787 0.857 0.833 0.857 0.833 0.500 0.604 0.684 0.780 0.658 0.769 0.657 0.772
22 0.727 0.760 0.846 0.915 0.775 0.867 0.603 0.609 0.935 0.949 0.711 0.784 0.708 0.723
23 0.771 0.792 0.824 0.885 0.775 0.825 0.676 0.755 0.787 0.881 0.813 0.793 0.769 0.784
24 0.697 0.783 0.889 0.826 0.846 0.809 0.649 0.745 0.691 0.841 0.678 0.768 0.729 0.791
25 0.720 0.738 0.850 0.850 0.810 0.785 0.522 0.375 0.780 0.795 0.692 0.747 0.713 0.703
26 0.626 0.691 0.851 0.840 0.869 0.792 0.586 0.702 0.769 0.776 0.647 0.702 0.632 0.680
27 0.595 0.605 0.780 0.894 0.648 0.894 0.690 0.642 0.816 0.844 0.625 0.622 0.578 0.571
28 0.725 0.821 0.868 0.875 0.816 0.814 0.574 0.636 0.777 0.861 0.752 0.886 0.701 0.783
29 0.700 0.676 0.738 0.806 0.666 0.757 0.623 0.554 0.820 0.800 0.714 0.714 0.597 0.590
30 0.645 0.667 0.728 0.785 0.655 0.747 0.574 0.623 0.750 0.784 0.653 0.700 0.635 0.657
Avg 0.693 0.734 0.826 0.851 0.792 0.817 0.590 0.605 0.789 0.845 0.701 0.755 0.675 0.706

Bold indicates the best classification approach among all the presented classifiers in the machine learning approach for each road

Table 10 The obtained F1 score

ML classifiers
from the seven ML classifiers
evaluated on the second dataset DT GB RF NB NN KNN SVM QFB-DTW [27]

F1 score
F1 score (SSW) 0.885 0.887 0.907 0.871 0.885 0.901 0.885 0.83
F1 score (DSW) 0.967 0.917 0.922 0.888 0.904 0.973 0.922 –
Bold indicates the optimal result between the DSW and SSW for each classifier

Table 11 The obtained F1 score

Heuristics
from the seven heuristic
algorithms evaluated on the P2 STDEV(Z) Swarm Z-DIFF G-ZERO Z-THRESH Nericell
second dataset
F1 score
F1 score (SSW) 0.557 0.846 0.805 0.463 0.815 0.676 0.585
F1 score (DSW) 0.725 0.856 0.853 0.600 0.852 0.728 0.715
Bold indicates the optimal result between the DSW and SSW for each classifier

Table 10 and Fig. 11a, the KNN classifier combined with score from 0.885 to 0.967. With the proposed DSW, the
DSW ranked in the first place with the best results and SVM and RF classifiers both produced an average F1 score
achieved an average F1 score of 0.973, compared to 0.901 of 0.922, rather than 0.885 and 0.907, respectively. Addi-
obtained with SSW. DT classifier increases the average F1 tionally, the GB classifier achieves an average F1 score of

123
Neural Computing and Applications (2022) 34:19015–19033 19029

(a) (b)
(c)

(d)
(e) (f)

(g) (h) (i)

(j) (k) (l)

Fig. 9 Comparison between DSW and SSW applied to ML and threshold-based approaches

0.917 when using DSW, compared to 0.887 when using as mentioned in Sect. 5.3.2. According to the conducted
SSW. Finally, the proposed DSW approach results in an experiments, the STDEV(Z) [33] gives the best results of
average F1 score of 0.904 and 0.888 for the NN and NB all the heuristic detectors presented in this paper, using the
classifiers, respectively, compared to 0.885 and 0.871 for DSW technique. The average F1 score using the SSW is
the classical SSW approach. raised from 0.826 to 0.851 with the DSW and performs
In conclusion, we found that KNN with DSW using better on 19 roads.
vibration-based approach is the most effective classifier for The G-ZERO heuristic detector [33] was ranked in the
detecting road anomalies using both datasets. second position before using the DSW technique. We note
that when this detector is combined with the DSW tech-
5.4.2 Threshold-based approach evaluation nique, the average F1 score is enhanced from 0.789 to
0.845, as shown in Fig. 10. Also, it performs better on 23
In the second part of the evaluation of the proposed DSW roads out of 30, and it can be considered as robust as shown
technique, we have implemented seven existing heuristics in Figs 9l. The only downside of this heuristic is the

123
19030 Neural Computing and Applications (2022) 34:19015–19033

Fig. 10 Average F1 score for 30 roads for each studied approaches, between the DSW and SSW

increase in the run-time compared to the other heuristic Besides, DTW is tailored to SSW and not to DSW due to
detectors, which is caused by the use of the three-axis (X, the fact that the first precondition for implementing this
Y, and Z) of the accelerometer, instead of using only the method is that the two-time series (the model window and
Z-axis by other heuristics. the test data window) need to have the same length.
Swarm [29] gives a good and stable result with DSW The QFB-DTW method proposed by Zengwei et al. [27]
and achieves an average F1 score of 0.817 against 0.792 has been implemented in our experiment using tenfold
with SSW and performs better on 19 roads. Z-THRESH cross-validation, which is ranked last among the ML
gets the max F1 score of 1 is obtained for both techniques, methods presented in this paper, with an F1 score average
DSW and SSW, on roads 11 and 12 as their data contain of 0.830. Despite the fact that we used QFB-DTW exclu-
speed bumps and potholes, respectively. The 100% detec- sively with anomaly data (as done in [27]), it takes a large
tion rate is due to the generated upward vertical signals that amount of time to learn, then generate a final model, and
reach the high threshold. However, the Z-THRESH finally to predict the test data.
detector did not perform well on road 13 as the data con-
tained only metal bumps as anomalies. According to 5.4.4 Time consumption evaluation
Fig. 8b, we need a low-threshold instead of a high-
threshold to detect metal bumps because of their downward In the case of the DSW technique, the variations in the
vertical signal. For this reason, road 13 using G-ZERO window length values are between 5 and 200 depending on
with the DSW technique, which uses the strategy of the the speed, unlike in the SSW technique where the window
low-threshold, performs better with metal bumps. length is fixed at 30.
The DSW technique combined with the P2 detector Figure 12 shows that most of the window length values
increases the average F1 score from 0.693 to 0.734. The for the DSW technique vary between 10 and 50, and the
last group of heuristic detectors includes Nericell [35] and mean value is about 30 samples. In this paper, the DSW
Z-DIFF [33], both of which increase the average F1 score technique obtained better results than the SSW one simply
after using DSW, as shown in the corresponding columns because the length is dynamic and, consequently, the
of Table 9. window only contained anomaly data without the smooth
In the case of the second dataset, Table 11 and Fig. 11b one.
illustrate the results obtained using both DSW and SSW After the adoption of the DSW technique on the training
techniques for the seven heuristics. As demonstrated, all and test data set (for 30 roads), the fluctuation in window
heuristic methods performed better when using DSW rather length reduced the total number of windows, as can be seen
than SSW. from Table 12.
For instance, it reduced training data by 13,320 windows
5.4.3 Dynamic time warping approach evaluation and test data by 22,969 windows, which in fact reduced the
run time of road anomaly detector algorithms for both ML
The DTW approach has been used only with the second and Threshold-based approaches (see Table 13). To con-
dataset due to the requirements of the large training data. clude this part, it is worthy to note that the used dataset [29]
presented two major limitations.

123
Neural Computing and Applications (2022) 34:19015–19033 19031

(a) (b)

Fig. 11 Comparison between DSW and SSW applied to ML and threshold-based approaches with the second dataset

To conclude this part, it is worthy to note that the used

dataset [29] presented two major limitations. The first one
is the illogical abrupt increase or decrease in speed as
shown in Fig. 2. In reality, the speed values change regu-
larly in a time interval. However, in the used dataset, speed
values at time Ti and Tiþ1 increase by 40 km/h (from 6 to
46 km/h). The second limitation is the absence of anomaly
data at high speeds (speeds superior to 50 km/h), which has
a negative influence on the DSW performance.

6 Conclusion and future work

The purpose of our study is to come up with a new

approach in the field of road anomaly detection that could
save people’s lives. Nowadays, scholars and researchers
Fig. 12 The variation in window length using the DSW algorithm
take advantage of the widely used smartphones, especially
those embedded with various sensors, such as GPS,
accelerometer, gyroscope, and cameras to detect road sur-
Table 12 The number of generated windows training and test data
face anomalies. The main contribution of this paper is
with DSW and SSW
based on using the proposed DSW technique instead of the
Number of windows: training data Number of windows: test data classical SSW in road anomaly detection. To evaluate the
DSW SSW DSW SSW proposed DSW technique, we compared it with the classic
SSW technique using seven classifiers and seven heuristic
13,656 26,976 23,327 46,296
algorithms under the same conditions. Also, we used a
Bold indicates the lower number of generated windows dataset consisting of 30 roads with multiple anomalies. As

Table 13 The average

Exec time of KNN (s) Exec time of DT (s) Exec time of SVM (s)
execution times for training and
test data using DSW and SSW DSW SSW DSW SSW DSW SSW
on 30 roads
3.620 4.239 0.243 0.334 2.402 5.470
Bold indicates the best execution times
Exec Times execution times

123
19032 Neural Computing and Applications (2022) 34:19015–19033

a result, the proposed DSW technique has shown its effi- profiling from smartphone sensors. Pervasive Mob Comput
ciency for all the used classifiers and heuristic algorithms 61:101103
11. Egaji OA, Evans G, Griffiths MG, Islas G (2021) Real-time
previously discussed. Indeed, the DSW technique reduces machine learning-based approach for pothole detection. Expert
also the number of windows by half compared to the SSW Syst Appl 184:115562
technique and decreases the time complexity for the 12. Pandey AK, Palade V, Iqbal R, Maniak T, Karyotis C, Akuma S
majority of classifiers and heuristics. In future work, we (2022) Convolution neural networks for pothole detection of
critical road infrastructure. Comput Electr Eng 99:107725
plan to improve the DSW technique by including other 13. Lekshmipathy J, Velayudhan S, Mathew S (2021) Effect of
parameters and filters in order to reduce the time com- combining algorithms in smartphone based pothole detection. Int
plexity and better deal with useless data. Also, we plan to J Pavement Res Technol 14(1):63–72
enhance the classification stage by adopting a fusion 14. Agrawal H, Gupta A, Sharma A, Singh P (2021) Road pothole
detection mechanism using mobile sensors. In: 2021 International
method for the ML approach. We also target to develop a conference on technological advancements and innovations
real-time centralized road anomaly detection system (cli- (ICTAI). IEEE, pp 26–31
ent/server application), to show people instantaneously, 15. Perttunen M, Mazhelis O, Cong F, Kauppila M, Leppänen T,
from one hand, the location of road surface anomalies on Kantola J, Collin J, Pirttikangas S, Haverinen J, Ristaniemi T et al
(2011) Distributed road surface condition monitoring using
online maps, at the same time of receiving accelerometer, mobile phones. In: International conference on ubiquitous intel-
gyroscope, and GPS data filtered by smartphones, and on ligence and computing. Springer, pp 64–78
the other hand, inform authorities about those anomalies. 16. Seraj F, van der Zwaag BJ, Dilo A, Luarasi T, Havinga P (2015)
Roads: a road pavement monitoring system for anomaly detection
using smart phones. In: Big data analytics in the social and
ubiquitous context. Lecture Notes in Computer Science, vol
9546. Springer, pp 128–146
Declarations 17. Singh G, Bansal D, Sofat S, Aggarwal N (2017) Smart patrolling:
an efficient road surface monitoring using smartphone sensors
Conflict of interest The authors declare that they have no known and crowdsourcing. Pervasive Mob Comput 40:71–88
competing financial interests or personal relationships that could have 18. Souza VM, Giusti R, Batista AJ (2018) Asfault: a low-cost sys-
appeared to influence the work reported in this paper. tem to evaluate pavement conditions in real-time using smart-
phones and machine learning. Pervasive Mob Comput
51:121–137
References 19. Valero M, Li F, Zhao L, Zhang C, Garrido J, Han Z (2021)
Vibration sensing-based human and infrastructure safety/health
monitoring: a survey. Digital Signal Process 114:103037
1. Organization WH (2018) Global status report on road safety
20. Turner A (2022) How many smartphones are in the world.
2018. World Health Organization, Geneva
Bankmycell. https://www.bankmycell.com/blog/how-many-
2. Qiao Y, Santos J, Stoner AM, Flinstch G (2020) Climate change
phones-are-in-the-world. Accessed 15 March 2022
impacts on asphalt road pavement construction and maintenance:
21. Wang J, Yang Y, Wang T, Sherratt RS, Zhang J (2020) Big data
an economic life cycle assessment of adaptation measures in the
service architecture: a survey. J Internet Technol 21(2):393–405
state of Virginia, United States. J Ind Ecol 24(2):342–355
22. Laguna JO, Olaya AG, Borrajo D (2011) A dynamic sliding
3. Sattar S, Li S, Chapman M (2018) Road surface monitoring using
window approach for activity recognition. In: International con-
smartphone sensors: a review. Sensors 18(11):3845
ference on user modeling, adaptation, and personalization.
4. Hou Z, Wang KC, Gong W (2007) Experimentation of 3D
Springer, pp 219–230
pavement imaging through stereovision. In: International con-
23. Ni Q, Zhang L, Li L (2018) A heterogeneous ensemble approach
ference on transportation engineering 2007, pp 376–381
for activity recognition with integration of change point-based
5. Moazzam I, Kamal K, Mathavan S, Usman S, Rahman M (2013)
data segmentation. Appl Sci 8(9):1695
Metrology and visualization of potholes using the microsoft
24. Xiao G, Li K, Zhou X, Li K (2016) Queueing analysis of con-
kinect sensor. In: 16th International IEEE conference on intelli-
tinuous queries for uncertain data streams over sliding windows.
gent transportation systems (ITSC 2013). IEEE, pp 1284–1291
Int J Pattern Recognit Artif Intell 30(09):1660001
6. Koch C, Georgieva K, Kasireddy V, Akinci B, Fieguth P (2015)
25. Smrithy G, Balakrishnan R, Sivakumar N (2019) Anomaly
A review on computer vision based defect detection and condi-
detection using dynamic sliding window in wireless body area
tion assessment of concrete and asphalt civil infrastructure. Adv
networks. In: Data science and big data analytics. Lecture Notes
Eng Inform 29(2):196–210
on Data Engineering and Communications Technologies, vol 16.
7. Yan WY, Yuan X-X (2018) A low-cost video-based pavement
Springer, pp 99–108
distress screening system for low-volume roads. J Intell Transp
26. Xiao G, Li J, Chen Y, Li K (2020) Malfcs: an effective malware
Syst 22(5):376–389
classification framework with automated feature extraction based
8. Alfarraj O (2020) Internet of things with bio-inspired co-evolu-
on deep convolutional neural networks. J Parallel Distrib Comput
tionary deep-convolution neural-network approach for detecting
141:49–58
road cracks in smart transportation. Neural Comput Appl 32:1–16
27. Zheng Z, Zhou M, Chen Y, Huo M, Sun L, Zhao S, Chen D
9. Patra S, Middya AI, Roy S (2021) Potspot: participatory sensing
(2022) A fused method of machine learning and dynamic time
based monitoring system for pothole detection using deep
warping for road anomalies detection. IEEE Trans Intell Transp
learning. Multimed Tools Appl 80(16):25171–25195
Syst 23(2):827–839
10. Alam MY, Nandi A, Kumar A, Saha S, Saha M, Nandi S,
28. Sun L, Ge C, Huang X, Wu Y, Gao Y (2019) Differentially
Chakraborty S (2020) Crowdsourcing from the true crowd:
private real-time streaming data publication based on sliding
device, vehicle, road-surface and driving independent road

123
Neural Computing and Applications (2022) 34:19015–19033 19033

window under exponential decay. Comput Mater Continua 35. Mohan P, Padmanabhan VN, Ramjee R (2008) Nericell: rich
58(1):61–78 monitoring of road and traffic conditions using mobile smart-
29. Carlos MR, Aragón ME, González LC, Escalante HJ, Martı́nez F phones. In: Proceedings of the 6th ACM conference on embedded
(2018) Evaluation of detection approaches for road anomalies network sensor systems, pp 323–336
based on accelerometer readings-addressing who’s who. IEEE 36. Silva N, Soares J, Shah V, Santos MY, Rodrigues H (2017)
Trans Intell Transp Syst 19(10):3334–3343 Anomaly detection in roads with a data mining approach. Pro-
30. Fox A, Kumar BV, Chen J, Bai F (2015) Crowdsourcing cedia Comput Sci 121:415–422
undersampled vehicular sensor data for pothole detection. In: 37. Yadav Munshi, Alam M Afshar (2018) Dynamic time warping
2015 12th annual IEEE international conference on sensing, (DTW) algorithm in speech: a review. Int J Res Electron Comput
communication, and networking (SECON). IEEE, pp 515–523 Eng 6(1):524–528
31. Eriksson J, Girod L, Hull B, Newton R, Madden S, Balakrishnan 38. Noor MHM, Salcic Z, Kevin I, Wang K (2017) Adaptive sliding
H (2008) The pothole patrol: using a mobile sensor network for window segmentation for physical activity recognition using a
road surface monitoring. In: Proceedings of the 6th international single tri-axial accelerometer. Pervasive Mob Comput 38:41–59
conference on mobile systems, applications, and services, 39. Li Y, Cheng Y (2019) Streaming feature selection for multi-label
pp 29–39 data with dynamic sliding windows and feature repulsion loss.
32. Jain M, Singh AP, Bali S, Kaul S (2012) Speed-breaker early Entropy 21(12):1151
warning system. In: NSDR 40. Dalmazo BL, Vilela JP, Curado M (2014) Online traffic predic-
33. Mednis A, Strazdins G, Zviedris R, Kanonirs G, Selavo L (2011) tion in the cloud: a dynamic window approach. In: 2014 Inter-
Real time pothole detection using android smartphones with national conference on future internet of things and cloud. IEEE,
accelerometers. In: 2011 International conference on distributed pp 9–14
computing in sensor systems and workshops (DCOSS), pp 1–6. 41. Zhou X, Ishida T, Murakami Y (2015) Dynamic sliding window
IEEE model for service reputation. In: 2015 IEEE international con-
34. Mohamed A, Fouad MMM, Elhariri E, El-Bendary N, Zawbaa ference on services computing. IEEE, pp 25–32
HM, Tahoun M, Hassanien AE (2015) Roadmonitor: an intelli-
gent road surface condition monitoring system. In: Intelligent Publisher’s Note Springer Nature remains neutral with regard to
systems’ 2014. Springer, pp 377–387 jurisdictional claims in published maps and institutional affiliations.

123