Received 18 October 2021; revised 10 February 2022 and 4 April 2022; accepted 20 April 2022.

Date of publication 25 April 2022;

date of current version 4 May 2022.
Digital Object Identifier 10.1109/OJITS.2022.3169700

A Data-Driven Model for Pedestrian Behavior

Classification and Trajectory Prediction
VASILEIA PAPATHANASOPOULOU, IOANNA SPYROPOULOU, HARRIS PERAKIS,
VASSILIS GIKAS, AND ELENI ANDRIKOPOULOU
School of Rural, Surveying and Geoinformatics Engineering (SRSE), National Technical University of Athens, 15780 Zografou, Greece
CORRESPONDING AUTHOR: V. PAPATHANASOPOULOU (e-mail: [email protected])
This research is co-financed by Greece and the European Union (European Social Fund- ESF) through the Operational Programme Human Resources
Development, Education and Lifelong Learning 2014–2020 in the context of the project “Advanced WiFi-RTT Based Localization Techniques
for the Development and Testing of Pedestrian behavior Classification” (MIS 5049177).

ABSTRACT Pedestrian modeling remains a formidable challenge in transportation science due to the
complicated nature of pedestrian behavior and the irregular movement patterns. To this extent, accurate and
reliable positioning technologies and techniques play a significant role in the pedestrian simulation studies.
The objective of this research is to predict pedestrian movement in various perspectives utilizing historical
trajectory data. The study features considered in this research are pedestrian class, speed and position.
The ensemble of these features provides a thorough description of pedestrian movement prediction, whilst
contributes to the context of pedestrian modeling and Intelligent Transportation Systems. More specifically,
pedestrian movement is grouped into different classes considering gender, walking pace and distraction by
employing random forest algorithms. Then, position and speed prediction is computed employing suitable
data-driven methods, in particular, the locally weighted regression (LOESS method), taking into account
the individual pedestrian’s profile. An LSTM-based (Long Short-Term Memory) model is also applied for
comparison. The methodology is applied on pedestrian trajectory data that were collected in a controlled
experiment undertaken at the Campus of the National Technical University of Athens (NTUA), Greece.
Prediction of pedestrian’s movement is achieved, yielding satisfactory results.

INDEX TERMS Behavior classification, distraction, pedestrian speed prediction, pedestrian trajectory
prediction, random forests, GNSS, position fix.

I. INTRODUCTION key aspect of sustainable mobility and is essential for the

T HE ECONOMIC growth and technological advance-

ments of the last decades resulted in a significant
increase in transportation needs while paving the way to
implementation of C-ITS.
Pedestrian movement is quite complex, as it is sudden with
random or occasional changes [2]. At the same time colli-
the concept of future cities. This requires all transporta- sion avoidance systems should accommodate for pedestrian
tion modes to be integrated in connected and cooperative movement predictions at a millisecond level [3]. Besides, it
intelligent transport systems (C-ITS), necessitating the rep- is important to develop online solutions that would enable
resentation of all system actors’ interactions [1]. Although redirecting suitable information and guidance back to pedes-
walking is the most commonly used form of transport, trians when moving in public spaces and large facilities
until recently the majority of research efforts were focused (malls, airports, etc.), but also in cities within the con-
on vehicular transport, while analyses of pedestrian motion text of pedestrian- vehicle interactions [4]. The development
dynamics is limited. Pedestrian safety, though, comprises a of C-ITS necessitates for accurate, real-time detection of
pedestrian movement. The analysis of pedestrian behavior
The review of this article was arranged by Associate Editor Emmanouil is also crucial for security, marketing and urban planning
Chaniotakis. purposes as well as for infrastructure design [5]. Nowadays,
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/

328 VOLUME 3, 2022

the rapid advancements in localization technologies, coupled of the same system. The present research adopts a holistic
with the development of smart positioning algorithms, pave approach treating both these processes.
the way towards the provision of ubiquitous positioning data The structure of the paper is as follows. In the next sec-
of increased accuracy, robustness and availability [6]–[7], tion, the scientific background in the field of pedestrian
enabling the improvement of pedestrian trajectory prediction behavior classification and pedestrian movement prediction
tools. is described. Following this, the proposed methodological
In order to make the best of large volume of the pedes- framework and its partial components, including emerg-
trian data generated in the big data era, it is imperative ing data-driven methods such as random forests, locally
that efficient data management and analysis techniques are weighted regression (LOESS) method and Long Short-Term
followed. A classical approach includes the classification Memory (LSTM), are presented. The proposed methodology
of data into individual groups with common attributes. is demonstrated utilizing pedestrian trajectory data collected
Classification of pedestrian behavior enables a finer percep- in a field survey at the NTUA Campus, Greece. Distributions
tion of pedestrian motion, as it assists the differentiation and of the collected data are explored and a classification of
interpretation of the way pedestrians react to various distinct pedestrian behavior is performed. Pedestrian speed and posi-
situations. Moreover, the acquisition of different pedestrian tion are then predicted using LOESS method and considering
profiles facilitates the training of detailed pedestrian mod- each pedestrian’s heterogeneous profile. An LSTM based
els. Furthermore, classification approaches provide further model is also used for comparison with the proposed method.
insight on the contribution of different factors in shaping up The analysis of the data is presented followed by the discus-
pedestrian behavior, thus proving to be a significant tool in sion of the results. The conclusions, limitations and some
movement prediction. future prospects are discussed in the last section.
Pedestrian behavior refers to a multitude of motion
attributes (speed, acceleration, direction, etc.) and it is II. SCIENTIFIC BACKGROUND
affected by various factors [8]. Firstly, pedestrian behav- Several researchers have dealt with the identification and
ior is highly correlated with the physical and anatomical classification of pedestrian behavior under various con-
characteristics of each individual (e.g., height, age, fitness ditions, while different types of parameters have been
level, state of mind). Secondly, it depends on spatial and employed for different group definitions such as gender,
temporal variations that pertain to environment type (mall, walking style, and so on. In 2001, Lee and Mase [15]
airport, metro, etc.), the time of day (rush hour or not), performed pedestrian behavior classification based on accel-
as well as the terrain type (flat area, multi-level area, eration measurements for navigation purposes. They char-
etc.) that all together contribute significantly to movement acterized each step as ‘level’, ‘up’ and ‘down’ of a stair-
dynamics and their variation. Furthermore, pedestrian-to- way. Jan et al. [16] utilized Modified Probabilistic Neural
pedestrian, pedestrian-to-vehicle, as well as, pedestrian-to- Networks in order to classify pedestrian behavior into normal
infrastructure interactions affect pedestrian behavior. It is and abnormal and identify security threats. Chen et al. [17]
therefore apparent that pedestrian behavior is characterized classified pedestrian trajectory data in classes of different
by great heterogeneity resulting in the generation of vari- motion patterns to use them a- priori for vehicle-to-
ous pedestrian behavior states. Distraction sources, including pedestrian collision avoidance. Okamoto et al. [5] classified
mobile phone use, affect also pedestrian behavior result- pedestrian behavior in a shopping mall into three categories:
ing, for example, to lower walking speeds [9]–[10]. The ‘going straight’, ‘finding the way,’ and ‘walking around,’
distraction effects incurred by mobile phone usage have based on each pedestrian’s walking speed, trajectory vari-
received limited attention [11]–[13], though distracted walk- ability, stopping rate and the movement of the pedestrian’s
ing entails high-risk, particularly in pedestrian-to-vehicle head. Keller and Gavrila [18] and Völz et al. [3] estimated
interactions [11], [14]. For instance, phone distraction and the next step of the pedestrian as ‘crossing’ or ‘not crossing’,
a fast-walking pace could raise serious safety issues as they also utilizing clustering techniques. Quintero et al. [19] clas-
lead in lack of attention. Also, they could lead to sudden sified pedestrian behavior into four distinct actions: walking,
change in route choice – for example, choose the shortest starting, stopping and standing. Raza et al. [20] recog-
path while in a rush or the quieter route while talking on a nized pedestrian gender based on body appearance using
mobile phone. deep learning technique. More recently, Fayyaz et al. [21]
In this research, we develop a simple and integrated employed both conventional and deep convolutional neural
methodology to predict pedestrian movement. Initially, we networks for pedestrian gender classification and achieved an
group the available data into classes of different gender accuracy of 89.3% and 82% respectively in their datasets.
(male/female), walking pace (fast/normal) and phone use Gender recognition is a key issue in several applications
(if pedestrians make use of a phone or not while walking). such as visual surveillance, demographics, human–computer
Subsequently, we predict pedestrian position and speed con- interaction [21].
sidering individual heterogeneity. Several researchers have Pedestrian speed prediction, utilizing machine learning
study pedestrian classification and trajectory prediction sep- techniques, comprises a topic that has attracted attention
arately, whilst a limited number of studies treat them as part especially in recent years. Tordeux et al. [22] showed that

VOLUME 3, 2022 329

PAPATHANASOPOULOU et al.: DATA-DRIVEN MODEL

FIGURE 1. Methodology for pedestrian movement prediction.

neural networks are able to provide better estimation of prediction. The proposed methodology employs a classifi-
pedestrian speed at corridor and bottleneck experiments, cation and a regression algorithm in order to assign the
while outperforming the use of classical pedestrian models. obtained data into classes with similar characteristics, as
Kouskoulis et al. [23] also suggested superior performance well as to estimate pedestrian behavior state in real time
of a locally weighted regression algorithm compared to a and predict pedestrian speed and position in a short time
social force model for speed modeling using real data. span. In the proposed approach, the individual pedestrian
Pedestrian position prediction comprises yet another sig- profile is formed utilizing observations from previous time
nificant element in future cities, as accurate predictions instants. New observations of the following time instants are
could prevent a considerable number of traffic injuries then classified into the appropriate classes enabling trajectory
and improve pedestrian safety when employed in collision prediction based on this personalized pedestrian profile. The
warning systems or embedded in autonomous vehicle path proposed methodological framework is presented in Fig. 1.
planning [24]. Pedestrian trajectory is usually characterized
as series of generation actions [25]. Several researchers have B. PEDESTRIAN BEHAVIOR CLASSIFICATION
used data-driven methods for pedestrian position prediction, For the needs of this study, a database with pedestrian
including novel LSTM algorithms, Recurrent Neural observations including position, speed, acceleration, etc. as
Networks and Gaussian processes [24]–[31]. Pedestrian well as individual characteristics such as age, gender and
movement involves classes of behavioral status, speed and height, is utilized. The associations between these parame-
traveled distance values, as well as movement direction. Most ters are identified for every pedestrian. Initially, mean values
existing pedestrian movement prediction models tend to deal for the attributes of each pedestrian are calculated. The
with just one of these aspects. In this paper, we propose an classification step that follows divides the available obser-
integrated methodological framework to consider them all. vations into different pedestrian classes (i.e., gender, pace
and distraction conditions). For the implementation of data
III. METHODOLOGY classification, the random forests classification algorithm is
A. METHODOLOGICAL FRAMEWORK applied to the relevant data. Speeds, accelerations and devi-
Capturing pedestrian behavior heterogeneity has the poten- ations from the mean pedestrian speed are considered in this
tial to offer accurate input in the development of Intelligent process. The final result is a detailed historical database with
Transportation Systems. This research is based on [32] in classified data, while new data are assigned to the class with
order to expand the existing methodology considering also the highest resemblance.
pedestrian trajectory prediction (i.e., position and speed fea- The efficiency of classification methods is usually assessed
tures) further to pedestrian behavior classification. Thus, via the True Positive and False Negative rates [33].
the present research develops an integrated methodology for Evaluation of the classification algorithm performance is
pedestrian behavior classification and pedestrian movement achieved using a confusion matrix for the test data. Values

330 VOLUME 3, 2022

on the diagonal correspond to true values (positives or prediction and the last one for direction angle prediction.
negatives), whereas the rest correspond to false values. The latter two are utilized to estimate the position of the
pedestrian. The model is trained using all the available obser-
C. PEDESTRIAN TRAJECTORY PREDICTION vations up to the time instant ti and it is then tested utilizing
Trajectory is defined as the time-profile of pedestrian motion the following time instants.
states, such as his/ her position and velocity [34]. In order to In general, the training period can vary per case. Two
develop a pedestrian data-driven model, the required explana- approaches can be adopted to determine the training period.
tory variables should be determined first, and the appropriate The first one is simpler and is usually applied in smaller
trajectory data should be collected. Data driven modeling datasets, while the second one is more customized and can
requires reliable data of high precision. Positioning accuracy, be applied when large datasets are available. In the first
availability and integrity are often considered as the most approach, the training period is defined as a fixed percentage
critical parameters in the quality of a position fix followed of the available dataset size, for example 50% of the obser-
by the update rate, position coverage, continuity and system vations are used for training and the rest for testing [35].
latency [4]. For example, in the case of Global Navigation In the second approach, the optimal training period is deter-
Satellite System (GNSS) positioning, the observation condi- mined by the desired prediction accuracy [36]. When the
tions of a receiver, such as signal blockage caused due to algorithm returns a low prediction error (lower than a set
high buildings nearby or steep gradients, signal attenuation threshold values), the training process is accomplished at a
resulting from foliage and multipath / interference due to satisfactory level. The training process can be repeated for
moving elements (vehicles and/or pedestrians), can result in better performance at suitable time periods, such as when the
poor positioning or even total inability to produce a naviga- pedestrian stops moving. In our case study, the first approach
tion solution [34]–[35]. Therefore, appropriate tools should is chosen due to limitations in the available data.
be utilized and appropriate measures need to be taken into Model training can be achieved with the employment of
account when collecting GNSS positioning data in order to machine learning techniques. In our research we apply the
minimize environmental effects. locally weighted regression algorithm (LOESS) and Long
In the adopted approach, model training is applied sep- Short-Term Memory (LSTM). The first method was selected
arately for each pedestrian in order to capture individual as it has been utilized for vehicle trajectory prediction
heterogeneity and develop more detailed and representative and the results were promising [37]. Thus, in our research
models. Since the algorithm is trained for each individual the validity of the method is assessed on pedestrian tra-
pedestrian/road user, it is trained in real-time conditions for jectory prediction. Furthermore, LOESS has been used
the behavior of that individual pedestrian and in the prevail- for our analysis, as it combines the simplicity of linear
ing environmental conditions. Thus, it becomes custom-made least squares regression with the flexibility of nonlinear
for the particular context of the training dataset that is fed regression. LSTM is utilized, as a reference method, as it
to the model. constitutes a fast-evolving field and is widely applied in
Such an approach is useful within the context of ITS trajectory prediction [38].
applications, the operation of which requires a personalized A number of different explanatory variables can be
profile of each user/pedestrian. The application will collect selected as input in the prediction algorithm. For instance, the
the required trajectory data of the user during a training social force model, a well-known classical pedestrian model,
period. The trained algorithm will in turn propose a more expresses the directed pedestrian’s acceleration as a function
accurate prediction of the individual’s path for the following of the deviation of the pedestrian’s velocity from a desired
time instants. Overfitting can be avoided by using data from velocity (which varies between individuals) and the distance
an extensive training period and by setting the appropriate from obstacles and other pedestrians [39]–[40]. In this study,
hyperparameters of the algorithm. As the training process we focus on the movement of each individual. In order to
requires a small amount of time, the training of the algo- predict the speed vi_t+τ and distance dsi_t+τ of pedestrian i
rithm can be iterated utilizing the newly collected data when at the time instant t + τ , where τ is the prediction time span,
the pedestrian stops moving. Naturally, the exact amount of we utilize speed vit and acceleration ait of the previous time
time depends on hardware constrains and on the size of the instant t as explanatory variables (Eq. (1) and (2)). In order
train data. Indicatively, for a dataset of a duration of about to predict the angle θi_t+τ of pedestrian i at the time instant
one minute processed on an average PC (RAM 16GB), the t + τ , the angle θit and the transition on x and y axis of the
required training time is about 1 sec. The trained algorithm previous time instant, dXit and dYit respectively, are used
can give an immediate real-time response of the pedes- as explanatory variables (Eq. (3)). As function f, various
trian’s movement prediction. This response will be useful data-driven methods can be used. In our research, LOESS
in pedestrian-to-vehicle (ex. danger alert for possible col- method is proposed and LSTM is used for comparison.
lision) and pedestrian-to-infrastructure communication (ex.
vi_t+τ ∼ f (vit + ait ) (1)
synchronization of traffic lights).
In this research, three data-driven sub-models are dsi_t+τ ∼ f (vit + ait ) (2)
developed, one for speed prediction, one for distance θi_t+τ ∼ f (θit + dXit + dYit ) (3)

VOLUME 3, 2022 331

PAPATHANASOPOULOU et al.: DATA-DRIVEN MODEL

TABLE 1. Scenarios of the conducted experiment.

The performance and efficiency of the algorithms
are evaluated using the Normalized Root Mean Square
Error (RMSN) metric. RMSN assesses the overall error and
performance of the proposed method on speed prediction
estimating the normalized difference between the observed
(Yobs ) and simulated values (Ysim ) (Eq. (4)) [41].

  obs 
sim 2
N· N n=1 Yn − Yn
RMSN = N (4)
obs values of future time-steps according to the observed values
n=1 Yn
The Final Displacement Error (FDE) metric is also used of past time-steps by minimizing a loss function. An LSTM
in order to estimate the difference between the predicted network contains one or more LSTM layers that take as
coordinates (Xsim , Ysim ) and true final points (Xobs , Yobs ) input a time series and produce as output another time series
(pedestrian’s positions). for future prediction. More specifically, LSTM architecture
 2  2 consists of three parts, known as gates: the input gate, the
N
n=1 Xnobs − Xnsim + Ynobs − Ynsim forget gate and the output gate [49]–[50]. Each LSTM has
FDE = . (5) a cell state through which the information is carried to the
N
gates. In the input gate, new information is stored in the cell.
D. METHODOLOGICAL COMPONENTS In the forget gate, the information is filtered and unnecessary
For the classification step of the methodology, random for- information is dismissed. Finally, in the output gate, the final
est algorithms [42] are used. They adhere to a machine output, namely the prediction for the next time instant is
learning method which allows data classification consider- produced. The whole process is run for a number of epochs
ing both numerical and categorical variables [43]. Also, they (also known as iterations) in order to ensure more accurate
have been used for classification of pedestrian trajectory prediction. The equations of the gates in LSTM are the
data [44]–[45]. Decision trees are built during the training following:
process. Associations and interactions among the input data    
are identified. In each tree, the optimal prediction is selected it+1 = σ wi · ht , xt+1 + bi (6)
   
among a subset of random predictions on each node [46]. ft+1 = σ wf · ht , xt+1 + bf (7)
This randomization reduces the correlation between the trees    
ot+1 = σ wo · ht , xt+1 + bo (8)
and contributes to overfitting avoidance.
For speed, distance and direction angle prediction, data- where, it+1 is the input gate, ft+1 is the forget gate, ot+1
driven models are developed using the LOESS method and is the output gate, σ a sigmoid function, ht the state of the
LSTM. LOESS was firstly introduced by Cleveland [47] and current time step and xt+1 the input at the next timestamp,
the following analysis is based on [48]. Locally weighted wi , wf and wo the weights of the respective gates, bi , bf and
regression yi = g(xi ) + εi , where i = 1, . . . , n is the index bo bias of the respective gates.
of observations, g is the regression function and εi are resid- The equations for the final output are the following:
ual errors, provides an estimate g(x) of each regression    
surface at any value x in the d-dimensional space of the c̃t+1 = tanh wc · ht , xt+1 + bc (9)
independent variables. The LOESS method identifies corre- ct+1 = ft+1 · ct + it+1 · c̃t+1 (10)
lations between the observed response variable yi and the ht+1 = ot+1 · tanh(ct+1 ) (11)
explanatory variables xi . In particular, it estimates a function
g(x) at the point x = x0 considering the parameter values where, c denotes the cell state, c̃t+1 is the candidate for the
in a parametric category. A regression surface is adapted cell state at timestamp t + 1, ct+1 is the cell state memory
to the data in the neighborhood of x0 . Hyperparameters of at timestamp t + 1 and ht+1 represents the hidden state.
the LOESS method include the ‘span’ and the ‘degree’. The
span is a smoothing parameter of the surface which deter- IV. CASE STUDY
mines the percentage of the data that are taken into account A. EXPERIMENTAL DATA
for each local fit. Each local regression uses either a first or To develop and evaluate our approach, a field experiment
a second-degree polynomial that is specified by the value of took place at the NTUA Campus, Greece [12]. More specif-
the “degree” parameter of the method. The data are weighted ically, during a controlled experiment participants were
according to their distance from the center of neighborhood instructed to walk along a 220 m path under six scenar-
x, therefore a distance and a weight function are required. ios (Table 1). These involved two pace conditions: normal
In the application of the LOESS method, the Euclidean dis- and fast, and three distraction conditions: no distraction (base
tance has been utilized as the distance function p to weight scenario), talking on the phone (distracted scenario) and tex-
the data. ting (distracted scenario) while walking. This resulted in a
LSTM constitutes a Recurrent Neural Network (RNN) total of six walks for each participant, the order of which
which is widely used in regression analysis, e.g., estimating differed between the participants to avoid order effects.

332 VOLUME 3, 2022

 FIGURE 3. Density plot for fast and normal walking.

bag and the GNSS patch antenna was placed on the pedes-
trian’s cap (Fig. 2(c)). The Rover Station raw GNSS data
logging was handled by a smartphone device connected to
the receiver via USB and running the RTKGPS+ application.
Notwithstanding the selected test site is generally free of
obstacles, partial satellite obstruction and signal multipath
effect due to nearby buildings - mainly due to the high-
rise building located east of the area - pose a challenge
to GNSS positioning. Preliminary evaluation highlighted the
limitations of a single-point positioning (SPP) solution pro-
viding average accuracy at the order of ∼3 m while the
differential post-processing kinematic (PPK) solution pro-
vided accuracy levels of a few decimeters (Fig. 2(b)) leading
to the selection of this data processing approach. The extrac-
tion of pedestrians’ trajectories was implemented using the
RTKLIB software resulting in an average position trueness
of 0.2 m. The sampling rate of the data collection was set
at 5 Hz. During analysis observations were down-sampled
at 1 sec. Pedestrians’ coordinates were computed during
FIGURE 2. Experimental set-up. (a) Route followed by participants, (b) SPP (red) the experiment including computation of pedestrian speed
and PPK (blue) solutions trajectory comparison, (c) GNSS equipment employed for and acceleration. From this point onwards the data analysis
pedestrians.
and classification, as well as the development of data-
driven algorithms, have been performed in R programming
The walking path of the experiment is indicated with the language [51].
red line in Fig. 2a. In total, trajectories of 36 pedestrians
were recorded utilizing low cost GNSS technology. As this B. DATA PRESENTATION
is ongoing research, about 15.000 observations (17 pedes- This section presents the available data in order to draw
trians) have been processed and studied for the six walking a few preliminary conclusions. Regarding the walking pace
scenarios: base, talking on the phone and texting, all of them two trends can be identified in Fig. 3. The curves that exhibit
both at normal and fast pace. a sharp peak correspond to “fast” observations while those
The low cost GNSS receivers u-blox EVK-M8 /NEO- shown relatively flattened curves correspond to “normal”
M8T, u-blox C94-M8P, the high-quality geodetic type observations. The first set of curves associates with high
Pinwheel 702-GG satellite antenna of NovAtel, as well as speed values and deviations from pedestrian’s mean speed,
the RTKLIB software and the RTKGPS+ mobile applica- while the second one covers a wider area with lower speed
tion were used for the collection of the pedestrian data. values and deviations from pedestrian’s mean speed.
The Base Station GNSS data were collected utilizing the Nowadays, the use of smartphones has been inherent part
u-blox C94-M8P receiver connected with the geodetic grade of our life, and thus, constitutes a new distraction on pedes-
satellite antenna (NovAtel Pinwheel 702-GG) placed on a trians’ movement that takes various forms (e.g., talking or
geodetic pilar of known coordinates. At the same time, it texting). As shown in Fig. 4, a consequence of phone related
was ensured that the pedestrian localization sensor place- distraction while walking is slowing down the speed. From
ment was as non-intrusive as possible to minimize potential the same plot it is apparent that pedestrian speed decreases
effects on pedestrian motion. For this purpose, a u-blox more rapidly when texting. Table 2 summarizes the average
EVK-M8/NEO-M8T GNSS receiver was placed in a small speed observed for different pace scenarios.

VOLUME 3, 2022 333

PAPATHANASOPOULOU et al.: DATA-DRIVEN MODEL

FIGURE 4. Speed density plot for different types of phone distraction.

(a)

(b)
FIGURE 5. Speed density plot for pedestrian groups with different age and gender
(base scenario and texting).
FIGURE 6. (a) Indicative routes (X and Y coordinates) under the three different
scenarios: base, talking on the phone, texting. (b) Indicative routes (X and Y
TABLE 2. Mean speed for all the experimental scenarios. coordinates) under the three different scenarios: base, talking on the phone, texting.

TABLE 3. Gender recognition- confusion matrix.

Speed densities are plotted for different pedestrian groups

for different conditions (base scenario and texting scenario)
in Fig. 5. Results indicate that speed reduction is more
intense for women. available data were used as train data and the remaining
Also, analysis proved that in addition to the walking speed, as test data. At first, we develop a random forest for gen-
distraction can affect pedestrian route choice. Few pedestri- der classification using the train data and the explanatory
ans may not turn on time due to lack of attention resulting by variables per time instant. In this case the objective of the
phone distraction. Fig. 6 illustrates this effect for a number classification is to classify all observations of all datasets to
of indicative routes and different distraction states. gender classes, based on the movement characteristics. The
number of decision trees is selected to be 500 as defined
C. PEDESTRIAN BEHAVIOR CLASSIFICATION by the algorithm which ensures the optimal classification of
Following the preliminary analysis we classify the available the available data. Then, the trained algorithm is applied on
data according to gender, walking pace and distraction state. the test data leading up to 97% gender recognition accuracy.
For all the classifications, the following variables have been The performance of the algorithm is evaluated considering
considered for every pedestrian: walking speed and accel- the confusion matrix (Table 3).
eration, deviation from the mean speed, traveled distance At a second stage, a random forest is developed using
(Euclidean distance) as well as pedestrian age and height. the train data in order to identify relationships between the
The height variable has been omitted for gender classifi- explanatory variables and the predictor variable “walking
cation as this variable was highly correlated. Half of the pace” at a time instant.

334 VOLUME 3, 2022

TABLE 4. Walking pace recognition- confusion matrix.

FIGURE 8. RMSN (%) of speed prediction for the pedestrians under three different
scenarios using LOESS method.

For the aforementioned classifications, all the variables are

significant, as p-values of less than 2.2e-16 were estimated.

D. SPEED PREDICTION
In this case study, pedestrian speed is computed every
time instant (time step of 1 sec) considering the speed
and the acceleration values obtained at the previous time
FIGURE 7. Variable relationships for “fast” and “normal” walking pace. instant (vt+1 ∼ vt + at ). A LOESS algorithm is trained
separately for each pedestrian, so that personalized pro-
TABLE 5. Phone use recognition- confusion matrix. files are developed. For every pedestrian, the first half of
the dataset is used for training purposes and the remain-
ing set of data for testing. The optimal hyperparameters
have been chosen using the ‘train’ function, which estimates
algorithm performance through resampling [52]. The train-
ing step involves the selection of the optimal span (between
0.1 and 0.9) and degree (1 or 2). ’Span’ at a value of 0.75
and ‘degree’ at a value of 1 were chosen as the optimal
values of hyperparameters by this function.
Fig. 8 presents the results obtained using RMSN as a
Then, the trained algorithm is applied on the test data quality metric. The results are promising, as the algorithm
leading to 81% (Table 4) accuracy for pedestrian’s walking provides an error less than 1% of the complete dataset. The
pace prediction, either normal or fast. less accurate speed prediction seems to correspond to walk-
Fig. 7 plots two classes of pedestrians, those walking at ing while texting situations. This can be attributed to the
fast pace and those at normal pace. Evidently, the higher the fact that texting causes the highest reduction on pedestrian
walking speed the higher the acceleration and the deviation speed compared to the other scenarios, namely the base and
from pedestrian’s mean speed. In the same plot, the class the walking while talking on the phone scenario.
with high values of these features is depicted with grey
color and corresponds to the fast motion. When walking E. POSITION PREDICTION
at a fast pace, pedestrians tend to move at higher speeds The position prediction includes the computation of the dis-
and perform more abrupt maneuvers than they usually do, tance (ds) between sequential position fixes and the direction
resulting in higher values of accelerations and deviations angle (θ ). The distance dst+1 at the next time instant is esti-
from their mean speed. mated considering the speed vt and the acceleration at of the
Finally, the development of a random forest aiming at previous time instant (dst+1 ∼ vt + at ). The direction angle
investigating the effects of phone use (texting or talking) was θt+1 at the next time instant is estimated considering the
attempted. The trained random forest recognized the type of pedestrian’s transitions dXt and dYt and the direction angle
phone use with 73 % accuracy. The phone use recognition θt of the previous time instant (θt+1 ∼ θt + dXt + dYt ).
confusion matrix is presented in the Table 5. As this is an In order to estimate the predicted pedestrian coordinates
ongoing research, the last of the three random forests should at the next time instant based on the predicted distance and
be further improved, by exploring alternative explanatory direction angle (Fig. 9), the following formulas are used:
variables through increasing the number of previous time
instants and using a larger number of data series in general. Xt+1 = Xt + dst+1 ∗ cos θt+1 (12)
Deep learning could also be tested in this context. ϒt+1 = ϒt + dst+1 ∗ sin θt+1 (13)

VOLUME 3, 2022 335

PAPATHANASOPOULOU et al.: DATA-DRIVEN MODEL

FIGURE 9. Estimation of predicted coordinates.

TABLE 6. Final displacement error per pedestrian, per scenario and method.

FIGURE 10. Position prediction with high accuracy for pedestrian No. 34 and the
scenario walking while texting.

to situations such as walking while talking on the phone.

This is accounted to distractions resulting in bypasses on
the pedestrian’s route as opposed to other (base and texting)
scenarios – see (Fig. 6a, b).
Alahi et al. [26] have managed to achieve a pedes-
trian trajectory forecasting for a fixed period of 4.8 sec
with an average Final Displacement Error at the order of
0.60 m using the Social Force Model and 0.61 m using
their proposed Social- LSTM model. Li et al. [55] obtained
final point errors 0.12-0.28 m in normal crossing scenar-
ios at 1 sec ahead using data-driven methods including
Gaussian Process (GP), LSTM, GP-LSTM, Character-based
LSTM, Sequence-to-Sequence (Seq2Seq) and attention-
based Seq2Seq. In addition, Rasouli et al. [30] propose an
algorithm based on an RNN encoder-decoder architecture for
trajectory prediction. Their model outperforms state-of-the-
art by 26% on their dataset. More specifically, they achieved
a prediction error of 1.44 km/h (equals to 0.40 m/sec) for
a time span of 1 sec. They also adopted an LSTM-based
model that produced a prediction error of 1.91 km/h (equals
to 0.53 m/sec) for a time span of 1 sec.
In this study, considering the findings of previous research,
we exercise the LOESS method leading to an average Final
As LSTM approach is widely used in trajectory prediction, Displacement Error at the order of 0.26 m at a time span
it is also utilized in this study for prediction performance 1 sec, and an LSTM model resulting to an average Final
comparison (Table 6). We built an LSTM Model in R pro- Displacement Error at the order of 0.37 m for the tex-
gramming language using TensorFlow 2.7.0 [53] and Keras ting scenario. Generally, our proposed model offers good
2.7.0 [54]. The hyperparameters were tuned based on the performance in comparison to the existing baselines.
validation error. An LSTM layer is followed by two dense Indicatively, Fig. 10 and Fig. 11 show an accurate and
layers whilst the Rectified Linear Units (ReLU) are used as a less accurate prediction of a pedestrian’s route. Even for
an activation function. the less accurate prediction, the route is very similar to the
The loss function to minimize is the Mean Squared Error. observed one. This is confirmed in Fig. 12 and Fig. 13 where
We train the model with a batch size of 60 for a time span the observed versus predicted coordinates are plotted for all
of 20 epochs. The learning rate is 0.001 and the validation pedestrians and scenarios.
split is 0.5.
Table 6 contains the results of the aforementioned anal- V. CONCLUSION
yses. In our case study, both algorithms offer good results. In this research, a methodology for pedestrians’ speed and
However, the LOESS method seems to produce slightly bet- trajectory prediction, adopting pedestrian classification as
ter results. The results of less accurate prediction correspond an intermediate step, has been proposed. The proposed

336 VOLUME 3, 2022

 available data. Gender recognition, walking pace and mobile
phone usage recognition have also been predicted satis-
factorily, though further investigation is necessary. Speed
and position prediction has been achieved successfully for
the majority of the data. The identification of personalized
profiles can contribute to more robust pedestrian mod-
els for various regimes. The proposed methodology can
predict pedestrian characteristics (gender) and characteris-
tics of pedestrian movement (walking pace, distraction) as
well as pedestrian trajectory. This type of information may be
extremely useful in the context of Intelligent Transportation
Systems and can be used as input for smart applications
which are now being developed for pedestrian and vehicle
FIGURE 11. Position prediction with lower accuracy for pedestrian No. 33 and the movement prediction such as smart traffic lights, application
scenario walking while talking on the phone. for danger proximity, interactions with autonomous vehicles,
etc.
This research contributes to pedestrian modeling as it
offers an integrated methodology for classification and tra-
jectory prediction, utilizing data driven methods. However,
there are some limitations. At first, the proposed methodol-
ogy is validated on experimental data considering a limited
number of participants, and therefore, limited variability.
In addition, the data were produced through an experiment
undertaken under controlled conditions leading to discrepan-
cies in pedestrian behavior between real life and experimental
conditions. Nevertheless, despite the inherent limitations in
data collection, the experimental data set offers a great oppor-
tunity to apply and test the proposed methodology, as well
as a firm basis for further development. As far as the limi-
tations in the application of the methodology are concerned,
FIGURE 12. X observed coordinates versus X predicted coordinates of the
pedestrians. the methodology was applied for the prediction of pedestrian
movement in a constrained time span. Any future efforts
should extend the time span. Finally, the study does not
deal with interactions between pedestrians, but focuses on
the movement of one individual.
In order to cope with these limitations, further testing
is necessary to study the discrepancies between realistic
and controlled conditions either through incorporating pedes-
trian interactions at high density scenarios or scenarios that
would adopt vehicle-pedestrian interactions. In this study
we estimated the speed and the trajectory at the subsequent
time instances using the values of the previous time instant.
A multi-step ahead prediction can also be applied utiliz-
ing the estimated values for the prediction of additional
time instant(s), aiming at a longer prediction time hori-
zon, a methodology that has already been applied in other
studies (e.g., for vehicles in [56]). Furthermore, in order to
FIGURE 13. Y observed coordinates versus Y predicted coordinates of the
pedestrians. demonstrate the effectiveness of the proposed methodology,
it should be tested on multiple datasets.
As data-driven approaches are becoming increasingly pop-
methodology utilizes random forests and the LOESS method, ular within the ITS domain, the utilization of novel and more
while an LSTM algorithm is used to provide further efficient positioning systems is making it possible to adopt
evidence. more detailed models for such applications [1]. Finally,
The results indicate the potential of the proposed method- provided that microscopic modeling reliability depends on
ology and a slightly better performance of LOESS technique the positioning quality and by extension on the effective-
in comparison to LSTM for trajectory prediction on the ness of the underlying pedestrian detection technologies, the

VOLUME 3, 2022 337

PAPATHANASOPOULOU et al.: DATA-DRIVEN MODEL

collection of high-quality datasets in a minimally-intrusive [19] R. Quintero, I. Parra, J. Lorenzo, D. Fernández-Llorca, and
manner still remains a widely open issue. M. A. Sotelo, “Pedestrian intention recognition by means of
a hidden Markov model and body language,” in Proc. IEEE
20th Int. Conf. Intell. Transp. Syst. (ITSC), 2017, pp. 1–7,
doi: 10.1109/ITSC.2017.8317766.
REFERENCES [20] M. Raza, M. Sharif, M. Yasmin, M. A. Khan, T. Saba, and
[1] V. Papathanasopoulou, C. Antoniou, and H. N. Koutsopoulos, “Data- S. L. Fernandes, “Appearance based pedestrians’ gender recognition
driven traffic simulation models: Mobility patterns using machine by employing stacked auto encoders in deep learning,” Future Gener.
learning techniques,” in Mobility Patterns, Big Data and Transport Comput. Syst., vol. 88, pp. 28–39, Nov. 2018.
Analytics: Tools and Applications for Modeling, C. Antoniou, [21] M. Fayyaz, M. Yasmin, M. Sharif, and M. Raza, “J-LDFR: Joint
L. Dimitriou, and P. F. Paperback, Eds. Amsterdam, The Netherlands: low-level and deep neural network feature representations for pedes-
Elsevier, 2019, pp. 263–295. trian gender classification,” Neural Comput. Appl., vol. 33, no. 1,
pp. 361–391, 2021.
[2] J. Guo, “Accuracy improvement of pedestrian trajectory prediction
by an extended Kalman filter and pedestrian behavior classification,” [22] A. Tordeux, M. Chraibi, A. Seyfried, and A. Schadschneider,
M.S. thesis, Dept. Electr. Comput. Eng., Rose-Hulman Institute of “Prediction of pedestrian speed with artificial neural networks,” in
Technology, 2018. Proc. Int. Conf. Traffic Granular Flow, Jul. 2017, pp. 327–335.
[23] G. Kouskoulis, I. Spyropoulou, and C. Antoniou, “Pedestrian simu-
[3] B. Völz, H. Mielenz, G. Agamennoni, and R. Siegwart, “Feature
lation: Theoretical models vs. data driven techniques,” Int. J. Transp.
relevance estimation for learning pedestrian behavior at crosswalks,”
Sci. Technol., vol. 7, no. 4, pp. 241–253, 2018.
in Proc. IEEE 18th Int. Conf. Intell. Transp. Syst., 2015, pp. 854–860,
[24] R. Quan, L. Zhu, Y. Wu, and Y. Yang, “Holistic LSTM for pedes-
doi: 10.1109/ITSC.2015.144.
trian trajectory prediction,” IEEE Trans. Image Process., vol. 30,
[4] V. Gikas, G. Retscher, and A. Kealy, “Collaborative positioning for
pp. 3229–3239, 2021, doi: 10.1109/TIP.2021.3058599.
urban intelligent transportation systems (ITS) and personal mobil-
[25] X. Li, Y. Liu, K. Wang, and F.-Y. Wang, “A recurrent atten-
ity (M): Challenges and perspectives,” in Mobility Patterns, Big
tion and interaction model for pedestrian trajectory prediction,”
Data and Transport Analytics: Tools and Applications for Modeling,
IEEE/CAA J. Autom. Sinica, vol. 7, no. 5, pp. 1361–1370, Sep. 2020,
C. Antoniou, L. Dimitriou, and Pereira F. Paperback, Eds. Amsterdam,
doi: 10.1109/JAS.2020.1003300.
The Netherlands: Elsevier, 2019, pp. 381–414.
[26] A. Alahi, K. Goel, V. Ramanathan, A. Robicquet, L. Fei-Fei, and
[5] K. Okamoto et al., “Classification of pedestrian behavior in a shop-
S. Savarese, “Social LSTM: Human trajectory prediction in crowded
ping mall based on LRF and camera observations,” in Proc. MVA,
spaces,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR),
Jun. 2011, pp. 1–5.
2016, pp. 961–971, doi: 10.1109/CVPR.2016.110.
[6] A. Masiero et al., “A case study of pedestrian positioning with UWB [27] J. Bock, T. Beemelmanns, M. Klösges, and J. Kotte, “Self-learning
and UAV cameras,” Int. Archiv. Photogrammetry Remote Sens. Spat. trajectory prediction with recurrent neural networks at intelligent
Inf. Sci., vol. XLIII-B1-2021, 2021, pp. 111–116. intersections,” in Proc. VEHITS, Apr. 2017, pp. 346–351.
[7] G. Retscher, V. Gikas, H. Hofer, H. Perakis, and A. Kealy, “Range [28] A. Vemula, K. Muelling, and J. Oh, “Modeling cooperative navigation
validation of UWB and Wi-Fi for integrated indoor positioning,” Appl. in dense human crowds,” in Proc. ICRA, 2017, pp. 1685–1692.
Geomatics vol. 11, no. 2, pp. 187–195, 2019. [29] Y. Xu, Z. Piao, and S. Gao, “Encoding crowd interaction with
[8] G. Kouskoulis and C. Antoniou, “Systematic review of pedestrian deep neural network for pedestrian trajectory prediction,” in Proc.
simulation models with a focus on emergency situations,” Transp. IEEE Conf. Comput. Vis. Pattern Recognit., 2018, pp. 5275–5284,
Res. Rec., vol. 2604, no. 1, pp. 111–119, 2017. doi: 10.1109/CVPR.2018.00553.
[9] New York Pedestrian Study, New York City Pedestrian Level of Service, [30] A. Rasouli, I. Kotseruba, T. Kunic, and J. K. Tsotsos, “PIE: A
Phase I, Department of City Planning, Transp. Division, New York, large-scale dataset and models for pedestrian intention estimation and
NY, USA, 2006. trajectory prediction,” in Proc. IEEE/CVF Int. Conf. Comput. Vis.,
[10] J. Hatfield and S. Murphy, “The effects of mobile phone use on pedes- 2019, pp. 6261–6270, doi: 10.1109/ICCV.2019.00636.
trian crossing behaviour at signalised and unsignalised intersections,” [31] X. Song et al., “Pedestrian trajectory prediction based on deep con-
Accid. Anal. Prevent., vol. 39, no. 1, pp. 197–205, 2007. volutional LSTM network,” IEEE Trans. Intell. Transp. Syst., vol. 22,
[11] J. Nasar, P. Hecht, and R. Wener, “Mobile telephones, distracted atten- no. 6, pp. 3285–3302, Jun. 2021, doi: 10.1109/TITS.2020.2981118.
tion, and pedestrian safety,” Accid. Anal. Prevent., vol. 40, no. 1, [32] V. Papathanasopoulou, I. Spyropoulou, H. Perakis, V. Gikas,
pp. 69–75, 2008. and E. Andrikopoulou, “Classification of pedestrian behavior
[12] E. Andrikopoulou, I. Spyropoulou, H. Perakis, and V. Gikas, using real trajectory data,” in Proc. 7th Int. Conf. Models
“Extracting pedestrian behavior using low cost GNSS receiver data,” Technol. Intell. Transp. Syst. (MT-ITS), Jun. 2021, pp. 1–6,
in Proc. 8th Transport Res. Arena TRA, Helsinki, Finland, Apr. 2020, doi: 10.1109/MT-ITS49943.2021.9529266.
pp. 1–6. [33] D. Ridel, E. Rehder, M. Lauer, C. Stiller, and D. Wolf, “A literature
[13] R. Alsaleh, T. Sayed, and M. H. Zaki, “Assessing the effect of pedes- review on the prediction of pedestrian behavior in urban scenarios,” in
trians’ use of cell phones on their walking behavior: A study based on Proc. 21st Int. Conf. Intell. Transp. Syst. (ITSC), 2018, pp. 3105–3112,
automated video analysis,” Transp. Res. Rec. J. Transp. Res. Board, doi: 10.1109/ITSC.2018.8569415.
vol. 2672, no. 35, pp. 46–57, 2018. [34] P. Kothari, S. Kreiss, and A. Alahi, “Human trajectory forecasting
[14] P. D. Loeb and W. A. Clarke, “The cell phone effect on pedes- in crowds: A deep learning perspective,” IEEE Trans. Intell. Transp.
trian fatalities,” Transp. Res. E, Logist. Transp. Rev., vol. 45, no. 1, Syst., early access, Apr. 19, 2021, doi: 10.1109/TITS.2021.3069362.
pp. 284–290, 2009. [35] T. Lin, Z. Jin, S. Choi, and H. Yeo, “A framework for pedestrian
[15] S.-W. Lee and K. Mase, “Recognition of walking behaviors for pedes- sub-classification and arrival time prediction at signalized intersection
trian navigation,” in Proc. IEEE Int. Conf. Control Appl. (CCA), using preprocessed Lidar data,” 2022, arXiv:2201.05877.
Sep. 2001, pp. 1152–1155, doi: 10.1109/CCA.2001.974027. [36] C. Amisse, M. E. Jijón-Palma, and J. A. S. Centeno, “Fine-tuning
[16] T. Jan, M. Piccardi, and T. B. Hintz, “Detection of suspicious pedes- deep learning models for pedestrian detection,” Boletim de Ciências
trian behaviour using modified probabilistic neural networks,” in Geodésicas, vol. 27, no. 3, 2021, Art. no. e2021017. [Online].
Proc. Image Vis. Comput. New Zealand Image Vis. Comput. Conf. Available: https://doi.org/10.1590/s1982-21702021000300017
Image Vision Comput. New Zealand, Auckland, New Zealand, 2002, [37] V. Papathanasopoulou and C. Antoniou, “Towards data-driven car-
pp. 237–241. following models,” Transp. Res. C, Emerg. Technol., vol. 55,
[17] Z. Chen, D. C. Ngai, and N. H. C. Yung, “Pedestrian behavior pp. 496–509, Jun. 2015.
prediction based on motion patterns for vehicle-to-pedestrian colli- [38] L. Zhu, H. Fan, Y. Luo, M. Xu, and Y. Yang, “Temporal
sion avoidance,” in Proc. 11th Int. IEEE Conf. Intell. Transp. Syst., cross-layer correlation mining for action recognition,” IEEE
Oct. 2008, pp. 316–321, doi: 10.1109/ITSC.2008.4732644. Trans. Multimedia, vol. 24, no. 51, pp. 668–676, Feb. 2021,
[18] C. G. Keller and D. M. Gavrila, “Will the pedestrian cross? A study on doi: 10.1109/TMM.2021.3057503.
pedestrian path prediction,” IEEE Trans. Intell. Transp. Syst., vol. 15, [39] D. Helbing, “A mathematical model for the behavior of pedestrians,”
no. 2, pp. 494–506, Apr. 2014, doi: 10.1109/TITS.2013.2280766. Behav. Sci., vol. 36, no. 4, pp. 298–310, 1991.

338 VOLUME 3, 2022

[40] D. Helbing and P. Molnár, “Social force model for pedestrian dynam- IOANNA SPYROPOULOU received the Diploma
ics,” Phys. Rev. E, Stat. Phys. Plasmas Fluids Relat. Interdiscip. Top., degree in civil engineering from the National
vol. 51, no. 5, p. 4282, 1995. Technical University in Athens (NTUA), Greece,
[41] C. Antoniou, H. N. Koutsopoulos, and G. Yannis, “Dynamic data- the M.Sc. degree in transport from Imperial
driven local traffic state estimation and prediction,” Transp. Res. C, College, and the Ph.D. degree in transport from
Emerg. Technol., vol. 34, pp. 89–107, Sep. 2013. University College London. He is an Assistant
[42] L. Breiman, “Random forests,” Mach. Learn., vol. 45, no. 1, pp. 5–32, Professor with NTUA. She has authored several
2001. scientific papers published in international jour-
[43] A. Verikas, A. Gelzinis, and M. Bacauskiene, “Mining data with nals and presented in research conferences, in the
random forests: A survey and results of new tests,” Pattern Recognit., field of transport. She has participated in several
vol. 44, no. 2, pp. 330–349, 2011. national and international research projects and
[44] E. Šikudová, K. Malinovská, R. Škoviera, J. Škovierová, M. Uller, and acts as an external expert evaluator of research proposals and projects
V. Hlaváć, “Estimating pedestrian intentions from trajectory data,” in for the EU and national organizations in the field of transport and intelli-
Proc. IEEE 15th Int. Conf. Intell. Comput. Commun. Process. (ICCP), gent transport systems. Her principal research interests are traffic modeling
2019, pp. 19–25, doi: 10.1109/ICCP48234.2019.8959707. and simulation, intelligent transport systems, traveler behavior, and active
[45] G. Fuhr and C. R. Jung, “Collective behavior recognition using transport modes.
compact descriptors,” 2018, arXiv:1809.10499.
[46] A. Liaw and M. Wiener, “Classification and regression by random-
Forest,” R News, vol. 2, pp. 18–22, Dec. 2002.
HARRIS PERAKIS received the Diploma degree
[47] W. S. Cleveland, “Robust locally weighted regression and smoothing
from the National Technical University of Athens,
scatterplots,” J. Amer. Stat. Assoc., vol. 74, no. 368, pp. 829–836,
where he is currently pursuing the Ph.D.
1979.
degree with the School of Rural and Surveying
[48] W. S. Cleveland and S. J. Devlin, “Locally weighted regression: An
Engineering. He has coauthored multiple peer-
approach to regression analysis by local fitting,” J. Amer. Stat. Assoc.,
reviewed journal and conference scientific publi-
vol. 83, no. 403, pp. 596–610, 1998.
cations. He has been participated as a Research
[49] H. Xue, D. Q. Huynh, and M. Reynolds, “PoPPL: Pedestrian trajectory
Associate with NTUA since 2015 in various
prediction by LSTM with automatic route class clustering,” IEEE
research projects (e.g., EMPARCO: Efficient
Trans. Neural Netw. Learn. Syst., vol. 32, no. 1, pp. 77–90, Jan. 2021,
management of parking under constraints and
doi: 10.1109/TNNLS.2020.2975837.
PEGASUS: Development of a self-trained truck
[50] H. Xue, D. Q. Huynh, and M. Reynolds, “SS-LSTM: A hierarchi-
driver-coaching system based on geolocation, IoT and BI techniques). He
cal LSTM model for pedestrian trajectory prediction,” in Proc. IEEE
participates as a teaching assistant in four under-graduate and one graduate
Winter Conf. Appl. Comput. Vis. (WACV), Mar. 2018, pp. 1186–1194,
courses with NTUA and he has assisted the supervision of eight diploma
doi: 10.1109/WACV.2018.00135.
theses. His research interests focus at indoor/ hybrid positioning algorithms
[51] R Core Team (R Found. Stat. Comput., Vienna, Austria). A
for groups of moving nodes utilizing multi-sensor approaches. He was a
Language and Environment for Statistical Computing. (2021).
member of the Management Committee of the COST Action SaPPART for
[Online]. Available: http://www.R-project.org/
which he has also performed a Short-Term Scientific Mission at EPFL. He
[52] M. Kuhn, “Building predictive models in R using the caret package,”
is active within the IAG organization and has participated to the IAG-FIG
J. Stat. Softw., vol. 28, no. 1, pp. 1–26, 2008.
sub-commissions collaborative field-tests for cooperative indoor/ outdoor
[53] M. Abadi et al., “TensorFlow: A system for large-scale machine learn-
positioning with the Ohio State University.
ing,” in Proc. 12th USENIX Symp. Oper. Syst. Des. Implement. (OSDI),
2016, pp. 265–283.
[54] F. Chollet, A. Yee, and R. Prokofyev. “Keras: Deep Learning
for Humans.” 2015. [Online]. Available: https://github.com/keras- VASSILIS GIKAS received the Diploma degree in
team/keras (Accessed: 2020). surveying engineering from the National Technical
[55] Y. Li, L. Xin, D. Yu, P. Dai, J. Wang, and S. E. Li, “Pedestrian University of Athens (NTUA), Greece, and the
trajectory prediction with learning-based approaches: A comparative Ph.D. degree in geodesy from the University of
study,” in Proc. IEEE Intell. Veh. Symp. (IV), Jun. 2019, pp. 919–926, Newcastle upon Tyne, U.K. He is a Professor
doi: 10.1109/IVS.2019.8814183. with NTUA. In the past he worked for the oil
[56] V. Papathanasopoulou, I. Markou, and C. Antoniou, “Online cal- exploration industry in the U.K. and USA as a
ibration for microscopic traffic simulation and dynamic multi-step positioning and navigation scientist and supervi-
prediction of traffic speed,” Transp. Res. C, Emerg. Technol., vol. 68, sor in seismic acquisition programs in Europe and
pp. 144–159, Jul. 2016. West Africa. His principal research interests are
in geodesy and GNSS, methods and multi-sensor
systems for positioning and navigation as well as dynamic monitoring and
analysis of structures and physical processes. He is a member of several pro-
VASILEIA PAPATHANASOPOULOU received the
fessional and scientific organizations. He is currently the Co-Chair of Com.
Diploma degree in surveying engineering, the
4 “Positioning and Applications”, International Association of Geodesy and
M.S. degree in geoinformatics, and the Ph.D.
the Vice-Chair of WG 6.1 “Deformation Measurement and Analysis” of
degree in transportation engineering from the
the International Federation of Surveyors.
National Technical University in Athens (NTUA),
Greece, in 2010, 2012, and 2018, respectively,
where she has produced auxiliary work with
the Department of Transportation Infrastructure, ELENI ANDRIKOPOULOU received the Diploma
School of Surveying Engineering from 2014 to degree in surveying engineering, the first M.S.
2018 and has participated in the research project degree in geoinformatics, and the second M.S.
EMPARCO concerning the efficient management degree in urban and regional planning from
of parking under constraints from 2014 to 2015. In addition, she worked the National Technical University of Athens in
with the concessionaire “Nea Odos S.A. and Central Greece S.A.” as a 2017, 2019, and 2019, respectively, where she is
Traffic and Transport Engineer and participating in the update of the traf- currently pursuing the Ph.D. degree in pedestri-
fic model of the motorways managed by the concessionaire from 2018 to ans’ movement characteristics with the School of
2021. She has published her work in several peer-reviewed scientific jour- Rural and Surveying Engineering. She has per-
nals, proceedings of international conferences and two chapters in books, formed research on driver and pedestrian behavior
while at the same time she is a regular reviewer in scientific journals. Her (diploma and 1st M.S. thesis) and walkability (2nd
research focuses on traffic simulation models, traffic management, machine M.Sc. thesis) and in her Ph.D. research she attempts to combine issues of
learning, autonomous vehicles, and intelligent transportation systems. walkability and pedestrian behavior.

VOLUME 3, 2022 339