National Technical University of Athens (NTUA)
Data-driven pedestrian simulation: an alternative to theory-based pedestrian simulation?
Abstract
dc:descriptionThe aim of the research is to examine whether data-driven pedestrian simulation models can outperform the theoretical ones and provide a robust model framework for pedestrian simulation. Initially, an extended literature review was performed to identify the existing pedestrian simulation models and the main parameters utilized in pedestrian simulation. To achieve the aim of the study, a comparative analysis of a well-known and widely applied theoretical pedestrian simulation model (i.e. the social force mode) and four data-driven techniques: the Artificial Neural Networks, the Support Vector Regression, the Gaussian Processes and the Locally Weighted Regression was conducted. A suitable methodological framework for the comparative analysis was designed. Initially, appropriate data (i.e. pedestrian trajectories) were collected from two different area types: a metro station during peak hours and a shopping mall during afternoon hours, via video recordings. Then, with the aid of an appropriate software, pedestrian trajectories were extracted. Due to the fact that the collected data include white noise, an algorithm for noise elimination was developed as a combination of existing smoothing filters. Subsequently, an appropriate pedestrian simulation model setup for the data-driven techniques was developed, as they do not cater specifically for pedestrian simulation framework. In order to conduct a fair comparison the variables of the theoretical model were employedin the data-driven models. Cross-validation was applied as the appropriate method for examining each model’s performance and to cater for data overfitting, while a combination of goodness-of-fit measures the models’ accuracy were estimated to assess the models in a holistic manner. The results indicate that data-driven methods have higher capability of simulating pedestrian movement as they perform better according to all of goodness-of-fit measures. Following the first level of comparison (compare models with the same parameters), additional parameters (agent’s characteristics and time parameter) have been included in the data-driven models in order to examine the possibility of improving (and its magnitude) the performance of these models. Results of this analysis indicate that the employment of the selected variables can improve data-driven pedestriansimulation models performance (they performed better for almost every goodness-of-fit measure).
Degree
thesis:*- Grantor dc:publisher
- National Technical University of Athens (NTUA)
- Year dc:date
- 2022
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Kouskoulis, Georgios
- Κουσκουλής, Γεώργιος
Subjects
dc:subject × 31- Προσομοίωση της πεζής κίνησης
- Μέθοδοι ανάλυσης δεδομένων
- Social force μοντέλο
- Τεχνητά νευρωνικά δίκτυα
- Μηχανές διανυσμάτων υποστήριξης
- Gaussian processes
- Τοπικά σταθμισμένη μη παραμετρική παλινδρόμηση
- Cross-validation
- Δείκτες καλής προσαρμογής
- Unscented kalman filter
- Κινητός μέσος όρος
- Ελαχιστοποίηση θορύβου δεδομένων
- Pedestrian simulation
- Data-driven techniques
- Social force model
- Artificial neural networks
- Support vector regression
- Locally weighted regression
- Goodness-of-fit measures
- Moving average
- Data noise elimination
- Άλλες Επιστήμες Μηχανικού και Τεχνολογίες
- Επιστήμες Μηχανικού και Τεχνολογία
- Μηχανική και Τεχνολογίες, άλλοι τομείς
- Επιστήμη Πολιτικού Μηχανικού
- Επιστήμη και Τεχνολογία των μεταφορών
- Other Engineering and Technologies
- Engineering and Technology
- Engineering and Technologies, miscellaneous
- Civil Engineering
- Transportation science and Technology
Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/51549