Predicting Pedestrian Crossing Behavior in Germany and Japan: Insights into Model Transferability

Chi Zhang, Janis Sprenger, Zhongjun Ni, Christian Berger · 2024

Read Predicting Pedestrian Crossing Behavior in Germany and Japan: Insights into Model Transferability by Chi Zhang, Janis Sprenger, Zhongjun Ni, Christian Berger (2024) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2412-03689.txt.

Introduction

Cars that drive themselves must guess when a person on foot will step into the road. Chi Zhang, Janis Sprenger, Zhongjun Ni, and Christian Berger compare that guessing task across two lands in a 2024 study. They use simulator data from Germany and Japan, where street habits differ.

Four kinds of learning models were trained to predict gap choice, zebra use, and walking paths. Walkers in the Japan study waited for larger gaps than those in Germany. Neural networks won at gap and zebra calls, while random forests won at path calls. A grouping trick, added without labels, lifted scores when a model moved to the other land. The results can help builders make safety tools that travel well.

Catalog record

CREATORChi Zhang, Janis Sprenger, Zhongjun Ni, Christian Berger
YEAR2024
COLLECTIONPAPERS
FORMATTXT
FILE SIZE71 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2412-03689

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Predicting Pedestrian Crossing Behavior in Germany and Japan: Insights into Model Transferability Chi Zhang, Janis Sprenger, Zhongjun Ni, Christian Berger Abstract Predicting pedestrian crossing behavior is important for intelligent traffic systems to avoid pedestrian-vehicle collisions. Most existing pedestrian crossing behavior models are trained and evaluated on datasets collected from a single country, overlooking differences between countries. To address this gap, we compared pedestrian road-crossing behavior at unsignalized crossings in Germany and Japan. We presented four types of machine learning models to predict gap selection behavior, zebra crossing usage, and their trajectories using simulator data collected from both countries. When comparing the differences between countries, pedestrians from the study conducted in Japan are more cautious, selecting larger gaps compared to those in Germany. We evaluate and analyze model transferability. Our results show that neural networks outperform other machine learning models in predicting gap selection and zebra crossing usage, while random forest models perform best on trajectory prediction tasks, demonstrating strong performance and transferability. We develop a transferable model using an unsupervised clustering method, which improves prediction accuracy for gap selection and trajectory prediction. These findings provide a deeper understanding of pedestrian crossing behaviors in different countries and offer valuable insights into model transferability. I Introduction Intelligent driving systems and smart traffic infrastructures are developed to enable safer and smarter travel. As pedestrians are vulnerable, understanding and predicting their behavior is crucial, especially at unsignalized crossings without traffic lights, where pedestrians constantly interact with traffic due to the ambiguity of the right-of-way. Predicting pedestrian behavior is challenging, as many factors influence pedestrian behavior [ 1 , 2 ] . Previous research studies focused on tasks including crossing intention prediction [ 3 , 4 , 5 , 6 , 7 , 8 ] and trajectory prediction [ 9 , 10 , 11 , 12 , 13 , 14 ] , but many of them did not explicitly consider pedestrian-vehicle interactions. When pedestrians interact with vehicles at unsignalized crossings, intelligent traffic systems should be able to predict their behavior, such as the time gap they will choose to cross, whether they will use zebra crossings, and their crossing trajectories. Zhang et al. [ 15 ] has predicted and analyzed pedestrian gap selection and zebra crossing usage using data collected by a simulator study conducted in Germany. Their study provides insights into the key factors that influence pedestrian behavior when interacting with multiple vehicles. However, Zhang et al.’s study [ 15 ] and most other existing models only focused on single datasets, neglecting the differences between countries. An intelligent traffic system should be functional and safe in different countries. In other words, the model trained in one country should be transferable to other countries. Therefore, in this study, we investigate and compare the differences between Germany and Japan, providing insights for bridging differences between countries and improving model transferability. Simulator studies have been widely used by researchers (e.g., [ 16 , 17 , 15 ] ) to investigate pedestrian-vehicle interactions, as it is safe and avoids real pedestrian-vehicle conflicts in near-crash events. In this context, the term “pedestrian-vehicle interactions” refers to the exchange of actions and reactions between pedestrians and vehicles that influence each other’s behavior during encounters. These interactions can be explicitly represented and modeled by measurable variables such as time to arrival (TTA) or time gap, pedestrian waiting time, pedestrian walking speed, and vehicle speed. In this study, we use simulator data collected in a virtual reality (VR) simulator by Sprenger et al. [ 17 ] from both Germany and Japan. The head-mounted VR systems ensure a consistent and controlled environment across different countries, reducing the impact of irrelevant factors and allowing us to focus on behavioral differences between the two countries. This study extends Zhang et al.’s research [ 15 ] , which explored scenarios where pedestrians interact with multiple vehicles in traffic flow using machine learning models. Compared to their research that only focused on data collected from Germany, we investigate the differences between data collected from Germany and Japan, and explore the transferability of these models. Furthermore, we expand the scope to include pedestrian crossing trajectories in addition to predicting gap selection behavior and zebra crossing usage. Particularly, for each prediction task, our research questions are: RQ1 What are the similarities and differences in pedestrian crossing behavior observed between studies conducted in Germany and Japan? RQ2 How transferable are models trained on data collected from Germany and Japan when tested on the other dataset? RQ3 How do unsupervised learning clustering methods enhance performance in building transferable models? The main contributions of this study include: • Proposing and evaluating machine learning models to predict pedestrian gap selection behavior, zebra crossing usage, and crossing trajectories in Germany and Japan, comparing behavioral patterns between the data collected from both countries. • Evaluating the transferability of models trained on the data collected from Germany and Japan to each other. • Proposing and comparing methods based on unsupervised learning-clustering to enhance model transferability and performance on datasets collected from Germany and Japan. II Related Work In this section, we review related studies, focusing on three key areas: pedestrian crossing behavior prediction, analysis of pedestrian crossing behavior, and differences in pedestrian behavior between countries. II-A Pedestrian Crossing Behavior Prediction Studies such as [ 9 , 10 , 11 , 12 , 13 , 14 , 5 ] focused on pedestrian trajectory prediction, using past spatial and temporal information to predict future movements. These methods typically use a 3.2 s time window for observation and output a 4.8 s prediction. However, they only rely on pedestrians’ past trajectories and do not explicitly consider pedestrian-vehicle interaction factors, failing to capture the complexities of crossing decisions that can extend beyond 8 s. To bridge this gap, this study considers interaction factors to predict pedestrians’ crossing paths without relying on their past trajectories. Some studies focused on pedestrian crossing intentions and applied deep learning on naturalistic data to determine whether a pedestrian intends to cross. Studies such as [ 3 , 4 , 6 , 8 ] mainly considered only pedestrian-related features, including pedestrian appearance, their surroundings, and pedestrian skeleton information. Other studies implicitly considered pedestrian-vehicle interactions by using relative pedestrian-vehicle distances [ 7 ] or their trajectories [ 5 ] as inputs to deep learning networks, or by modeling interactions through graph convolutional networks [ 18 , 19 ] . However, these approaches rely on implicit interaction models, which lack explainability. An explicitly defined, interpretable model of pedestrian-vehicle interactions is often missing due to the complexity of controlling the environment in naturalistic data and the difficulty in collecting near-crash scenarios for safety reasons. Studies [ 20 , 21 , 22 , 16 ] investigated pedestrian-vehicle interaction, focusing on scenarios where pedestrians interact with a single vehicle, considering factors such as time to arrival, position, velocity, pedestrian waiting time, and personal traits. These studies neglected the complex situations where pedestrians interact with multiple vehicles, which is common in real-world environments. Simulator studies provide a controlled environment to ensure pedestrian safety and collect detailed interaction information, avoiding real-world near-crash scenarios and are utilized by many researchers. Jayaraman et al. [ 23 ] analyzed and predicted gap acceptance behavior at zebra crossings,

Read the complete text