Pedestrian Behavior Prediction Using Machine Learning Methods
Doctoral thesis, 2024
Background: Accurate pedestrian behavior prediction is essential for reducing fatalities from pedestrian-vehicle collisions. Machine learning can support automated vehicles to better understand pedestrian behavior in complex scenarios.
Objectives: This thesis aims to predict pedestrian behavior using machine learning, focusing on trajectory prediction, crossing intention prediction, and model transferability.
Methods: We identified research gaps by reviewing the literature on pedestrian behavior prediction. To address these gaps, we proposed deep learning models for pedestrian trajectory prediction using real-world data, considering social and pedestrian-vehicle interactions. We integrated spectral features to improve model transferability. Additionally, we developed machine learning models to predict pedestrian crossing intentions using simulator data, analyzing interactions in both single and multi-vehicle scenarios. We also investigated cross-country behavioral differences and model transferability through a comparative study between Japan and Germany.
Results: For trajectory prediction, incorporating social and pedestrian-vehicle interactions into deep learning models improved accuracy and inference speed. Integrating spectral features using discrete Fourier transform improved motion pattern capture and model transferability. For crossing intention prediction, neural networks outperformed other machine learning methods. Key factors that influence pedestrian crossing behavior included the presence of zebra crossings, time to arrival, pedestrian waiting time, walking speed, and missed gaps. The cross-country study revealed both similarities and differences in pedestrian behavior between Japan and Germany, providing insights into model transferability.
Conclusions: This thesis advances pedestrian behavior prediction and the understanding of pedestrian-vehicle interactions. It contributes to the development of smarter and safer automated driving systems.
Objectives: This thesis aims to predict pedestrian behavior using machine learning, focusing on trajectory prediction, crossing intention prediction, and model transferability.
Methods: We identified research gaps by reviewing the literature on pedestrian behavior prediction. To address these gaps, we proposed deep learning models for pedestrian trajectory prediction using real-world data, considering social and pedestrian-vehicle interactions. We integrated spectral features to improve model transferability. Additionally, we developed machine learning models to predict pedestrian crossing intentions using simulator data, analyzing interactions in both single and multi-vehicle scenarios. We also investigated cross-country behavioral differences and model transferability through a comparative study between Japan and Germany.
Results: For trajectory prediction, incorporating social and pedestrian-vehicle interactions into deep learning models improved accuracy and inference speed. Integrating spectral features using discrete Fourier transform improved motion pattern capture and model transferability. For crossing intention prediction, neural networks outperformed other machine learning methods. Key factors that influence pedestrian crossing behavior included the presence of zebra crossings, time to arrival, pedestrian waiting time, walking speed, and missed gaps. The cross-country study revealed both similarities and differences in pedestrian behavior between Japan and Germany, providing insights into model transferability.
Conclusions: This thesis advances pedestrian behavior prediction and the understanding of pedestrian-vehicle interactions. It contributes to the development of smarter and safer automated driving systems.
trajectory prediction
intention prediction
machine learning
Pedestrian behavior
pedestrian-vehicle interaction
deep learning
Jupiter 520, Lindholmen, Chalmers | University of Gothenburg
Opponent: Dr. He Wang, Associate Professor, Department of Computer Science, University College London, UK
Author
Chi Zhang
Software Engineering 2
Pedestrian Behavior Prediction Using Deep Learning Methods for Urban Scenarios: A Review
IEEE Transactions on Intelligent Transportation Systems,;Vol. 24(2023)p. 10279-10301
Journal article
Social-IWSTCNN: A social interaction-weighted spatio- temporal convolutional neural network for pedestrian trajectory prediction in urban traffic scenarios
IEEE Intelligent Vehicles Symposium, Proceedings,;Vol. 2021-July(2021)p. 1515-1522
Paper in proceeding
Learning the Pedestrian-Vehicle Interaction for Pedestrian Trajectory Prediction
2022 8th International Conference on Control, Automation and Robotics (ICCAR),;Vol. 2022-April(2022)p. 230-236
Paper in proceeding
Spatial-Temporal-Spectral LSTM: A Transferable Model for Pedestrian Trajectory Prediction
IEEE Transactions on Intelligent Vehicles,;(2023)
Journal article
Cross or Wait? Predicting Pedestrian Interaction Outcomes at Unsignalized Crossings
IEEE Intelligent Vehicles Symposium, Proceedings,;(2023)
Paper in proceeding
Predicting and Analyzing Pedestrian Crossing Behavior at Unsignalized Crossings
IEEE Intelligent Vehicles Symposium, Proceedings,;(2024)
Paper in proceeding
Predicting Pedestrian Crossing Behavior in Germany and Japan: Insights into Model Transferability
IEEE Transactions on Intelligent Vehicles,;(2024)
Journal article
Pedestrian safety is a critical global concern, with an urgent need to reduce fatalities caused by pedestrian-vehicle collisions. Machine learning offers a powerful tool for reducing these collisions by improving the prediction of pedestrian behavior. This thesis focuses on predicting and analyzing pedestrian behavior in complex traffic scenarios using machine learning methods.
After reviewing existing studies, we propose deep learning models to improve pedestrian trajectory prediction. These models consider social interactions between pedestrians and their interactions with vehicles, enhancing both accuracy and inference speed. Additionally, we improve model transferability by including spectral features.
We also investigate pedestrian crossing intentions when interacting with vehicles using machine learning methods. Key factors such as the presence of zebra crossings, waiting time, walking speed, and missed crossing chances strongly influence pedestrian behavior. A cross-country comparison between Japan and Germany reveals both similarities and differences in pedestrian behavior, providing valuable insights into model transferability.
Overall, this research advances the prediction of pedestrian behavior, providing insights for safer pedestrian-vehicle interactions in complex scenarios. These findings can guide the development of smarter, safer automated driving systems.
After reviewing existing studies, we propose deep learning models to improve pedestrian trajectory prediction. These models consider social interactions between pedestrians and their interactions with vehicles, enhancing both accuracy and inference speed. Additionally, we improve model transferability by including spectral features.
We also investigate pedestrian crossing intentions when interacting with vehicles using machine learning methods. Key factors such as the presence of zebra crossings, waiting time, walking speed, and missed crossing chances strongly influence pedestrian behavior. A cross-country comparison between Japan and Germany reveals both similarities and differences in pedestrian behavior, providing valuable insights into model transferability.
Overall, this research advances the prediction of pedestrian behavior, providing insights for safer pedestrian-vehicle interactions in complex scenarios. These findings can guide the development of smarter, safer automated driving systems.
Supporting the interaction of Humans and Automated vehicles: Preparing for the Environment of Tomorrow (Shape-IT)
European Commission (EC) (EC/H2020/860410), 2019-10-01 -- 2023-09-30.
Areas of Advance
Information and Communication Technology
Transport
Subject Categories
Robotics
Computer Science
Computer Vision and Robotics (Autonomous Systems)
ISBN
978-91-8069-818-4
Publisher
Chalmers
Jupiter 520, Lindholmen, Chalmers | University of Gothenburg
Opponent: Dr. He Wang, Associate Professor, Department of Computer Science, University College London, UK