Date of Award
5-16-2025
Publication Type
Dissertation
Degree Name
Ph.D.
Department
Civil and Environmental Engineering
Keywords
Autonomous Vehicle; Car-following behaviour; Causal Inference; Machine learning; Safety
Supervisor
Chris Lee
Rights
info:eu-repo/semantics/embargoedAccess
Creative Commons License
This work is licensed under a Creative Commons Attribution-NonCommercial-No Derivative Works 4.0 International License.
Abstract
Car-following behaviour is important for understanding and modelling traffic flow and safety. Recently, data-driven models have been developed to predict car-following behaviour due to their capability of capturing complex behavioural patterns in varying road conditions. These models have also been applied to autonomous vehicle (AV) controls to replicate human-like car-following behaviour. However, existing data-driven car-following models often overlook the dynamic influence of road geometry (curvature and slope) and have limitations in balancing safety, comfort, and computational efficiency. While some data-driven car-following models include road geometry as input variables, their internal feature-ranking processes do not guarantee that road geometry variables significantly influence predictions of car-following behaviour. Computational inefficiency also remains a challenge because advanced data-driven models like reinforcement learning algorithms require high computational resources and are unsuitable for resource-constrained environments. Furthermore, data-driven car-following models occasionally predict the behaviours which lead to a collision, which does not guarantee the safety of car-following, unlike collision-free physics-based models, such as the Gipps model. To address this limitation, hybrid car-following models, which combine physics-based and data-driven models to calculate the weighted average of model predictions, have been developed. However, since existing hybrid car-following models assume fixed or binary dynamic weights of physics-based and data-driven models, they may predict car-following behaviours that are less sensitive to varying driving conditions, such as changes in road geometric conditions and negatively affect driver and passenger comfort. Thus, these existing models have limitations in controlling AVs to ensure safety and comfort with the reduced computational power. To address these challenges, this thesis develops a new hybrid car-following model which predicts more comfortable and safer car-following behaviour in different driving conditions compared to the existing models. The OpenACC dataset, which contains real-world vehicle trajectories under varying geometric conditions across Italy, Sweden, and Hungary, was used to develop the model. The proposed hybrid car-following model determines the weights of physics-based and data-driven models based on drivers' risk perception, vehicle dynamics and road geometry, such as curvature and slope. In this study, the Knowledge Distillation Neural Network (KDNN) model was used as a data-driven model since it can balance between predictive accuracy and computational efficiency unlike more complex models (e.g. Long short-term memory (LSTM)) and light-weight models (e.g., Multi-layer perception (MLP)). Also, the Gipps' car-following model was used as a physics-based model since it ensures collision-free behaviour. It was found that the proposed hybrid model showed superior performance in terms of safety compared to the LSTM, MLP, and the hybrid model with binary dynamic weights. The model always maintained an acceptable minimum time-to-collision (TTC) value of above 1.5 seconds. Additionally, the proposed model had the lowest Deceleration Rate to Avoid Collision (DRAC) compared to the baseline KDNN and the binary-weighted hybrid model, indicating smoother and safer deceleration behaviour. It was also found that the proposed hybrid model predicted more gradual transitions in car-following behaviours during changes in driving states (e.g., free-flow and car-following) and road geometry conditions (e.g. straight and curved roads). Thus, the proposed hybrid model can be used for controlling longitudinal movements of AVs to increase passengers' comfort and safety in varying driving conditions.
Recommended Citation
Adewale, Ayobami, "A New Hybrid Data-Driven and Physics-Based Car-Following Model Considering Road Geometry and Safety" (2025). Electronic Theses and Dissertations. 9700.
https://scholar.uwindsor.ca/etd/9700