The Dynamics of Lane-changing Maneuvers in Motorway Traffic Flow: Data Fusion, Analysis, and Modeling

Motorway traffic flow, at the heart of modern transportation networks, is a complex and dynamic system. It encompasses vehicle movement, driver interactions, and the multifaceted factors influencing road travel efficiency, safety, and sustainability. At its core, motorway traffic flow entails the study of vehicle operations within road networks and their interactions. Driver maneuvers encompass both longitudinal actions, guiding vehicles along their path, and lateral maneuvers, typically involving lane changes within motorway networks.

Lane-changing maneuvers (LCMs) involve gap acceptance decisions and require that the vehicle temporarily occupies two lanes which results in reduced capacity, decreased speeds, and increased accident risk. LCM maneuvers are especially prevalent in complex motorway segments like weaving areas, making such areas congestion and accident hotspots.

The emergence of connected and automated vehicles presents promising opportunities to enhance traffic flow through real-time guided and cooperative management strategies. An in-depth empirical analysis of LCMs can unveil the circumstances under which drivers perform these maneuvers, aiding researchers in modeling these decisions and understanding the relationship between LCMs, congestion formation, congestion discharge (capacity drop), and accidents. This deeper understanding can inform the development of more effective traffic management strategies based on well-defined hypotheses. Furthermore, while microsimulation software plays a significant role in traffic flow studies, uncalibrated simulations lack accuracy and may lead investigations astray. Research efforts have predominantly focused on calibrating these microsimulators for longitudinal movements, while calibration for lateral movements has been largely overlooked. Comprehensive empirical studies can address this gap. The development of an LCM prediction model relies on a thorough understanding of why and when LCMs occur and their associated descriptive variables. Such insights can only be gained through empirical analysis, making lane-change and lateral maneuver studies essential research areas.

Despite their significance, there is a noticeable gap of research on lateral vehicle movements in the existing literature. This research gap can be attributed to the scarcity of appropriate empirical datasets, particularly continuous vehicle trajectories with sufficient lateral accuracy.

Understanding drivers' lane-changing behavior, a fundamental aspect of this thesis, constitutes a crucial stride in these studies and endeavors. However, the cornerstone is procuring appropriate data, underpinning the primary mission of this thesis. It endeavors to provide not only the requisite data for its LCM study but also to pave the way for future research of this type.

The principal aim of this study is to establish a data set for investigating lane-changing maneuvers (LCMs) within motorway networks that avoids the limitations in temporal scope and network coverage of existing datasets.

The contributions of this research can be categorized into five domains. The primary contributions are two independent platforms for furnishing data on lane-changing maneuvers (LCMs) and the subsequent demonstration of the investigative possibilities related to drivers' LCM behavior.

The foremost contribution of this thesis is a four-step methodology that rectifies lateral bias in trajectory data from smartphones. This approach incorporates data fusion, combining trajectory data with information from loop detectors to accurately determine lateral positions, thereby revealing driving lanes and lane changes. The methodology's efficacy is assessed using drone and closed-circuit television (CCTV) data, revealing a rate exceeding 94% in correctly matching trajectory data with loop detector records. Notably, this approach rectifies lateral position errors between detector stations, with over 90% of trajectory points correctly positioned within lanes. The algorithm's streamlined calibration process, reliant on just two parameters, renders it easily applicable to various test networks.

The subsequent contribution is a probabilistic methodology for re-identifying vehicles at sequential lane-specific loop detectors. This approach encompasses four distinct modules: calibration, lane-restricted re-identification, non-lane-restricted re-identification, and a mechanism for addressing inconsistencies while finalizing approximated trajectories and their associated lane changes. Bayesian estimation plays a pivotal role in calculating similarity probabilities based on three criteria, facilitating the identification of optimal matches between upstream and downstream observations. Recursive re-identification occurs within predefined boundaries established by previously identified lane-keeping vehicles. A lane-change filter is implemented to validate cross-lane matches. The methodology's validation, conducted through CCTV video data, demonstrates success rates of 99.15% and 96.97% across diverse traffic flow scenarios.

The final contribution of this thesis is an exhaustive exploration of LCM behavior within complex weaving segments. First, we empirically examine various facets of drivers' LCM behavior within an extensive weaving area. The results showed different LCM behavior during peak and off-peak hours, including selecting different locations inside the weaving segment for the maneuvers. Then, a multiclass macroscopic lane change model is presented. This discrete choice model was trained on the dataset of reconstructed continuous trajectories from this thesis and predicts lane change rates based on macroscopic traffic conditions within lanes. The models incorporate diverse traffic flow characteristics, stimuli, and inhibiting variables, showcasing robust predictive capabilities through rigorous estimation, validation, and demonstration trials.