Development of A Trajectory Population Data and its Application in CAV Research

TR Number
Journal Title
Journal ISSN
Volume Title
Virginia Tech

Vehicle trajectory data has played a critical role in the recent history of traffic flow and CAV operations-related studies. However, available trajectories have limited coverage, either spatial or temporal. The implementation of CAV technology is expected to produce a large-scale trajectory dataset. However, at the initial implementation level, the trajectory data produced is expected to have gaps in terms of completeness. This research develops a data model for large-scale trajectory data that can be built on CAV-collected trajectories and easily manipulated to produce traffic parameters for CAV control and operation research. A benchmarking process has been applied to test a trajectory reconstruction approach to develop a population database from partial trajectories to fill the expected data gap in CAV feedback. The large-scale trajectory data is then used in CAV operations-related studies focusing on CAV's integration with human drivers and developing performance matrices for CAV-controlled optimized trajectories.

This research used large-scale vehicle trajectory data from Wide Area Motion Imagery (WAMI) developed by PVLabs for modeling and analyzing traffic characteristics as a surrogate of CAV-collected trajectories. This timestamped location data capture provides trajectory information at an interval of one second. Trajectories from an approximate area of four-square kilometers in downtown Hamilton, Canada, are used to develop a data model to extract and store traffic characteristics. The video data was collected for two three-hour continuous periods, one in the morning and one in the evening of the same day. Like other moving object detection-based algorithms, this data suffers from false-positive detection, false-negative detection, and other positional inaccuracies caused by faulty image registration. A context-based trajectory filtering algorithm has been developed and validated against ten minutes of vehicle counts from actual WAMI images. The filtered data provides a sample of trajectories over the area, including complete and partial vehicle trajectories, excluding undetected ones.

The missing trajectory reconstruction process using a dynamic state estimation process is developed to reconstruct partial and missing trajectories. A data analytics approach predicts the number of missing trajectories between two successive detections in the traffic stream on a roadway lane. A benchmarking test of the performance of the missing trajectory prediction algorithm is conducted using the NGSIM I80 database. A frame-by-frame learning method is developed to join the identified missing trajectories. This data analytics approach preserves the naturalistic property of the trajectory, which was a concern of previous traffic-flow model-based approaches. Joining partial/split trajectories provides a more comprehensive picture of the trajectory population. Due to data structure similarities, including the nature of the split and missing trajectories, the methods developed in this study to recover trajectories can be adopted for future CAV feedback data in a mixed traffic scenario.

The applicability of using the large-scale trajectory data model is explored in two performance areas of CAV operations. The first is a scenario-based testing process, which evaluates the "intelligence" of a CAV in handling interactions with Human driven Vehicles (HV) by artificially replacing an HV in the traffic stream with a CAV. Scenario-based testing is conducted for a particular Operational Design Domain (ODD). The ODD is defined as operating conditions under which particular driver assistance or automated control systems are designed to function. Existing literature on scenario-based testing primarily focuses on CAV-HV interaction on highways as large-scale naturalistic trajectory data are available to facilitate such studies. This research explores car-following and lane-changing aspects of arterial CAV testing. The large-scale trajectory data model generates testing scenarios and calibrates the surrogate model for CAV operation. The modification to the trajectory data model to accommodate the scenario-based testing is illustrated. The second consists of using the large-scale trajectory data model to estimate a new trajectory smoothness parameter that can indicate the impact of intersection stop-and-go movement on the smoothness of the entire trajectory. This smoothness parameter can be applied as an optimization variable in future trajectory control-based intersection management. Long-duration trajectories from the large-scale trajectory data are used to estimate the spectral arc length parameter for trajectory smoothness. This research only estimates smoothness parameters for human-driven vehicles to illustrate its applicability for vehicle trajectories.

This research developed a framework for applying expected partial trajectories from CAV technology in estimating near-complete trajectories. The large-scale data application process in two CAV operations-related studies is also provided.

Vehicle Trajectory, Trajectory Validation, Trajectory Reconstruction, Trajectory Smoothness, Connected and Automated Vehicles