Institute of Transportation, MOTC

         According to the observation perspective, traffic flow theory can be divided into the macroscopic and microscopic levels. Macroscopic traffic flow theory describes the overall behavior of vehicles, whereas microscopic traffic flow theory describes the behavior between two vehicles. In traffic data, vehicle trajectory data, which contains the spatial positions of vehicles at each time instance, is the most detailed form of traffic data. The most recognized vehicle trajectory dataset is the Next Generation SIMulation (NGSIM), a program initialized by the Federal Highway Administration (FHWA) of the US Department of Transportation in 2005. Up to now, the database has been widely used in numerous traffic studies to explore for the traffic characteristic and used to derive traffic flow models. However, there is no commonly recognized methodology to evaluate the quality and accuracy of NGSIM. The accuracy of the database was constrained by the video collection method and the image recognition technology at the time of the NGSIM program. Several synchronized cameras collected the videos with different monitoring angle mounting on top of high buildings adjacent to the roadway, resulting in parallax and shadows issues in the images. The accuracy of the NGSIM dataset has been criticized in several recent studies, advocating that trajectory reconstruction is needed before the trajectory database is used for microscopic studies. The purpose of trajectory reconstruction is to clean out the inconsistencies in the vehicle trajectory pairs and make sure the trajectory data can be more reasonable before it is used for further analysis at the microscopic level.
     Each location of the NGSIM dataset covers a time period of only 45 minutes, and traffic scenarios contained in the data are limited. Coupled with the limitation of image resolution, the quality of trajectory data is difficult to improve further. Due to the rapid advancement of emerging technologies such as Unmanned Aerial Vehicle (UAV) andComputer Vision (CV) in recent years, there have been many successful traffic survey cases in collecting vehicle trajectories similar to the NGSIM program. Still, the accuracy of trajectories remains to be assessed.
       Therefore, this study develops a quality assessment method and trajectory reconstruction algorithm for a vehicle trajectory dataset, which is obtained with aerial videos collected by UAV on highway sections and extracted by a deep-learning-based computer vision technique. The quality of reconstructed trajectory is further verified by embedding the reconstructed four-corner bounding boxes of vehicles onto the aerial videos, so as one can efficiently reconfirm the accuracy of the reconstructed trajectory by visualization. This research further calculates the macroscopic and microscopic traffic flow characteristics and compares the differences in traffic characteristics with the datasets before and after trajectory reconstruction. The results show that, regardless of high or low traffic density scenarios in the traffic stream, the proposed trajectory reconstruction methodology can significantly improve the quality of trajectory data.