Enhanced-NGSIM
Description
The Next Generation Simulation (NGSIM) is among the most widely utilized traffic datasets. Although it is frequently used, a number of comprehensive analysis have surfaced, highlighting intrinsic problems with the dataset’s trajectory data, from non-existent collisions between vehicles to impossible acceleration values. In this project, we introduce a fully automated deep learning pipeline for re-extracting all vehicle trajectories from the NGSIM United States Highway 101 (US-101) dataset. Our core Artificial Intelligence (AI) contribution is the deployment of a state-of-the-art super-resolution network coupled with transformer-based object detection (RT-DETR) and homography-based tracking. This AI methodology is applied to the engineering challenge of accurately extracting microscopic traffic trajectory data from degraded, uncalibrated video feeds. The findings indicate a notable improvement compared to the original dataset, with a 20% rise in the quantity of trajectory points, a nearly 50% decrease in the number of vehicles impacted by unattainable acceleration values, and the removal of vehicle collisions. The final dataset will be publicly accessible for its usage.