Institute of Transportation, MOTC

         Vehicle and pedestrian trajectories are important references for many traffic analysis applications. In this study, we focus on the complex traffic environment at intersections and use deep learning to automatically locate the positions of vehicles and pedestrians in aerial images, and then extract a large amount of rich information on vehicle and pedestrian trajectories. There are three main contributions in this study: (1) Combining aerial camera with several cross-domain technologies such as deep learning and image processing, a highly automated human/vehicle trajectory extraction technology is accomplished using real intersections as test data. (2) Solve the problem of fitting bounding boxes to vehicles and detecting small objects such as pedestrians by using Mask R-CNN and YOLOv3, respectively; (3) Expanding traditional trajectory information from lines to travel area information to further enrich trajectory information. By automatically extracting the trajectory information on 2D maps, it helps to enrich the information needed for advanced traffic analysis.