Ferramenta de extração de trajetórias de usuários do sistema de transporte em vídeos de monitoramento do tráfego usando técnicas de visão computacional

Abstract

The traffic studies at the microscopic scale focus on the individual movements of transportation system users. Examples of such studies include behavioral modeling and road safety assessments. In this setting, trajectory data of vehicles and pedestrians are extremely useful for the development of these studies, but their acquisition poses a challenge due to the vast amount of observations that need to be collected at each moment in time. However, with the growing success of the field of Artificial Intelligence, specifically Computer Vision, new and more efficient models and algorithms have emerged capable of automating the extraction of object trajectories in videos. Given this context, this undergraduate thesis aims to construct a tool for extracting trajectories of transportation system users from traffic monitoring videos using well-established, open-source computer vision models and algorithms available in the literature. The users referred to here are pedestrians, cyclists, cars, trucks, and buses. The constructed tool was based on the YOLOv7 object detection model (WANG et al., 2022) and the StrongSORT object tracking algorithm (DU et al., 2023). After validating the tool in the tasks of classificatory vehicle counting, determination of the instants when vehicles enter and exit the pedestrian crosswalk, pedestrian crossing count and estimation of pedestrian delay and crossing time, promising outcomes emerged from the tasks related to vehicles, however, was acknowledged greater room for improvement in the tool's quality given the results obtained in the tasks related to pedestrian crossings.