From Artificial Intelligence to Traffic Analysis: Automatic Video Detection and Tracking of Cyclists, Pedestrians, Mopeds and Vehicles
Open AccessWith the improvements in video recording, storage, and communication technologies, there is an opportunity to utilize digital images for different real-world applications. Such applications include roadway traffic detection and control where intelligent and effective methods are needed to transform visual information into useful mobility and safety metrics. In the transportation domain, there is a wealth of video files/data already gathered. The objective of this thesis is to adopt Artificial Intelligence (AI) to utilize video information and automatically detect and trace the movements of objects such as pedestrians, cyclists, mopeds, and vehicles. Toward realizing such an objective, there is a need to analyze the traced trajectories to efficiently get useful information such as flows, densities, and speeds. For additional validation purposes and in order to evaluate the proposed automatic video detection (AVD) methodology, this thesis compares some extracted AVD results with those obtained from an existing video detection and tracing exercise performed manually. There are two types of video data used in this research. The first video data provided by the Delft University of Technology (TU Delft) shows a mixed non-directional non-motorized traffic in a shared space next to a transit station in Amsterdam, the Netherlands. The second video data provided by the Korea Advanced Institute of Science and Technology (KAIST) shows a freeway environment administered by the Korea Expressway Corporation (KEC), South Korea. The adopted AVD approach adapts the YOLOv5 tool to detect the objects and Deep Sorting to trace locations across time. Moreover, Bezier curve and wedge transformation are adopted to smooth the trajectories while transforming the visual data into real world useful traffic information data. For the mixed city environment (i.e., video data provided by TU Delft), comparing the automatically extracted data with the manually extracted data led to a detection accuracy of 100% for pedestrians and mopeds (i.e., 100% of the pedestrians are detected) and 91.11% for cyclists. The average relative error speed is 0.07 for cyclists and 0.03 for pedestrians. According to the fundamental diagrams estimated based on the trajectory data, the free flow speed is around 5 m/s, and the jam density is 1.94 pedestrian-cyclist-moped/m2. The maximum reported flow at the study zone is 2.42 pedestrian-cyclist-moped/s-m. These measures are plausible and comparable to those reported in the literature. On the other hand, for the freeway environment (i.e., video data provided by KAIST), the average travel time for the monitored congested lanes (with a length of 100 m) is 14.15 s and for non-congested lanes is 3.94 s. The maximum recorded density is 95 veh/km-lane; the maximum capacity is 2387 veh/h/ln; and the maximum speed is 125 km/h. The results obtained are consistent with traffic metrics recorded in the literature and the corresponding fundamental diagram shape is consistent with the different traffic flow regions recorded.
- All rights reserved
Notice to Authors
If you are the author of this work and you have any questions about the information on this page, please use the Contact form to get in touch with us.