Real-Time Risk Prediction at Signalized Intersection Using Graph Neural Network (06-012)

Project Description: The project concerns with Real Time Risk Prediction at Signalized Intersection Using Graph Neural Network. We have primarily shown how existing infrastructure cameras and computer vision methods can be leveraged to study real time risk at every intersection. We use the kinematics, and relative behavior of each of the actors and use them in a graph neural network framework. However, infrastructure cameras often do not cover the full 360 degree view of the intersection. Hence, we do not gather a full picture of the intersection for risk prediction. Risk prediction related research can be accelerated if we use better bird’s eye view and better quality videos to develop algorithms that can measure risks. These learning can be transferred to infrastructure cameras later. In that effort we have leveraged a new drone dataset that captures four intersection in Virginia providing a detailed picture of the intersection, entry and exit of each vehicle in the intersection. Data Scope: Drone-based trajectory dataset would be useful for analyzing the behavior at various intersections. Traffic forecasting and safety metric calculation are some of the things that can be performed on this kind of dataset. Drone-based data collection allows for precise localization and trajectory extraction of traffic participants. Data Specification: The description of the annotated file is as follows: For the original video, I extracted 1000 frames of images, including 500 on May 31st and 500 on June 15th respectively; Each image corresponds to a CSV file. The format of the CSV file is shown in the right figure：Each row represents the annotation information of a target in the image. Label_id：is equivalent to object id x: The value of the center position of the rectangle on the u-axis of the image y: The value of the center position of the rectangle on the V-axis of the image width & length：The length and width of the rectangle along the longitudinal direction of the target(Sorry, it's actually reversed here) heading: The angle at which the forward direction of the target rotates counterclockwise relative to the u-axis (0-360 degrees) class：Categories of annotated targets 1：pedestrian 2：bicycle/motorcycle 3：tricycle 4：animal 5：car 6：bus 7: Truck/carrier/van(Excluding pickup trucks, we think they are closer to car)