AUTOMATED TRAFFIC SAFETY ASSESSMENT TOOL UTILIZING MONOCULAR 3-D DETECTION ALGORITHM

Traffic conflict analysis has gained significant attention over the last few years. Recent research has made significant progress in the development of detection and tracking systems by utilizing various technologies such as closed-circuit televisions, unmanned aerial vehicles, and sensor fusion exploiting light detection and ranging. By examining the challenges hinder the widespread adoption of traffic conflicts analysis using artificial intelligence in safety assessment techniques, it is found that the limitations are broken into three categories: monitoring device, detection algorithm, and outputs processing related limitations. In a trial to overcome these limitations, this study utilized Convolutional Neural Network (CNN) based algorithm, that is essentially developed for autonomous vehicle systems, to detect and track road users at signalized intersections. The primary advantage of this algorithm lies in its ability to depict road users through the utilization of three-dimensional bounding cuboids. In an effort to refine the precision of the tracking process, this study further incorporates a suite of post-processing algorithms. These are specifically designed for data augmentation, the reconstruction of seamless trajectories, and the accurate estimation of vehicle states and indicators. This representation method simplifies the entire traffic data processing procedure and improves the detection for better establishment of an automated safety assessment tool utilizing Artificial Intelligence. The algorithm has been examined on various video footages and proved its detection superiority for high traffic volumes at intersections with an overall precision of 95.08%, and a recall of 92.81%. It exhibited exceptional accuracy and efficacy in identifying a range of crucial conflict scenarios, achieved through the application of four unique traffic conflict indicators.