Precise speed monitoring is vital for law enforcement, crash reduction, and roadway safety. This study applies deep learning to extract precise speed profiles of pedestrians and vehicles to evaluate Surrogate Safety Measures (SSMs) from dynamic (video) data. The YOLOv11, combined with Deep SORT, was implemented to detect and track pedestrians and vehicles. To enhance detection accuracy, this study refined object identification and implemented class filtration within the framework. Additionally, an Occlusion Handling and Re -Identification Module was integrated to mitigate misclassification and spatial overlap, ensuring high predictive reliability. The framework demonstrates high reliability, achieving an accuracy of 0.951 and a precision of 0.956 across all categories. These results indicate that the integration of YOLOv11 and Deep SORT is a robust method for extracting trajectories and speeds, even under occlusion. The overall estimation accuracy is notably promising, with the Mean-Absolute-Error (MAE) and Root-Mean-Square-Error (RMSE) recorded at 0.71 m/s and 1.7 m/s, respectively, for pedestrian velocity. Two SSMs, Time to Collision (TTC) and Post-Encroachment Time (PET), were calculated to assess intersection safety. Analysis of 30 observations revealed 5 hazardous situations based on TTC and 8 based on PET thresholds. The outcomes offer a more accurate computation of TTC and PET values, enhancing clarity on the time required for a vehicle or pedestrian to stop or avoid a collision and the recovery duration following an evasive action or a pedestrian-vehicle conflict. This approach provides a comprehensive toolkit for safety engineers to identify high-risk zones and complex interaction patterns that traditional methods fail to capture. • Integrating occlusion handling and Re -ID into the YOLO+Deep SORT framework. • Detection accuracy and precision exceeded 0.9 for all categories. • TTC and PET analyses identified 5 and 8 hazardous situations, respectively. • Classifying SSM thresholds quantified pedestrian-vehicle interaction severity.
Parvez et al. (Wed,) studied this question.