Understanding pedestrian behavior in urban traffic has paramount significance in the field of road safety, especially in the cases of intersections where the pedestrian–vehicle interaction is frequent. The objective of this study is to investigate whether vehicle movement features can predict pedestrian non-compliance (illegal crossings) using a LightGBM classifier, which is a gradient boosting framework based on decision trees, known for its efficiency and accuracy in structured data classification tasks, contributing to the development of intelligent crosswalk monitoring systems. To create the dataset used in behavioral classification, a computer vision algorithm with YOLOv8 (You Only Look Once) and ResNet-50 models, with Kalman Filtering and homography transformations mapping image coordinates to ground plane positions, was used. The results indicate that vehicle speed and coordinates are good predictors of pedestrian crossing action, even without pedestrian-side features. The model performed particularly well in identifying legal crossings (high precision and recall), while maintaining moderate recall for illegal events. The absolute value of velocity was the most important feature, meaning that oncoming speed is the most important factor in the probability of non-compliance by pedestrians.
| ID | pc590 |
| Presentation | |
| Full Text | |
| Tags |
