Pedestrian Crossing Direction Prediction at Intersections for Pedestrian Safety

Pedestrians are among the most vulnerable road users, with significant risks arising at intersections due to potential conflicts with vehicular traffic. With advancements in deep learning, many approaches have sought to address this issue by understanding pedestrian intentions based on their behaviors. However, at intersections, a pedestrian's crash risk is highly influenced by their crossing direction, yet no prior work has specifically focused on predicting crossing directions, leaving a critical gap in ensuring pedestrian safety in these high-risk areas. This study introduces a novel camera-invariant framework for predicting pedestrian crossing directions at intersections using keypoints and trajectory data extracted from CCTV footage. To address challenges posed by varying intersection geometries and camera perspectives, we developed a global coordinate system that standardizes spatial features. The framework leverages Transformer-based models, Graph Convolutional Networks (GCNs), and a hybrid Transformer+GCN approach to extract spatial and temporal features from the pedestrian behaviors. Our proposed framework with Transformer-based model achieved accuracy of 94.10% and F1-Score of 92.35%, demonstrating its effectiveness in capturing pedestrian intentions across diverse scenarios. With its demonstrated high accuracy and effectiveness, our framework predicts pedestrians' crossing directions, which not only can enhance pedestrian safety by reducing potential conflicts with vehicular traffic at intersections but also can be applied to signal optimization systems to improve traffic flow. Furthermore, the geometric-invariant model ensures that the system is easily transferable across intersections. It reduces the burden of collecting more data and training it from several camera perspectives. The implementation of this work is publicly available at https://github.com/Kimyounggun99/CrossingDirectionPrediction.