Predicting Pedestrian Crossing Intentions at Unsignalized Intersections Using Machine Learning and Real-World Trajectories

Predicting pedestrian crossing intentions at unsignalized intersections remains a critical yet underexplored challenge for autonomous vehicles and Advanced Driver Assistance Systems (ADAS). Although prior research on pedestrian intent prediction exists, much of it relies heavily on image-based datasets and computationally intensive deep learning models, limiting their applicability in resource-constrained real-time environments. This study explores whether classical machine learning and lightweight models, using only trajectory and simple contextual features, can achieve competitive performance in predicting pedestrian crossing intentions. Using the inD drone dataset, we transformed bird's eye trajectories into a car-centric perspective, incorporating occlusion, sensing range, and sensing frequency to simulate realistic ADAS conditions. A novel feature, namely the cutting momentum, is introduced to capture pedestrian movement toward the ego vehicle's projected trajectory. Combined with ego velocity and time-to-collision, these features are used to train a Random Forest classifier. The model achieves a mean accuracy of 91.8\% when evaluated at the observation level, and 38\% recall with precision 100\% when evaluated at the event level, demonstrating its potential for deployment on cost-effective ADAS hardware.