Machine Learning-Based Pedestrian Intention Prediction Models for Collision Warning at Unsignalized Crosswalks
Bildiriler > Tam Metin Bildiri
Yazarlar: Sami Haktan Cangut, Yalçın ALVER
International Conference on Road and Rail Infrastructure, , , Uluslararası, Tam Metin Bildiri, cilt 9, ss 791-798
DOI: 10.5592/co/cetra.2026.1849
Özet
Pedestrian safety remains a critical challenge in urban environments and pedestrians account for approximately 21% of all traffic fatalities in Turkey. A major contributor to this risk is the uncertainty inherent in pedestrian–driver interactions at unsignalized crosswalks where drivers yield in only 29% of observed cases. Misinterpretation of pedestrian intentions in such ambiguous situations can lead to hazardous outcomes. This study addresses this problem by developing and comparing machine learning models that predict pedestrian crossing intention with high reliability. A comprehensive dataset containing 4, 835 pedestrian observations was created by combining the widely used JAAD, PIE, and PSI datasets. For each pedestrian, fifteen consecutive video frames were analyzed to extract kinematic features such as position and acceleration along with social interaction cues derived from nearby pedestrians. Three neural network architectures were evaluated using a two-stage transfer learning procedure. The models included a Bidirectional Long Short-Term Memory network (BiLSTM), a Social-LSTM incorporating social pooling mechanisms, and a Transformer model designed to capture complex temporal dependencies. Performance evaluation demonstrated that the Social-LSTM architecture achieved the highest overall performance with a precision of 85.02% and an F1-score of 84.57% by explicitly modeling the social context. Conversely, the Transformer model recorded an F1-score of 83.44% and suggested that architectural complexity does not guarantee superior performance when training data is limited. The BiLSTM provided a strong, kinematic-focused baseline with an F1-score of 84.06%. These findings confirm that sequential machine learning models can effectively anticipate pedestrian crossing intentions. The results highlight that incorporating social behavioral theories into engineering solutions is essential for developing robust collision warning systems aimed at reducing pedestrian–vehicle conflicts.