Objective Accurately predicting pedestrian crossing behavior is crucial for ensuring the safety of both vehicles and pedestrians in the field of intelligent driving. Methods A pedestrian crossing behavior prediction model was designed which combined various computer vision technologies. The model accurately assessed pedestrian intentions by analyzing their positions postures actions and environmental features. To enhance the model?? s perception capability for pedestrians at different distances preprocessing with multi-scale magnification and post-processing for data filtering and smoothing were employed. A two-stage method called prediction after condition PAC was proposed to achieve more effective prediction of pedestrian crossings. Results Testing results based on the JAAD dataset indicated that the proposed model achieved an average accuracy of 89. 31% representing an improvement of 8. 76% compared with traditional singlestage methods. Conclusion Further analysis of feature importance shows that after incorporating the road surface area feature the prediction accuracy significantly increases from 68. 43% to 85. 06% highlighting the importance of the relationship between pedestrian location and the road profile in the study of pedestrian crossing behavior. This has significant implications for reducing vehicle-pedestrian collision accidents and enhancing the safety of intelligent driving vehicles.