An online adaptive learning approach for predicting multi-type traffic participants' microscopic behavior

Predicting multi-type traffic participant behavior in dynamic transportation hubs remains challenging. Existing deep-learning frameworks lack robust online learning and autonomous error-correction, limiting their reliability under evolving conditions. Most state-of-the-art models exhibit poor cross-scenario generalization, requiring intensive retraining for new settings. Prediction errors also propagate over time due to absent real-time correction. To address these limitations, this paper introduces an adaptive framework integrating online learning with probabilistic error correction. Key innovations include: (1) an Extended Kalman Filter-based module for real-time trajectory correction; (2) a hierarchical graph encoder enabling transfer learning with minimal retraining; and (3) unified node-edge-plane modeling for multimodal context fusion. Validated using real-world hub data and redesigned benchmark experiments, our method significantly outperforms existing approaches in unseen scenarios, positioning it as a promising solution for real-time behavioral prediction in modern traffic systems.