STeF-LSTM: A Hybrid Framework Integrating Periodic and Sequential Modeling for Mapping Motion Patterns

Human awareness is crucial for autonomous robotic systems to safely and efficiently interact with people in shared environments. However, reliably modeling human movements over extended periods poses substantial challenges due to human behavior’s inherent complexity. This paper introduces STeF-LSTM, a novel hybrid predictive framework that integrates the strengths of Frequency Map Enhancement (FreMEn) for periodic pattern modeling and Long Short-Term Memory (LSTM) networks for capturing sequential dependencies. By combining FreMEn’s frequency-domain representation to encode long-term rhythmic patterns and LSTM’s robust short-term sequential learning capabilities, the proposed method addresses the limitations of each standalone approach. Evaluations with real-world pedestrian datasets demonstrate significant improvements in prediction accuracy and generalization. STeF-LSTM reduces pattern prediction divergence error by approximately 50%. Further, if the model is applied for motion prediction, the displacement errors decrease by around 7% compared to state-of-the-art methods like CLiFF-LHMP. These results highlight the value of integrating periodic and sequential modeling for robust long-term human motion pattern forecasting, improving navigation, planning, and collaboration in autonomous robotic deployments.