<p>Accurate pedestrian trajectory prediction is a crucial capability for advanced autonomous driving systems. In free-walking environments, trajectories are often influenced by social interactions among individuals, and effectively capturing these interactions can significantly enhance prediction accuracy. While some methods achieve strong results using complex neural networks, they often struggle to filter out task-irrelevant information, leading to suboptimal learning of interaction dynamics. To alleviate this, we propose a novel dual-path framework that integrates dynamic interactions with individual global movement trends. At each observed time step, dynamic interactions are modeled alongside a rule-based pruning strategy that refines interaction relationships, allowing the model to focus on task-relevant information. The main contributions of this study are twofold: (1) a dual-path framework that seamlessly combines interaction dynamics with global motion information through kinematic feature encoding, and (2) a temporal second-order pruning strategy for constructing dynamic interaction graphs, enabling the model to capture evolving interaction patterns more effectively. Experimental results demonstrate that the proposed method achieves competitive performance compared to baseline approaches, with further comparative studies validating its ability to enhance both interaction modeling and trajectory prediction accuracy.</p>