Dual-Branch Architecture Model Enhanced by Wavelet Transform Network for Progressive Pedestrian Trajectory Prediction

Pedestrian trajectory prediction seeks to forecast future positions by jointly modeling time-domain interactions and frequency-domain dynamics. Prior work often focuses on localized temporal dependencies while ignoring spectral cues and global context. We introduce a dual-branch wavelet–attention framework that addresses these limitations: (1) The first branch incorporates a wavelet transform network into a Self-Attention-based Transformer to capture local features by learning the complex dependencies of input trajectories at different spatial locations. Wavelet-Enhanced Self-Attention Branch decomposes observed trajectories using Discrete Wavelet Transform into low-frequency approximation coefficients and multi-level detail coefficients. Each coefficient set is embedded and attended independently via causal multi-head self-attention. (2) Global Attention Mechanism (GAM) Branch applies channel-wise and spatial-wise attention to refine embeddings. We employ an adaptive gating fusion to balances local spectral and global contextual features. A three-stage progressive training schedule (short-term, destination, full trajectory) further enhances convergence. Through experimental validation on four commonly used public datasets, our approach surpasses existing state-of-the-art methods in terms of prediction accuracy.