Title: MUSCLE-NET: A Predicted-Multiscale-Aware Network for Pedestrian Trajectory Forecasting

Abstract: Precise forecasting of pedestrian trajectories is a critical requirement for ensuring safety in autonomous driving and intelligent transportation systems. Although recent advancements have yielded significant progress, current methodologies often fail to fully leverage the variety of available observations. Furthermore, they frequently neglect the scale dependency inherent in future motion, applying uniform processing to multiscale features despite the varying dynamics of movement. This oversight compromises robustness across a wide spectrum of pedestrian behaviors. To overcome these limitations, we introduce the Predicted-MUltiSCale-Aware Network (MUSCLE-NET), a framework designed for pedestrian trajectory forecasting that merges complementary multimodal signals with mechanisms for scale-adaptive prediction. The architecture centers on a Multiscale Multimodal Feature Extraction (MMFE) module. This component synthesizes multiscale representations, employs modality-aware recalibration, and utilizes directional cross-modal fusion to generate semantically aligned features derived from bounding boxes, velocities, and pose data. Leveraging these extracted features, a Multiscale Enhanced Hierarchical Prediction (MEHP) module refines future motion predictions through a probabilistic coarse predictor, scale-aligned fusion, and progressive refinement steps. This process adaptively identifies scale-relevant cues to reduce spatial drift. Comprehensive evaluations on the JAAD and PIE benchmarks indicate that MUSCLE-Net delivers competitive results and demonstrates consistent improvements over existing state-of-the-art trajectory prediction techniques.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC