Abstract
Human trajectory prediction is inherently challenging due to the stochastic and multi-modal nature of pedestrian movement. We propose DifTraj, a novel diffusion-based framework that bridges the gap between extrinsic social interactions and intrinsic destination-based intentions. By explicitly modeling these two factors and utilizing a "leapfrog" initialization for the diffusion process, DifTraj achieves significant performance gains, including a 19.0% improvement in accuracy on standard benchmarks like ETH-UCY. Our approach ensures that predicted paths are not only socially plausible but also aligned with the latent goals of the agents.
Introduction
Predicting where a pedestrian will walk requires understanding both how people move (socially aware navigation) and why they move (goal-oriented behavior). Existing models often focus heavily on extrinsic interactions—such as avoiding collisions—while ignoring the intrinsic intentions that drive movement. Furthermore, while diffusion models offer powerful generative capabilities, they often suffer from slow convergence and sensitivity to initialization. DifTraj addresses these limitations by integrating explicit intention modeling with a refined, feature-guided diffusion process.
Methodology
DifTraj employs an end-to-end architecture that fuses two primary feature sets to guide a denoising diffusion process: extrinsic interaction features and intrinsic intention features.
Extrinsic Interaction and Intrinsic Intention
The model captures temporal and spatial dependencies using LSTMs with Time-Attention and Graph Attention Networks (GAT). This ensures that predicted paths are "socially acceptable." Simultaneously, it predicts the likely destination point using a transformer-based encoder. This predicted destination acts as a "north star," providing a clear signal for the person's latent intent.
Diffusion with Leapfrog Initialization
Instead of starting the diffusion process from pure Gaussian noise, DifTraj uses the fused intention and interaction features to estimate an initial "noised trajectory distribution." This "leapfrog" approach maintains the generative power of diffusion while significantly improving inference speed and precision. Additionally, a Sample Consistency Loss ($L_{sc}$) is introduced to constrain the variance between multiple generated samples, forcing them to remain high-quality and close to the ground truth.
Experimental Evaluation
The framework was evaluated on two gold-standard benchmarks: ETH-UCY and the Stanford Drone Dataset (SDD).
Quantitative Performance
DifTraj sets a new state-of-the-art across multiple metrics:
- ETH-UCY: Achieved a 19.0% improvement in Average Displacement Error (ADE) and a 24.2% improvement in Final Displacement Error (FDE) compared to the previous best model (LED).
- SDD: Outperformed existing methods by 5.2% in ADE and 8.3% in FDE.
Qualitative Analysis
Visualizations across complex scenes, such as the "UNIV" dataset where people merge into groups or avoid obstacles, show that DifTraj produces paths significantly closer to the actual ground truth compared to previous diffusion-based methods.
Conclusion and Future Work
DifTraj demonstrates that combining explicit intention and interaction modeling within a diffusion framework significantly enhances trajectory prediction accuracy. While the model handles human-to-human interactions effectively, future work could better incorporate static environmental constraints, such as walls or crosswalks, directly into the intention module to further refine destination estimation.
Summary updated