Abstract
Pedestrian trajectory prediction is a critical task for autonomous systems, requiring an understanding of complex social interactions and temporal dynamics. Traditional methods often decouple spatial and temporal modeling, leading to information dilution and a failure to capture cascading interaction effects. We present UniEdge, a framework that unifies spatial and temporal modeling into a single patch-based graph structure. By treating social interactions and temporal dependencies simultaneously and modeling edge-to-edge influence, UniEdge achieves state-of-the-art performance on major benchmarks, including ETH, UCY, and the Stanford Drone Dataset.
Introduction
Predicting human movement is challenging because pedestrians constantly adjust their paths in response to others. Most current models address this by first analyzing spatial interactions (proximity in a single frame) and then temporal trends (individual history). However, this decoupled approach suffers from two major flaws: information dilution, where the interaction signal weakens across multiple steps, and the neglect of "cascading" effects—how the relationship between two pedestrians influences a third party over time.
Methods
UniEdge introduces three primary innovations to overcome these limitations:
- Unified ST Graph (Patch-based): Instead of frame-by-frame processing, trajectories are segmented into overlapping patches. Within each patch, spatial and temporal connections are modeled as first-order relationships in a single graph. This "short-circuits" the path information must travel, reducing effective resistance and improving sensitivity to long-range social cues.
- E2E-N2N-GCN (Dual-Graph Convolution): UniEdge employs a two-pronged convolution strategy. While a standard graph handles Node-to-Node (N2N) social interactions, a secondary "edge graph" models Edge-to-Edge (E2E) influence. This captures how interaction patterns themselves propagate through a crowd. This is implemented using Hodge-Laplacian Laguerre Convolution (HLLConv) for spectral filtering.
- Transformer Encoder Predictor: To avoid the limitations of auto-regressive (step-by-step) prediction, UniEdge uses a Transformer encoder with learnable placeholders for future positions. This allows the model to refine the entire predicted trajectory globally by attending to the full historical context.
Experimental Evaluation
The model was evaluated on three standard benchmarks: ETH (subsets ETH, HOTEL), UCY (UNIV, ZARA1, ZARA2), and the Stanford Drone Dataset (SDD). These datasets provide a variety of real-world scenarios, from sparse pedestrian environments to complex university campus interactions.
Results and Discussion
UniEdge consistently outperformed previous state-of-the-art (SOTA) methods in both Average Displacement Error (ADE) and Final Displacement Error (FDE). On the ETH/UCY datasets, it achieved a 5.3% improvement in average ADE and a 24.2% improvement in average FDE compared to MSRL. On the SDD, it outperformed HighGraph by 5.9% in ADE.
Qualitative analysis reveals that UniEdge is particularly effective at predicting non-linear maneuvers, such as collision avoidance and gradual convergence between individuals. By treating edges as "first-class citizens," the model captures the "cascade effects" where one person's reaction triggers a chain of movements in others.
Conclusion
UniEdge demonstrates the power of unified spatial-temporal modeling and edge-enhanced graph convolutions for trajectory prediction. While the model shows some sensitivity to crowd density—struggling slightly in very sparse scenes like ETH—future work on adaptive patch segmentation and the integration of environmental context (e.g., sidewalk boundaries) promises to further enhance its predictive capabilities.
Summary updated