Pedestrian trajectory prediction based on social force-guided soft attention graph convolutional network
Year: 2026
Type: article
Abstract: The accuracy of pedestrian trajectory prediction is crucial for enhancing traffic management, refining autonomous driving systems, and bolstering public safety. To accurately predict the future movements of pedestrians, it is imperative to account for the complexity of their interactions with the environment, while individual intentions are also indispensable. To address these critical limitations, this paper presents a novel pedestrian trajectory prediction framework (SF-SAGCN) incorporating a dynamic adjustment method for pedestrian interaction soft attention weights and a pedestrian-environment interaction confidence factor. This framework transforms the purely data-driven soft attention mechanism into a dynamic, intent-aware topological representation guided by the Social Force Model. This deeply coupled architecture is designed to mitigate the risk of the model becoming trapped in short-sighted local optima due to an over-reliance on instantaneous speed and direction. Specifically, the interaction soft attention weights are dynamically rectified by coupling the history-derived desired-motion driving force with multi-pedestrian interaction forces. Concurrently, a pedestrian-environment interaction confidence factor enriches the semantic representation of spatial scene nodes. Finally, Temporal Convolutional Networks are employed to generate pedestrian movement prediction trajectories. Extensive tests on public datasets demonstrate a clear quantitative advantage over previous methods, and ablation and leave-one-scene-out experiments further demonstrate the model’s component effectiveness and cross-scene generalization capability.
Source: Complex & Intelligent Systems
Authors: Li Hui, Luo Mingyue, Luo Wanbo, Li Hewei, Li Jianan, Qin Wei
Language: English
Subfield: Safety Research
Field: Social Sciences
Domain: Social Sciences
Open Access status: gold
Estimated APC paid: $2,490
OpenAlex ID: W7236200081
