Scholarlink

<p>In the field of pedestrian trajectory prediction, existing models often rely on simplistic spatiotemporal modeling techniques, which frequently fail to adequately capture complex social interactions and temporal dependencies. Common issues among most models include the neglect of local temporal information, the inability to effectively model complex interactions in dynamic environments, and performance degradation in multi-step predictions. To address these challenges, this paper proposes a spatiotemporal graph convolutional network (ST-GCN) framework enhanced by contrastive learning and a temporal attention mechanism. The model aims to improve prediction accuracy by strengthening the learning of spatiotemporal features. Specifically, a graph contrastive learning module is introduced to enhance the model’s ability to distinguish latent spatiotemporal patterns through contrastive learning strategies. Meanwhile, a temporal attention mechanism is designed to allow the model to focus on crucial time steps, thereby improving its responsiveness to dynamic changes. In addition, the model integrates multi-layer ST-GCNs with temporal convolutional networks to simultaneously extract spatial relationships and capture temporal dependencies. The main innovation lies in the combination of contrastive learning and temporal attention, which not only reinforces the spatiotemporal modeling capabilities but also improves the adaptability and generalization of the model in complex scenarios. Experimental results on the ETH and UCY datasets demonstrate that the proposed model achieves significant improvements over traditional baseline models, achieving an average displacement error of 0.39 and a final displacement error of 0.68, indicating superior performance in predicting pedestrian trajectories across various complex scenarios.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Integrating contrastive representation learning and temporal attention into spatiotemporal graph convolution for pedestrian motion prediction

  • 期刊论文

  • 发表日期：

    2026年5月25日

查看全文
  • Shengrui Liu,
  • Xiangyu Li,
  • Jiaxuan Zhao,
  • Xiaolan Yang,
  • Chunhong Yuan,
  • Lei Ren

摘要

In the field of pedestrian trajectory prediction, existing models often rely on simplistic spatiotemporal modeling techniques, which frequently fail to adequately capture complex social interactions and temporal dependencies. Common issues among most models include the neglect of local temporal information, the inability to effectively model complex interactions in dynamic environments, and performance degradation in multi-step predictions. To address these challenges, this paper proposes a spatiotemporal graph convolutional network (ST-GCN) framework enhanced by contrastive learning and a temporal attention mechanism. The model aims to improve prediction accuracy by strengthening the learning of spatiotemporal features. Specifically, a graph contrastive learning module is introduced to enhance the model’s ability to distinguish latent spatiotemporal patterns through contrastive learning strategies. Meanwhile, a temporal attention mechanism is designed to allow the model to focus on crucial time steps, thereby improving its responsiveness to dynamic changes. In addition, the model integrates multi-layer ST-GCNs with temporal convolutional networks to simultaneously extract spatial relationships and capture temporal dependencies. The main innovation lies in the combination of contrastive learning and temporal attention, which not only reinforces the spatiotemporal modeling capabilities but also improves the adaptability and generalization of the model in complex scenarios. Experimental results on the ETH and UCY datasets demonstrate that the proposed model achieves significant improvements over traditional baseline models, achieving an average displacement error of 0.39 and a final displacement error of 0.68, indicating superior performance in predicting pedestrian trajectories across various complex scenarios.