Scholarlink

<p>Pedestrian trajectory prediction plays a crucial role in enabling fully autonomous driving in urban environments, this paper proposes a GRU-Transformer network with multi-information integration (MGFormer) to predict pedestrian trajectories from a driving perspective. Firstly, MGFormer integrates pose and optical flow information with observed trajectory information. It uses skeleton sequence reorganization and local optical flow division techniques to reduce information distortion caused by occlusion. Subsequently, a cross-information fusion attention mechanism based on information evaluation is introduced. This mechanism comprehensively considers the importance of various information sources and the relevance of different features within each information source. Finally, MGFormer utilizes an innovative trajectory decoding network that combines GRU and Transformer models to enhance the effectiveness of decoding and predicting fused features. On the JAAD and PIE datasets, this method outperforms other approaches in predicting the displacement error within a 1.5 second, while also significantly reducing inference time compared to the latest model.</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

MGFormer: a multi-information-based GRU-transformer network for pedestrian trajectory prediction

  • 期刊论文

  • 发表日期：

    2025年5月3日

查看全文
  • Quankai Liu,
  • Haifeng Sang,
  • Jinyu Wang,
  • Wangxing Chen

摘要

Pedestrian trajectory prediction plays a crucial role in enabling fully autonomous driving in urban environments, this paper proposes a GRU-Transformer network with multi-information integration (MGFormer) to predict pedestrian trajectories from a driving perspective. Firstly, MGFormer integrates pose and optical flow information with observed trajectory information. It uses skeleton sequence reorganization and local optical flow division techniques to reduce information distortion caused by occlusion. Subsequently, a cross-information fusion attention mechanism based on information evaluation is introduced. This mechanism comprehensively considers the importance of various information sources and the relevance of different features within each information source. Finally, MGFormer utilizes an innovative trajectory decoding network that combines GRU and Transformer models to enhance the effectiveness of decoding and predicting fused features. On the JAAD and PIE datasets, this method outperforms other approaches in predicting the displacement error within a 1.5 second, while also significantly reducing inference time compared to the latest model.