Scholarlink

<p>Multi-person motion forecasting is challenging due to the complex relationship between individual autonomy and social interactions. Existing methods model joint movements assuming universal social engagement, neglecting autonomous individuals and causing noise or errors in independent mover predictions. To solve this issue, this paper proposes a novel fine-grained behavior interaction-aware network (FBINet) to learn spatial-temporal and effective social relationships for efficient multi-person motion forecasting. We claim that foreknowledge of people’s potential interactions with other individuals is paramount for the efficacious prediction of subsequent behaviors. Consequently, we design an interaction perceptron based on human dynamics to classify individuals into two categories: independent individuals and social groups, depending on the presence of interactive behaviors. Based on the foreknowledge of people’s potential interactions with other individuals, we then develop two analytical modules: autonomous individual analysis module to model the long-term spatio-temporal dependencies of independent individuals; multi-person interaction dependency parsing module to capture the interactions among social groups and effectively integrate them into individual joint movements. Experiments demonstrate that our approach outperforms the latest multi-person prediction models, achieving an average improvement of 18.9% in JPE, 23.7% in APE, and 28.4% in FDE on the CMU Mocap (UMPM) 3-person dataset. Code is available at <a href="https://github.com/Xinyi-L205/FBINet">https://github.com/Xinyi-L205/FBINet</a>.</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

Fine-grained behavior interaction-aware network for efficient multi-person motion forecasting

  • 期刊论文

  • 发表日期：

    2026年2月3日

查看全文
  • Wenming Cao,
  • Xinyi Liu,
  • Jianqi Zhong

摘要

Multi-person motion forecasting is challenging due to the complex relationship between individual autonomy and social interactions. Existing methods model joint movements assuming universal social engagement, neglecting autonomous individuals and causing noise or errors in independent mover predictions. To solve this issue, this paper proposes a novel fine-grained behavior interaction-aware network (FBINet) to learn spatial-temporal and effective social relationships for efficient multi-person motion forecasting. We claim that foreknowledge of people’s potential interactions with other individuals is paramount for the efficacious prediction of subsequent behaviors. Consequently, we design an interaction perceptron based on human dynamics to classify individuals into two categories: independent individuals and social groups, depending on the presence of interactive behaviors. Based on the foreknowledge of people’s potential interactions with other individuals, we then develop two analytical modules: autonomous individual analysis module to model the long-term spatio-temporal dependencies of independent individuals; multi-person interaction dependency parsing module to capture the interactions among social groups and effectively integrate them into individual joint movements. Experiments demonstrate that our approach outperforms the latest multi-person prediction models, achieving an average improvement of 18.9% in JPE, 23.7% in APE, and 28.4% in FDE on the CMU Mocap (UMPM) 3-person dataset. Code is available at https://github.com/Xinyi-L205/FBINet.