Scholarlink

<p>Pedestrian trajectory prediction is critical in fields such as autonomous driving and robot navigation. It enables autonomous vehicles and robots to make proactive decisions, avoid collisions with pedestrians, and ensure pedestrian safety. Some methods utilize Generative Adversarial Networks (GAN) for trajectory prediction. However, there are still two major drawbacks that hinder the improvement of trajectory prediction accuracy. 1) The complex interactions among multiple pieces of information have not been fully captured, preventing the information from being fully utilized. 2) The uncertainty of network outputs has not been fully modeled and processed. To address these challenges and enhance the model’s adaptability to unknown data, especially to provide an effective solution for managing complex tasks, a novel multi-feature fusion trajectory prediction method is introduced. Firstly, this method designs a new feature fusion submodule. It considers the correlation between multi-feature information through the Power Average (P-A) operator. It also enables the model to better learn the differences among multiple feature information and effectively capture the complex interactions among them. Then, the uncertainty of the network’s output is modeled by incorporating a conservative output state through triangular fuzzy numbers, enhancing the flexibility of the network. Evidence theory is crucial in practical applications due to its effectiveness in representing and processing uncertain information. Therefore, we adopt the belief measure of evidence theory as the uncertainty output of the network. Finally, considering that evidence distance can accurately assess the differences between uncertain information, we embed the evidence distance formula into the loss function of the model as a loss term. This loss term is used to address the discrepancy between the model’s predicted output and the actual distribution, thereby optimizing the information loss of the model and enhancing prediction accuracy. Experiments conducted on the public ETH/UCY datasets reveal that the proposed method achieves higher accuracy, particularly for trajectory prediction in complex situations. Additionally, we have not significantly reduced the model’s efficiency.</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

A new multi-feature fusion trajectory prediction method

  • 期刊论文

  • 发表日期：

    2025年9月9日

查看全文
  • Tian Yang,
  • Gang Wang,
  • Jian Lai,
  • Yang Wang

摘要

Pedestrian trajectory prediction is critical in fields such as autonomous driving and robot navigation. It enables autonomous vehicles and robots to make proactive decisions, avoid collisions with pedestrians, and ensure pedestrian safety. Some methods utilize Generative Adversarial Networks (GAN) for trajectory prediction. However, there are still two major drawbacks that hinder the improvement of trajectory prediction accuracy. 1) The complex interactions among multiple pieces of information have not been fully captured, preventing the information from being fully utilized. 2) The uncertainty of network outputs has not been fully modeled and processed. To address these challenges and enhance the model’s adaptability to unknown data, especially to provide an effective solution for managing complex tasks, a novel multi-feature fusion trajectory prediction method is introduced. Firstly, this method designs a new feature fusion submodule. It considers the correlation between multi-feature information through the Power Average (P-A) operator. It also enables the model to better learn the differences among multiple feature information and effectively capture the complex interactions among them. Then, the uncertainty of the network’s output is modeled by incorporating a conservative output state through triangular fuzzy numbers, enhancing the flexibility of the network. Evidence theory is crucial in practical applications due to its effectiveness in representing and processing uncertain information. Therefore, we adopt the belief measure of evidence theory as the uncertainty output of the network. Finally, considering that evidence distance can accurately assess the differences between uncertain information, we embed the evidence distance formula into the loss function of the model as a loss term. This loss term is used to address the discrepancy between the model’s predicted output and the actual distribution, thereby optimizing the information loss of the model and enhancing prediction accuracy. Experiments conducted on the public ETH/UCY datasets reveal that the proposed method achieves higher accuracy, particularly for trajectory prediction in complex situations. Additionally, we have not significantly reduced the model’s efficiency.