Scholarlink

<p>Crowd counting plays a crucial role in analyzing and understanding crowd behavior. Existing models generally rely on large parameters to achieve high counting accuracy. This increases computational demands and limits deployment on mobile edge devices. On the other hand, lightweight networks often face difficulties in managing scale variation and show poor performance in complex crowd counting tasks because of their simplified design. To tackle these challenges, we propose a crowd counting model, termed&#xa0;Distillation Hierarchical Mixture of Experts&#xa0;(DHMoE). It is composed of two primary components. The first is a knowledge distillation training model. It transfers fine-grained knowledge from the pre-trained teacher model to the lightweight student model and improves counting accuracy. Second, to solve the problems of scale variation and complex environments, a hierarchical mixture of experts&#xa0;(HMoE) is proposed. The four stages of the student model are organized into four experts, where each network handles crowd features at a different scale. This approach effectively addresses scale variation and improves counting accuracy in diverse environments. Experimental results on four crowd and four vehicle datasets demonstrate that the proposed DHMoE achieves excellent counting accuracy while maintaining a lightweight design.The code is available at <a href="https://github.com/sdut-jacheng/DHMoE">https://github.com/sdut-jacheng/DHMoE</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

Towards trustworthy crowd counting by distillation hierarchical mixture of experts for edge-based cluster computing

  • 期刊论文

  • 发表日期：

    2025年8月14日

查看全文
  • Jing-an Cheng,
  • Qilei Li,
  • Alireza Souri,
  • Xiang Lei,
  • Chen Zhang,
  • Mingliang Gao

摘要

Crowd counting plays a crucial role in analyzing and understanding crowd behavior. Existing models generally rely on large parameters to achieve high counting accuracy. This increases computational demands and limits deployment on mobile edge devices. On the other hand, lightweight networks often face difficulties in managing scale variation and show poor performance in complex crowd counting tasks because of their simplified design. To tackle these challenges, we propose a crowd counting model, termed Distillation Hierarchical Mixture of Experts (DHMoE). It is composed of two primary components. The first is a knowledge distillation training model. It transfers fine-grained knowledge from the pre-trained teacher model to the lightweight student model and improves counting accuracy. Second, to solve the problems of scale variation and complex environments, a hierarchical mixture of experts (HMoE) is proposed. The four stages of the student model are organized into four experts, where each network handles crowd features at a different scale. This approach effectively addresses scale variation and improves counting accuracy in diverse environments. Experimental results on four crowd and four vehicle datasets demonstrate that the proposed DHMoE achieves excellent counting accuracy while maintaining a lightweight design.The code is available at https://github.com/sdut-jacheng/DHMoE.