Scholarlink

<p>Current automated piglet counting systems struggle in modern farming environments. This is due to piglets’ rapid movement, highly similar appearance, and tendency to crowd, pile up, and climb over each other. This study proposes a vision-based counting framework specifically optimized for high-density piglet settings. Our core innovations include three key elements. First, we enhance the recall rate of YOLOv8 detection by using a reweighted loss function. This improves detection completeness, laying a critical foundation for subsequent tracking. Second, we design a small trajectory fusion (STF) algorithm. STF merges short, fragmented trajectories into longer ones if they are spatially and temporally adjacent. Third, we replace the traditional single-line counting method with a multi-line counting module (MLCM). This increases the robustness of the count. By emphasizing operational continuity, our system reduces reliance on appearance-based re-identification. This is particularly important given the piglets’ highly similar looks. Testing on over 29&#xa0;h of annotated video data demonstrates strong performance. Our system achieves a counting error rate (CER) of 0.5%. This represents a 3.3<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\times \)</EquationSource> <EquationSource Format="MATHML"><math> <mo>×</mo> </math></EquationSource> </InlineEquation> improvement over DeepSORT (1.746% CER) and a 20.1<InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(\times \)</EquationSource> <EquationSource Format="MATHML"><math> <mo>×</mo> </math></EquationSource> </InlineEquation> improvement over StrongSORT (10.658% CER). Notably, the re-identification modules of both baseline trackers received basic training on piglet datasets. Ablation studies confirm the effect of our components. The integrated approach reduces the error rate by 54.9% (from 1.1% to 0.5%). The related code is available at <a href="https://github.com/mengjiexu/PigletCounting_STF_MLCM">https://github.com/mengjiexu/PigletCounting_STF_MLCM</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

Robust piglet counting in crowded environments via small trajectory fusion and multi-line counting module

  • 期刊论文

  • 发表日期：

    2025年12月5日

查看全文
  • Mengjie Xu,
  • Lei Dai,
  • Hailing Wu,
  • Zhihua Chen

摘要

Current automated piglet counting systems struggle in modern farming environments. This is due to piglets’ rapid movement, highly similar appearance, and tendency to crowd, pile up, and climb over each other. This study proposes a vision-based counting framework specifically optimized for high-density piglet settings. Our core innovations include three key elements. First, we enhance the recall rate of YOLOv8 detection by using a reweighted loss function. This improves detection completeness, laying a critical foundation for subsequent tracking. Second, we design a small trajectory fusion (STF) algorithm. STF merges short, fragmented trajectories into longer ones if they are spatially and temporally adjacent. Third, we replace the traditional single-line counting method with a multi-line counting module (MLCM). This increases the robustness of the count. By emphasizing operational continuity, our system reduces reliance on appearance-based re-identification. This is particularly important given the piglets’ highly similar looks. Testing on over 29 h of annotated video data demonstrates strong performance. Our system achieves a counting error rate (CER) of 0.5%. This represents a 3.3 \(\times \) × improvement over DeepSORT (1.746% CER) and a 20.1 \(\times \) × improvement over StrongSORT (10.658% CER). Notably, the re-identification modules of both baseline trackers received basic training on piglet datasets. Ablation studies confirm the effect of our components. The integrated approach reduces the error rate by 54.9% (from 1.1% to 0.5%). The related code is available at https://github.com/mengjiexu/PigletCounting_STF_MLCM.