Scholarlink

Background <p>Efficient monitoring of sow nursing behavior is relevant to animal welfare monitoring and management of preweaning mortality (PWM) risk, which remains a significant challenge in intensive swine production systems. Traditional observation methods are labor-intensive and inadequate for large-scale farms, requiring automated solutions. Precision livestock farming (PLF) technologies, particularly computer vision and deep learning models, offer opportunities for continuous monitoring of sow posture–nursing-state classes. The YOLO (you only look once) architecture has demonstrated high accuracy and speed for object detection in different environments. This study developed and evaluated a modified YOLO11n model optimized with TensorRT to monitor sow posture–nursing-state classes in farrowing crates. In this study, “nursing” was operationalized from top-view images as visible piglet snout/mouth contact oriented toward the udder/teat line; frames without visually confirmable contact were labeled not nursing. The model’s lightweight architecture and inference acceleration address the computational constraints of real-time applications in farrowing-room monitoring.</p> Results <p>The modified YOLO11n achieved an mAP@50 of 98.90%, while reducing complexity to 207 layers and 5.0 GFLOPs by removing the small-object detection head. This enabled inference times of 4.6 ms on an NVIDIA A100 GPU and 6.1 ms on a T4 GPU with TensorRT optimization. The framework detected and classified ten posture–nursing-state classes (sitting, standing, and three lying postures, each labeled as nursing or not nursing) on representative test images, including frames with partial occlusion and variable lighting. Misclassifications were observed primarily among visually similar classes, such as Sow_Sitting_Nursing and Sow_Sitting_Not_Nursing, consistent with limited visibility of piglet-to-teat contact in some postures from the top-view angle. TensorRT optimization further reduced inference latency by 58.56% (A100) and 44.55% (T4), demonstrating consistent inference-latency reductions on the evaluated GPU platforms.</p> Conclusions <p>This study demonstrates the potential of leveraging lightweight deep learning architectures for real-time behavioral monitoring in PLF. The modified YOLO11n and TensorRT optimization provide an efficient framework for automated monitoring of sow posture–nursing-state classes, supporting welfare-oriented monitoring workflows relevant to PWM risk management. Future work will evaluate lightweight temporal post-processing and embedded deployment to reduce frame-to-frame label flicker and assess performance under on-farm compute and I/O constraints.</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

Accelerating sow nursing behavior monitoring with modified YOLO11n architecture and TensorRT integration

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年4月22日

查看全文
  • Mamunur Rahman,
  • Victor Hugo Silva Souza,
  • Tami M. Brown-Brandl,
  • Gary A. Rohrer,
  • Yeyin Shi,
  • Isabella C. F. S. Condotta

摘要

Background

Efficient monitoring of sow nursing behavior is relevant to animal welfare monitoring and management of preweaning mortality (PWM) risk, which remains a significant challenge in intensive swine production systems. Traditional observation methods are labor-intensive and inadequate for large-scale farms, requiring automated solutions. Precision livestock farming (PLF) technologies, particularly computer vision and deep learning models, offer opportunities for continuous monitoring of sow posture–nursing-state classes. The YOLO (you only look once) architecture has demonstrated high accuracy and speed for object detection in different environments. This study developed and evaluated a modified YOLO11n model optimized with TensorRT to monitor sow posture–nursing-state classes in farrowing crates. In this study, “nursing” was operationalized from top-view images as visible piglet snout/mouth contact oriented toward the udder/teat line; frames without visually confirmable contact were labeled not nursing. The model’s lightweight architecture and inference acceleration address the computational constraints of real-time applications in farrowing-room monitoring.

Results

The modified YOLO11n achieved an mAP@50 of 98.90%, while reducing complexity to 207 layers and 5.0 GFLOPs by removing the small-object detection head. This enabled inference times of 4.6 ms on an NVIDIA A100 GPU and 6.1 ms on a T4 GPU with TensorRT optimization. The framework detected and classified ten posture–nursing-state classes (sitting, standing, and three lying postures, each labeled as nursing or not nursing) on representative test images, including frames with partial occlusion and variable lighting. Misclassifications were observed primarily among visually similar classes, such as Sow_Sitting_Nursing and Sow_Sitting_Not_Nursing, consistent with limited visibility of piglet-to-teat contact in some postures from the top-view angle. TensorRT optimization further reduced inference latency by 58.56% (A100) and 44.55% (T4), demonstrating consistent inference-latency reductions on the evaluated GPU platforms.

Conclusions

This study demonstrates the potential of leveraging lightweight deep learning architectures for real-time behavioral monitoring in PLF. The modified YOLO11n and TensorRT optimization provide an efficient framework for automated monitoring of sow posture–nursing-state classes, supporting welfare-oriented monitoring workflows relevant to PWM risk management. Future work will evaluate lightweight temporal post-processing and embedded deployment to reduce frame-to-frame label flicker and assess performance under on-farm compute and I/O constraints.