Authors

Abstract

With the advancement of intelligent farming technology, computer vision-based animal behavior recognition has become an important tool for improving production efficiency and animal welfare in modern farming management. To overcome the challenge of balancing computational efficiency and accuracy in existing behavior recognition systems, this study proposes an optimized model based on YOLOv8-EAPNet for accurately recognizing four main pig behaviors: standing, sitting, lateral lying, and prone lying. The framework adopts a multi-level lightweight design, incorporating three advanced technologies—C2f-ECA, SPPELAN, and Detect_AFPN—to enhance joint feature response, resolve spatial differences between sitting and lateral lying, and reconstruct semantics in occluded areas. This strengthens the model's robustness in complex farming environments and significantly improves the accuracy of pig behavior recognition. Validated on farm data, the model achieved an average precision improvement of 1.5% on a self-built dataset, with specific accuracy increases of 0.9% for standing, 1.7% for sitting, 3.0% for prone lying, and 0.3% for lateral lying. This technology provides an automated tool for early warning of limb injuries and respiratory diseases in pigs, promoting the upgrade of intelligent health management in the livestock industry and supporting the modernization of large-scale pig farming.

Abstract in Chinese

随着智能化养殖技术的进步，基于计算机视觉的动物行为识别成为现代养殖管理中提高生产效率与动物福利的重要工具。为了克服现有行为识别系统在处理计算效率与精确度之间的平衡问题，本研究提出了一种基于YOLOv8-EAPNet的优化模型，用于精确识别猪的四种主要行为：站立、坐姿、侧卧和趴卧。该框架采用了多层次轻量化设计，结合C2f-ECA、SPPELAN和Detect_AFPN三项先进技术，分别用来强化关节特征响应，解析坐姿与侧卧空间差异和重建遮挡区域语义。增强了模型在复杂养殖环境下的鲁棒性，显著提升了猪行为识别的精度。基于猪场数据验证，模型在自建猪场数据集上平均精度提升1.5%，模型显著提升四类行为判别精度，站立行为的识别精度较基准模型提升0.9%。坐立行为的识别精度较基准模型提升1.7%。趴卧行为的识别精度较基准模型大幅提升3.0%，侧卧行为的识别精度较基准模型提升提升0.3%。该技术为猪只肢蹄损伤、呼吸道疾病早期预警提供自动化工具，推动畜牧业智能化健康管理升级，为畜牧业提供了精确的行为识别和生理健康评估工具，通过趴卧行为监测关联呼吸道疾病风险，为主动健康防控提供核心技术支撑，从而促进了大规模农业的发展和猪只养殖方式的现代化。