中国农机化学报2026，Vol.47Issue(1)：94-99,117,7.DOI:10.13733/j.jcam.issn.2095-5553.2026.01.014
基于多尺度融合的牛行为识别方法
A cattle behavior recognition method based on multi-scale fusion
摘要
Abstract
In the current rapid development of animal husbandry and the trend towards refined management of pastures,non-contact and high-precision identification of cattle behavior is of great significance for the construction of smart pastures.Aiming at the problem of low accuracy in cattle behavior recognition,a multi-scale fusion model is proposed to recognize the standing,lying,feeding,drinking,excreting,calving,and crawling behaviors of cattle.In the feature extraction stage,the CNN branch focuses on capturing fine-grained features of local behavior,while the Swin Transformer branch uses a hierarchical window self attention mechanism to establish cross regional global semantic associations for extracting global features.In response to the key requirement of multi-scale feature expression in behavior recognition,the PAN module is introduced to achieve progressive fusion of deep and shallow features through a bidirectional feature pyramid architecture.This effectively solves the problem of synchronous recognition of small target behavior and large-scale behavior in complex backgrounds,overcomes the limitations of traditional single modal methods in long-distance dependency modeling,significantly improves the robustness of the model to interference factors such as occlusion and lighting changes through dynamic complementarity of local global features,and achieves accurate recognition of cattle behavior through end-to-end training.The ablation experiments were adopted to analyze and compare five models,namely YOLOv8,Swin Transformer,Vision Transformer,EfficientNet and MobileNet,the results showed that the multi-scale fusion model achieved an accuracy rate of 98.8%in recognizing cattle behavior,with a frame rate FPS of 64.78 f/s.Multi scale fusion has a high recognition accuracy for cattle behavior recognition,which basically meets the needs of breeding farms in real-time monitoring and can provide technical support for achieving precise breeding.关键词
牛/行为识别/自注意力机制/多尺度融合/智慧牧场Key words
cattle/behavior recognition/self attention mechanism/multi-scale fusion/smart farms分类
农业科技引用本文复制引用
Li Xin,Wang Liying,Wang Yueming,Chu Yanhua,Zhang Zhirong..基于多尺度融合的牛行为识别方法[J].中国农机化学报,2026,47(1):94-99,117,7.基金项目
内蒙古自治区科技计划项目(2023YFSW0014) （2023YFSW0014）
内蒙古自治区直属高校基本科研业务费项目(2023QNJS193,2024XKJX018) （2023QNJS193,2024XKJX018）
内蒙古自治区自然科学基金项目(2024MS06026) （2024MS06026）
