基于YOLO系列算法的牦牛目标检测与盘点系统研究

Research on Yak Object Detection and Counting System Based on YOLO Series Algorithms

摘要: 在高原牧区，传统牦牛养殖方式受限于资源竞争、疾病传播及管理困难等问题，智慧畜牧业逐渐成为提升养殖效率的关键。本研究面向成年牦牛与幼牦牛混圈混牧场景下的自动盘点需求，构建了规模约18,000张的自建数据集，并基于YOLOv11、YOLOv12引入SAConv、SCConv、ECA、DySample、PSA及Slide Loss等多种改进模块，通过系统的消融实验和热力图可视化全面评估了各改进策略的有效性。在此基础上结合K230嵌入式平台和上位机数据分析软件，实现了一套面向实际部署需求的智能盘点系统。实验结果表明，所提出的改进模型在自建数据集上取得了mAP50为0.994、mAP50-95为0.930的检测效果，并能稳定运行于−40℃~70℃的极端环境，为高原牧区智能化养殖管理提供了切实可行的技术方案。

Abstract: In high-altitude pastoral areas, traditional yak farming is limited by resource competition, disease transmission and management difficulties, and intelligent animal husbandry has gradually become a key approach to improving farming efficiency. Targeting the automatic counting of mixed-grazing scenarios with adult and young yaks, this paper builds a self-collected dataset of approximately 18,000 images and introduces SAConv, SCConv, ECA, DySample, PSA and Slide Loss into YOLOv11 and YOLOv12 to construct multiple improved detectors. Ablation studies and heatmap visualization comprehensively verify the effectiveness of each module. Combined with a K230 embedded platform and a host-side analysis tool, an intelligent counting system is implemented. Experimental results show that the improved models achieve mAP50 of 0.994 and mAP50-95 of 0.930, and the system can stably operate from −40 to 70 degrees Celsius, providing a feasible solution for intelligent management of high-altitude pastures.