基于EL-YOLO v8n的轻量化群养猪只姿态识别与跟踪方法

Lightweight Method for Group-housed Pig Pose Recognition and Tracking Based on EL-YOLO v8n

摘要:

针对复杂群养环境下不同日龄生猪姿态复杂、个体相似度高及相互遮挡导致的检测精度低和跟踪难等问题，本文提出一种基于EL-YOLO v8n的轻量化群养生猪姿态识别跟踪方法。在YOLO v8n模型基础上引入EfficientRepBiFusion结构增强多尺度特征融合，设计LSDGCD轻量化检测头提升特征提取能力，同时采用优化的EIoU损失函数提高正样本的学习效果。在轨迹跟踪方面，结合ByteTrack算法实现了复杂场景下猪只身份保持和连续性跟踪，有效减少因遮挡和姿态切换带来的ID切换和漏检。试验结果表明，该模型对群养猪只的站立、躺卧、犬坐和趴卧等姿态识别精确率分别为98.7%、93.2%、99.3%及92.3%，与YOLO v8n相比，平均检测精确率提升6.24%，召回率提升7.61%，mAP@0.5提升8.49%，参数量降低15.28%，模型内存占用量仅5.45 MB；与ByteTrack结合，IDF1达到90.7%，MOTA和HOTA分别为88.6%和66.2%，优于DeepSORT和BoT-SORT。该模型基于Jetson Nano在猪场侧边缘部署，实现了猪只姿态的持续检测跟踪，为生猪智慧养殖管理提供了技术支撑。

Abstract:

Aiming to address the problems of low detection accuracy and difficult tracking caused by complex postures, high individual similarity and mutual occlusion of group-housed pigs, a lightweight pose recognition and tracking method for pigs based on an EL-YOLO v8n model was proposed. On the basis of YOLO v8n, EfficientRepBiFusion (ERB) structure was introduced to enhance multi-scale feature fusion, a lightweight LSDGCD detection head was designed to improve feature extraction capability, and an optimized EIoU loss function was adopted to enhance the learning effect of positive samples. In trajectory tracking, ByteTrack was integrated to achieve identity maintenance and tracking continuity of pigs, effectively reducing ID switches and missed detections caused by occlusion and pose switching. Experimental results showed that in the pose detection of pigs, the recognition accuracies of the proposed model for standing, lying, sitting and sprawling postures were 98.7%, 93.2%, 99.3% and 92.3%, respectively. Compared with YOLO v8n, the average detection accuracy of the model was improved by 6.24%, the recall rate was increased by 7.61%, the mean average precision (mAP@0.5) was raised by 8.49%, the number of parameters was reduced by 15.28%, and the model memory footprint was only 5.45 MB. When combined with ByteTrack, the model attained an IDF1 of 90.7%, with multiple object tracking accuracy (MOTA) and higher order tracking accuracy (HOTA) reaching 88.6% and 66.2% respectively, outperforming that of DeepSORT and BoT-SORT. The model realized edge deployment at pig farms by Jetson Nano, enabling continuous detection and tracking of the pig postures, which provided technical support for group-housed pig intelligent breeding.