基于改进HRNet的牛体关键点检测算法  

Cattle body keypoint detection algorithm based on improved HRNet

作　　者：赵雪莲 张继凯 何一豪 曾翔皓 庄琦 ZHAO Xuelian;ZHANG Jikai;HE Yihao;ZENG Xianghao;ZHUANG Qi(Information Engineering School,Inner Mongolia University of Science and Technology,Baotou 014010,China;Engineering Training Centre,Inner Mongolia University of Science and Technology,Baotou 014010,China)

机构地区：[1]内蒙古科技大学信息工程学院,内蒙古包头014010 [2]内蒙古科技大学工程训练中心,内蒙古包头014010

出　　处：《内蒙古科技大学学报》2024年第2期172-177,共6页Journal of Inner Mongolia University of Science and Technology

基　　金：内蒙古自治区科技重大专项(2019ZD025);内蒙古自治区自然科学基金(2021MS06007);内蒙古自治区科技计划项目(2019GG138)。

摘　　要：针对现有关键点检测算法在复杂背景下检测精度低、高运算量等问题,提出一种轻量级关键点检测模型SE-HRNet。首先设计2种轻量型模块:SECAneck模块和SECAblock模块,在保持网络性能的同时减低计算参数,加快训练速度。其次,整合空间注意力机制于多分辨率融合阶段,使得模型对于不易检测到的关键点的定位和识别更为敏感。在自制牛体关键点数据集上进行实验评估,结果表明:改进后的HRNet网络比原网络参数量和运算浮点数分别减少了18.8 M和5.2 G,平均精度达到了93.2%,平均召回率达到了91.5%,每秒帧数(FPS)达到了36.3。In response to the low accuracy and high computational demands of existing keypoint detection algorithms in complex environments,the SE-HRNet,a lightweight model was developed.This was accomplished by first creating two lightweight modules,the SECAneck and the SECAblock,which allowed for a reduction in computational parameters and faster training without sacrificing performance.In addition,a spatial attention mechanism was incorporated with the multi-resolution fusion stage,improving the model’s sensitivity to difficult-to-detect keypoints.Experiments on custom-built cattle body keypoint dataset show that the refined HRNet network decreases parameter count by 18.8 million and computational operations by 5.2 billion.The model now achieves an average precision of 93.2%and the average recall rate of 91.5%,while running at 36.3 frames per second.

关 键 词：关键点检测 高分辨率网络 注意力机制 多分辨率融合阶段 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]