RTMPose-MCA:一种改进的瞳孔和眼角点定位模型  

RTMPose-MCA:An Improved Model for Pupil and Eye Corner Localization

作　　者：张丁玮 万亚平[1] 邹刚 罗扬[1] 张璇[3] ZHANG Ding-wei;WAN Ya-ping;ZOU Gang;LUO Yang;ZHANG Xuan(School of Computer,University of South China,Hengyang 421001,China;Hunan Zhongke Help Innovation Research Institute,Changsha 410000,China;Xiangya Hospital,Central South University,Changsha 410000,China)

机构地区：[1]南华大学计算机学院,湖南衡阳421001 [2]湖南中科助英智能科技研究院,湖南长沙410000 [3]中南大学湘雅医院,湖南长沙410000

出　　处：《计算机技术与发展》2025年第4期164-171,共8页Computer Technology and Development

基　　金：湖南省自然科学基金(2024JJ7428)。

摘　　要：精确定位瞳孔和眼角点有助于准确测量认知障碍评估中的视觉反应时间这一关键指标,由于形状、光照等图像噪声和干扰问题,实现高精度定位仍存在挑战。为此,提出了一种基于改进RTMPose的瞳孔和眼角点定位模型RTMPose-MCA,以进一步提升定位的准确性。首先,设计了多尺度融合卷积模块MSFCM替换原模型的第一个卷积模块,增强了对区域细节信息的提取能力。其次,设计了通道方差融合注意力模块CVFAM替换Backbone部分模块,增强了对全局和局部特征的捕捉能力,减弱光照等噪声干扰。最后,设计了空洞融合卷积模块AFCM替代Head中的7×7卷积,减少了参数量。实验结果表明,在BioID和GI4E数据集上,RTMPose-MCA模型在瞳孔和眼角点的定位精度方面优于其他对比算法,平均像素距离误差分别为0.82像素和1.08像素,且模型的参数量处于较低水平。这些结果验证了该模型在复杂环境下能够有效定位瞳孔和眼角点。Accurate localization of the pupil and eye corners contributes to precisely measuring the visual reaction time,a key indicator in cognitive impairment assessments.However,achieving high-precision localization remains challenging due to shape variations,lighting conditions,and image noise interference.To address these issues,we propose an improved RTMPose-based model,named RTMPose-MCA,for pupil and eye corner localization to enhance accuracy.Firstly,a Multi-Scale Fusion Convolution Module(MSFCM)is designed to replace the first convolution module of the original model,which enhances the ability of extracting regional details.Secondly,the Channel Variance Fusion Attention Module(CVFAM)is designed to replace specific Backbone modules,which strengthens the capture of global and local features and weakens the noise interference such as light.Lastly,the Atrous Fusion Convolution Module(AFCM)substitutes the 7×7 convolution in the Head,which reduces the number of parameters and improves the scalability of the model.Experimental results on the BioID and GI4E datasets show that the RTMPose-MCA model outperforms other comparison algorithms in localization accuracy,achieving mean pixel distance error of 0.82 pixels and 1.08 pixels,respectively,while maintaining a relatively low parameter count.These findings demonstrate the model's effectiveness in accurately localizing the pupil and eye corners in complex environments.

关 键 词：瞳孔定位 眼角定位 RTMPose 特征融合 注意力模块 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]