摘要
针对手势识别背景的多样化、复杂化问题,为了排除复杂的背景对手势识别的干扰,更好地在移动边缘设备上应用手势识别方法。基于改进YOLOv5+MobileNetV3的动态手势识别方法,将YOLOv5中的主干神经网络替换成轻量级神经网络MobileNetV3。同时,MobileNetV3作为特征提取器的主干网络对手势骨骼的21个关键点进行特征提取,并使用Wing_Loss损失函数对训练过程加以监督,以提高不同类别间的可分性和紧凑性。实验结果表明,该算法在检测精度上由78.66%提升至84.64%,相比较原始YOLOv5算法,有6%的提升,从而为轻量化模型的部署奠定了基础,提升了人机交互的体验感。
Aiming at the diversity and complexity of gesture recognition background,in order to eliminate the interference of complex background on gesture recognition,we can better apply gesture recognition methods on mobile edge devices.Based on the improved dynamic gesture recognition method of YOLOv5+MobileNetV3,the backbone neural network in YOLOv5 is replaced by the lightweight neural network MobileNetV3.At the same time,MobileNetV3 acts as the backbone network of feature extractor to extract the 21 key points of gesture skeleton,and uses Wing_loss function supervises the training process to improve the separability and compactness between different categories.The experimental results show that the detection accuracy of this algorithm is improved from 78.66%to 84.64%,which is 6%higher than the original YOLOv5 algorithm.This lays the foundation for the deployment of lightweight models and improves the experience of human-computer interaction.
作者
霍英
张嘉明
陈台兴
林鸿森
彭季雨
HUO Ying;ZHANG Jiaming;CHEN Taixing;LIN Hongsen;PENG Jiyu(School of Information Engineering,Shaoguan University,Shaoguan Guangdong 512005)
出处
《软件》
2023年第11期47-52,共6页
Software
基金
广东省自然科学基金(2021A1515011803)
攀登计划广东省科技创新战略专项资金资助项目(pdjh2023a0471)。
作者简介
霍英(1975-),女,博士,教授,研究方向:大数据与物联网、社会网络。
