摘要
针对传感器对手势识别存在范围小、鲁棒性弱等痛点,采用MediaPipe机器学习框架对捕获的手势图像实时遍历,利用卷积神经网络进行高斯平滑滤波,并结合21个特征关节点手掌模型,根据欧氏空间距离判别阈值和单个手指曲率对指间动作做出分类,通过坐标关系建立指尖和模型特征点之间的实时映射。经测试,目标区域内手势识别准确精度达到98%,并实现了对控制音量大小等操作类手势的准确识别。
In response to the pain points of small range and weak robustness of sensors in gesture recognition,the MediaPipe machine learning framework is used to traverse the captured gesture images in real-time,and a convolutional neural network is used for Gaussian smoothing filtering.Combined with a palm model of 21 feature joint points,fingertip movements are classified based on Euclidean distance discrimination threshold and individual finger curvature,and real-time mapping between fingertips and model feature points is established through coordinate relationships.According to the test,the accuracy of gesture recognition in the target area has reached 98%,and the accurate recognition of operation gestures such as control volume is realized.
作者
李银银
陈磊
杨罡
赵静
LI Yin-yin;CHEN Lei;YANG Gang;ZHAO Jing(School of Computer Science,Huainan Normal University,Huainan,232038,Anhui;School of Chemical&Material Engineering,Huainan Normal University,Huainan,232038,Anhui)
出处
《蚌埠学院学报》
2024年第2期71-76,共6页
Journal of Bengbu University
基金
全国重点实验室开放课题(COGOS-2023HE02)
安徽省高校优秀青年科研项目(2022AH030143)
淮南师范学院自然科学研究项目(2022XJYB056)。
关键词
人机交互
手势识别
卷积神经网络
human-computer interaction
gesture recognition
convolutional neural network