基于深度学习的动态手势检测与识别算法研究  

Research on Dynamic Gesture Detection and Recognition Algorithm Based on Deep Learning

作　　者：陈吴东 CHEN Wudong(Chizhou University,Chizhou 247000,China)

机构地区：[1]池州学院,安徽池州247000

出　　处：《现代信息科技》2025年第8期54-60,共7页Modern Information Technology

摘　　要：手势识别对实现人机交互具有重要意义。为实现动态情况下算法高精度目标检测识别,首先以YOLOv5目标检测为基础,利用算法内部特征金字塔结构和多尺度融合的结构特征,确定目标手势的坐标信息。再通过MediaPipe模型对手部关键点进行检测,确定手部关节的向量角度,分析手指弯曲情况,从而判断手势的具体动作。采用位置确定和动作分类分模型实现的方式,有效改善了手势在动态情况下因旋转、遮挡等因素导致的识别率下降问题。训练样本选取HaGRID开源手势数据集中的6种类别,实验测试结果表明,结合后的算法在数据中单手识别检测精度均值达mAP@0.5,检测速度达到40 FPS,模型大小为88.5 MB。Gesture recognition is of great significance to realize human-computer interaction.In order to realize high-precision target detection and recognition under dynamic conditions,this paper is based on YOLOv5 target detection firstly,and determines the coordinate information of the target gesture by using the feature pyramid structure and multi-scale fusion structural features inside the algorithm.Then it uses the MediaPipe model to detect the key points of the hand,determines the vector angle of the hand joints,and analyzes thefinger bending situation,so as to judge the specific gesture.Using the methods of position determination and implementation by using separate models for action classification effectively improves the problem that the reduced recognition rate of gestures caused by factors such as rotation and occlusion in dynamic conditions.The training samples are selected from six categories in the open-source gesture dataset HaGRID.The experimental test results demonstrate that the mean value of one-hand recognition detection accuracy of the combined algorithm is up to mAP@0.5 and the detection speed is up to 40 FPS,and the model size is 88.5 MB.

关 键 词：手势识别 YOLOv5 MediaPipe 手部关节点检测 手势数据集 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]