基于深度学习的动态手势检测与识别算法研究

打开文本图片集

中图分类号：TP391.4 文献标识码：A 文章编号：2096-4706（2025）08-0054-07

Abstract： Gesture recognition is of great significance torealize human-computer interaction.In order torealize highprecisiontarget detection and recognition under dynamicconditions，this paper is based on YOLOv5 target detection firstly and determines the coordinate informationof the target gesture byusing thefeature pyramid structureand multi-scale fusion structuralfeaturesinsidethealgorithm.ThenitusestheMediaPipemodeltodetectthekeypointsofthehand，deterinesthe vectorangleofthehand joints，andanalyzes thefingerbendingsituation，soas tojudge the specific gesture.Using themethods of positiondeterminationand implementationbyusingseparate models foractionclasficationeffectivelyimproves the problem that thereduced recognitionrateof gesturescaused byfactors suchasrotationandoccusionin dynamicconditions.The training samplesare selected fromsixcategories intheopen-source gesturedataset HaGRID.Theexperimentaltestresults demostrate that the mean value of one-hand recognition detection accuracy of the combined algorithm is up to and the detection speed is up to 40 FPS，and the model size is 88.5 MB.

Keywords： gesture recognition; YOLOv5;MediaPipe; hand joint point detection; gesture dataset

0 引言

人机交互是指人与计算机之间通过某种特殊方式实现信息交换的过程，传统人机交互采用穿戴传感器的方式，由于传感器的感知范围有限且信号不具有普适性的问题，无法满足人们日常生活的实际要求。（剩余5938字）