基于手部骨架数据的非特定连续手语深度识别  

Nonspecific Continuous Sign Language Deep Recognition Based on Hand Skeleton Data

作　　者：李柯景[1] 鲁光男[1] LI Ke-jing;LU Guang-nan(College of Computer Science and Technology,Changchun University,Changchun Jilin 130022,China)

机构地区：[1]长春大学计算机科学技术学院,吉林长春130022

出　　处：《计算机仿真》2025年第3期244-248,共5页Computer Simulation

基　　金：吉林省教育厅科学技术研究规划项目(2022LY505L38)。

摘　　要：非特定连续手语是指不限定手语类型和手势数量,由于手部骨架数据可能存在遮挡、变形等情况,会导致手部骨架提取和识别精度降低。为了有效提升非特定连续手语深度识别结果的准确性,提出一种基于手部骨架数据的非特定连续手语深度识别方法。采用距离场和分水岭算法,获得含有手势骨架的骨架潜在图,通过主动轮廓线模型确定骨架端点,并利用A*算法对其修剪,获取手部骨架特征。将手部骨架特征对应的背景作为负样本,得到手势方向梯度图特征,引入卷积神经网络展开训练,获取非特定连续手语检测分类器,确定目标手势区域,实现非特定连续手语深度识别。实验结果表明,所提方法能够准确提取手部骨架特征,非特定连续手语深度识别准确率在90%以上,且识别时间短。Continuous signer-independent sign language refers to unlimited sign language types and gesture quantities.At present,some hand skeleton data may have occlusion or deformation,leading to low accuracy in hand skeleton extraction and recognition.In order to effectively improve the accuracy of depth recognition,this paper proposes a depth recognition method for continuous signer-independent sign language based on hand skeleton data.Firstly,distance field and watershed algorithm were adopted to obtain a potential skeleton map containing gestures.Then,skeleton endpoints were determined by the active contour model.Meanwhile,they were pruned by using the A*algorithm,so that the hand skeleton feature could be obtained.Moreover,the background corresponding to the hand skeleton feature was used as a negative sample to obtain a direction gradient feature of the gesture.Furthermore,convolutional neural networks were introduced for training to obtain a detection classifier for continuous signer-independent sign language,thus determining the region of the target gesture.Finally,depth recognition was achieved.Experimental results show that the proposed method could accurately extract hand skeleton features.The accuracy of depth recognition is over 90%and the recognition time is short.

关 键 词：手部骨架数据 非特定连续手语 深度识别 骨架潜在图 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]