融合人体骨架和姿势信息特征的轻量级人体动作识别方法  

A lightweight human motion recognition method incorporating both skeletal and postural information features

作　　者：王振宇[1] 向泽锐[1,2] 支锦亦[1,2] 叶浩航 丁铁成 WANG Zhenyu;XIANG Zerui;ZHI Jinyi;YE Haohang;DING Tiecheng(School of Design and Art,Southwest Jiaotong University,Chengdu 611730,China;Institute of Human-Machine Environment System Design,Southwest Jiaotong University,Chengdu 611730,China)

机构地区：[1]西南交通大学设计艺术学院,四川成都611730 [2]西南交通大学人机环境系统设计研究所,四川成都611730

出　　处：《应用科技》2024年第2期135-144,共10页Applied Science and Technology

基　　金：国家重点研发计划项目(2022YFB4301203);教育部2022年第二批产学合作协同育人项目(220705329291641);津发科技−功效学会“人因与功效学”项目.

摘　　要：针对人体动作识别任务中特征值选取不当导致识别率低、使用多模态数据导致训练成本高等问题,提出一种轻量级人体动作识别方法。首先使用OpenPose、PoseNet提取出人体骨架信息,使用BWT69CL传感器提取姿势信息;其次对数据进行预处理、特征融合,对人体动作进行深度学习分类识别;最后,为验证此方法的有效性,在公开数据集WISDM、UCIHAR、HASC和自建的人体动作数据集上进行实验验证,并使用改进的目标引导注意力机制(target-guided attention,TGA)–长短期记忆(long short term memory,LSTM)网络输出最终的分类结果。实验结果表明,在自建数据集下融合姿势和骨架特征达到99.87%准确率,相比于只使用姿势信息特征,识别准确率提高了约5.31个百分点;相比于只使用人体骨架特征,识别准确率提高了约1.87个百分点;在识别时间上相比于只使用姿势信息,识别时间降低了约29.73 s;相比于只使用人体骨架数据,识别时间降低了约9 s。使用该方法能及时有效地反映人体的运动意图,有助于提高人体动作和行为的识别准确率和训练效率。A lightweight human action recognition method is proposed to solve the problems of low recognition rate due to improper selection of feature values and high training cost due to the use of multimodal data in the human action recognition task.Firstly,the human skeleton information was extracted using OpenPose and PoseNet,and the pose information was extracted using the BWT69CL sensor;And next,the data was preprocessed,the features were fused,and the human actions were recognized by deep learning and classification.Finally,in order to verify effectiveness of this method,experimental validation was performed on the public datasets WISDM,UCI-HAR,HASC,and selfconstructed human action dataset,and the improved target-guided attention(TGA)-long short-term memory(LSTM)network was used to output the final classification results.The experimental results show that the fusion of pose and skeleton features under the self-constructed dataset achieves 99.87%accuracy,which improves the recognition accuracy by about 5.31 percentage points compared with that using only pose information features,and improves the recognition accuracy by about 1.87 percentage points compared with that using only human skeleton features,and reduces the recognition time by about 29.73 s compared with that using only pose information in the recognition time.Compared with that using only human skeleton data,the recognition time is reduced by about 9 s.Using this method can reflect the human body's movement intention in a timely and effective manner,which helps to improve the recognition accuracy and training efficiency of human movements and behaviors.

关 键 词：人体骨架 姿势信息 轻量级 人体动作识别 目标引导注意力机制 数据集 多模态 特征提取 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]