摘要
为了进一步提高居家监护场景下人体动作识别的可靠性与实时性,更好地辅助出院后的卒中患者进行康复训练,提出一种基于单目视觉的在线人体动作识别算法。融合姿态估计OpenPose与最近邻匹配算法对监控视频流中的目标人体生成动作序列。通过滑动窗口选取原始姿态特征并对其预处理转化为鲁棒性特征,输入到多层LSTM长短时记忆网络中进行康复动作识别。实验结果表明,该方法对活动背景、人体穿着、无关人员的干扰等具有较强的适应能力,能够在线识别连续的康复动作且准确率达90.66%,在居家康复训练场景中有一定的应用价值。
In order to further improve the stability and real-time of human action recognition in the home monitoring scene,and better assist discharged stroke patients to perform rehabilitation training,we propose an online action recognition algorithm based on monocular vision.OpenPose and nearest neighbor matching algorithm were combined to generate action sequences for target human body in surveillance video streams.The original features were selected by sliding window and transformed into robustness features.The features were input into multi-layer LSTM for rehabilitation action recognition.The experimental results show that the proposed method has strong adaptability to the background of activities,human body wear,interference from unrelated personnel,etc.It can recognize continuous rehabilitation actions online with an accuracy rate of 90.66%,which has certain application value in home rehabilitation scene.
作者
闫航
陈刚
崔莉亚
张乐芸
胡北辰
Yan Hang;Chen Gang;Cui Liya;Zhang Leyun;Hu Beichen(College of Information Engineering,Zhengzhou University,Zhengzhou 450001,Henan,China;Internet Medical and Health Service Collaborative Innovation Center,Zhengzhou University,Zhengzhou 450052,Henan,China;College of Nursing and Health,Zhengzhou University,Zhengzhou 450001,Henan,China)
出处
《计算机应用与软件》
北大核心
2021年第2期171-178,共8页
Computer Applications and Software
基金
国家重点研发计划项目(2017YFB1401200)
河南省科技攻关项目(182102310137)
赛尔网络下一代互联网技术创新项目(NGII20170716)
2019年河南省高等学校重点科研项目(19A520039)。