江汉大学学报（自然科学版）

采用计算机视觉技术，利用Kinect相机设计开发面向智慧医疗的无穿戴式动作检测系统，对人体姿态、行为和动作进行准确识别，辅助医生进行临床诊断与患者恢复情况评估。系统分为前端、后端和数据传输三个模块。前端使用Qt框架构建人机交互界面采集人体关节点数据，通过数据传输模块发给后端进行人体动作的智能识别；后端利用基于特征增强的时空图卷积神经网络，对人体动作进行智能识别，并返回至前端；前端根据识别结果计算关节活动范围并进行展示。经实验测试，该检测系统动作识别准确率达到99.02%，在使用GPU加速时，完成一次动作识别和关节活动范围计算仅需500 ms，校准后人体关节活动范围测量误差在±2°以内，符合临床诊断标准要求。该系统有助于推动医疗行业向更加智能、高效、便捷的方向发展，提升医疗服务的智能化、远程化、精准化。在慢病管理、康复训练、老年人健康监护、远程医疗等领域具有较好的应用价值。

In this paper,computer vision technology was used to design a non-wearable human action detection system for smart healthcare,which can accurately recognize human posture,behavior,and actions,and assist doctors in clinical diagnosis and evaluation of patients′ recovery. The system consists of three modules:the front end,the back end,and data transmission. The front end uses Qt framework to build a human-computer interaction interface for collecting human joint-point data,which are then transmitted to the back end through the data transmission module for intelligent human action recognition. The back end module employs a feature-enhanced spatiotemporal graph convolutional neural network to recognize human actions intelligently and return the results to the front end. Based on the recognition results,the front end calculates and displays the joint range of motion(ROM).Experimental results show that the proposed system achieves an action recognition accuracy of 99. 02%. With GPU acceleration, one action recognition and joint range of motion calculation can be completed in only 500 ms. After calibration,the measurement error of the human joint range of motion is within ± 2°,meeting the requirements of clinical diagnostic standards. The system helps promote the development of the medical industry toward greater intelligence, efficiency, and convenience, and improves the intelligence, remote accessibility,and precision of medical services. It also has good application value in chronic disease management,rehabilitation training,elderly health monitoring,and telemedicine.