基于视频时空特征提取分类的动作分析评估模型  

An action analysis and evaluation model based on video spatiotemporal feature extraction and classification

作　　者：陈迪 李焱芳 毕卫云 李朗 蒲珊珊 CHEN Di;LI Yanfang;BI Weiyun;LI Lang;PU Shanshan(School of Basic Medical Sciences,Air Force Medical University,Xi’an 710032,China;The First Affiliated Hospital of AFMU,Xi’an 710032,China)

机构地区：[1]空军军医大学基础医学院,陕西西安710032 [2]空军军医大学第一附属医院,陕西西安710032

出　　处：《现代电子技术》2024年第8期160-164,共5页Modern Electronics Technique

基　　金：陕西省“十四五”教育科学规划课题(SGH22Y1356)。

摘　　要：为拓展机器视觉技术在医工结合场景下的应用,文中基于改进的时空Transformer模型,提出一种动作规范识别模型。该模型由数据嵌入层、时空Transformer层、决策融合层组成。数据嵌入层利用Openpose模型从sRGB图像中提取人体骨骼数据,降低环境部署成本;时空Transformer层使用时空模块和块间模型对图像数据特征进行训练和分类,提升原模型的分类精度;决策融合层实现对应用场景的规范性判别。实验测试结果表明:所提算法的TOP1和TOP5精度指标在所有对比算法中均为最优;在以心肺复苏术为例进行的实际应用测试中,该算法的综合性能较为理想,能够满足工程需要。In order to expand the application of machine vision technology in medical and industrial integration scenarios,an action specification recognition model based on an improved spatiotemporal Transformer model is proposed.The model is composed of data embedding layer,spatiotemporal Transformer layer,and decision fusion layer.In the data embedding layer,the Openpose model is used to extract human bone data from sRGB images,so as to reduce environmental deployment costs.In the spatiotemporal Transformer layer,the spatiotemporal modules and inter block models are used to train and classify image data features,so as to improve the classification accuracy of the original model.The decision fusion layer is used to realize the normative discrimination for application scenarios.The experimental testing results show that the TOP1 and TOP5 accuracy indicators of the proposed algorithm are the best among all comparative algorithms.In practical application testing using cardiopulmonary resuscitation as an example,the comprehensive performance of the algorithm is relatively ideal and can meet engineering needs.

关 键 词：计算机视觉 时空Transformer模型 骨骼模型 决策融合 动作识别 多头注意力机制 

分 类 号：TN919-34[电子电信—通信与信息系统] TP391[电子电信—信息与通信工程]