基于改进YOLO V7的校园人体姿态识别  

The Identify of Human Postures Based on an Improved YOLO V7

作　　者：刘思默[1] 马瑞军 何建华 方凤玲[1] LIU Simo;MA Ruijun;HE Jianhua;FANG Fenging(Fujian Polytechnic of Information Technolougy,Fuzhou,Fujian 350003,China)

机构地区：[1]福建信息职业技术学院,福建福州350003

出　　处：《福建农机》2023年第4期19-24,共6页Fujian Agricultural Machinery

基　　金：福建信息职业技术学院院级课题(Y21109)。

摘　　要：校园人体姿态检测对于监管人员快速识别人体行为,保证校园稳定具有重要意义。针对目前校园人体姿态识别主要依赖人工,效率低下的问题,提出一种基于改良的YOLO V7快速检测模型。首先引入尺寸大小自适应输入模块对输入图像进行标准化和归一化处理,提高模型的适应能力;然后将YOLO V7模型进行剪枝处理提高检测效率;再定义了19个关节点作为人体姿态的特征点,最后使用关节点匹配模块提高关节点匹配的准确率。经过实验,模型可以每秒检测44张图像,检测精度在mAP@0.5时可以达到0.955,相较于原始的YOLO V7检测速率和精度均有所提升。The detection of human postures on campus is of significant importance for supervisory staff to quickly identify human behaviors and ensure campus stability.In response to the current issue that the recognition of human postures on campuses mainly relies on manual work,which is inefficient,a fast detection model based on an improved YOLO V7 is proposed.Firstly,a size-adaptive input module is introduced to standardize and normalize the input images,enhancing the adaptability of the model.Then,the YOLO V7 model is pruned to improve detection efficiency.Furthermore,19 key points are defined as the feature points of human postures,and finally,a key point matching module is employed to enhance the accuracy of key point matching.Experimental results show that the model can detect 44 images per second,and the detection accuracy can reach 0.955 at mAP@0.5,which is an improvement in both detection speed and accuracy compared to the original YOLO V7.

关 键 词：人体姿态识别 尺寸自适应 姿态检测 关节点匹配 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]