联合信息增强和特征融合的人体摔倒检测算法  

Human fall detection algorithm combining information enhancement and feature fusion

作　　者：王凤随 邵凯丽 杨海燕 WANG Fengsui;SHAO Kaili;YANG Haiyan(School of Electrical Engineering,Anhui Polytechnic University,Wuhu 241000,China;Key Laboratory of Advanced Perception and Intelligent Control of High-end Equipment,Ministry of Education,Wuhu 241000,China)

机构地区：[1]安徽工程大学电气工程学院,芜湖241000 [2]高端装备先进感知与智能控制教育部重点实验室,芜湖241000

出　　处：《中国惯性技术学报》2024年第8期771-778,共8页Journal of Chinese Inertial Technology

基　　金：安徽省自然科学基金(2108085MF197);安徽高校省级自然科学研究重点项目(KJ2019A0162);安徽工程大学国家自然科学基金预研项目(Xjky2022040)。

摘　　要：为了提高多场景下人体摔倒姿势的实时检测能力,提出了一种基于信息增强模块和注意力特征融合的YOLOv7-tiny目标检测改进算法。首先,针对重要区域特征信息的敏感度欠缺问题,在主干网络嵌入对比感知全局信息增强模块,可有效学习特征权重,增强网络对人体姿势的判别能力;其次,为有效利用上下文信息,引入了一种密集坐标注意力特征融合结构,在通道维度上融合浅层和深层语义信息,保留有用特征信息的位置权重,促进网络对人体姿势信息的充分表达。最后,在人体摔倒姿势数据集上对所提算法进行验证。实验结果表明所提算法的平均检测精度达到了77%,较基线网络提高了3.7%,有效改善了人体摔倒行为检测的识别能力。同时,在学生课堂行为数据集SCB1、SCB2和PASCAL VOC测试集上对所提算法进行验证,其平均检测精度较基线网络分别提高了0.6%、0.5%和2.1%,验证了算法的通用性。In order to improve the real-time detection capability of human fall pose in multiple scenarios,an improved algorithm for YOLOv7-tiny target detection based on information enhancement module and attention feature fusion is proposed.Firstly,to address the lack of sensitivity of feature information in important regions,a contrast-aware global information enhancement module is embedded in the backbone network to effectively learn feature weights and enhance the network's ability to discriminate human poses.Secondly,in order to effectively utilize contextual information,a dense-coordinate attentional feature fusion structure is introduced by introducing a channel-dimensional fusion of shallow and deep semantic information that retains the position weight of useful feature information and facilitate the adequate representation of human pose information in the network.Finally,the proposed algorithm is validated on the human fall pose dataset.The experimental results show that the proposed algorithm achieves an average accuracy of 77%,which is 3.7% higher than that of the baseline network,and effectively improves the recognition ability of human fall behavior detection.Meanwhile,the proposed algorithm is validated on the student classroom behavior datasets SCB1,SCB2 and PASCAL VOC test sets,and the average detection accuracy over the baseline network is improved by 0.6%,0.5% and 2.1% respectively,validating the versatility of the algorithm.

关 键 词：摔倒检测 YOLOv7轻量网络 信息增强 特征融合 注意力机制 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]