摘要
智能监控中,由于存在环境复杂、监控目标多、画质质量差、人员尺寸不同等因素,从而给人体异常行为识别带来很多挑战。为了提高视频中人员异常行为识别的准确率和识别效率,提出了人员异常行为识别方法YOLO-CARAFE。该方法首先利用轻量级上采样算子CARAFE代替最近邻插值上采样算子,CARAFE不仅利用相邻像素进行工作,还会对相邻像素进行加权融合,可以在大感受野中聚合上下文信息,从而提高在复杂场景下人体异常行为识别时神经网络的特征提取和融合能力;其次,利用Focal-EIOU损失函数的难易样本学习策略,使得模型更加关注难以分类的目标对象,有效减小预测框与真实框之间的差异,提高人体异常行为识别的准确度,有效解决异常行为样本数据量少的问题。通过在自建数据集上的实验表明,YOLO-CARAFE在人体异常行为识别上具有良好的识别效果,提出的YOLO-CARAFE算法在R不变的情况下mAP@0.5,P分别为96.9%,97.6%,提高了1.9百分点,7.4百分点,能够满足监控视频中人员异常行为识别对于准确度的需求。
In intelligent monitoring,there are many factors such as complex environment,multiple monitoring targets,poor picture quality,and different personnel sizes,which bring many challenges to human abnormal behavior recognition.In order to improve the accuracy and efficiency of human abnormal behavior recognition in video,we propose YOLO-CARAFE for human abnormal behavior recognition.In this method,CARAFE,a lightweight up-sampling operator,is first used to replace the nearest neighbor interpolation up-sampling operator.CARAFE not only works with adjacent pixels,but also performs weighted fusion of adjacent pixels,which can aggregate context information in the large sensing field,thereby improving the feature extraction and fusion capability of neural networks in recognizing human abnormal behaviors in complex scenes.Secondly,using the difficulty sample learning strategy of Focal-EIOU loss function,the model pays more attention to the target objects that are difficult to classify,effectively reduces the difference between the prediction frame and the real frame,improves the accuracy of human abnormal behavior identification,and effectively solves the characteristics of small amount of abnormal behavior sample data.Experiments on self-built data sets show that YOLO-CARAFE has a good recognition effect on human abnormal behavior recognition.When R is unchanged,mAP@0.5 and P of the proposed YOLO-CARAFE algorithm are 96.9%and 97.6%respectively,increasing by 1.9 percentage points and 7.4 percentage points.It can meet the accuracy and real-time requirements of abnormal behavior identification in surveillance video.
作者
李嘎
加云岗
王志晓
张九龙
闫文耀
高昂
薛尧
LI Ga;JIA Yun-gang;WANG Zhi-xiao;ZHANG Jiu-long;YAN Wen-yao;GAO Ang;XUE Yao(School of Computer Science,Xi’an Polytechnic University,Xi’an 710600,China;School of Computer Science and Engineering,Xi’an University of Technology,Xi’an 710048,China;Xi’an Innovation College,Yan’an University,Xi’an 710100,China;National Satellite Meteorological Center,Beijing 100081,China;Xi’an Jiaotong University,Xi’an 710049,China)
出处
《计算机技术与发展》
2024年第6期185-191,共7页
Computer Technology and Development
基金
陕西省重点研发区域创新引领计划(2022QFY01-17)
教育部人文社会科学研究青年基金(16YJCZH109)
西安市科技计划(2023JH-RGZNGG-0011)
陕西省科技厅重点研发计划项目(2023-YBGY-217)
咸阳市科技局“揭榜挂帅”科技项目(L2022-JBGS-GY-07)。
