文献类型：期刊文献

中文题名：基于条件轻量级神经网络的视频入侵检测算法

英文题名：Video Intrusion Detection Algorithm Based on Conditional Lightweight Neural Network

作者：陈天宇[1];楚程钱[1];万思远[1];万永菁[1];孙静[2]

机构：[1]华东理工大学信息科学与工程学院,上海200237;[2]上海卓希智能科技有限公司研发部,上海201611

年份：2023

卷号：49

期号：12

起止页码：152

中文期刊名：计算机工程

外文期刊名：Computer Engineering

收录：CSTPCD;;Scopus;北大核心:【北大核心2020】；CSCD:【CSCD_E2023_2024】；

基金：国家自然科学基金(61872143)。

语种：中文

中文关键词：视频入侵检测;高斯混合模型;信息熵;轻量级神经网络;条件计算;模型剪裁

外文关键词：video intrusion detection;Gaussian mixture model;information entropy;lightweight neural network;conditional computation;model tailoring

摘要：机场安防场景需要快速准确地对入侵目标进行检测及报警,现有的算法计算量大,无法满足实时处理需求。针对该问题,结合传统高斯混合模型的前景提取算法与轻量级神经网络,设计基于条件计算的轻量级入侵检测算法。在前景提取阶段使用基于信息熵的自适应学习因子更新算法动态更新高斯混合模型的学习因子,减小高斯混合模型面对镜头突入时造成模型失效的概率。目标检测阶段以ResNeXt作为主干网络,将小型ResNet作为策略网络,使混合感受野的深度可分离卷积作为残差块,设计基于条件计算的轻量级神经网络,降低网络推理时的计算量。实验结果表明,该算法在监控视频数据集和OTB100数据集上的误检率分别4.4%、9.2%,漏检率分别为2.3%、9.8%,与Faster-YOLO等传统目标检测算法相比,该算法在保证检测精度的情况下使检测速度平均提高了2.6倍。
Airport security needs to quickly and accurately detect and alarm an intrusion target.Existing algorithms require a large number of calculations and cannot meet real-time processing requirements.To address this problem,a lightweight intrusion detection algorithm based on conditional computation is designed by combining the traditional Gaussian mixture model foreground extraction algorithm with a lightweight neural network.In the foreground extraction stage,an adaptive learning factor update algorithm based on information entropy is used to dynamically update the learning factors of the Gaussian mixture model,thereby reducing the probability of model failure caused by the Gaussian mixture model in case of lens break-ins.In the target detection stage,ResNeXt is used as the backbone network;a small ResNet is used as the policy network;and a depth-wise separable convolution of the mixed receptive field is used as the residual block.A lightweight neural network based on conditional calculations is designed to reduce the number of calculations required for network inference.In the experimental results,the false detection rates of the algorithm on the surveillance video and OTB100 datasets were 4.4%and 9.2%,respectively,and the missed detection rates were 2.3%and 9.8%,respectively.Compared with traditional target detection methods,the proposed algorithm increases detection speed by an average of 2.6 times while ensuring detection accuracy.