基于上下文信息和MLE机制的视频异常检测算法  

Video anomaly detection algorithm based on contextual information and MLE mechanism

作　　者：焦雪 高立青 JIAO Xue;GAO Liqing(School of Economics and Management,Taiyuan University of Technology,Jinzhong 030600,China)

机构地区：[1]太原理工大学经济与管理学院,山西晋中030600

出　　处：《现代电子技术》2025年第9期8-14,共7页Modern Electronics Technique

基　　金：国家自然科学基金项目(72004154)。

摘　　要：针对现有的视频异常检测方法缺乏对前景对象的自主选择性,使模型对背景等无关信息的敏感性增加,进而导致重构或预测误差与前景目标数量成正比,造成误报的问题,提出一种基于上下文信息和最大局部误差机制的视频异常检测算法。该算法设计了一种基于生成对抗网络的检测框架,并提出SSPCAB-UNet生成器模型,通过在UNet中加入自监督预测卷积关注块来增强模型对局部特征和全局上下文信息的理解能力,减少对无关信息的关注,从而降低了误报的可能性。此外,使用最大局部误差机制来关注局部异常区域的预测程度,缓解因前景目标数量增多造成的较大预测误差问题。通过双模块的协同工作可以有效减少由于缺乏对前景对象的自主选择性而产生的误报问题。所提方法在CUHK Avenue、UCSD Ped2和Shanghai Tech三个数据集上的检测精度分别达到87.5%、98.2%和75.4%,验证了所提模型的有效性。The existing video anomaly detection methods lack autonomous selectivity for foreground objects,which increases the model′s sensitivity to irrelevant information such as background,and thus leads to reconstruction or prediction errors proportional to the number of foreground objects,resulting in false alarms,so a video anomaly detection algorithm based on contextual information and maximum local error(MLE)mechanism is proposed.In the algorithm,a detection framework based on generative adversarial networks(GANs)is designed and a generator model SSPCAB-UNet is proposed,which enhances the model′s ability to understand local features and global contextual information by adding a self-supervised predictive convolutional attentive block(SSPCAB)to the UNet,so as to reduce the attention to irrelevant information,and reduce the possibility of false alarms.In addition,the MLE mechanism is used to focus on the prediction of local anomalous regions,so as to alleviate the large prediction errors caused by the increased number of foreground objects.The false alarm due to lack of autonomous selectivity of foreground objects can be reduced effectively by the synergy of the dual modules.The method achieves detection accuracies of 87.5%,98.2%and 75.4%on the datasets CUHK Avenue,UCSD Ped2 and Shanghai Tech,respectively,which validates the effectiveness of the proposed model.

关 键 词：异常检测 自监督预测卷积关注块 最大局部误差机制 自动编码器 深度学习 生成对抗网络 

分 类 号：TN911.73-34[电子电信—通信与信息系统] TP391[电子电信—信息与通信工程]