基于深度学习和高斯混合模型的异常识别算法  

Anomaly Detection Algorithm Based on Deep Learning and Gaussian Mixture Model

作　　者：范冰 李鹏[1] 金舒[1] 王志心[1] 王媛媛[1] FAN Bing;LI Peng;JIN Shu;WANG Zhixin;WANG Yuanyuan(Institute of System Automation Technology and Application,Guodian Nanjing Automation Co.,Ltd.,Nanjing 210000)

机构地区：[1]国电南京自动化股份有限公司研究院系统自动化技术与应用研究所,南京210000

出　　处：《计算机与数字工程》2024年第4期989-994,共6页Computer & Digital Engineering

摘　　要：现有的异常识别方法无法兼顾与人有关的异常行为识别和与人无关的异常现象识别,且在目标有遮挡的情况下识别率较低。针对上述问题,提出基于深度学习的网络框架对正常帧的人体姿态的结构特征片段进行训练,片段化训练使得算法具有抗遮挡能力。此外,为了识别与人无关的异常情况(如汽车、自行车、遗留物等),提出使用改进的自适应K值的高斯混合模型对视频背景进行建模,从而对与人无关的异常情况进行提取。实验表明,算法在公开数据集Avenue和ShanghaiTech上的异常识别AUC分别提升了0.025和0.034。Current approaches for anomaly detection cannot tackle both human-related abnormal behavior recognition and non-human anomaly recognition.Moreover,the detection accuracy is low especially in the case of occlusion.To solve this problem,the network based on GRU is proposed to train the structural feature segments of the human body posture in the normal frame.In doing so,this approach is able to anti-occlusion.This paper further introduces an improved Gaussian mixture model with adaptive K value to extract abnormal situations(such as cars,bicycles,remnants,etc.)that are not related to human.Experiments show that the anomaly detection AUC of proposed method increased by 0.025 and 0.034 on Avenue dataset and ShanghaiTech dataset respectively.

关 键 词：高斯混合模型 异常行为识别 片段化训练 抗遮挡 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]