面向人体异常行为识别的FDS-ABPG-GoogLeNet模型研究  

Research on FDS-ABPG-GoogLeNet model for human abnormal behavior recognition

作　　者：李一帆 李聪聪 李亚南[1] 王斌[1] LI Yifan;LI Congcong;LI Yanan;WANG Bin(College of Information Science and Technology,Hebei Agricultural University,Baoding 071001,China)

机构地区：[1]河北农业大学信息科学与技术学院,河北保定071001

出　　处：《现代电子技术》2025年第6期136-146,共11页Modern Electronics Technique

基　　金：河北省教育厅科学研究重点项目(ZD2021056);河北省高等学校科学研究项目(203777119D);2023河北省引进海外留学人员计划(C20230333)。

摘　　要：随着人口老龄化的加剧,老年人异常行为的识别技术已成为医疗保健领域亟需解决的关键问题。目前的异常行为识别算法面临一个挑战,即无法确保在识别多种异常行为的同时提高模型的识别准确率与计算效率。为解决此问题,提出一种FDS-ABPG-GoogLeNet模型。该模型采用了三种不同层级的改进Inception模块,并将这些模块在网络深层和浅层结构中并行连接,在中层结构中引入残差结构,通过特征融合的方式显著提高了网络的计算效率和识别准确率。同时,针对异常行为数据集中动作单一的问题,自建了包含多种异常动作的数据集,并通过将一维动作时序数据二维图形化处理后使得行为动作特征更易于提取。实验结果表明,所提FDS-ABPG-GoogLeNet模型的准确率、灵敏度和特异性分别达到99.40%、99.49%和99.93%。With the exacerbation of population aging,the identification technology of abnormal behaviors in the elderly has become a critical issue urgently needing to be addressed in the healthcare field.The current abnormal behavior recognition algorithm is faced with a challenge,that is,it cannot ensure the recognition accuracy and computational efficiency of the model while recognizing various abnormal behaviors.To address this issue,the FDS-ABPG-GoogLeNet model is proposed.In this model,three improved Inception modules at different levels are incorporated,and they are connected in parallel in both deep and shallow network structures.The residual structure is introduced in the middle structure,which significantly improves the computational efficiency and recognition accuracy of the network by means of the feature fusion.In order to solve the problem of single action in abnormal behavior data set,a dataset containing multiple abnormal actions is self built.By graphically processing one-dimensional action time series data in two dimensions,it makes it easier to extract behavioral action features.The experimental results demonstrate that the proposed FDS-ABPG-GoogLeNet model can realize an accuracy,senstivity,and specificity of 99.40%,99.49%,and 99.93%,respectively.

关 键 词：异常行为识别 Inception模块 残差结构 特征融合 特征提取 卷积神经网络 

分 类 号：TN925-34[电子电信—通信与信息系统] TP391.9[电子电信—信息与通信工程]