基于小样本学习的老年人异常行为识别模型研究  

Research on Recognition Model of Abnormal Behavior of the Elderly Based on Few-Shot Learning

作　　者：王聪敏 周艳聪[2] 

机构地区：[1]天津商业大学理学院,天津 [2]天津商业大学信息工程学院,天津

出　　处：《计算机科学与应用》2023年第3期465-471,共7页Computer Science and Application

摘　　要：当今社会老龄化程度逐步加重,老年人的看护问题日益凸显。基于视频识别的老年人异常行为识别旨在助力老年人看护问题。但该领域相关公开的老年人异常行为数据稀缺。针对这种情况,本文收集了老年人异常行为数据,并结合小样本学习设计了老年人异常行为识别模型(RAAE, Recognition model of Abnormal Action of the Elderly)。该网络通过patch级富集模块获取patch级空间信息,后接通道注意力模块获取帧间时间语义信息,最后通过CrossTransformer进行时间建模。实验表明该网络和以往模型相比,对老年人异常行为的识别准确率得到了4.7%的提升。Nowadays, the aging of society is gradually increasing, and the care of the elderly is increasingly prominent. The recognition of abnormal behaviors of the elderly based on video recognition aims to help the elderly care problems. However, relevant public data on abnormal behavior of the elderly in this field are scarce. In response to this situation, this paper collected the data of abnormal behavior of the elderly, and designed a recognition model of abnormal behavior of the elderly (RAAE, Recognition model of Abnormal Action of the Elderly) based on few-shot learning. The network obtains the patch-level spatial information through the patch-level enrichment module, and then ob-tains the inter-frame temporal semantic information through the channel attention module. Finally, time modeling is performed through Cross Transformer. The experiment shows that compared with previous models, the recognition accuracy of the network for abnormal behavior of the elderly has been improved by 4.7%.

关 键 词：异常行为识别 小样本学习 深度学习 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]