时域模型对视频行人重识别性能影响的研究  

Research on Impact of Different TemporalModelingMethods on Video-Based Person Re-identification

作　　者：项俊[1] 林染染 黄子源 侯建华[1] XIANG Jun;LIN Ranran;HUANG Ziyuan;HOU Jianhua(College of Electronics and Information Engineering,South-Central University for Nationalities,Wuhan 430074,China)

机构地区：[1]中南民族大学电子信息工程学院,武汉430074

出　　处：《计算机工程与应用》2020年第20期152-157,共6页Computer Engineering and Applications

基　　金：国家自然科学基金(No.61671484,No.61701548);湖北省自然科学基金(No.2018CFB503);中南民族大学中央高校基本科研业务专项(No.CZY19039)。

摘　　要：行人重识别是计算机视觉领域一个重要的研究方向。近年来,随着视频监控需求的日益增长,基于视频序列的行人重识别研究受到了广泛的关注。典型的视频序列行人重识别系统由三部分构成:图片特征提取器(例如卷积神经网络)、提取时域信息的时域模型、损失函数。在固定特征提取器和损失函数的前提下,研究不同时域模型对视频行人重识别算法性能的影响,包括时域池化、时域注意力、循环神经网络。在Mars数据集上的实验结果表明:与基于图像的行人重识别基准算法相比,采用时域池化模型、时间注意力模型可以有效改善识别精度,但采用循环神经网络后识别效果比基准算法有所下降。As an important task in computer vision field,video-based person re-identification(Re-ID)has attracted significant attention in recent years due to the increasing demand of video surveillance.Typically,a video-based person Re-ID system is composed of an image-level feature extractor(e.g.CNN),a temporal modeling method to capture temporal information and a loss function.In this paper,it explores the impact of different temporal modeling methods on video-based person Re-ID by fixing the image-level feature extractor and the loss function to be the same.Three temporal models are considered,including temporal pooling,temporal attention and Recurrent Neural Network(RNN).Experimental results conducted on the MARS dataset demonstrate that compared with the image-based Re-ID baseline,both temporal pooling and temporal attention models can improve the recognition accuracy,and RNN’s performance has dropped to some extent.The conclusions achieved in the paper can provide helpful insight for the design of video-based person re-identification algorithms.

关 键 词：视频行人重识别 深度神经网络 特征提取 时域模型 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]