无人机场景下基于Transformer的轻量化行人重识别  

Transformer based lightweight person re-identification in unmanned aerial vehicle images

作　　者：胡海峰[1] 倪宗煜 赵海涛[1,2] 张红 沐勇 吴建盛[3] HU Haifeng;NI Zongyu;ZHAO Haitao;ZHANG Hong;MU Yong;WU Jiansheng(School of Communications and Information Engineering,Nanjing University of Posts and Telecommunications,Nanjing 210003,China;Ministry of Education Ubiquitous Network Health Service System Engineering Research Center,Nanjing University of Posts and Telecommunications,Nanjing 210003,China;College of Geographic and Biological Information,Nanjing University of Posts and Telecommunications,Nanjing 210023,China)

机构地区：[1]南京邮电大学通信与信息工程学院,江苏南京210003 [2]南京邮电大学教育部泛在网络健康服务系统工程研究中心,江苏南京210003 [3]南京邮电大学地理与生物信息学院,江苏南京210023

出　　处：《南京邮电大学学报（自然科学版）》2024年第3期48-62,共15页Journal of Nanjing University of Posts and Telecommunications：Natural Science Edition

基　　金：国家自然科学基金(62071242,61571233,61901229,61872198)资助项目。

摘　　要：针对无人机场景下行人重识别所呈现的多视角多尺度特点,以及传统的基于卷积神经网络的行人重识别算法受限于局部感受野结构和下采样操作,很难对行人图像的全局特征进行提取且图像空间特征分辨率不高。提出一种无人机场景下基于Transformer的轻量化行人重识别(Lightweight Transformer-based Person Re-Identification,LTReID)算法,利用多头多注意力机制从全局角度提取人体不同部分特征,使用Circle损失和边界样本挖掘损失,以提高图像特征提取和细粒度图像检索性能,并利用快速掩码搜索剪枝算法对Transformer模型进行训练后轻量化,以提高模型的无人机平台部署能力。更进一步,提出一种可学习的面向无人机场景的空间信息嵌入,在训练过程中通过学习获得优化的非视觉信息,以提取无人机多视角下行人的不变特征,提升行人特征识别的鲁棒性。最后,在实际的无人机行人重识别数据库中,讨论了在不同量级主干网和不同剪枝率情况下所提LTReID算法的行人重识别性能,并与多种行人重识别算法进行了性能对比,结果表明了所提算法的有效性和优越性。Person re-identification(ReID)in the unmanned aerial vehicle(UAV)surveillance scenario has multi-viewpoints and multi-scales,and traditional convolutional neural network(CNN)-based methods for ReID suffer from the distribution of receptive fields and downsampling operators.These methods are hard to extract global features of pedestrian images,resulting in low image resolution in the feature space.A lightweight transformer-based person Re-Id(LTReID)method is proposed to utilize the attention-based method to extract multiple diversified body parts,and adopt the circle loss and the margin sample mining loss to enhance the ability to extract the discriminative features and provide the fine-grained retrieval.Furthermore,the Transformer model is pruned by using the post-training lightweight mask search algorithm,which can facilitate the deployment of the model on the platform of UAVs.On this basis,a learned spatial information embedding method is proposed to obtain an optimized non-visual information during the training,which can extract invariant features in the multi-viewpoints scenario of UAVs and improve the robustness of ReID.Finally,in the real person ReID dataset in aerial imagery,the factors that affect the performance of the LTReID method are discussed,such as multi-scale transformer networks and pruning ratios,and compared with those of the state-of-the-art ReID methods.Experimental results demonstrate the effectiveness and superiority of the proposed LTReID.

关 键 词：无人机场景 行人重识别 Transformer轻量化 空间信息嵌入 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]