《電子技術(shù)應(yīng)用》
您所在的位置:首頁(yè) > 其他 > 設(shè)計(jì)應(yīng)用 > 融合外觀特征的行人重識(shí)別方法
融合外觀特征的行人重識(shí)別方法
信息技術(shù)與網(wǎng)絡(luò)安全
彭玉青1,2,李 偉1,2,郭永芳1
(1.河北工業(yè)大學(xué) 人工智能與數(shù)據(jù)科學(xué)學(xué)院,天津300401; 2.河北省大數(shù)據(jù)計(jì)算重點(diǎn)實(shí)驗(yàn)室,天津300401)
摘要: 針對(duì)行人重識(shí)別中由于姿勢(shì)變化、視角改變、遮擋等引起的識(shí)別率不高的問(wèn)題,提出了融合外觀特征的行人重識(shí)別方法。該方法通過(guò)兩個(gè)網(wǎng)絡(luò)分支的設(shè)計(jì),分別提取行人的全局特征和局部特征,二者融合后得到行人的外觀特征。同時(shí)結(jié)合分類(lèi)損失和度量學(xué)習(xí)損失,通過(guò)多任務(wù)學(xué)習(xí)策略對(duì)兩個(gè)網(wǎng)絡(luò)分支進(jìn)行模型優(yōu)化。此外,該模型設(shè)計(jì)了隨機(jī)擦除算法,在數(shù)據(jù)集中加入噪聲,增強(qiáng)模型的魯棒性。實(shí)驗(yàn)結(jié)果表明:融合外觀特征的行人重識(shí)別方法大大提高了行人重識(shí)別的準(zhǔn)確率,在Market-1501數(shù)據(jù)集上rank1達(dá)到了92.82%、mAP 達(dá)到了80.51%,在DukeMTMC-reID數(shù)據(jù)集上rank1達(dá)到了85.06%、mAP達(dá)到了72.72%。
中圖分類(lèi)號(hào): TP391
文獻(xiàn)標(biāo)識(shí)碼: A
DOI: 10.19358/j.issn.2096-5133.2021.06.006
引用格式: 彭玉青,李偉,郭永芳. 融合外觀特征的行人重識(shí)別方法[J].信息技術(shù)與網(wǎng)絡(luò)安全,2021,40(6):33-37,49.
Person re-identification incorporating appearance feature
Peng Yuqing1,2,Li Wei1,2,Guo Yongfang1
(1.School of Artificial Intelligence and Data Science,Hebei University of Technology,Tianjin 300401,China; 2.Hebei Province Key Laboratory of Big Data Calculation,Tianjin 300401,China)
Abstract: Aiming at the problem of low recognition rate caused by pose, viewpoint and occlusion in person re-identification, a method incorporating appearance feature is proposed. The method designs two network branches to extract global feature and local feature of pedestrians respectively, and the two are fused to obtain the appearance features of pedestrians. Simultaneously the model is optimized by a multi-task learning strategy for both network branches through combining classification loss and metric learning loss. In addition, the model combines with random erasing algorithm to add noise to the dataset for enhancing the robustness of the model. The experimental results show that the proposed method incorporating appearance feature greatly improves the accuracy of person re-ID, with rank-1 reaching 92.82% and mAP reaching 80.51% on the Market1501 dataset, and rank-1 reaching 85.06% and mAP reaching 72.72% on the DukeMTMC-reID dataset.
Key words : person re-identification;feature incorporating;random erasing;multi-task learning

0 引言

行人重識(shí)別(Person Re-identification,Person re-ID)是指跨監(jiān)控設(shè)備下的行人檢索問(wèn)題,即給定一個(gè)監(jiān)控行人圖像,利用計(jì)算機(jī)視覺(jué)技術(shù)在其他監(jiān)控?cái)z像頭拍攝產(chǎn)生的大型圖片庫(kù)中準(zhǔn)確找到該行人圖片,在智能安防、智能監(jiān)控以及智能商業(yè)等領(lǐng)域具有廣泛應(yīng)用。但由于圖片分辨率低、行人姿勢(shì)變化較大、視角變化、遮擋、光照變化、背景雜亂干擾等問(wèn)題,行人重識(shí)別當(dāng)前面臨巨大的挑戰(zhàn)。

行人重識(shí)別方法分為特征提取和相似性度量?jī)蓚€(gè)步驟,傳統(tǒng)的行人重識(shí)別方法將兩個(gè)步驟分開(kāi)研究,只對(duì)其中一個(gè)步驟改進(jìn)、優(yōu)化。特征提取方法主要采用顏色、形狀和紋理等低維視覺(jué)特征來(lái)表達(dá)行人外觀,如RGB直方圖等。模型提取特征后通過(guò)學(xué)習(xí)距離度量函數(shù)進(jìn)行相似性度量。近年來(lái),隨著卷積神經(jīng)網(wǎng)絡(luò)(CNN)的發(fā)展,許多深度學(xué)習(xí)的方法應(yīng)用到行人重識(shí)別中,將特征提取和相似性度量整合為一個(gè)統(tǒng)一的整體,同時(shí)改進(jìn)兩個(gè)模塊學(xué)習(xí)有辨別力的特征。ZHENG L等人[1]提出IDE(ID Embedding)網(wǎng)絡(luò),利用行人身份標(biāo)簽訓(xùn)練ResNet50網(wǎng)絡(luò),進(jìn)行微調(diào)后獲取行人全身特征進(jìn)行身份識(shí)別。SUN Y F等人[2]提出PCB(Part-based Convolutional Baseline)網(wǎng)絡(luò),采用統(tǒng)一分割策略提取細(xì)節(jié)特征,設(shè)計(jì)了RPP(Refined Part Pooling)模塊調(diào)整偏差,增強(qiáng)劃分模塊的一致性,解決了分割導(dǎo)致的行人身體部位不對(duì)齊等問(wèn)題。李聰?shù)热薣3]提出多尺度注意力機(jī)制(Multi-Scale Attention,MSA)的行人重識(shí)別方法,將多尺度特征融合與注意力方法相結(jié)合,使網(wǎng)絡(luò)能自適應(yīng)地調(diào)節(jié)感受野的大小,但此方法忽略相似性度量對(duì)模型的有效性。



本文詳細(xì)內(nèi)容請(qǐng)下載:http://theprogrammingfactory.com/resource/share/2000003597



作者信息:

彭玉青1,2,李  偉1,2,郭永芳1

(1.河北工業(yè)大學(xué) 人工智能與數(shù)據(jù)科學(xué)學(xué)院,天津300401;

2.河北省大數(shù)據(jù)計(jì)算重點(diǎn)實(shí)驗(yàn)室,天津300401)


此內(nèi)容為AET網(wǎng)站原創(chuàng),未經(jīng)授權(quán)禁止轉(zhuǎn)載。