Computer Engineering and Applications ›› 2018, Vol. 54 ›› Issue (20): 219-222.DOI: 10.3778/j.issn.1002-8331.1804-0027

Previous Articles     Next Articles

Pedestrian re-identification based on fine-tuned pre-trained convolutional neural network model

LI Jinming,QU Yi,PEI Yuhao,YI Zejiang   

  1. College of Information Engineerng, Engineering University of PAP, Xi’an 710086, China
  • Online:2018-10-15 Published:2018-10-19

预训练卷积神经网络模型微调的行人重识别

李锦明,曲  毅,裴禹豪,扆泽江   

  1. 武警工程大学 信息工程学院,西安 710086

Abstract: In order to solve the problem of the low robustness and slow convergence in the way of hand-crafted features extraction due to the influence of the different camera angle and distinct illumination, it uses the pedestrian database to adjust the pre-trained convolutional neural network model, and then calculates the similarity degree by Euclidean distance. Experimental results show that the deep pedestrian features are 9.51%, 11.12%, 16.63%, 16.96% better than the hand-crafted features in the mean Average Precision(mAP) and the speed of convergence is faster, it indicates that the deep feature can improve the performance of the pedestrian re-identification.

Key words: pedestrian re-identification, convolutional neural network, pre-trained model, deep features

摘要: 针对行人重识别中传统的人工提取的行人浅层特征因受摄像机角度、光照等外界环境的影响,鲁棒性不好,收敛速度慢的问题,研究使用预训练卷积神经网络模型在行人数据库上进行微调的方法,对行人图片进行特征提取,从而得到高维的深层行人特征,最后通过欧氏距离进行相似性的度量。实验结果证明,深层的行人特征在平均准确度评估标准上,相比于传统的人工设计特征,分别得到了9.51%、11.12%、16.63%、16.96%的提高,收敛速度也变得更快,说明深层特征的行人识别能力更强。

关键词: 行人重识别, 卷积神经网络, 预训练模型, 深层特征