Attribute-Based Fused Feature for Person Re-identification

2018 3rd International Conference on Insulating Materials, Material Application and Electrical Engineering

材料科学;无线电电子学;电工学

Zhao, Qingqing^1 ; Chen, Changhong^1 ; Song, Wanru^1 ; Liu, Feng^1

School of Nanjing University of Posts and Telecommunications, Nanjing, China^1

关键词: Attribute-based;    Correction mechanism;    Foreground images;    Low-level features;    Person re identifications;    State-of-the-art methods;    Sub classifiers;    Time complexity;   

Others  :  https://iopscience.iop.org/article/10.1088/1757-899X/452/4/042014/pdf
DOI  :  10.1088/1757-899X/452/4/042014

学科分类：材料科学（综合）

来源: IOP

【 摘 要 】

In this paper, we propose an attribute-based fused feature which combined low-level and attribute features for person re-identification. Pedestrian parsing and different low-light image enhancement methods are adopted before feature extraction and Random Forest (RF) classifier is used as the basic classifier. Firstly, each of the image is divided into eight strips, and the strips are composed into four parts: head, up-body, low-body and bag areas. Above these four parts, the gradient-LOMO (Local Maximal Occurrence) features are extracted separately to form sub-classifiers features. The sub-classifiers of the independent and different dimensional attributes are respectively trained and then fused to generate a classifier, which describes 21-dimensional attribute features and is combined with the correction mechanism. Then, this classifier can be used to predict attributes on unmarked datasets. In this way, the time complexity of manual marking can be reduced effectively. Finally, the low-level features, which extracted from original and foreground images, are combined with attributes as the fused feature. The results of experiments on three common person re-identification datasets (VIPeR, PRID450S and GRID), indicate that our attribute-based fused feature exhibits prominently performance which outperforms the state-of-the-art methods for person re-identification.