DDN前景提取结合映射模型学习的行人再识别  

Person Re-Identification Based on DDN Foreground Extraction and Mapping Model Learning

作　　者：胡正平[1] 张敏姣 李淑芳 HU Zheng-ping;ZHANG Min-jiao;LI Shu-fang(School of Information Science and Engineering, Yanshan University, Qinhuangdao, Hebei 066004, Chin)

机构地区：[1]燕山大学信息科学与工程学院,河北秦皇岛066004

出　　处：《信号处理》2018年第5期572-582,共11页Journal of Signal Processing

基　　金：国家自然科学基金面上项目(61771420);河北省自然科学基金项目(F2016203422)

摘　　要：随着视频监控设备的广泛应用,行人再识别成为智能视频监控中的关键任务,具有广阔的应用前景。该文提出一种基于深度分解网络前景提取和映射模型学习的行人再识别算法。首先利用DDN模型对行人图像进行前景分割,然后提取前景图像的颜色直方图特征和原图像的Gabor纹理特征,利用提取的行人特征,学习不同摄像机之间的交叉映射模型,最后通过学习的映射模型将查寻集和候选集中的行人特征变换到一个特征分布较为一致的空间中,进行距离度量和排序。实验证明该算法能够提取较为鲁棒的行人特征,可克服背景干扰问题,行人再识别匹配率得到有效的提高。With the wide application of video surveillance equipment,person re-identification has become a critical task in intelligent video surveillance and has a wide application prospect. This paper presents a person re-identification algorithm based on deep decompositional network foreground extraction and mapping model learning. Firstly,the DDN model is used to segment the pedestrian images,then the color histogram of the foreground images and the Gabor texture feature of the whole original images are extracted. The cross-view mapping model between the cameras will be studied by using the extracted pedestrian features. Finally,the person features of the search set and the candidate set are transformed into a same space where the feature distribution is more consistent,after that the distance metric learning and sorting are performed. Experiments show that the model can extract more robust person features,overcoming the back ground interference,and the person re-identification matching rate has been effectively improved.

关 键 词：行人再识别 映射模型学习 深度分解网络 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]