DGD卷积神经网络行人重识别  

Person re-identification on DGD convolutional neural networks

作　　者：杨忠桃 章东平 杨力 井长兴 

机构地区：[1]中国计量大学信息工程学院,浙江杭州310018

出　　处：《中国计量大学学报》2017年第4期504-508,共5页Journal of China University of Metrology

基　　金：浙江省自然科学基金资助项目(No.LY15F020021);浙江省科技厅公益性项目(No.2016C31079)

摘　　要：行人重识别在视频监控领域是一个非常具有挑战性的问题,不同的摄像头位置角度、光照等因素会使同一行人的图像差异较大.文章提出一种DGD(Domain Guided Dropout)卷积神经网络(CNN)与样本相对距离结合的行人重识别算法:首先,通过卷积神经网络来提取来自多个域的数据中具有一般性及鲁棒性的特征;其次,通过计算各个特征样本之间的相对距离来筛选出更具有一般性及鲁棒性的特征;最后,比较筛选出的特征间的欧氏距离进行重识别.实验结果表明,该算法能够提高行人重识别的效率.Person re-identification is challenging task in intelligent video surveillance because of the various differences from illumination to camera angles.We proposed an algorithm combining multiple domain guided dropout（DGD）convolutional neural networks（CNN）with sample relative distance.The representative and robust feature of multiple domain data was learned by using DGD convolutional neural networks and filtered by calculating the relative distance between features.By comparing the Euclidean distance between feature samples,re-identification was processed.The experimental results show that the proposed algorithm can improve the efficiency of human re-identification.

关 键 词：卷积神经网络 样本相对距离 欧氏距离 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]