摘要
当前基于稀疏表示的行人再识别都是通过松弛l_0正则项为l_1正则项以达到逼近l_0范数稀疏性的目的.在满足有限等距性质(RIP)条件下,l_1和l_0具有等价性,然而在具有杂乱背景、物体遮挡等众多干扰因素的行人再识别任务中,却很难满足RIP条件.因此,文中提出混合l_2/l_(1/2)范数的组稀疏表示方法,通过将gallery集中同一行人图像序列视为一组,利用l_2范数约束组内结构,l_(1/2)范数约束组间结构,对遮挡和杂乱背景等干扰因素具有更高的鲁棒性.为了进一步增强模型的判别性,引入人体结构约束,将行人图像划分为若干近邻块区域,针对每一区域分别构造适应性的混合l_2/l_(1/2)范数的组稀疏模型,最终融合全部稀疏模型得出再识别结果.在当前具有挑战性的2个多行人图像序列数据集PRID 2011和iLIDS-VID上的实验验证文中方法的有效性.
In the existing person re-identification approaches based on sparse representation, l1 regularization is generally utilized to approximate lo-norm sparsity. Under the restricted isometry property （RIP） conditions, the l0 regularization is equivalent to the l1 regularization. However, it is difficult to meet the RIP conditions in person re-identification with many disturbing factors, such as cluttered background and object occlusion. In this paper, a group sparse representation method with mixed l2/l1/2- norm is proposed. The identical person image sequence in the gallery is regarded as a group, the intra- group structure is constrained by l2-norm, and the inter-group structure is constrained by l1/2-norm. The resulting model is more robust to the occlusion and cluttered backgrounds. The human body structure constraint is introduced to further enhance the discriminability of the proposed model. The person image is divided into several neighboring block regions. An adaptive sparse model of mixed l2/l1/2 norm is constructed for each region. Finally, the several sparse models are merged to identify persons. Experiments on PRID 2011 and iLIDS-VID datasets verify the effectiveness of the proposed model.
作者
李小宝
郭立君
张荣
洪金华
LI Xiaobao;GUO Lijun;ZHANG Rong;HONG Jinhua(Faculty of Electrical Engineering and Computer Science,Ningbo University,Ningbo 315211)
出处
《模式识别与人工智能》
EI
CSCD
北大核心
2018年第9期773-785,共13页
Pattern Recognition and Artificial Intelligence
基金
浙江省自然科学基金项目(No.LY17F030002)
浙江省公益技术研究计划项目(No.LGF18F020007)资助~~
关键词
行人再识别
稀疏表示
组稀疏
范数
Person Re-identification
Sparse Representation
Group Sparse
Norm
