高级检索

    基于距离中心化与投影向量学习的行人重识别

    Person Re-Identification Based on Distance Centralization and Projection Vectors Learning

    • 摘要: 现有的基于投影的行人重识别方法具有训练时间长、投影矩阵维数高、识别率低等问题.此外在建立训练集时,还会出现类内样本数目远少于类间样本数目的情况.针对这些问题,提出了基于距离中心化的相似性度量算法.在构建训练集时,将同一组目标群体特征值中心化,利用中心特征值来构建类间距离,而类内距离保持不变.这样使得类内类间样本数目接近,可以很好地缓解类别不平衡所带来的过拟合风险.另外在学习投影矩阵时,利用训练集更新策略,学习若干组投影向量,使得到的投影向量近似正交,这样既可以有效减少运算复杂度和存储复杂度,又可以使得学习到的投影向量能够通过简单的相乘近似得到原来的投影矩阵.最后,在学习投影向量时采用共轭梯度法,该方法具有二次收敛性,能够快速收敛到目标精度.实验结果表明:提出的算法具有较高的效率,在不同数据集上的识别率都有明显的提升,训练时间也比其他常用的行人重识别算法要短.

       

      Abstract: Existing projection-based person re-identification methods usually suffer from long time training, high dimension of projection matrix, and low matching rate. In addition, the intra-class samples may be much less than the inter-class samples when a training data set is built. To solve these problems, this paper proposes a distance-centralization based algorithm for similarity metric learning. When a training data set is to be built, the feature values of a same target person are centralized and the inter-class distances are built by these centralized values, while the intra-class distances are still directly built from original samples. As a result, the number of intra-class samples and the number of inter-class samples can be much closer, which reduces the risk of overfitting because of class imbalance. In addition, during learning projection matrix, the resulted projection vectors can be approximately orthogonal by using a strategy of updating training data sets. In this way, the proposed method can significantly reduce both the computational complexity and the storage space. Finally, the conjugate gradient method is used in the projection vector learning. The advantage of this method is its quadratic convergence, which can promote the convergence. Experimental results show that the proposed algorithm has higher efficiency. The matching rate can be significantly improved, and the time of training is much shorter than most of existing algorithms of person re-identification.

       

    /

    返回文章
    返回