基于迭代投影向量学习的行人重识别

Person re-identification based on iterative projection vectors learning

为降低行人重识别问题中投影矩阵学习的时间复杂度和过拟合风险,提出一种带有正则化因子的迭代投影向量学习的算法。利用行人图像特征空间分布的迭代更新策略,结合随机梯度下降方法,对更新后的特征进行学习,得到若干列满足目标精度的相互正交的投影向量;将这些投影向量相乘得到最终的投影矩阵,有效降低运算复杂度。将该算法与其它常用的方法在现有公共数据集上进行比较,比较结果表明,该算法的识别效率明显高于其它方法,训练时间少于其它常用方法。

To release the risk of overfitting and time complexity of projection matrix learning in person re-identification,an algorithm based on stochastic gradient descent iterative projection vectors learning was proposed.The iterative update strategy of person’s image feature space distribution was designed,and with the method of stochastic gradient descent,several orthogonal projection vectors were learnt,which met the objective accuracy.By multiplying these vectors to get final projection matrix,this method can effectively reduce the computational complexity.By comparing to the other common methods on the existing public datasets,experimental results show that the proposed method has higher recognition efficiency and shorter training time than other common methods.