Abstract: In this paper, a local region selection and local feature extraction algorithm based on deep learning is proposed for human pose variation, alignment and partial occlusion in the person re-identification problem. The algorithm firstly obtains the basic features by the residual convolutional neural network, then extracts the features of different candidate local regions by the multi-scale sliding windows, and groups them according to their coverage area. Each group selects an optimal local feature and merges the global features to obtain the final feature representation. The experimental results show that the local features extracted by this method are more representative and improve the accuracy of person re-identification.

Key words: person re-identification, deep learning, local feature, multi-scale sliding window

摘要： 针对行人重识别问题中人体姿态变化、对齐及部分遮挡等情况，提出了一种基于深度学习的局部区域选择和局部特征提取算法。算法首先利用残差卷积神经网络获取基本特征，然后利用多尺度的滑动窗口提取不同候选局部区域特征，并按照覆盖区域进行分组，每组选择一个最优局部特征，并融合整体特征得到最终特征表达。实验结果表明，通过该方法提取的局部特征具有更好的表达能力，提高了行人重识别的精确度。

关键词: 行人重识别, 深度学习, 局部特征, 多尺度滑动窗口