摘要： 行人重识别是利用计算机视觉技术判断图像或者视频序列中是否存在特定行人的技术。受行人姿态、遮挡、光照变化等因素的影响，传统的行人重识别方法中特征的表达能力有限，导致准确率降低，提出一种融合不同尺度对比池化特征的行人重识别方法。利用残差网络ResNet50提取行人图像的多尺度特征，在网络的不同层次上，通过对输入的特征进行全局平均池化和最大平均池化，将每组平均池化特征和最大池化特征相减，对相减得到的差异特征与最大池化特征进行相加，获得具有强判别性的对比池化特征。在此基础上，利用三元组损失和交叉熵损失联合优化模型，提高模型的泛化能力，同时采用重排序技术优化网络性能。实验结果表明，该方法在Market1501和DukeMTMC-reID数据集上的首位命中率分别达到96.41%和91.43%，平均精度均值为94.52%和89.30%，相比SVDNet、GLAD和PCB等方法，其行人重识别的准确率较高。

关键词: 行人重识别, 多尺度特征, 对比池化特征, 特征融合, 深度学习

Abstract: Person re-identification is a technology that uses computer vision to identify whether there are specific people in images or video sequences.Owing to the influence of the person's posture, occlusion, illumination change, and other factors, the expression ability of features in traditional person re-identification methods is limited, resulting in reduced accuracy.A person re-identification method that combines the contrast pooling feature at different scales is proposed.The residual network ResNet50 is used to extract the multi-scale features of the images of the people.At different levels of the network, through the global average pooling and maximum average pooling of the input features, each group of average pooling features and maximum pooling features are subtracted, and the subtracted difference features and maximum pooling features are added to obtain highly discriminative constrast pooling fusion features.On this basis, the triplet loss and cross entropy loss joint optimization model is used to improve the generalization ability of the model;the reordering technology is used to optimize the network performance.The experimental results show that the first ranking of this method on the Market1501 and DukeMTMC-reID datasets are 96.41% and 91.43%, respectively, and the average accuracies are 94.52% and 89.30%, respectively.Compared with SVDNet, GLAD, and PCB, this method has a higher person re-identification accuracy.

Key words: person re-identification, multi-scale feature, contrast pooling feature, feature fusion, Deep Learning(DL)

中图分类号: 

• TP391.41