利用局部监督的跨模态行人重识别研究

跨模态行人重识别技术旨在从非重叠视域不同模态的摄像头捕获的行人图像中，识别出特定行人，行人图像间存在巨大的跨模态差异以及模态内部差异，导致识别率不高。为此，提出了一种利用局部监督的跨模态行人重识别方法(LSN)。首先将可见光图像转换成与红外图像更为接近的灰度图像，在图像层面缓解跨模态的差异，并使用共享参数的双流网络，提取具有判别性的共享特征，在特征层面缓解跨模态差异；其次，设计了局部监督网络，增强了对背景、遮挡等噪声的鲁棒性，缓解了模态内部差异；最后，设计了跨模态分组损失、联合身份损失对网络进行约束。实验结果显示，在SYSU-MM01数据集上，评价指标rank-1和mAP分别达到了53.31%、50.88%;在RegDB数据集上，达到了73.51%、68.55%,实验结果优于同类方法，验证了该方法的有效性和先进性。

Cross-modality person re-identification using local supervision

Cross-modality person re-identification is a key technology to achieve 24 h×7 intelligent security. The technique aims to identify specific pedestrians from pedestrian images captured by cameras with different modalities in non-overlapping fields of view. There are huge cross-modality differences between pedestrian images as well as intra-modality differences, resulting in poor recognition rates. In order to solve this problem, this paper proposed a cross-modality person re-identification method using local supervision(LSN). Firstly, it converted the visible images into grayscale images that are closer to the infrared images to mitigate the cross-modality differences at the image level, and extracted discriminative shared features using a two-stream network with shared parameters to mitigate the cross-modality differences at the feature level. Secondly, it designed a local supervision network to enhance the robustness to background, occlusion and other noises and mitigate the intra-modality differences. Finally, it designed a cross-modality group loss in combination with the identity loss to constrain the network. The experimental results show that the evaluation metrics rank-1 and mAP reach 53.31% and 50.88% on the SYSU-MM01 dataset, and 73.51% and 68.55% on the RegDB dataset, respectively. The experimental results outperform similar methods, which verify the effectiveness and advancement of the proposed method.