基于全局和局部特征的跨模态行人再识别方法

Cross-modal pedestrians re-identification method based on global and local features

跨模态行人再识别是对同一行人的可见光图像和红外图像之间进行匹配和识别,相对于单模态行人再识别的难度进一步加大。本文采用循环生成对抗网络(Cycle GAN)转换和扩充数据集,尽可能减少行人体态变化带来的影响;在ResNet50网络的基础上引入全局特征对比模块和局部特征模块,减少图像噪声和行人遮挡带来的影响;将交叉熵损失和改进的三元组损失以比例加和的形式作为多损失联合函数,对网络监督训练。实验结果表明,该方法的平均精确度均值和前十位命中率都达到了较高的水平,优于当前多数方法。

Cross-modal pedestrians re-identification is to match and identify the visible light image and infrared image of the same pedestrian, which is more difficult than single-modal pedestrian re-identification. In this paper, Cycle GAN is used to transform and expand the dataset to minimize the impact of pedestrian posture changes. Based on the ResNet50 network, a global feature comparison module and a local feature module are introduced to reduce images noise and the impact of pedestrians occlusion. The cross-entropy loss and the improved triplet loss in the form of proportional summation are used as a multi-loss joint function to supervise the training of the network. The experimental results show that the average precision and the top ten hit rate of the proposed method have reached a high level, which is better than most of the current methods.