基于图卷积的跨模态行人重识别方法

A cross⁃modality pedestrian re⁃identification method via graph convolution

近年来,红外-可见光的行人重识别在视频监控、网络刑侦等领域应用广泛,这项任务的目的是实现RGB摄像机和红外摄像机下出现的同一行人的匹配。由于行人图像在RGB模态和红外模态下存在较大差异,因而使得该项任务具有一定的挑战性。文中提出一种基于图卷积的跨模态行人重识别方法,同时提出一种新颖的异心三元组损失函数,用于更好表征行人特征。该方法首先对水平切割方法进行改进,在此基数上以局部特征和全局特征为节点构建图卷积神经网络,并利用构建的图卷积神经网络学习图像结构化特征;然后,引入了一种全新的异心三元组损失函数,并结合Softmax损失函数进一步提高模型性能。两个公开数据集上进行的对比实验、消融实验以及可视化实验结果验证了文中所提方法的卓越性能。

In recent years, infrared and visible pedestrian re⁃identification has been widely used in manyfields, like video surveillance, network criminal investigation, etc. The purpose of infrared and visiblepedestrian re⁃identification is to match the images of the same pedestrian that exists in both a RGB camerafootage and an infrared camera footage. It is a challenging task due to the large discrepancy between RGBimages and infrared images. In this paper, a cross⁃modality pedestrian re⁃identification method based ongraph convolution is proposed, and a novel heterogeneous triplet loss function is proposed to obtain personcharacteristics. First, the network improves the horizontal cutting method, constructs a graphconvolutional neural network with local features and global features as nodes, and utilizes the constructedgraph convolutional neural network to obtain image structural characteristics. Second, a newheterogeneous triplet loss is introduced as a new function that is combined with the Softmax loss functionto further improve the network performance. The results of comparative experiments, ablation experimentsand visualization experiments conducted on two public datasets demonstrate the excellent performance ofthe proposed method.