Abstract:

To solve the difficulties brought by illumination and viewpoint varieties in video-based person re-identification（re-ID）, a network combining the region-based quality estimation and attribute classification is proposed. Some images are pre-processed by cutting the bottom part. the images are input into convolutional network after being divided into three-part-division, which will predict the quality of each part. the network is trained by combining with the attribute labels of pedestrians to finish the whole re-ID process. Through learning global features and local features of pedestrians, the network can effectively handle occlusions and misalignments, thus achieve comparable results in three public datasets.

Key words: person re-identification, computer vision, video-based, quality estimation, attribute classification, alignment

摘要：

针对基于视频的行人重识别中由于光照与视角变化带来的问题，提出了一种结合局域质量评估网络与行人属性特征的网络。对部分行人图像进行预处理，裁掉部分行人图像的底部；将行人分割成三段通过卷积神经网络对其进行质量评估；结合事先人工标注的行人属性标签，进行训练从而完成重识别的过程。通过学习行人的全局特征和局部特征，能够有效解决行人图像中出现的遮挡和不对齐问题，通过在三个数据集上的结果对比表明方法实现了准确率上的提升。

关键词: 行人重识别, 计算机视觉, 基于视频, 质量评估, 属性识别, 对齐