多姿态图像生成的行人重识别算法研究

Research on Person Re-Identification Algorithm Based on Multi-Pose Image Generation

不同行人的高度相似性以及相同行人外观姿态的差异性,使得不同摄像头下的行人重识别面临严峻的挑战。生成对抗网络可以合成新的图像,被认为是解决行人姿态变化的主要技术手段。提出一种基于多姿态图像生成的行人重识别算法,利用生成对抗网络生成不同姿态的行人图像,通过归一化消除姿态的影响,从而大幅度提升行人重识别的整体性能。该行人重识别算法包括多姿态行人图像生成、不同姿态的行人特征提取与融合、距离度量和重排序三部分内容。在Market-1501数据集和DukeMTMC-ReID数据集上的实验证实了所提出算法的有效性,通过与state-of-the-art行人重识别方法比较,展示了多姿态图像生成方法在行人重识别任务中的优越性,同时表明生成行人图像的特征与原始图像的特征是相互补充的。

The high similarities of different person and great diversities in appearance of the same person pose grand challenges to person re-identification(ReID).Generative adversarial networks(GAN)can synthesize new images and are considered to be the main technical means to solve the problem of pose variations.This paper proposes a multi-pose image generation algorithm for person ReID,which adopts GAN to generate person images with different poses.By eliminating the influence of pose through normalization,the overall performance of person ReID can be improved greatly.The proposed person ReID algorithm includes three parts:multi-pose person image generation,feature extraction and fusion for person in different pose,distance measurement and reranking.Experiments on the Market-1501 dataset and Duke MTMC-ReID dataset demonstrate the validity of the proposed algorithm and show the superiority of multi-pose image generation for improving person ReID through comparisons with the state-of-the-art methods.At last,it shows that the features of the generated images and the original images are complementary.