传统行人重识假设行人在监控场景内不换衣服，但这在实际中是无法保证的。为此，本项目提出研究换装行人重识别。.本项目基于具有强大图像生成能力的生成对抗网络，研究一个联合学习框架，提取具有判别性的衣服无关、身份有关特征，进行鲁棒的换装行人重识别。具体内容包括：(1)利用从RGB和灰度图像提取的“与”特征图，通过生成对抗网络，实现特征图超分辨率化，使得新生成的特征图具有更好的分类效果，从而获得鲁棒的细粒度特征；(2)在随机噪声作用下，利用从灰度图像获得的行人特征生成保持身份的RGB换衣行人图像，从而实现训练图像扩充，促使模型学习到衣服无关的体态特征，并从同一行人图像的体态特征池中提取行为特征；(3)将以上内容整合到一个联合学习框架，采用合适的优化方案，使得各部分彼此协调并互为利用，仅在身份信息的指导下完成换装行人重识别。.本项目将为换装行人重识别提供一个新的研究思路，具有重要的理论意义和应用价值。

Traditional person re-identification assumes that pedestrians do not change their clothes in a surveillance scene. But in practice it is impossible to be guaranteed. For this reason, this project proposes to study the cross-clothes person re-identification..Based on the Generative Adversarial Net(GAN) which has a strong image generation ability, this project will study a joint learning network to extract discriminative features independent with clothes information for identifying an individual, and to carry out robust cross-clothes person re-identification. There are three specific contents are included. Firstly, by using the "and" feature maps extracted from RGB and the corresponding gray images, the feature maps with super-resolution can be obtained with GAN. The newly generated feature maps have a better classification effect and the robust fine-grained features will be extracted. Secondly, the RGB pedestrian images maintaining identity but changing clothes are generated for augmenting the train dataset with the help of the random noise and pedestrian feature maps obtained from the gray images. It promotes the model to learn the body features independent with clothes. Then, we can extract the behavior features from the pool of body features with the same identity. Thirdly, the above two parts are integrated into a joint learning framework. By adopting some appropriate optimization scheme, it will make each part coordinate with each other and make use of each other. Only by the guidance of identity information, the cross-clothes person re-identification can be conducted. .This project will provide a new research idea for the cross-clothes person re-identification. And this study has very important theoretical significance and application value.