基于多源跨域信息融合的行人再识别关键算法研究

Person Re-identification focuses on the how to precisely retrieve the surveillance image frames containing the same person, and has been broadly used in security surveillance, customer identification, personal behavior analysis, and etc. Traditional person Re-Identification algorithms have the following limitations: 1) They have a poor ability in cross-domain adaptation and generalization; 2) The performance of the person Re-ID algorithms based on image processing is easy to be affected by lightness, view angle, occlusion, and image quality; 3) There exists performance bottle-neck while processing big video data in the centralized mode. Regarding the above challenging problems, this research project aims to solve the person Re-ID problems in large-scale camera networks. Specifically, we will focus on the research of the unsupervised transfer learning algorithm based on fusion of multi-source information, which performs the incremental training based on the unlabeled video data collected in target domains to improve cross-domain adaptation ability and integrates with multi-modal information to eliminate the limitation of image processing. Meanwhile, we will propose the pedestrian image augmentation algorithm based on generative adversarial nets to eliminate the difference of image styles in different domains. Furthermore, we will also propose the distributed person Re-ID algorithm based on the collaboration of multiple intelligent terminals with autonomous learning ability to solve the centralized bottle-neck problem.

“行人再识别”主要研究如何在多个监控视频中准确检索包含某个行人的图像帧，在城市安防监控、商场客户辨识、个体行为分析等方面有广泛的应用价值。传统的行人再识别算法存在以下局限性：1）跨域适配性较低，泛化能力较弱；2）存在图像处理的局限性，容易受光线、角度、遮挡、图像质量的影响；3)集中处理海量监控视频，存在性能瓶颈，扩展代价昂贵。围绕上述挑战问题，本课题面向大规模摄像头网络下的行人再识别应用，重点研究融合多源多模态信息的无监督迁移学习算法，充分利用目标域的海量无标签视频数据进行增量训练，从而提升行人分类器的跨域适配性，并通过融合多模态信息来弥补图像处理的局限性。同时我们将研究基于生成对抗网络的跨域行人图像增强算法，减少不同摄像头网络所获取的行人图像的风格差异性。在此基础上，我们还将研究具有自主学习能力的多智能终端分布式协同行人再识别算法，解决集中处理的性能瓶颈问题。

本课题面向大规模摄像头网络下的行人再识别应用场景，从如下三个方面开展研究：1）研究融合多源多模态信息的无监督迁移学习算法，提升行人分类器的跨域适配性；2）研究基于生成对抗网络的跨域行人图像增强算法，减少跨域数据差异性；3）研究具有自主学习能力的多智能终端分布式协同行人再识别算法，解决集中处理的性能瓶颈问题。实验表明，融合多模态信息可显著提高行人再识别算法的性能，较大幅度提升域适应性。基于该项目研究的核心算法，我们设计并实现了融合多模态时空大数据的视频监控系统原型，相对于现有的单纯依赖图像检索技术的智能监控系统而言，该系统通过融合时空数据较大程度提高监控视频的应用效率。