基于对抗多目标学习的行人重识别方法研究

Person re-identification is one of the key technologies in video surveillance and is a hot topic in the field of computer vision. Existing person re-identification methods are limited by the characteristics of large data and small samples and are susceptible to factors such as resolution, occlusion, pose, illumination, and similar person interference, resulting in false alarm, missed detection, and poor generalization performance. This project intends to study person re-identification methods based on adversarial multi-objective learning. The main research contents are as follows: in order to solve the problem that person re-identification is easily influenced by resolution, occlusion, pose and illumination, this project studies the siamese generative adversarial network for diversity samples generation; to solve the problem that person re-identification is easily disturbed by similar pedestrians, the person re-identification siamese generative adversarial network will be developed; to solve the difficulty of generative adversarial network training, this project studies the method of network parameter learning and structure pruning based on multi-objective optimization. The research results can theoretically realize the construction of siamese generative adversarial networks model for person re-identification and propose a model optimization method based on multi-objective learning. It can also reduce the false alarm rate and missed detection rate of person re-identification while improving the generalization performance of person re-identification in technically, so as to provide technical support for application of person re-identification technology in the fields of video surveillance and security.

行人重识别是视频监控中的关键技术之一，是当前计算机视觉领域的研究热点。现有的行人重识别方法受大数据小样本特性的制约以及分辨率、遮挡、姿态、光照、相似行人等因素的干扰存在虚警、漏检和泛化性能差的问题。本项目拟研究基于对抗多目标学习的行人重识别方法，主要研究内容包括：针对行人重识别受分辨率、遮挡、姿态、光照变化影响的问题，研究面向多样性样本生成的孪生生成式对抗网络模型；针对行人重识别易受相似行人干扰的问题，研究面向相似样本学习的行人重识别孪生生成式对抗网络模型；针对生成式对抗网络训练困难的问题，研究基于多目标优化的参数学习和结构修剪方法。研究结果可在理论上构建针对行人重识别任务的孪生生成式对抗网络模型和提出基于多目标学习的模型优化方法，技术上降低行人重识别的虚警率、漏检率同时提升行人重识别方法的泛化性能，从而为视频监控和安防等领域中行人重识别应用提供技术保障。

本项目以视频安防为应用背景，针对行人重识别易受分辨率、遮挡和相似行人干扰等问题进行研究，具体包括基于生成式对抗网络模型的多样性样本扩充方法、基于深度判别特征学习的目标识别方法、基于多目标优化的参数和结构联合学习方法等内容。在多样性样本扩充方面：针对图像生成存在外观和纹理信息丢失的问题，提出了高分辨图像生成的孪生生成式对抗网络模型；针对相似性人干扰问题，提出了基于GAN以及中心约束三元损失的行人搜索方法；针对监控视频相邻帧信息强相关特性，提出了基于半监督自适应自步学习的行人重识别方法。在深度判别特征学习方面：针对行人图像模态不一致问题，提出了基于跨模态注意力一致性网络的行人重识别方法；针对行人图像姿态不统一问题，提出了基于可解释注意力部件模型的行人重识别方法；针对目标背景复杂多变问题，提出了基于互补特征学习的目标重识别方法；针对视频监控数据标注困难和摄像头不一致问题，提出了基于注意和联合优化的无监督跨域行人重识别。在多目标参数和结构联合学习方面：针对残差网络计算复杂度高问题，提出了基于进化多目标优化的残差网络结构修剪方法；针对深度卷积神经网络处理小数据集容易过拟合问题，提出了基于多目标优化的深度神经网络结构修剪方法；针对深度神经网络模型训练过程中超参数难以设置的问题，提出了基于协同进化深度神经网络参数学习方法。本项目通过采用多样性样本扩充、深度判别特征学习、深度神经网络模型修剪等一系列技术手段有效地提升了行人重识别性能，从理论丰富了行人重识别的方法，在实验上提升了行人重识别的准确率，为视频监控和安防等领域中行人重识别应用提供技术支撑。