行人重识别目标中心编码外观模型的研究

Person re-identification is a challenging task in computer vision due to the significant variations on viewpoints, poses and illuminations among different cameras. In addition, background clutters and occlusions increase the difficulties. Therefore, it is critical important to design good performance appearance model for improving the accuracy of person re-identification. However, the traditional appearance models rely on excessive vision features in complicated styles and the accuracy of person re-identification is not very high. With this motivation, in this project we present a new Object Centric Coding (OCC) appearance model for person re-identification. It includes the following aspects..First, the mask containing the genuine body is obtained for each person via a foreground extraction tragedy. The purpose of this phase is to separate the appearance of body parts from the rest of the scene, which enables the following coding descriptor to focus on the sole person..Second, by removing the background region, discriminant vision features are learnt followed by Locality-constrained Linear Coding (LLC), which could significantly improve the performance of low level features..Third, to obtain a proper evaluation for the OCC, we validate the OCC model on the challenging public datasets using multi metric learning methods. We believe that a good appearance model is likely to be validated by different metrics and produces consistent and high performance..Finally, this project aims to design a robust and discriminant appearance model for boosting the accuracy (matching rates) in person re-identifications. Also it presents a new solution for how to improve the performance of person re-identification.

行人重识别是计算机视觉领域中具有挑战性的问题，主要表现在不同摄像头下的行人在视角、姿态、光照上具有较大的差异以及目标区域容易受背景的干扰。因此为目标行人建立良好的外观模型对提高行人重识别的准确性具有重要的意义。针对目前传统外观模型方法过多依赖于复杂视觉特征的组合且识别率不高的情况。本项目提出目标中心编码的外观模型以进一步提高行人重识别的准确性。主要研究内容包括：（1）研究目标行人前景区域的提取策略，以解决行人目标“不对齐”现象和减少背景的干扰。（2）在前景目标区域提取的基础上，提出使用目标中心编码的方法以提高行人重识别的准确性。（3）针对目前对外观模型评价方法单一的问题，提出使用多种距离学习的方法来评价外观模型的性能。本项目的研究为提高行人重识别准确性提供了新的思路。

本项目拟从外观模型建立的角度出发以减少视角、姿态、光照上的差异和背景干扰，从而进一步提高行人重识别的准确性。在项目的执行过程中，（1）研究了行人前景区域的提取策略，利用SCA分析提取前景区域，在前景区域图像上按分块的方式提取视觉特征，并进行局部约束编码以提升特征的性能；（2）研究了基于块的行人重识别方法，尝试用卷积神经网将前景区域和背景区域分离以减少背景干扰，同时采用二维高斯模型的方法针对行人图像主要区域进行特征提取，以提高行人重识别的准确性；（3）考虑到距离学习在一定程度上可以看作是特征学习，因此提出了基于最小二乘判别分析距离学习的行人重识别方法。与传统距离学习方法不同，它只需要采用匹配的行人对进行训练，首先利用相似的训练样本对进行子空间学习，得到低维空间的一个有限目标点集合。然后将训练样本通过线性回归的方式与目标点进行拟合，并采用最小二乘的方法得到线性回归的投影矩阵，该矩阵与其转置的乘积就是最后的距离矩阵，该方法具有较高的效率且可以扩展到增量学习以及核学习。所有的研究实验均在公开发布的数据集上进行，并表现出较好的实验结果，在一定程度上提高了行人重识别的准确性。