基于二值特征描述符的目标表示及其应用

The robust and fast feature descriptor is the key problem for the object detection,tracking and recognition, which is the hot topic in the field of computer vision and pattern recognition. The object is be represented by the real number vector in the classic real-value feature descriptor. The binary feature descriptor explores the binary vector to descriptor the object, which can matched fast by the hamming distance. The binary feature descriptor is real-time and low memory, which is the development trend of the feature descriptor.The feature descriptor based on binary coding will be researched in this project. The hashing projection is learned by keeping the object structure information in the Euclidean space. The object is mapped to the hamming space and the memory is decreased for the feature descriptor. The distance metric is learned by the machine learning and fast approximate nearest neighbor searching is be researched. The binary feature descriptor is used to object detection and tracking, which will play an important role in the real system. Therefore, the implementation and the expected results of the project will provide new ideas and new methods for the effective and efficient object representation model, which has a broad theoretical basis for the application of prospect research.

建立鲁棒高效的目标特征描述符，是实现目标检测、跟踪与识别等视觉任务系统的关键问题，目前是计算机视觉和模式识别领域的研究热点。与传统的实值特征描述符用实数向量表示目标相比，二值特征描述符用一串二值向量表示目标，通过汉明距离实现快速匹配，因此它具有快速、低存储的特点，是特征描述符的发展趋势。本项目将研究基于二值编码的特征描述符，寻找保持目标内部结构信息的哈希变换，将目标从欧式空间投影到汉明空间，降低特征描述符的存储空间，结合不同视觉任务系统的先验知识和约束条件，学习与目标语义相关的距离度量，研究汉明空间的快速最近邻搜索策略，实现实时匹配，并将其应用于目标检测和目标跟踪中，为实际应用奠定良好的基础。因此，本项目的开展和预期成果将为高效的目标表示模型提供新思路和新方法，是一项既有理论意义又有广阔应用前景的课题。

按照要求，在原有研究的基础上，实现了基于判别式跟踪的方法，在图像目标特征提取上，采用全局积分直方图特征，并融合局部特征，频域特征和梯度特征构建基于深度学习的特征模型，基于多特征和目标提议框给出的多模型进行尺寸变化下的特征模型初步构建。随后基于局部特征加权方式进行二值域的映射，形成多特征多尺寸的多表征模型。再利用滤波方式构建频域或核相关滤波进行回归计算。在跟踪过程中，基于提议框，采用自适应支持向量机进行样本分类和模型更新，适应尺寸，光照以及遮挡等变化。