基于流形的多尺度特征人脸检测和跟踪方法研究

人脸检测跟踪系统在公共安全维护、疫情防控等领域中发挥着十分重要的作用，现有的检测跟踪方法存在数据需求大、特征维度高、非结构性强等不足，无法实现对人脸的快速准确检测。本项目针对公共安全视频数据非结构化、先验知识匮乏、特征稀疏等特点，围绕人脸流形的空间构建、人脸图像数据点集聚类以及流形上人脸目标运动轨迹预测，具体开展以下内容研究：1）多尺度特征维度下的人脸图像流形空间表示，以降低人脸流形空间构建复杂度；2）高维人脸图像流形空间数据点集降维，通过描述流形上数据点局部概率分布特征信息，提高人脸图像数据点映射到低维空间时的准确性；3）黎曼子流形中的人脸目标运动轨迹预测，利用循环神经网络实现多平台公共安全视频的人脸轨迹预测，提升目标轨迹预测准确率与效率。本项目的研究成果可大大提高公共安全领域的人脸检测效率，为整体提升社会治安防控能力水平提供可靠的技术支撑。

Face detection tracking system plays a very important role in many fields, such as public security maintenance, epidemic prevention and control. Existing detection and tracking methods cannot detect and recognize human faces fast and accurately due to their defects of high data demand, high feature dimension, and strong non-structure. Therefore, Focusing on the problem of public security video data being unstructured, short of prior knowledge and inconspicuous features, this project will be carried out from three aspects of spatial construction of face manifolds, data points clustering of face images and prediction of the movement trajectory of face targets on manifolds, which are specified as follows: 1)Manifold space representation of face images in multi-scale feature dimensions to reduce the complexity of face manifold space construction. 2)Dimensionality reduction of data point set of high-dimensional face image manifold space to improve accuracy of mapping face image data points to low-dimensional space by describing local probability distribution feature information of data points on manifolds.3)Prediction of facial target motion trajectory in Riemannian manifolds, in which the recurrent neural network is applied to the face trajectory prediction of multi-platform public security video for improving accuracy and efficiency of target trajectory prediction. The research results of this project can greatly improve the efficiency of face detection in the field of public security and provide reliably technical support for the overall improvement of social security prevention and control capabilities.