基于视频的人脸检测与对齐算法研究

发布时间：2018-06-07 11:18

本文选题：人脸检测 + 人脸对齐　；参考：《吉林大学》2017年硕士论文

【摘要】：人脸对齐主要用来定位人脸关键部位的特征点,即在人脸区域内,对人脸的眼睛、鼻子等局部关键位置和轮廓进行描述。人脸检测是人脸对齐的基础,其主要作用是在图像中找到人脸位置。人脸对齐可以给出比较细致的人脸特征点分布,因此被广泛应用到人脸识别、人脸动画和姿态估计等方面。在实际应用中,不仅要求算法要具备较高的精确性,还需要算法具有较好的实时性。为了更好的将人脸检测与对齐算法应用到生活当中,本文将提升算法的处理速度与准确性作为主要研究内容。本文的主要研究工作有以下两个方面:1.在人脸检测方面,传统的Viola-Jones人脸检测算法(简称:VJ人脸检测算法)需对视频的每帧都进行整体的遍历来检测人脸,为了减少无效遍历的时间消耗,本文利用了视频帧间的连续性和相似性对其进行改进,提出了人脸位置预测算法。本算法根据人脸的多少将视频划分为少人脸视频和多人脸视频,并提出了不同的人脸位置预测方法,少人脸视频采用回归算法预估出将要检测的视频帧中人脸的可能出现范围,然后在预估范围内检测人脸,这样检测所需遍历的区域由整个视频帧缩小到预测的位置周围,从而降低时间消耗。多人脸部分采用SVM响应图跟踪视频中通过VJ算法检测得到的人脸目标,并得出目标的跟踪位置。虽然跟踪算法的时间复杂度也相对较高,但是仍远低于对视频整体进行遍历的耗时。最后通过本文给出的自适应阈值方程将两种位置预测方法进行整合,使其可应用到实时视频中进行人脸检测。2.在人脸对齐方面,本文对CLM算法框架进行了优化。CLM算法中利用SVM构造响应图并对响应图的数据进行拟合,需要进行多次迭代,这样虽然提升了算法的准确率,但也同时增大了算法的时间复杂度。为了减少耗时,本文对CLM算法模型中形状模型部分进行优化,使用速度快,但准确度略差的FPS3000算法构造初始形状来替代原本的平均形状,从而减少CLM算法中数据拟合的迭代次数。本文改进的算法在保证了算法准确度的同时,还有效降低了算法的耗时。通过本文提出的算法,使人脸检测部分的运算速度提升了50%左右,改进的人脸对齐算法可以实时并准确地对视频中的人脸进行处理,本文除了给出改进算法的性能对比实验,还在最后利用本文提出的算法完成了一个实时的视频人脸动画应用案例。
[Abstract]:Face alignment is mainly used to locate the feature points of the key parts of the face, that is, to describe the local key positions and contours of the face, such as eyes, nose and so on, in the face region. Face detection is the basis of face alignment, and its main function is to find the face position in the image. Face alignment can give a more detailed distribution of face feature points, so it is widely used in face recognition, face animation and pose estimation. In practical application, the algorithm not only needs high accuracy, but also needs better real-time performance. In order to better apply face detection and alignment algorithm to life, this paper will improve the processing speed and accuracy of the algorithm as the main research content. The main research work of this paper has the following two aspects: 1. In the aspect of face detection, the traditional Viola-Jones face detection algorithm (called: VJ face detection algorithm) needs to traverse every frame of video to detect face, in order to reduce the time consumption of invalid traversal. In this paper, the continuity and similarity of video frames are used to improve it, and a face location prediction algorithm is proposed. According to the number of human faces, this algorithm divides the video into less face video and multi-face video, and proposes different face location prediction methods. Less face video uses regression algorithm to estimate the possible range of human face in the video frame to be detected. Then the face is detected in the prediction range, so that the region that the detection needs to traverse is reduced from the whole video frame to the area around the predicted position, thus reducing the time consumption. In the part of multi-face, the SVM response graph is used to track the face target in video, and the location of the target is obtained by using VJ algorithm. Although the time complexity of the tracking algorithm is relatively high, it is still far less time consuming to traverse the whole video. Finally, the adaptive threshold equation presented in this paper integrates two kinds of position prediction methods, which can be applied to face detection in real-time video. In the aspect of face alignment, this paper optimizes the framework of CLM algorithm by using SVM to construct response graph and fit the data of response graph, which requires several iterations, which improves the accuracy of the algorithm. But it also increases the time complexity of the algorithm. In order to reduce the time consuming, this paper optimizes the shape model part of the CLM algorithm model, and uses the FPS3000 algorithm, which has high speed but little accuracy, to construct the initial shape instead of the original average shape. In order to reduce the number of iterations of data fitting in CLM algorithm. The improved algorithm not only ensures the accuracy of the algorithm, but also effectively reduces the time consuming of the algorithm. The algorithm proposed in this paper improves the speed of human face detection by about 50%, and the improved human face alignment algorithm can deal with the face in video in real time and accurately. In addition to the performance comparison experiment of the improved algorithm, Finally, a real-time video face animation application case is completed by using the proposed algorithm.
【学位授予单位】：吉林大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【相似文献】