基于多层随机森林分类的人脸姿态估计算法研究

发布时间：2019-02-11 17:38

【摘要】：近年来,对于人脸图像方面的研究日益增多,人脸检测、人脸特征点定位、人脸跟踪及识别发展迅速,国内外的许多学者已经研究出了相对有效的方法和技术,在应用于人的正脸方向时,效果显著且准确率较高。对人脸进行姿态估计,能够为人脸图像的进一步研究提供技术支持。由于本课题是在具体的研究项目上所提出的,为了能够将视频中感兴趣的人脸用拍摄的普通人脸进行替换,使获得的新视频替换效果显著且自然,本文主要是针对人脸左右偏转在㧟30°到30°,上下俯仰在㧟30°到30°的情况,通过训练分类器对图像或者视频中人脸不同朝向的离散角度进行估计。本文采用将主动形状模型(Active Shape Models,ASM)特征点检测方法与随机森林分类算法相结合的方式,首先利用自行设计的样本采集装置采集人脸样本,以用于随机森林分类器的训练。用ASM算法检测得到68个特征点并对其进行归一化处理,通过实验对比,最终利用选取的7个关键点与其余特征点之间的距离作为特征来训练分类器。由于所得距离特征个数较多,即7×67(忽略点到自身的距离)=469,包含了较多的冗余信息,因此通过主成分分析算法(Principal Component Analysis,PCA)进行择优选取,可以降低约90%的数据量。利用随机森林数据处理的高效性,将获得的距离特征作为训练随机森林的输入数据,训练得到不同朝向的分类器,构成多层随机森林。实验结果表明,本文的算法在人脸左右偏转范围为㧟30°到30°,俯仰范围为㧟30°到30°时,能够较为准确的得到视频中人脸不同姿态的角度值。实验表明,提取灰度特征、Gabor特征训练得到的随机森林分类器,其准确率要低于本文选取距离特征训练所得的分类器;同时与其他姿态估计算法相比,本文基于多层随机森林分类的姿态估计算法所得到的结果较为准确,并且效率较高。
[Abstract]:In recent years, the research on face image is increasing day by day, face detection, face feature point location, face tracking and recognition are developing rapidly. Many scholars at home and abroad have developed relatively effective methods and techniques. When applied to face orientation, the effect is remarkable and the accuracy is high. Face pose estimation can provide technical support for the further study of face image. Because this subject is put forward on the specific research project, in order to be able to replace the interested face in the video with the ordinary face taken, so that the new video replacement effect is remarkable and natural. This paper mainly aims at the situation that the face deflects from-30 掳to 30 掳and pitches up and down from-30 掳to 30 掳, and estimates the discrete angles of different face orientations in image or video by training classifier. In this paper, the active shape model (Active Shape Models,ASM) feature point detection method is combined with the random forest classification algorithm. Firstly, a self-designed sample acquisition device is used to collect face samples for the training of the random forest classifier. 68 feature points are detected by ASM algorithm and normalized. Finally, the distance between the selected seven key points and the other feature points is used as the feature to train the classifier. Due to the large number of distance features, that is, 7 脳 67 (distance from neglect point to itself) = 469, which contains a lot of redundant information, the amount of data can be reduced by 90% by selecting the optimal distance by principal component analysis (Principal Component Analysis,PCA) algorithm. Based on the high efficiency of random forest data processing, the obtained distance feature is used as input data of training random forest, and the classifier with different orientations is trained to form multi-layer random forest. The experimental results show that the proposed algorithm can accurately get the angle values of different face pose when the range of face deflection is -30 掳to 30 掳and pitch range is -30 掳to 30 掳. The experimental results show that the accuracy of the random forest classifier which is obtained by extracting gray feature and Gabor feature training is lower than that obtained by distance feature training in this paper. At the same time, compared with other attitude estimation algorithms, the results obtained by this algorithm based on multi-layer stochastic forest classification are more accurate and efficient.
【学位授予单位】：河北工业大学
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：TP391.41

【参考文献】