基于卡口监控视频的人脸特征点定位关键技术研究

发布时间：2018-11-15 07:56

【摘要】：人脸识别技术是当前计算机视觉、机器学习等领域的研究热点,在安防、信息安全等领域有十分重要的应用背景。其中,人脸特征点定位算法作为人脸识别与验证的关键步骤,与人脸识别准确率息息相关。为此,本文以火车站、公交等监控卡口为研究对象,针对人脸视角、分辨率变化等特点开展了特征点定位方法的研究。主要工作包括:(1)建立了卡口监控人脸数据库。火车站、机场等监控卡口人脸识别是安防领域的难点,其图像通常具有俯视角、分辨率低、人脸姿态多变、光照条件复杂等特点,导致现有的主流算法难以高效地定位卡口监控人脸面部特征点。目前的公开人脸数据库一般是通过爬虫在互联网上采集,与卡口监控系统中人脸图像特点差异较大。为此,本文通过处理大量的火车站和快速公交站监控视频数据,利用半自动的方法对卡口监控人脸图像进行检测、标定、筛选处理,最终得到了卡口监控人脸数据库。该数据库包括火车站卡口监控人脸图像6647张,快速公交车站卡口监控中人脸图像1287张。本文中涉及到的人脸特征点定位实验大多利用该卡口监控人脸数据库中数据进行训练、测试和评估。(2)针对局部二值特征算法(Local Binary Features,LBF)的在卡口监控中人脸特征点定位存在的问题,提出了基于LBF增量学习的卡口人脸特征点定位算法。该方法的主要的思想是在LBF算法级联回归训练的最后一级,利用增量学习的方法,向已经回归得到的模型中导入一部分卡口监控人脸数据库中的人脸图像,对新数据回归得到新的形状增量,从而对现有模型进行修正以达到预期效果。本文所选取的新加入人脸图像的数量是原训练集人脸图像数量的十分之一。实验结果表明,该方法在卡口监控人脸图像特征点定位上更优于时下主流的监督梯度下降法(Supervised Descent Method,SDM)、回归树集合算法(Ensemble of Regression Trees,ERT)和LBF算法。(3)针对卡口监控人脸的俯视视角、分辨率低和运动模糊等特点,提出了基于权重自学习的多任务级联卷积神经网络(Multi-task Cascaded Convolutional Networks,MTCNN)算法。该算法首先调整MTCNN中多任务的权重分布,使MTCNN的网络结构侧重解决人脸特征点定位问题。然后在MTCNN网络结构中加入权重自学习模块,使其能自动学习得到多任务协调计算的最佳权重分布,从而进一步提高对卡口监控中人脸图像特征点定位的精度。实验结果表明,该方法在卡口监控中人脸图像特征点定位的准确率高于SDM、ERT、LBF、MTCNN和基于LBF增量学习卡口人脸特征点定位算法。最后对本文工作进行了总结,并对本文后续工作进行了展望。
[Abstract]:Face recognition is a hot topic in the field of computer vision and machine learning. It has a very important application background in the field of security and information security. As a key step of face recognition and verification, face feature location algorithm is closely related to the accuracy of face recognition. For this reason, this paper takes the monitoring bayonet such as railway station and bus as the research object, and carries out the research on the feature point location method according to the features of the human face angle of view and the change of the resolution. The main work is as follows: (1) the face database of bayonet monitoring is established. Face recognition of railway station, airport and other monitoring bayonets is a difficult problem in the field of security. The image is usually characterized by low resolution, variable face pose, complex illumination conditions and so on. As a result, the existing mainstream algorithms are difficult to locate the facial feature points efficiently. At present, the open face database is generally collected on the Internet by crawlers, which is different from the features of face images in the bayonet monitoring system. Therefore, through processing a large number of monitoring video data of railway station and bus rapid transit station, this paper uses semi-automatic method to detect, calibrate, filter and process the face image of the bayonet monitoring. Finally, the face database of the bayonet monitoring is obtained. The database includes 6647 face images from railway station bayonets and 1287 face images from bus rapid transit stations. Most of the experiments of facial feature point localization in this paper use the bayonet to monitor the data in face database for training, testing and evaluation. (2) the local binary feature algorithm (Local Binary Features,. In this paper, the problem of face feature point location based on LBF) is discussed. A face feature point location algorithm based on LBF incremental learning is proposed. The main idea of this method is that at the last level of cascade regression training of LBF algorithm, the incremental learning method is used to import a part of the face image in the face database to the model that has been regressed. New shape increments are obtained from the new data regression, and the existing models are modified to achieve the desired results. The number of newly added face images selected in this paper is 1/10 of the original training set. The experimental results show that the proposed method is better than the mainstream supervised gradient descent method (Supervised Descent Method,SDM) and the regression tree set algorithm (Ensemble of Regression Trees,) in the feature point location of face images monitored by the bayonet. ERT) and LBF algorithm. (3) in view of the features of top-down view, low resolution and motion blur, a multi-task concatenated convolution neural network (Multi-task Cascaded Convolutional Networks,MTCNN) algorithm based on weight self-learning is proposed. The algorithm firstly adjusts the weight distribution of multi-task in MTCNN to make the network structure of MTCNN focus on the problem of facial feature point location. Then the weight self-learning module is added to the MTCNN network structure so that it can automatically learn the optimal weight distribution of multi-task coordination calculation and further improve the accuracy of facial image feature point location in the bayonet monitoring. The experimental results show that the accuracy of this method in facial image feature point location is higher than that in SDM,ERT,LBF,MTCNN and LBF incremental learning. Finally, the work of this paper is summarized, and the future work of this paper is prospected.
【学位授予单位】：集美大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】