视频监控中基于iOS平台的人脸检测与识别

发布时间：2019-01-05 16:51

【摘要】：iOS,即苹果公司开发的移动操作系统,主要应用于iPhone手机、iPad平板等。现在越来越多的人在iOS平台上聊天、上网、观看视频等等,iPhone已成为很多人的随身必备物品。人脸与指纹、虹膜一样与生俱来,具有唯一性和不易被复制性,是身份检测的重要指标。人脸检测与识别可以帮助视频监控实现智能化。再结合基于iOS的移动互联网平台,可以将监控视频通过无线网络传输到iPhone手机或iPad平板,就可以在手机或平板上对监控视频进行实时的人脸检测与识别,既方便快捷又节约成本。本文的工作主要包括以下四个方面：(1)实现iOS平台下监控视频的传输。首先利用摄像头采集到视频图像,然后采用Flash Media Live Encoder对视频进行编码,并利用Flash Media Server作为服务器,通过HLS(HTTP Live Stream)协议将视频流传输到iPhone上并显示。(2)实现监控视频的预处理。本文对视频的预处理主要有视频去噪、直方图均衡化以及白平衡。首先对视频去噪,主要采用形态学滤波消除视频中的椒盐噪声。接着对视频进行直方图均衡化,以增加图像的对比度,有利于人脸检测。最后通过完美反射法对视频进行白平衡调整,从而削弱光照对人脸检测带来的的影响。预处理有助于人脸检测的准确率以及提高人脸检测的速度。(3)实现改进Adaboost算法的人脸检测。本文分析讨论了Viola-Jones在2004年提出的基于Adaboost人脸检测算法,并在iOS平台下对Viola-Jones人脸检测算法提出了三点改进：首先提出了基于权值更新的Adaboost算法改进；然后后提出了基于查询子窗口大小的Adaboost算法改进；最后通过肤色检测来加快人脸检测。(4)实现改进LBP算法的人脸识别。首先采用Gabor变换实现多尺度、多方向的特征提取。然后通过改进的LBP算法提取纹理信息。最后通过Fisherfaces算法实现向量投影并降维,并通过余弦相似度对图像进行分类。实验结果表明,该算法不仅对光照有较好的鲁棒性,还提高了人脸识别速度和准确率。总的来说,本文主要实现基于iOS平台下监控视频的人脸检测与识别,首先对监控视频做一系列的预处理,然后对预处理后的视频图像进行人脸检测,最后在人脸检测的基础上实现人脸识别。实验结果表明,本文设计的系统通过对人脸检测的改进加快了Adaboost分类器训练速度,降低了人脸检测的频率,降低了程序的能耗；通过对人脸识别算法的改进,在增强对光照鲁棒性的同时提高人脸识别准确率。
[Abstract]:IOS, the mobile operating system developed by Apple, is mainly used in iPhone phones, iPad tablets and so on. Now more and more people chat, surf the Internet, watch videos and so on the iOS platform. IPhone has become a necessary item for many people. Face and fingerprint are the same as iris, unique and hard to be duplicated, which is an important index of identity detection. Face detection and recognition can help video surveillance to achieve intelligent. Combined with the mobile Internet platform based on iOS, surveillance video can be transmitted to iPhone mobile phone or iPad tablet through wireless network, and real-time face detection and recognition can be carried out on mobile phone or tablet. It is convenient and quick as well as cost saving. The work of this paper mainly includes the following four aspects: (1) realize the transmission of surveillance video on iOS platform. Firstly, the video image is captured by the camera, then the video is encoded by Flash Media Live Encoder, and Flash Media Server is used as the server. The video stream is transmitted to iPhone by HLS (HTTP Live Stream) protocol and displayed. (2) the preprocessing of surveillance video is realized. In this paper, video preprocessing mainly includes video denoising, histogram equalization and white balance. Firstly, the morphological filter is used to eliminate the salt and pepper noise in the video. Then the histogram equalization is used to increase the contrast of the image, which is beneficial to face detection. Finally, the white balance of video is adjusted by perfect reflection method, which weakens the influence of illumination on face detection. The preprocessing is helpful to the accuracy of face detection and the speed of face detection. (3) face detection based on improved Adaboost algorithm is implemented. This paper analyzes and discusses the Adaboost based face detection algorithm proposed by Viola-Jones in 2004, and proposes three improvements to the Viola-Jones face detection algorithm under the iOS platform. Firstly, an improved Adaboost algorithm based on weight updating is proposed. Then an improved Adaboost algorithm based on query sub-window size is proposed. Finally, skin color detection is used to speed up face detection. (4) face recognition based on improved LBP algorithm is implemented. Firstly, Gabor transform is used to realize multi-scale and multi-direction feature extraction. Then the texture information is extracted by the improved LBP algorithm. Finally, vector projection and dimensionality reduction are realized by Fisherfaces algorithm, and the images are classified by cosine similarity. Experimental results show that the algorithm not only has good robustness to illumination, but also improves the speed and accuracy of face recognition. In general, this paper mainly realizes the face detection and recognition of surveillance video based on iOS platform. Firstly, we do a series of preprocessing to the surveillance video, and then we detect the face of the pre-processed video image. Finally, face recognition is realized on the basis of face detection. The experimental results show that the system improves the training speed of Adaboost classifier, reduces the frequency of face detection and reduces the energy consumption of the program. The algorithm of face recognition is improved to improve the robustness of illumination and the accuracy of face recognition.
【学位授予单位】：浙江理工大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TN948.6

【相似文献】