【作者】 王铮；

【导师】 黄樟钦；

【作者基本信息】 北京工业大学 ， 工程硕士（专业学位）， 2017， 硕士

【摘要】 图像目标的监测和辨识一直都是计算机视觉范畴里非常关键的问题。人脸可以作为一个人身份验证的依据,它是人类的主要辨识特征之一。同时,人脸检测也是人机交互领域的重要技术,应用于监控系统、相关图像检索等各个方面。现在,使用最多的人脸检测算法是基于AdaBoost的检测算法。这种检测算法保证了检测的效率和准确率。人脸检测的研究,一直专注于设计新的算法或提高现有方法的准确性。而嵌入式设备资源有限,导致现有的人脸检测方法在嵌入式设备上处理效率低下。因此,人脸检测的效率逐渐成为了人脸检测的关键问题。针对此问题,本文对人脸检测实现了并行化计算,在Parallella平台上,利用集群和Epiphany多核处理器对基于AdaBoost的人脸检测算法进行了并行计算处理,并且设计了并行处理的人脸检测系统。本文从以下几个方面进行研究。首先,指出了人脸检测过程中扫描窗口部分占用极大的系统资源这一弊端,提出了多核并行人脸检测的解决方案。其次,在此基础之上,鉴于视频中每帧图像的分辨率相同,导致每帧图像的检测工作任务相似,本文对并行人脸检测系统进行了优化,对单帧图像检测过程中的相同部分只做一次处理,不再重复计算,从而提高了系统的性能。然后,本文采用MPI技术把单板的并行人脸检测拓展到集群,利用集群进一步提高计算速度。其中集群内每一块开发板都使用多核处理器并行计算。最终,采用Parallella平台实现并行人脸检测系统,Parallella平台包含ZYNQ和多核处理器,其中ZYNQ中的ARM A9处理器运行人脸检测系统,多核处理器协助ARM处理复杂计算。此外,本文对基于Parallella平台的人脸检测的并行计算系统的性能进行了验证。更多还原

【Abstract】 The object monitoring and identification based on image processing have been the key research topics in the field of computer vision.Face is the basis for human identity verification,which is one of the main characteristics of human identification.Meanwhile,face detection has become a new human-computer interaction technology and widely used in monitoring system,image retrieval and other aspects.Now,the most commonly used face detection method is Ada Boost algorithm,which ensures the efficiency and accuracy of detection.The research of face detection has been focused on designing new algorithms or improving the accuracy of existing methods.However,due to the limited resources of embedded devices,the existing face detection methods are inefficient in embedded devices.Therefore,the efficiency of face detection has become the key problem of face detection.To solve this problem,this paper realizes the parallel computing of face detection.Based on the Parallella platform,the AdaBoost based face detection algorithm is processed in parallel with the cluster and the Epiphany multicore processor,and the parallel processing system for face detection is designed.This paper takes the studies from the following aspects.Firstly,this paper points out the defect in face detection process,which is the scanning window occupies a great deal of system resources,and proposes a multicore parallel face detection method.Secondly,considering the detection task of each frame is similar as the resolution of each frame in the video is same,so this paper process the same part of the single frame image detection only once to optimize the parallel face detection system and improve system performance.Thirdly,using MPI technology to extend the parallel face detection of single board to the cluster,and improve the computing speed by deploying the cluster.Each of the boards in the cluster uses the multicore processor for parallel computation.Finally,this paper uses the Parallella platform to realize the parallel face detection system.The Parallella platform includes ZYNQ and multicore processor,the ARM A9 processor running the face detection system in ZYNQ and multicore processors help ARM deal with complex computations and the performance of the parallel computing system based on Parallella platform is verified.更多还原