【作者】 陈煜；

【导师】 周军；

【作者基本信息】 上海交通大学 ， 信号与信息处理， 2009， 硕士

【摘要】 随着人工智能和计算机技术的发展,机器视觉研究取得了长足的进步,并引起了越来越多的关注,其研究成果被广泛应用于国防建设、航空航海、医药卫生、安全监控等国民经济的各个领域。人脸检测与跟踪则是其中一个重要的应用方向。人脸检测的目的是检测某监控区域内是否存在人脸并确定人脸出现的位置、大小等信息;目标跟踪的目的是对选中的目标进行轨迹的描述,涉及到目标特征的提取、目标的匹配等多个方面。本文研究的主要目的是实现对视频序列中人脸的实时监控。系统由检测和跟踪两部分组成。检测算法主要基于AdaBoost算法,通过分析AdaBoost算法,训练得到正面人脸的级联检测器。本文的跟踪算法主要基于CamShift算法。通过分析影响跟踪不稳定的原因,选用HSV颜色空间的H和S分量产生目标的二维颜色直方图,并对其中的直方项进行归并和加权处理,降低了目标区域中的非目标像素点对跟踪结果造成的影响,使目标的二维颜色直方图能更加准确地描述目标模式,从而在保持CamShift算法快速性的同时提高其跟踪鲁棒性。然后,本文将检测和跟踪结合起来,在PC平台上实现出能对QVGA视频实时监控的人脸检测跟踪系统。另外,本文还研究了TI DaVinci处理器的软件开发,基于AdaBoost和改进的CamShift算法设计了基于DaVinci处理器的人脸检测与跟踪系统的模型。通过设计定点算法,将人脸检测程序移植到DaVinci的C64x+ DSP处理器上,利用编译器优化选项、Intrinsics、线性汇编、Cache对程序进行优化,并根据xDAIS-DM算法标准将经优化后的人脸检测算法封装成可重用的算法包。随后,本文分析了DaVinci处理器的Codec Engine软件架构及其远程方法调用的机制,最后,在DaVinci处理器的ARM端通过V4L2和FB驱动程序并利用VISA API调用在DSP端运行的人脸检测算法模块完成人脸检测应用程序,从而在DaVinci处理器上实现出完整的人脸检测系统,对QVGA和VGA视频分别达到6.4fps、1.63fps的检测速率。更多还原

【Abstract】 With the development of artificial intelligence and computer science, Machine Vision Technology has made considerable progress and attracts more and more attention, and it has been widely used in various economic aspects, such as national defense, aviation and navigation, medical treatment and health care, security monitoring, and so on. Face detecting and tracking is one of the most important applications. The purpose of face detecting is to detect if there exist faces in the surveillant area. If that, then determine the information of location and size of the faces. Object tracking is also an important application branch of Computer Vision. The goal of tracking is to recognize target objects from the background and extract features, then to depict the moving locus of them.The aim of this paper is to implement real-time surveillance on human faces appeared in video. The system can be divided into two parts of detecting and tracking. Detecting algorithm is mainly based on AdaBoost. By analyzing it, a cascaded detector of frontal face has been trained. Tracking method of the paper is mainly based on CamShift algorithm. By analyzing its instability during tracking process, H and S components of HSV color space have been chosen to generate a 2D color histogram which describes the object region. Besides, the 2D histogram is processed by merging and weighting operation to decrease the interference from non-object pixels in object region greatly. In this way, the object pattern can be more accurately described by 2D color histogram, and CamShift algorithm becomes much more robust while keeping its appealing processing speed. After that, Detecting and tracking algorithms are combined to implement a real-time face detecting and tracking system to QVGA video on PC.In addition, the software development of TI DaVinci processor is also researched in this paper, and a model of DaVinci based face detecting and tracking system has been designed by AdaBoost and Improved CamShift algorithms. By design of fix-pointed algorithms, face detecting and tracking program has been transplanted to the C64x+ DSP on DaVinci. Optimization options of compiler, intrinsics, linear assembly and cache have been utilized to optimize the program, and the optimized face detecting algorithm has been packaged according to the xDAIS-DM standard. After that, Codec Engine framework of DaVinci processor and its RPC mechanism have been analyzed in the paper. In the end, by V4L2 and FB driver and calling face detecting algorithm module run on DSP end by VISA APIs, a completed face detecting application has been finished on the ARM end of DaVinci. Detecting rate of 6.4fps and 1.63fps has been acquired on QVGA and VGA video respectively.更多还原