快速人脸跟踪系统

【作者】 刘强；

【导师】 蔡灿辉；

【作者基本信息】 华侨大学 ， 信号与信息处理， 2007， 硕士

【摘要】 人脸跟踪是计算机视觉、人工智能领域一个重要的研究方向。作为自动人脸识别、视频会议、基于内容的压缩与检索等领域中的一项关键技术,人脸跟踪具有广泛的应用前景,受到研究者的普遍重视。本文首先介绍了人脸跟踪的研究进展以及人脸检测技术和跟踪技术的主要算法并着重介绍了连续自适应均值移动(Continuously Adaptive Mean Shift, CAMShift)算法。该算法不受跟踪目标的形状、大小的影响,具有很强地抗亮度变化和噪声干扰的能力,而且CPU的占用率非常低,适合作为人机交互界面的接口,但抗肤色背景干扰和目标遮挡的能力较差。本文根据CAMShift算法进行单人脸跟踪中出现的问题,提出肤色增强、消除类肤色背景和引入辅助信息等方法对CAMShift算法进行改造,有效地解决了类肤色干扰和目标的遮挡引起丢失目标问题,大大提高了跟踪的鲁棒性;采用AdaBoost快速人脸检测算法初始化跟踪目标,实现了自动人脸跟踪系统。实验结果证明改进后的人脸跟踪算法,在静态背景下,单人脸跟踪可以达到每秒800帧的快速跟踪,具有很强的实时性和鲁棒性。在此基础上,提出基于CAMShift的多人脸跟踪系统,引入相同矩阵消除法、最优排序法和多辅助信息等方法解决多目标跟踪出现的跟踪窗口重复、目标丢失和个体对应等问题;采用对整帧图像进行颜色反向投影法解决目标更新问题。实验表明,本文提出的多人脸跟踪算法能够快速(38～156帧/秒)、准确(97%以上)地跟踪每个目标。其中,跟踪误差主要表现为多个人脸间的距离很小时,目标对应错误和受遮挡时目标暂时丢失。更多还原

【Abstract】 Facetracking is an important research field of computer vision and artificial intelligence. As a key technology, facetracking has received lots of attentions in automatic human face recognition, content based compression and retrieval, video conference, and other fields.Firstly, recent developments of facetracking and main algorithms of face detection & tracking are introduced in this thesis, and the Continuously Adaptive Mean Shift (CAMShift) algorithm is discussed in detail. Since the algorithm can robustly track target of different shape and size with the immunity against illuminant fluctuation and noise inference, and has low CPU load, it can serve as an efficient human and computer interface. However, CAMShift suffers when flesh-like interference and occlusion occur.Some methods, such as the enhanced color information, removing the flesh-like background and accessory information, are proposed to enhance the robust of CAMShift, and an AdaBoost fast face detector is used to initiate the searching window automatically. The improved CAMShhift algorithm can tra- ck an object robustly and automatically, which can track about 800 frames per second under the static background and can successfully resist the occlude and flesh-like interference. The experimental results demonstrate the robustness and efficiency of the proposed algorithm in real-time face tracking.The improved CAMShift algorithm is extended to multi-faces tracking. Same window elimination, maximum choice, and multiple accessory infor- mation methods are proposed to solve target-loss and individual identification problems. A whole frame PDI (Probabilistic Distribution Image) is produced to update the tracked objects. Empirical data have testified the proposed algor- ithm can fast (38~156fps), and accurately (above 97%) track multiple faces, where the tracking errors mainly happen in individual identification while the distance of two faces are too close and temporarily tracking failure when tra- cked face is occluded.更多还原