外文原文：
Face Detection and Tracking
Face detection is in several important, because the input face image taken by automated systems. Examples include automatic face recognition system, based on video surveillance and early warning systems, face / body tracking system and the perceptual user interface. Most face detection algorithms can be classified based on characteristics or appearance. In recent years, the appearance of the face detection algorithm, using machine learning and statistical estimation methods have shown that all the existing face detection methods on their achievements.
The appearance of the face detection examples include AdaBoost algorithm FloatBoost the S - AdaBoost algorithm, neural networks, support vector machine (SVM), Hidden Markov Model and Bayesian classifier. Viola and Jones, who made a strong AdaBoost face detection algorithm that can detect faces a higher detection rate with swift and powerful way. Li et al. Floating algorithm proposed to improve learning and increase the minimum classification error rate of AdaBoost algorithm an improved version. Float boost algorithm uses backtracking mechanism to improve after each iteration AdaBoost Face Detection rate. However, this approach is computationally more intensive than the AdaBoost algorithm. Jiang Luoti out of the S - AdaBoost algorithm model in dealing with outlier detection and classification of a variant of AdaBoost algorithm. Since it is S - AdaBoost algorithm is used to calculate the different stages of different classifiers, which suffer from low efficiency and lack of accuracy.
In a variety of neural network-based face detection method, Raleigh and others work. Particularly well known. Raleigh et al. Multi-layer neural network learning training sets of face and non face to face model composed of face and non face images. Their major drawback is that technology, which is mainly limited to the vertical test positive face. Although Raleigh et al. To further promote their method to detect rotation of the face image, the reported result is not because of the low detection rate of the resulting optimism. The face of support vector machines (SVMs) using detection technology, structural risk minimization to minimize the expected generalization error upper bound. The main disadvantage of support vector machines included in the learning process and high-memory intensive computing. The face of Hidden Markov Model (HMMs) for the detection of non-face-face mode of the assumptions and parameters can be used as a random Markov process, its use can learn a clear estimate of the output characteristics of the process parameters. A HMM model training goal is to estimate the corresponding parameters of HMM model, the maximum probability or observed the possibility of training data. Schneidman and Kanadepresent a Bayesian face detection, which is the local face appearance and pattern of the multi-scale joint probability estimate of the position classification. However, the naive Bayesian classifier performance reports are poor. To solve this problem, a face detection limit Shineideman Bayesian networkfor perform a search in the large cargo possible network structures to determine the optimal structure of Bayesian network based classifier.
Tracking objects in video have been widely studied computer because the computer, such as the various autonomous robot, video surveillance, the human eye tracking and face tracking algorithm has been widely used application of visual tracking vision researchers. Tracking algorithm designed to operate in a wider range of structures and less need to deal with the case of uncertainty and errors caused by occlusion of complex issues, in light of the changes, view and object size. Therefore, many techniques have been developed to different visual target tracking for the above problem, reported in the literature over the past 10 years.
A variety of visual object tracking technology can be divided into the image (region), based on contour-based or filter based. Image (region) of the tracking method is usually extracted from the input of the general characteristics of the image, then combine these features with high-level regionbased site information. Intille and so on. Propose a BLOB, which were used to track real-time tracking Extraction of foreground minus the background of the area. Reception area, and then split into spots on skin color. This method has a good run-time performance, but be materially and adversely from a spot in the combined calculation, the way in the field of objects to each other. Contour-based object tracking method assumes that the trace is bounded by the contour lines known performance. Be tracked from one contour pixel to the next predefined imageframe contour shape model. Contour model of power or flexibility to handle the shape of the object changes due to deformation or change in size. Filter-based prediction method is based on the object and updated features over time, that is, successive frames in the video sequence. Object shape and location of the object tracking over time is to fully solve the problem of where the traditional Kalman tracking can be effectively modeled as a linear dynamic system, the case filter. Extended Kalman filter (EKF) is a traditional extended Kalman filter to the nonlinear process of a single peak, but by the local linear non-linear behavior can be close. However, it has been widely accepted to be superior to the traditional particle filter tracking performance of the Kalman filter, since the particle filter provides a powerful and will not be tracked object is limited to a linear system model framework.
Particle filter, also known as sequential Monte Carlo filter, visual tracking has been widely used to solve linear and non-normal limit from the motion model generated. Particle filters approximate the basic idea is to use recursive Bayesian posterior density filter weights with some of the distribution of particle sets. The condensation algorithm, a simple particle filter, proposed by the Isard to solve the problem, from the non-linear and non-tracking the movement caused by the normal model. In the sampling step, the combination of the proposed algorithm uses a simple drawing of the particles are distributed, which defines the state of the particle in the conditions on the distribution of a frame. The distribution of the proposal, then the posterior distribution approximation goal. A universal adherence to the disadvantage of the condensation algorithm, the proposed distribution is not included in the longer term from the current convergence and thus to create the necessary framework for the posterior distribution of information will take time.
Various approaches have been developed to improve the traditional particle filter tracking performance. Li and other al.propose Kalman particle filter (KPF) and the unscented particle filter (UPF) of the modified particle sampling within the visual contour tracking procedures. This approach allows an unscented Kalman filter or a Kalman filter incorporated into the proposal distribution is estimated using the current observations. Kalman filter or unscented Kalman filter display in the search space to guide the particles in the region may be set to high, thus reducing the number of particles needed to estimate proposed distribution. In order to solve the occlusion problem, Wang and Zhang propose a Markov random field (MRF)-based dynamic Bayesian particle filter to track objects within the framework of representation. This method is converted into a variety of medium-term forecast regional composite objects, in order to improve modeling and tracking accuracy. Zhang et al. Currently the kernel particle filter to improve the system for multiple object tracking data association techniques, sampling efficiency. The program calls the kernel to approximate the posterior density continuously, in which the object representation and the localization of the kernel distribution, based on the density gradient generated from the kernel value. However, this method can not handle the situation, in which a sharp change in the movement of objects. Rathi and other al.formulate geometry technology to handle the object as active contour deformation. However, their poor technical performance, the tracked object is totally occluded in several frameworks. Isard and McCormick gram of Bayesian multi-drop out of the track (thorns), particle filters an early variant of the track in which the number of objects can not be changed, in the tracking process. However, this method relies on a fixed background model to determine the foreground object, this is not always the actual situation in the real world tracking. To solve this problem, Okuma and so on. Relaxed assumption of a fixed background, and allow different backgrounds, to address real-world image sequences. Okuma other al.also proposed a particle filter to improve target detection and tracking of multiple, staggered a simple particle filter algorithm of AdaBoost algorithm condensation (BPF) of the. However, this method does not systematically incorporated into an object model to ensure accurate approximation of the proposal distribution, nor solve the congestion problem.
In this paper, we propose a new particle filtering scheme as an adaptive particle filter (APF) called, so that more accurate estimate ofthe proposed distribution and the posterior distribution. Based Isard, Li et al previous work. , Vermaak and other al.and Okuma and so on. We also recommend that combine with AdaBoost Face Detection APF - 1 tracking algorithm based on face detection and face tracking algorithm for integration of new programs. Our long-term armed algorithm and AdaBoost algorithm proposed combination of drive as adaptive particle filter (BAPF). In the proposed BAPF based tracking program, AdaBoost algorithm is used to detect faces, but also validate the input of an image tracking algorithms face the armed police to survive, the purpose is to track the entire frame face detected in the video sequence. The BAPF algorithm proved to have a very good track to track objects in those circumstances is a serious obstruction of the results. Experimental results show that the BAPF program provides a powerful face detection and tracking accuracy in each case the face tracking.
外文翻译：
人脸检测与追踪
人脸检测是在几个重要的是，由于人脸图像输入采取自动化系统。例子包括全自动人脸识别系统，基于视频监控和预警系统，人脸/身体跟踪系统和感性的人机界面。大多数人脸检测算法可分为基于特征或外观为基础。近年来，外观的人脸检测算法，采用机器学习和统计估算方法已显示所有现有的人脸检测方法的优异成绩。
外观的人脸检测技术的例子包括AdaBoost算法，该算法FloatBoost的S - AdaBoost算法，神经网络，支持向量机（SVM），隐马尔可夫模型和贝叶斯分类器。中提琴和琼斯提出强有力的AdaBoost的人脸检测算法，该算法可以检测面临一个具有较高的检测率迅速和强有力的方式。李等人。提出的浮算法，提高学习与最小误差率提高了分类器的AdaBoost算法的一个改进版本。浮法升压算法使用回溯机制，提高后的每个迭代过程AdaBoost的人脸检测率。但是这种方法在计算上比更密集的AdaBoost算法。江洛提出了S - AdaBoost算法，在处理离群模式检测和分类的AdaBoost算法的一个变种。既然是S - AdaBoost算法用于计算不同阶段不同分类器，它患有计算效率低下和缺乏准确性。
在各种不同的神经网络为基础的人脸检测方法，罗利等人的工作。特别是众所周知的。罗利等人。采用多层神经网络学习训练的脸和非面对面模式套脸和非脸图像组成。他们的技术的重大缺点是，该检测主要是限于直立正面的面孔。虽然罗利等人。进一步推广他们的方法来检测旋转的人脸图像，报告的结果是不是因为由此产生的低破案率乐观。面对支持向量机（SVMs）的检测技术的使用结构风险最小化，尽量减少上层的预期推广误差的约束。支持向量机的主要缺点包括在学习过程和高内存需求密集型计算。面对隐马尔可夫模型（HMMs的）的检测技术所面临的假设和非面对面模式可以作为一个参数随机马尔可夫过程，其使用可以学到一个明确的估计过程参数输出的特点。一个HMM模型的训练目标是估计的HMM模型的相应参数，最大限度的概率或观察到的训练数据的可能性。施奈德曼和Kanadepresent一个朴素贝叶斯人脸检测，这是对当地的外观和在多尺度面临格局中的地位联合概率估计的分类。然而，朴素贝叶斯分类器性能的报道是穷人。为了解决这个问题，提出了限制史内德曼贝叶斯networkfor脸检测执行一个在大货舱搜索可能的网络结构，以确定最佳的构造分析贝叶斯网络为基础的分类器。
在视频跟踪对象也被广泛研究的计算机，因为计算机，例如各种自治机器人，视频监控，人眼跟踪和人脸跟踪算法广泛使用的跟踪视觉应用视觉研究人员。目标跟踪算法设计运行在更广泛的结构和较少的情况下需要处理的不确定性和误差闭塞引起的复杂问题，在光照的变化，观点和对象规模。因此，许多技术已经发展到针对不同的视觉目标跟踪上述问题，在过去10年文献报道。
对视觉对象跟踪的各种技术可以分为图像（地区）为基础，轮廓为基础或过滤为基础的。图像（地区）的跟踪方法通常从输入图像中提取的一般特征，然后结合regionbased这些功能采用高级别现场信息。 Intille等。提出一个BLOB，其中用于实时跟踪跟踪人的背景减去提取前景的地区。前台区域，然后分成基于肤色斑点。这种方法具有良好的运行时性能，但是从一个在受到重大不利的斑点合并计算时，在现场的方式彼此的对象。基于轮廓的跟踪方法假定跟踪对象是由等高线为界已知的性能。被跟踪的轮廓像素从一个imageframe到下使用预定义的轮廓形状模型。动力或弹性轮廓模型用来处理对象的形状因变形或规模变化的变化。过滤为基础的方法是基于对象的预测和更新的功能随着时间的推移，即对在视频序列连续帧。物体的形状和物体的位置随着时间的推移跟踪是充分解决了传统卡尔曼跟踪问题在哪里可以有效地作为一个线性动态系统建模案件过滤器。扩展卡尔曼滤波（EKF）是一个传统的卡尔曼滤波扩展到非线性的单峰过程，但其中非线性行为可以由当地线性接近。但是，它已被广泛接受的粒子过滤器要优于传统的卡尔曼滤波的跟踪性能方面，由于粒子过滤器提供了一个强大而不会被跟踪对象仅限于一个线性系统模型框架。
粒子过滤器，也被称为序贯蒙特卡罗过滤器，已广泛应用于视觉跟踪，以解决限制来自非线性和非正常的议案模型产生。粒子过滤器的基本思想是利用近似的后验密度递归贝叶斯过滤器在与某些分配权重的粒子集的。该冷凝算法，一个简单的粒子过滤器，由伊萨德提出的旨在解决问题，从跟踪的非线性和不正常所引起的运动模型。在抽样一步，凝聚算法使用一个简单的建议分布绘制的粒子集，它定义了粒子的状态在上一帧的条件分布。该提案的分布，然后用逼近目标后验分布。一个普遍遵守的凝结算法的缺点是，建议分布不纳入从目前从而在更长远的衔接到所需的后验分布造成的框架需要时间的信息。
已经制定了各种办法来改善传统的粒子滤波器的跟踪性能。 Li等al.propose卡尔曼粒子滤波器（KPF）和无味粒子滤波（UPF）的改进粒子采样轮廓跟踪的视觉范围内的程序。这种方法使一个或一无迹卡尔曼滤波卡尔曼滤波的使用纳入该提案分布估计目前的观测。卡尔曼滤波无迹卡尔曼滤波或引导显示在搜索空间中的粒子可能设置为高的地区，从而减少所需的粒子数目估计建议分布。为了解决遮挡问题，王，张提出一个具有马尔可夫随机域（MRF）的基于粒子滤波的动态贝叶斯框架内的跟踪对象的代表性。这种方法转换成多种中期预测区域复合对象，以提高建模和跟踪精度。张等人。目前内核粒子过滤器，以改善对多个物体跟踪系统中的数据关联技术，采样效率。这项计划调用内核来近似后验密度不断，其中对象的代表性和本地化的内核分配的基础上，从内核产生的密度梯度值。但是，此方法不能处理的情况，其中一组物体的运动规律变化急剧。 Rathi等al.formulate几何参数化技术来作为处理对象的主动轮廓变形。然而，他们的技术表现不佳时，被跟踪的对象是在几个框架完全闭塞。伊萨德和麦考密克提出了贝叶斯多一滴，跟踪（荆棘），粒子过滤器的一个早期的变体，在其中的跟踪对象的数量不得更改，在跟踪的过程。然而，这种方法依赖于一个固定的背景模型，以确定前景对象，这种情况并不总是在现实世界的实际情况跟踪。为了解决这个问题，大隈等。放松了一个固定的背景假设，并允许不同的背景，以处理现实世界的图像序列。大隈等al.also提出了一个提高粒子滤波器的多目标检测和跟踪，交错一个简单的微粒对冷凝过滤算法AdaBoost算法（BPF）的。但是，此方法并没有系统地纳入一个对象模型，以保证准确的建议分布近似，也没有解决阻塞问题。
在本文中，我们提出一个新的粒子滤波计划，作为一个自适应粒子滤波器（APF）的称为，让有更高的准确度估计ofthe建议分布和后验分布。基于伊萨德，李等人先前的工作。，Vermaak等al.and大隈等。，我们还建议结合起来，与AdaBoost的人脸检测的APF - 1为基础的跟踪算法的人脸检测和人脸追踪一体化新方案算法。我们长远的建议武警算法和AdaBoost算法相结合的方式带动作为自适应粒子过滤器（BAPF）。在建议BAPF为基础的追踪计划下，AdaBoost算法是用来检测的面孔，也验证输入的一帧图像跟踪面临生存而武警算法，目的是跟踪整个帧图像中检测到的面孔视频序列。该BAPF算法被证明产生非常良好的跟踪在那些跟踪的对象是严重阻塞的情况下的结果。实验结果表明，该BAPF计划提供了强大的人脸检测和跟踪在各种情况下准确的人脸追踪。
