CN106570471B - Scale-adaptive multi-pose face tracking method based on compression tracking algorithm - Google Patents

Scale-adaptive multi-pose face tracking method based on compression tracking algorithm Download PDF

Info

Publication number
CN106570471B
CN106570471B CN201610947868.1A CN201610947868A CN106570471B CN 106570471 B CN106570471 B CN 106570471B CN 201610947868 A CN201610947868 A CN 201610947868A CN 106570471 B CN106570471 B CN 106570471B
Authority
CN
China
Prior art keywords
face
algorithm
tracking
target
target face
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN201610947868.1A
Other languages
Chinese (zh)
Other versions
CN106570471A (en
Inventor
吴怀宇
陈镜宇
钟锐
何云
程果
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Wuhan University of Science and Engineering WUSE
Original Assignee
Wuhan University of Science and Engineering WUSE
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Wuhan University of Science and Engineering WUSE filed Critical Wuhan University of Science and Engineering WUSE
Priority to CN201610947868.1A priority Critical patent/CN106570471B/en
Publication of CN106570471A publication Critical patent/CN106570471A/en
Application granted granted Critical
Publication of CN106570471B publication Critical patent/CN106570471B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
    • G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
    • G06V40/16—Human faces, e.g. facial parts, sketches or expressions
    • G06V40/161—Detection; Localisation; Normalisation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/40—Scenes; Scene-specific elements in video content
    • G06V20/41—Higher-level, semantic clustering, classification or understanding of video scenes, e.g. detection, labelling or Markovian modelling of sport events or news items

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Multimedia (AREA)
  • Theoretical Computer Science (AREA)
  • Health & Medical Sciences (AREA)
  • General Health & Medical Sciences (AREA)
  • Oral & Maxillofacial Surgery (AREA)
  • Human Computer Interaction (AREA)
  • Computational Linguistics (AREA)
  • Software Systems (AREA)
  • Image Analysis (AREA)

Abstract

The invention discloses a kind of dimension self-adaption multi-pose Face trackings based on compression track algorithm, coarse localization is carried out to target face using compression track algorithm, to reduce the search range of Face datection algorithm, and then improve target person face detection algorithm accuracy and real-time；Using Face datection algorithm, the accurate positionin of target face is realized, while realizing the dimension self-adaption tracking of target face；Using target person face detection algorithm, solves the problem of that target face leaves and track failure when camera lens is again introduced into；Using the continuity of target face motion process in time, the tracking continuity problem in the case of the detection failure of target person face detection algorithm is realized.The present invention can guarantee that camera carries out accurately and effectively multi-pose to target face by above method, dimension self-adaption tracking, can be widely used among video monitoring, human-computer interaction, virtual reality and various security systems such as: ATM machine monitors access control system.

Description

Dimension self-adaption multi-pose Face tracking based on compression track algorithm
Technical field
The present invention relates to image steganalysis field more particularly to a kind of dimension self-adaption people based on compression track algorithm Face track algorithm.
Background technique
In recent years, scientific research personnel achieves huge progress in face tracking technology.In the Web-based instruction, video conference, prison Depending on requiring to track target face in real time with specific occasions such as monitoring, data transmitting and analysis.Remote teaching, view Frequency communication, videophone, identity validation, human-computer interaction etc. are all closely bound up with face tracking.
Currently, there is much more classical Face tracking algorithm, such as Camshift track algorithm, covered based on sequence Particle filter method, Mean shift algorithm of special Caro etc., although these algorithms can carry out accurately target face Tracking still when the posture of target face changes, tracks and is easy failure.
To solve the problems, such as in real time effectively to track multi-pose Face, Kaihua Zhang proposes one kind and is based on The compression of compressed sensing (compressive sensing, CS) theory tracks (Compressive Tracking, CT) algorithm. The problem of CT algorithm very good solution carries out real-time tracking to multi-pose Face, while having operand small, tracking velocity is fast, The advantages that strong real-time, still has that scale cannot be adaptive during tracking to target.Meanwhile when It after target leaves camera lens, is again introduced into, tracking failure.
Summary of the invention
The technical problem to be solved by the present invention is in view of the above technical problems, propose a kind of based on compression track algorithm Dimension self-adaption multi-pose Face tracking is effectively realized and is tracked to the dimension self-adaption of multi-pose target face.
The present invention is in order to solve the above technical problems, adopt the following technical scheme that
A kind of dimension self-adaption multi-pose Face tracking based on compression track algorithm, it is characterised in that in conjunction with face Detection algorithm and CT algorithm carry out detecting and tracking to target face；First on video display window, target face is examined It surveys, frame selects target face, then starts CT algorithm and carries out tracking and positioning to the target face selected by frame；On the basis of CT algorithm On, then by the Face datection algorithm based on Adaboost and Haar feature, target face is accurately positioned.
In above-mentioned technical proposal, including following key step:
Step S1: opening camera, reads in video data stream, meanwhile, starting target person face detection algorithm is realized to video The detection of target face in stream；
Step S2: on the basis of the target face scale detected, the generation window that a size is greater than target face is generated Mouthful；
Step S3: CT algorithm is initialized with the window generated, constructs the Bayes classifier of the positive negative sample of CT algorithm, together When, start CT algorithm, rough tracking is carried out to target face；
Step S4: using the tracking box of CT algorithm as the detection window of target face；
Step S5: judge whether detection window touches the boundary of video display window, if the judgment is Yes, return step S1；Conversely, starting the Face datection algorithm based on Adaboost and Haar feature in detection window, inside detection window Target face is detected；
Whether step S6: detecting face in judgment step S5, if the judgment is Yes, records in present frame, target person The size and location of face, and as display box；Conversely, being bold small and position using target person in the former frame recorded It sets, generates the display box of present frame；
Wherein, each parameter definition is as follows: detection window: the detection zone of target face, while being also the tracking of CT algorithm Frame；Tracking box: the tracking box of CT algorithm；Display box: the tracking effect of algorithm final output.
In above-mentioned technical proposal, the detection of step S1 target face refers to identifies specific face from complex background.
In above-mentioned technical proposal, step S2 generates window on the basis of the target face center detected, and length and width are pair 1.2~1.8 times for answering target face length and width scale, to guarantee CT algorithm to the accuracy of target face tracking.
Start CT algorithm in above-mentioned technical proposal, in step S3 and carry out rough tracking to target face, including generate with The building and update of machine calculation matrix, compression tracking Bayes classifier, specifically carry out as follows:
S31: t-th moment of step when t frame picture read in when, by the background to target face and its surrounding into Row sampling, to get the background negative sample around several target face positive samples and target face；Then pass through one A sparse calculation matrix carries out feature extraction to positive negative sample, then trains Bayes classifier with the feature extracted, quite In the Bayes classifier for initializing positive negative sample, lay the groundwork for the starting of next step CT algorithm；
Step S32: when t+1 frame picture is read in, using the position of target face in t frame picture and size as Benchmark is sampled around it, generates n detection block, then carries out feature extraction, feature extraction institute to this n detection block The sparseness measuring matrix of use and step S1 are identical to sparseness measuring matrix involved in the detection of target face in video flowing； Bayes classifier these features extracted to n detection block for reusing the initialization of t frame are classified, and are classified The window of maximum ratio is tracking box；New target window is thus got.
In above-mentioned technical proposal, the generating process of the n detection block are as follows: with rectangular area position where target face The upper left corner be that 45 positive samples using 4 pixels as radius are chosen in the center of circle；Using 8 pixels as inside radius, 25 pixels are outer Radius chooses 50 negative samples.
In above-mentioned technical proposal, the Face datection algorithm in step S5 based on Adaboost and Haar feature includes following tool Body step:
Step S51: face is described with Haar-Like feature, using the method for integrogram come the spy to face The characteristic value of sign is calculated；
Step S52: being classified using Adaboost algorithm, so that the feature that can most represent face is selected, namely These Weak Classifiers, are then combined by Haar-Like rectangular characteristic block, and then construct a strong classifier；
Step S53: the strong classifier that training obtains is connected, thus the stacking classification of one cascade structure of composition Device.
In above-mentioned technical proposal, the above-mentioned dimension self-adaption multi-pose Face tracking based on compression track algorithm, base The library open source OpenCV that Visual Studio 2010 and version under 7 operating system of windows are 2.4.4.
In above-mentioned technical proposal, the opening of camera, the reading of video and the formation of video frame are all based on The library function in the library OpenCV.
In compared with the existing technology to target face tracking during, although Compression alone track algorithm can be realized colourful State face is continuously tracked, but scale but cannot be adaptive；And simple Face datection algorithm, in the detection process, although It can accomplish the dimension self-adaption to target face tracking, still, since Face datection algorithm cannot be guaranteed that each frame can be examined Target face is measured, tracking picture will appear non-continuous event, and when occurring multiple faces in picture, non-targeted face pair There is the problems such as interference in the tracking of target face.It is proposed by the present invention based on compression tracking dimension self-adaption multi-pose Face with Track algorithm, continuous, quick, effective tracking can be carried out and based on Adaboost to multi-pose Face by combining compression track algorithm The advantages of Face datection algorithm of learning algorithm can quickly and effectively detect face realizes and carries out ruler to multi-pose Face Spend adaptive tracing.Coarse localization is carried out to target face using compression track algorithm, to reduce Face datection algorithm Search range, and then improve target person face detection algorithm accuracy and real-time；Using Face datection algorithm, target face is realized Accurate positionin, while realize target face dimension self-adaption tracking；Using target person face detection algorithm, target face is solved When leaving camera lens and being again introduced into, the problem of tracking failure；Using the continuity of target face motion process in time, mesh is realized Mark the tracking continuity problem in the case of the detection failure of Face datection algorithm.The present invention can guarantee camera by above method Accurately and effectively multi-pose is carried out to target face, dimension self-adaption tracking can be widely used in video monitoring, man-machine friendship Mutually, among virtual reality and various security systems such as: ATM machine monitors access control system.
Detailed description of the invention
Fig. 1 is the dimension self-adaption multi-pose Face tracking flow chart of the invention based on compression track algorithm.
Specific embodiment
Technical solution in order to further illustrate the present invention is described in detail this programme below in conjunction with attached drawing 1.
Dimension self-adaption multi-pose Face tracking of the invention based on compression track algorithm as shown in Figure 1, in reality During now, first on the display window of video, target face is detected, frame selects target face, then starting pressure Contracting track algorithm carries out tracking and positioning to the target face selected by frame.Know on the basis of compressing track algorithm, then through face Other algorithm, accurately positions target face.For convenience of the explanation of problem, now it is defined as follows:
Detection window: the detection zone of target face；
Tracking box: the tracking box of track algorithm；
Display box: the tracking effect of algorithm final output.
In above scheme, the technology platform needed is the Visual Studio 2010 under 7 operating system of windows, with And version is the library open source OpenCV of 2.4.4.
In above scheme, the opening of camera, the reading of video and the formation of video frame are all based on the library OpenCV Library function.
In above scheme, face recognition algorithms in the library OpenCV that face recognition algorithms are.
In above scheme, specifically comprise the following steps:
Step S1: opening camera, reads in video data stream, meanwhile, starting target person face detection algorithm is realized to video The detection of target face in stream.
Target Face datection refers to identifies specific face from complex background.
Step S2: on the basis of the target face detected, the window that a size is slightly larger than target face is generated.
To further illustrate the concrete methods of realizing for generating window, supplement following steps:
Step S21: it is found through experiment that, for CT algorithm during tracking to target, target face is bigger, and tracking is got over Accurately.Therefore during generating window, for the present invention on the basis of the target face center detected, length and width are corresponding mesh 1.2~1.8 times for marking face length and width scale, to guarantee CT algorithm to the accuracy of target face tracking.
Step S3: with initialization CT algorithm is generated, constructing the Bayes classifier of the positive negative sample of CT algorithm, meanwhile, starting CT algorithm carries out rough tracking to target face.It is as follows to compress track algorithm:
Step S31: when t frame picture is read in, being sampled by the background to target face and its surrounding, from And get several positive samples (target face) and negative sample (background around target face).Then sparse by one Calculation matrix positive negative sample is carried out feature extraction (i.e. realization dimensionality reduction), then with the trained Bayes's classification of the feature extracted Device is equivalent to the Bayes classifier for initializing positive negative sample, and the starting for compression track algorithm in next step is laid the groundwork.
Step S32: when t+1 frame picture is read in, using the position of target face in t frame picture and size as Benchmark is sampled around it, disposably generates n detection block, is then carried out feature extraction to this n detection block and (is used Sparseness measuring matrix and step S1 in it is identical).Bayes classifier these features for reusing the initialization of t frame are divided Class, the window for the maximum ratio classified are target window.New target window is thus got.
For the concrete methods of realizing for further illustrating compression track algorithm, following steps are supplemented:
Step S321: random measurement matrix is generated
One very typical random measurement matrix is random gaussian matrix, and matrix element meets N (0,1) distribution.But When needing to carry out dimensionality reduction to the higher image space of dimension, this matrix is not able to satisfy actual demand in sparse degree, A kind of very sparse mxn (m row, n column) random measurement matrix R, element r in matrix are used in CT algorithmijIt (indicates random to survey I-th row of moment matrix, jth column)
As s=2,1-1/s=1/2, that is to say has 1/2 element in random measurement matrix be zero, and calculation amount becomes former 1/2 come；
As s=3,1-1/s=2/3, that is to say has 2/3 element in random measurement matrix be zero, and calculation amount becomes former 2/3 come；
By this matrix, calculation amount just greatly reduces.S=m/4 is had chosen in CT algorithm, matrix R's is each Value that row only needs to calculate c element (number of sampling block that c expression generates at random, less than 4, generally 2 or 3).So Its computation complexity is O (cn).In addition, we only need to store the nonzero element of R, so required memory space is also very It is few.
Step S322: the building and update of compression tracking Bayes classifier
Compressing classifier employed in track algorithm is Bayes classifier, at the dimensionality reduction by random measurement matrix After reason, corresponding characteristic value is acquired, it is assumed that each element is all independently distributed.The then classification standard of classifier H (v) Are as follows:
Wherein, y ∈ (0,1) is classification samples label, and the value 0,1 of y respectively indicates positive sample and negative sample, viIndicate i-th A given sample, P (vi| y=1) and P (vi| y=0) respectively indicate the probability that given sample is positive sample and negative sample.It is assumed that two The priori conditions of class sample are identical, i.e. P (y=0)=0.5=P (y=1).It is random that Diaconis and Freedman demonstrates higher-dimension The accidental projection of vector is nearly all Gaussian Profile.Therefore it is presumed that the conditional probability P (v in classifier H (v)i| y=1) With P (vi| y=0) Gaussian Profile is also belonged to, and four parameters can be usedTo describe:
In formula,Positive negative sample is respectively indicated, 1 expression positive sample classifier is above designated as, is above designated as 0 expression Negative sample classifier,iIndicate theiIt is a,WithRespectively indicating mean value isVariance isGaussian Profile It is with mean valueVariance isGaussian Profile.The corresponding parameter for Bayes classifier updates, and has:
In formula: λ is Studying factors, and λ > 0.
Step S4: using the tracking box of CT algorithm as the detection window of target face.
Since the size of the tracking box of CT algorithm is slightly larger than target face, when target face is close to camera, according to It can so guarantee that target face is in the inside of detection window, to guarantee scale of the target face during camera certainly It adapts to.
Step S5: judging whether detection window touches the boundary of video window, if the judgment is Yes, return step S1； Conversely, starting the Face datection algorithm based on Adaboost algorithm and Haar-Like feature in detection window, to detection window Internal target face is detected.
By the judgement on boundary, target face tracking is reasonably switched with detection process, solve CT algorithm with The tracking Problem of Failure occurred during track target face (during tracking using CT algorithm to target item, works as mesh After mark leaves camera lens, when being again introduced into, it may appear that the problem of tracking is failed).
To further illustrate the Face datection algorithm based on Adaboost algorithm and Haar-Like feature, following step is supplemented It is rapid:
Step S51: face is described with Haar-Like feature, using the method for integrogram come the spy to face The characteristic value of sign is calculated.
Step S52: being classified (this is equivalent to a weak typing) using Adaboost algorithm, most can generation to select These Weak Classifiers, are then combined by the feature (Haar-Like rectangular characteristic block) of table face, and then construct one strong Classifier.
Step S53: the strong classifier that training obtains is connected, thus the stacking classification of one cascade structure of composition Device can effectively improve the speed of detection by cascade mode.
Whether step S6: detecting face in judgment step S5, if the judgment is Yes, records in present frame, target face Size and location, and as display box；Conversely, using target person face size and location in the former frame recorded, Generate the display box of present frame.
Dimension self-adaption multi-pose Face tracking based on compression track algorithm of the invention, it is characterised in that: two A algorithm --- the combination of compression tracking (Compressive Tracking) algorithm and Face datection algorithm, advantage obtain It is complementary.During algorithm is realized, the Visual Studio 2010 and version being mainly utilized under 7 platform of windows are 2.4.4 algorithm is realized in the library open source OpenCV.
Due in the library OpenCV, the front face based on Adaboost algorithm and Haar-Like eigenface detection algorithm Detection angle range substantially [- 20., 20.], when the deflection angle of face is more than this range, detection can fail, and then make Tracking process is obtained to occur discontinuously.
In view of the main reason for detection failure is that face deflection angle is excessive, and in deflection process, face is relative to mirror The distance of head does not occur more significantly to change, therefore, can be using the size of target face in former frame as in present frame The size of target face.And CT algorithm has preferable robustness, the i.e. variation of posture not for the attitudes vibration of target face Tracking effect more can be significantly influenced, thus it can be assumed that in adjacent two field pictures, target item and tracking box central point Between relative position be constant, and then by former frame, the relative position between target item and the central point of tracking box, with And the position of tracking box central point determines the position of target item in present frame in present frame.
To sum up, the dimension self-adaption multi-pose Face tracking proposed by the invention based on compression track algorithm, benefit Can be real-time with compression track algorithm, quickly target face is tracked, then on the basis of compressing tracking, people is added Face detection algorithm detects target face, to realize tracking.The algorithm, which substantially overcomes, will compress track algorithm application During face tracking, the problem and single Face datection algorithm that scale cannot be adaptive can not rule out non-targeted people in picture During face realizes tracking to the interference of target face tracking and single Face datection, due to not detecting target face So that tracking discontinuous problem.It is proposed of the invention has so that the tracking of target face is simpler quickly in camera Good expansion and practicability.

Claims (8)

1.一种基于压缩跟踪算法的尺度自适应多姿态人脸跟踪方法，其特征在于结合人脸检测算法和CT算法对目标人脸进行检测跟踪；首先在视频显示窗口上，对目标人脸进行检测，框选出目标人脸，然后启动CT算法对被框选的目标人脸进行跟踪定位；在CT算法的基础上，再通过基于Adaboost和Haar特征的人脸检测算法，对目标人脸进行精确的定位；包括如下主要步骤：1. a scale-adaptive multi-pose face tracking method based on a compression tracking algorithm, characterized in that the target face is detected and tracked in conjunction with a face detection algorithm and a CT algorithm; at first on the video display window, the target face is carried out. Detect, frame and select the target face, and then start the CT algorithm to track and locate the frame-selected target face; on the basis of the CT algorithm, the face detection algorithm based on Adaboost and Haar features is used to carry out the detection of the target face. Precise positioning; includes the following main steps: 步骤S1：打开摄像头，读入视频数据流，同时，启动目标人脸检测算法实现对视频流中目标人脸的检测；Step S1: turn on the camera, read in the video data stream, and at the same time, start the target face detection algorithm to detect the target face in the video stream; 步骤S2：以检测到的目标人脸尺度为基准，生成一个尺寸大于目标人脸的生成窗口；Step S2: generating a generation window with a size larger than the target face based on the detected target face scale; 步骤S3：用生成的窗口初始化CT算法，构建CT算法正负样本的贝叶斯分类器，同时，启动CT算法，对目标人脸进行粗略的跟踪；Step S3: initialize the CT algorithm with the generated window, construct a Bayesian classifier of positive and negative samples of the CT algorithm, and at the same time, start the CT algorithm to roughly track the target face; 步骤S4：将CT算法的跟踪框作为目标人脸的检测窗口；Step S4: use the tracking frame of the CT algorithm as the detection window of the target face; 步骤S5：判断检测窗口是否触碰到视频显示窗口的边界，如果判断为是，返回步骤S1；反之，在检测窗口内启动基于Adaboost和Haar特征的人脸检测算法，对检测窗口内部的目标人脸进行检测；Step S5: determine whether the detection window touches the border of the video display window, if it is determined to be yes, return to step S1; otherwise, start the face detection algorithm based on Adaboost and Haar features in the detection window, and detect the target person inside the detection window. face detection; 步骤S6：判断步骤S5中是否检测到人脸，如果判断为是，记录下当前帧中，目标人脸的大小及位置，并将其作为显示框；反之，利用记录下来的前一帧中目标人脸大小及位置，生成当前帧的显示框；Step S6: determine whether a human face is detected in step S5, and if yes, record the size and position of the target human face in the current frame, and use it as a display frame; otherwise, use the recorded target in the previous frame. The size and position of the face, generate the display frame of the current frame; 其中，各参数定义如下：检测窗口：目标人脸的检测区域，同时也是CT算法的跟踪框；跟踪框：CT算法的跟踪框；显示框：算法最终输出的跟踪效果。The parameters are defined as follows: detection window: the detection area of the target face, which is also the tracking frame of the CT algorithm; tracking frame: the tracking frame of the CT algorithm; display frame: the tracking effect of the final output of the algorithm. 2.根据权利要求1所述的基于压缩跟踪算法的尺度自适应多姿态人脸跟踪方法，其特征在于步骤S1目标人脸的检测指从复杂背景中识别出特定的人脸。2 . The scale-adaptive multi-pose face tracking method based on the compression tracking algorithm according to claim 1 , wherein the detection of the target face in step S1 refers to identifying a specific face from a complex background. 3 . 3.根据权利要求1所述的基于压缩跟踪算法的尺度自适应多姿态人脸跟踪方法，其特征在于步骤S2生成窗口以检测到的目标人脸中心点为基准，长宽为对应目标人脸长宽尺度的1.2～1.8倍，以保证CT算法对目标人脸跟踪的准确性。3. the scale-adaptive multi-pose face tracking method based on the compression tracking algorithm according to claim 1, is characterized in that step S2 generates the window with the detected target face center point as a benchmark, and the length and width are corresponding target faces 1.2 to 1.8 times the length and width to ensure the accuracy of the CT algorithm for tracking the target face. 4.根据权利要求1所述的基于压缩跟踪算法的尺度自适应多姿态人脸跟踪方法，其特征在于步骤S3中启动CT算法对目标人脸进行粗略的跟踪，包括生成随机测量矩阵、压缩跟踪贝叶斯分类器的构建与更新，具体按如下步骤进行：4. the scale adaptive multi-pose face tracking method based on compression tracking algorithm according to claim 1, it is characterized in that starting CT algorithm in step S3 carries out rough tracking to target face, including generating random measurement matrix, compression tracking The construction and update of the Bayesian classifier are as follows: 步骤S31：第t个时刻当第t帧图片读入的时候，通过对目标人脸及其周围的背景进行采样，从而获取到若干张目标人脸正样本以及目标人脸周围的背景负样本；然后通过一个稀疏的测量矩阵对正负样本进行特征提取，再用提取到的特征训练贝叶斯分类器，相当于初始化正负样本的贝叶斯分类器，为下一步CT算法的启动做铺垫；Step S31: at the t-th moment, when the t-th frame picture is read in, by sampling the target face and its surrounding background, several positive samples of the target face and negative background samples around the target face are obtained; Then use a sparse measurement matrix to extract the features of the positive and negative samples, and then use the extracted features to train the Bayesian classifier, which is equivalent to initializing the Bayesian classifier of the positive and negative samples, paving the way for the next step of the CT algorithm. ; 步骤S32：当第t+1帧图片读入的时候，将第t帧图片中目标人脸的位置与大小作为基准，在其周围进行采样，生成n个检测框，然后对这n个检测框进行特征提取，特征提取所采用的稀疏测量矩阵和步骤S1对视频流中目标人脸的检测中所涉及的稀疏测量矩阵相同；再使用第t帧初始化的贝叶斯分类器对n个检测框所提取的这些特征进行分类，分类得到的最大比例的窗口即为跟踪框；这样就获取到了新的目标窗口。Step S32: When the t+1th frame picture is read in, take the position and size of the target face in the tth frame picture as a benchmark, perform sampling around it, generate n detection frames, and then analyze the n detection frames. Perform feature extraction. The sparse measurement matrix used in feature extraction is the same as the sparse measurement matrix involved in the detection of the target face in the video stream in step S1; These extracted features are classified, and the window with the largest proportion obtained by classification is the tracking frame; thus, a new target window is obtained. 5.根据权利要求4所述的基于压缩跟踪算法的尺度自适应多姿态人脸跟踪方法，其特征在于所述的n个检测框的生成过程为：以目标人脸所在矩形区域位置的左上角为圆心，以4个像素为半径，选取45个正样本；以8个像素为内半径，25个像素为外半径选取50个负样本。5. the scale-adaptive multi-pose face tracking method based on compression tracking algorithm according to claim 4, it is characterized in that the generation process of described n detection frames is: with the upper left corner of the rectangular area position where the target face is located is the center of the circle, with 4 pixels as the radius, and 45 positive samples are selected; with 8 pixels as the inner radius and 25 pixels as the outer radius, 50 negative samples are selected. 6.根据权利要求1所述的基于压缩跟踪算法的尺度自适应多姿态人脸跟踪方法，其特征在于步骤S5中基于Adaboost和Haar特征的人脸检测算法包括如下具体步骤：6. the scale adaptive multi-pose face tracking method based on compression tracking algorithm according to claim 1, is characterized in that in step S5, the face detection algorithm based on Adaboost and Haar feature comprises following concrete steps: 步骤S51：将人脸用Haar-Like特征来进行描述，使用积分图的方法来对人脸的特征的特征值进行计算；Step S51: describe the face with Haar-Like features, and use the method of integral graph to calculate the eigenvalues of the features of the face; 步骤S52：使用Adaboost算法来进行分类，从而挑选最能代表人脸的特征，也即Haar-Like矩形特征块，然后将这些弱分类器进行组合，进而构建出一个强分类器；Step S52: Use the Adaboost algorithm to classify, so as to select the feature that best represents the face, that is, the Haar-Like rectangular feature block, and then combine these weak classifiers to construct a strong classifier; 步骤S53：将训练得到的强分类器进行串联，从而组成一个级联结构的层叠分类器。Step S53: Connect the strong classifiers obtained by training in series to form a cascaded classifier with a cascade structure. 7.根据权利要求1所述的基于压缩跟踪算法的尺度自适应多姿态人脸跟踪方法，其特征在于基于windows 7操作系统下的Visual Studio 2010，以及版本为2.4.4的开源OpenCV库。7. The scale-adaptive multi-pose face tracking method based on compression tracking algorithm according to claim 1, is characterized in that based on Visual Studio 2010 under Windows 7 operating system, and the open source OpenCV library of version 2.4.4. 8.根据权利要求1所述的基于压缩跟踪算法的尺度自适应多姿态人脸跟踪方法，其特征在于摄像头的打开、视频的读入、以及视频框的形成，都是基于OpenCV库的库函数。8. the scale adaptive multi-pose face tracking method based on compression tracking algorithm according to claim 1, it is characterized in that the opening of camera, the reading of video and the formation of video frame are all based on the library function of OpenCV library .
CN201610947868.1A 2016-10-26 2016-10-26 Scale-adaptive multi-pose face tracking method based on compression tracking algorithm Active CN106570471B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201610947868.1A CN106570471B (en) 2016-10-26 2016-10-26 Scale-adaptive multi-pose face tracking method based on compression tracking algorithm

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201610947868.1A CN106570471B (en) 2016-10-26 2016-10-26 Scale-adaptive multi-pose face tracking method based on compression tracking algorithm

Publications (2)

Publication Number Publication Date
CN106570471A CN106570471A (en) 2017-04-19
CN106570471B true CN106570471B (en) 2019-04-26

Family

ID=58535097

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201610947868.1A Active CN106570471B (en) 2016-10-26 2016-10-26 Scale-adaptive multi-pose face tracking method based on compression tracking algorithm

Country Status (1)

Country Link
CN (1) CN106570471B (en)

Families Citing this family (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109990893B (en) * 2017-12-29 2021-03-09 浙江宇视科技有限公司 Photometric parameter calculation method and device and terminal equipment
CN108805094A (en) * 2018-06-19 2018-11-13 合肥工业大学 Data enhancement methods based on artificial face
CN109271883A (en) * 2018-08-28 2019-01-25 武汉科技大学 A kind of method for tracking target merging study mechanism
CN110705478A (en) * 2019-09-30 2020-01-17 腾讯科技（深圳）有限公司 Face tracking method, device, equipment and storage medium
CN111259907B (en) * 2020-03-12 2024-03-12 Oppo广东移动通信有限公司 Content identification method and device and electronic equipment
CN113591607B (en) * 2021-07-12 2023-07-04 辽宁科技大学 Station intelligent epidemic situation prevention and control system and method

Citations (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103905733A (en) * 2014-04-02 2014-07-02 哈尔滨工业大学深圳研究生院 Method and system for conducting real-time tracking on faces by monocular camera

Patent Citations (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103905733A (en) * 2014-04-02 2014-07-02 哈尔滨工业大学深圳研究生院 Method and system for conducting real-time tracking on faces by monocular camera

Non-Patent Citations (3)

* Cited by examiner, † Cited by third party
Title
一种改进的实时压缩跟踪算法;钟权 等;《光电工程》;20140430;全文
基于Adaboost算法的驾驶员疲劳驾驶检测;熊池亮;《中国优秀硕士学位论文全文数据库 工程科技II辑》;20140115;第4页第2段，第17页最后一段，第18页
尺度自适应的多模型压缩跟踪算法;刘晴 等;《系统工程与电子技术》;20160430;全文

Also Published As

Publication number Publication date
CN106570471A (en) 2017-04-19

Similar Documents

Publication Publication Date Title
CN106570471B (en) Scale-adaptive multi-pose face tracking method based on compression tracking algorithm
Mahmood et al. WHITE STAG model: Wise human interaction tracking and estimation (WHITE) using spatio-temporal and angular-geometric (STAG) descriptors
Ji et al. Advances in view-invariant human motion analysis: A review
CN112801043B (en) Real-time video face keypoint detection method based on deep learning
Kamal et al. A hybrid feature extraction approach for human detection, tracking and activity recognition using depth sensors
Devanne et al. 3-d human action recognition by shape analysis of motion trajectories on riemannian manifold
Thome et al. A real-time, multiview fall detection system: A LHMM-based approach
TWI430185B (en) Facial expression recognition systems and methods and computer program products thereof
Gupta et al. Nose, eyes and ears: Head pose estimation by locating facial keypoints
Czupryński et al. High accuracy head pose tracking survey
Rani et al. Kinematic joint descriptor and depth motion descriptor with convolutional neural networks for human action recognition
Abate et al. Head pose estimation by regression algorithm
Tan et al. Real-time accurate 3D head tracking and pose estimation with consumer RGB-D cameras
Tian et al. Absolute head pose estimation from overhead wide-angle cameras
Zhong et al. DynaTM-SLAM: Fast filtering of dynamic feature points and object-based localization in dynamic indoor environments
Wang et al. A new hand gesture recognition algorithm based on joint color-depth superpixel earth mover's distance
Elias et al. Understanding the gap between 2D and 3D skeleton-based action recognition
Badave et al. Head pose estimation based robust multicamera face recognition
Hu et al. Human interaction recognition using spatial-temporal salient feature
Ren et al. An investigation of skeleton-based optical flow-guided features for 3D action recognition using a multi-stream CNN model
Ming Hand fine-motion recognition based on 3D Mesh MoSIFT feature descriptor
CN105740819A (en) Integer programming based crowd density estimation method
CN107230220A (en) A kind of new space-time Harris angular-point detection methods and device
Siriteerakul Advance in head pose estimation from low resolution images: A review
Jiang et al. Spatial and temporal pyramid-based real-time gesture recognition

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant