CN101477626B - Method for detecting human head and shoulder in video of complicated scene - Google Patents

Abstract

The invention relates to a method for detecting head and shoulders of human body in a video with complex scene, which belongs to the technical field of computer information mining. The method comprises the following steps: manually determining the head and shoulder picture, the background picture and the pictures of other parts of the human body in each frame of a video as the positive and the negative samples, and mirroring the pictures; extracting the gradient vectors of the positive and the negative samples, and training a first classifier; selecting the head and the shoulder picture as the new positive sample, and the pictures of other parts of the human body as the new negative sample; extracting the gradient vectors of the positive and the negative samples, and training a second classifier; determining the position and the size of a window to be detected in one frame of a video frame to be detected, and extracting the gradient vector; classifying the gradient vector by the first classifier, and stopping detection if the classification result is negative; classifying the gradient vector by the second classifier if the first classification result is positive, and determining that the window contains the head and the shoulder if the classification is positive; and changing the position and the size of the window to detect new windows. The method can improve the accuracy as well as the detection speed.

Description

Translated fromChinese

一种在复杂场景的视频中进行人体头肩检测的方法A Method for Human Head and Shoulders Detection in Video of Complex Scenes

技术领域technical field

本发明属于计算机信息挖掘技术领域，特别涉及一种在复杂场景的图片中对人体的头肩检测的方法，尤其涉及在真实世界的监控视频的帧中行人的头肩检测。The invention belongs to the technical field of computer information mining, and in particular relates to a method for detecting the head and shoulders of a human body in a picture of a complex scene, in particular to detecting the head and shoulders of a pedestrian in a frame of a monitoring video in the real world.

背景技术Background technique

近年来，在计算机视频分析领域，在视频中对人体的检测是一个热门的研究方向。在人体检测的各种方法中，通过检测身体的各部分来检测人体是一个重要辅助手段。而对这些身体各部分，头肩区域是一个非常显著的特征。由于视频中的经常会出现人体被部分遮挡的情况，导致检测困难，而此时头肩还有很高的概率被检测出来，所以检测头肩对检测人体很好的辅助作用。同时，在视频事件检测领域，人在头肩附近的许多动作往往包含一些隐含的事件信息，比如挥手或者打电话等。所以，复杂背景下的头肩检测具有重要的意义。In recent years, in the field of computer video analysis, human body detection in video is a hot research direction. Among the various methods of human body detection, detecting human body by detecting various parts of the body is an important auxiliary means. Of these body parts, the head and shoulders area is a very prominent feature. Since the human body is often partially blocked in the video, it is difficult to detect, and at this time the head and shoulders still have a high probability of being detected, so the detection of the head and shoulders is a good auxiliary function for detecting the human body. At the same time, in the field of video event detection, many actions of people near the head and shoulders often contain some implicit event information, such as waving or making a phone call. Therefore, the detection of head and shoulders in complex background is of great significance.

头肩检测属于目标检测，而在目标检测领域，方法可以分为两类，一是做背景提取或分割，分离出的前景目标作为检测结果。二是在图像中直接搜索目标。在视频中用背景提取的方法只能应用于静态摄像机，对于场景中静止不动的目标，检测非常困难，这限制了它的应用范围。所以现在一般采用在图像中直接搜索目标的方法。这些方法一般使用分类器根据目标的特征对目标进行分类。目标的特征是目标本身包含的特征信息，比如目标图像区域的颜色直方图，纹理，梯度等。提取目标的特征之后，分类器根据目标的特征对目标的类别进行判断。目前国际通用的分类器主要是支持向量机(以下简称SVM)，但是目前的单级SVM分类器往往由于只进行一次分类，准确率不高。Head and shoulders detection belongs to target detection. In the field of target detection, methods can be divided into two categories. One is to do background extraction or segmentation, and the separated foreground target is used as the detection result. The second is to directly search for the target in the image. The method of background extraction in video can only be applied to static cameras, and it is very difficult to detect stationary objects in the scene, which limits its application range. Therefore, the method of directly searching for the target in the image is generally used now. These methods generally use classifiers to classify objects according to their characteristics. The feature of the target is the feature information contained in the target itself, such as the color histogram, texture, gradient, etc. of the target image area. After extracting the features of the target, the classifier judges the category of the target according to the features of the target. At present, the internationally used classifier is mainly support vector machine (hereinafter referred to as SVM), but the current single-stage SVM classifier often has a low accuracy rate because it only performs one classification.

发明内容Contents of the invention

本发明的目的是为克服已有技术的不足之处，提出一种在复杂场景的视频中进行人体头肩检测的方法，以梯度方向直方图作为描述目标的特征。采用两级的SVM作为分类器，可以提高准确率，同时提高检测速度。The purpose of the present invention is to overcome the deficiencies of the prior art, and propose a method for detecting human head and shoulders in the video of complex scenes, using the gradient direction histogram as a feature to describe the target. Using a two-stage SVM as a classifier can improve the accuracy and improve the detection speed.

本发明是将一定数量的头肩图片和背景图片作为正负样本集，训练SVM作为第一级分类器。用头肩图片和身体非头肩部分的图片作为正负样本，训练SVM作为第二级分类器。这样就构成了一个两级的级联分类器。检测区域依次经过这两级SVM检测，以此检测结果作为最终结果。The present invention uses a certain number of head and shoulder pictures and background pictures as positive and negative sample sets, and trains SVM as a first-level classifier. Using head and shoulder pictures and pictures of non-head and shoulder parts of the body as positive and negative samples, train SVM as a second-level classifier. This constitutes a two-stage cascade classifier. The detection area is detected by the two-stage SVM in turn, and the detection result is taken as the final result.

本发明中使用的SVM是目前在国际比较流行的LibSVM分类器，不对它进行修改。The SVM used in the present invention is the currently popular LibSVM classifier in the world, and it will not be modified.

本发明提出的在复杂场景的视频中进行人体头肩检测的方法，主要包括以下步骤：The method that the present invention proposes carries out human body head and shoulders detection in the video of complex scene, mainly comprises the following steps:

(1)从待检测的一类视频中选择一个视频。从该视频的各帧中人工标定一定数量(至少1000张)的头肩图片、一定数量(至少1000张)背景图片和一定数量(至少1000张)的身体其他部分的图片，其中要求这些图片的边长至少1厘米。以头肩图片作为正样本图片，以背景图片作为负样本图片；；(1) Select a video from a class of videos to be detected. Manually mark a certain number (at least 1000) of head and shoulder pictures, a certain number (at least 1000) of background pictures and a certain number (at least 1000) of pictures of other parts of the body from each frame of the video. The sides are at least 1 cm long. Take the head and shoulders image as a positive sample image, and use the background image as a negative sample image;;

(2)将得到的正负样本图片进行左右镜像，增加样本的数量；(2) Mirror the obtained positive and negative sample pictures left and right to increase the number of samples;

(3)提取得到的正负样本图片的梯度方向直方图，并将梯度方向直方图转化为向量的形式，作为样本图片的梯度向量；(3) Extract the gradient direction histogram of the positive and negative sample pictures, and convert the gradient direction histogram into a vector form as the gradient vector of the sample picture;

(4)用从正负样本中提取出的梯度向量对第一级支持向量机(SVM)进行训练，生成一个用于分类的第一级模型；(4) Use the gradient vector extracted from the positive and negative samples to train the first-level support vector machine (SVM) to generate a first-level model for classification;

(5)用所述头肩图片作为新正样本，用所述身体其他部分的图片代替所述背景图片作为新负样本；(5) Use the head and shoulders picture as a new positive sample, and replace the background picture with pictures of other parts of the body as a new negative sample;

(6)提取所述新正负样本图片的梯度方向直方图，并将梯度方向直方图转化为1乘N的向量的形式，N为正整数，作为新样本的梯度向量；(6) Extract the gradient direction histogram of the new positive and negative sample pictures, and convert the gradient direction histogram into a vector form of 1 multiplied by N, where N is a positive integer as the gradient vector of the new sample;

(7)用从新的正负样本中提取出的梯度向量对第二级支持向量机(SVM)进行训练，生成一个用于分类的第二级模型；(7) Use the gradient vector extracted from the new positive and negative samples to train the second-level support vector machine (SVM) to generate a second-level model for classification;

(8)读入一个待检测视频，提取该视频的一帧图像；(8) read in a video to be detected, and extract a frame image of the video;

(9)确定该帧图像上某一个待检测窗口的位置和大小，采用步骤(3)的方法提取该窗口的梯度方向直方图，并得到该窗口的梯度向量；(9) determine the position and the size of a certain window to be detected on the frame image, adopt the method of step (3) to extract the gradient direction histogram of the window, and obtain the gradient vector of the window;

(10)将该梯度向量通过第一级分类器进行分类检测，如果分类结果为负(即此窗口不包含头肩头像)，则结束该窗口的检测，转步骤(11)；如果第一级分类结果为正(即第一级分类器判定此窗口包含头肩头像)，则将该梯度向量通过第二级分类器进行分类检测；如果第二级分类结果为负，转步骤(11)，如果分类结果为正，则确认该窗口包含头肩，将窗口的坐标保存下来，作为该窗口的检测结果；(10) Classify and detect the gradient vector through the first-level classifier, if the classification result is negative (that is, this window does not contain the head and shoulder head portrait), then end the detection of the window and go to step (11); if the first level The classification result is positive (that is, the first-level classifier determines that this window contains the head and shoulder head portrait), then the gradient vector is classified and detected by the second-level classifier; if the second-level classification result is negative, go to step (11), If the classification result is positive, it is confirmed that the window contains head and shoulders, and the coordinates of the window are saved as the detection result of the window;

(11)改变窗口的位置和大小，采用步骤(3)的方法提取该窗口的梯度方向直方图，并得到该窗口的梯度向量，转步骤(10)进行该窗口的分类检测，最终得到各窗口的检测结果。(11) Change the position and size of the window, use the method of step (3) to extract the gradient direction histogram of the window, and obtain the gradient vector of the window, turn to step (10) to classify and detect the window, and finally get each window test results.

上述方法的步骤(3)具体包括如下步骤：The step (3) of the above method specifically includes the following steps:

(31)将每一个样本图片作为一个窗口，将窗口分为MxN的块，块之间有30％-50％的重叠，M、N均为正整数；(31) each sample picture is used as a window, and the window is divided into MxN blocks, with 30%-50% overlap between the blocks, and M and N are positive integers;

(32)该每个块平分成多个单元；(32) Each block is equally divided into multiple units;

(33)对该每个单元内的像素计算梯度方向和大小；(33) Calculate the gradient direction and size of the pixel in each unit;

(34)将每个单元内的像素的梯度按方向统计成一个直方图，并将梯度方向直方图转化为向量的形式，作为该单元的梯度向量；(34) The gradient of the pixel in each unit is counted into a histogram according to the direction, and the gradient direction histogram is converted into a vector form, as the gradient vector of the unit;

(35)将各个单元的梯度向量连接成一个长向量，作为该块的梯度向量；(35) Connect the gradient vectors of each unit into a long vector as the gradient vector of the block;

(36)将窗口内各个块的梯度向量连接成一个长向量，作为该窗口的梯度向量；(36) The gradient vectors of each block in the window are connected into a long vector, as the gradient vector of the window;

(37)将该窗口的梯度向量进行归一化，作为该样本图片的梯度向量；(37) normalize the gradient vector of the window as the gradient vector of the sample picture;

上述方法的步骤(11)改变窗口的位置和大小，进行该窗口的分类检测，具体包括如下步骤：The step (11) of said method changes the position and the size of window, carries out the classification detection of this window, specifically comprises the following steps:

(111)移动该窗口的坐标，并保持30％-80％重叠度，进行该窗口的分类检测；(111) Move the coordinates of the window, and keep 30%-80% overlap, and carry out the classification detection of the window;

(112)改变窗口的尺寸大小(根据视频中头肩的大小确定取值范围)，并且依次移动窗口的位置，进行该窗口的分类检测。(112) Change the size of the window (determine the value range according to the size of the head and shoulders in the video), and move the position of the window in turn to perform classification detection of the window.

本发明的特点及效果：Features and effects of the present invention:

本发明提出的在复杂场景的视频中进行人体头肩检测的方法，用于对真实世界里的监控视频中的人进行头肩检测。选用了梯度方向直方图作为头肩的特征表示。梯度方向直方图最近几年开始用于目标检测领域，选用它表示头肩图像，能够保存目标的边缘走向特征，实验已经证明，它是一种鲁棒的特征，能提高检测的性能。同时，使用不同的样本训练两个支持向量机，将它们构成级联分类器。在检测过程中，第一级分类器去掉明显不含头肩的区域，通过第一级分类器的窗口再由第二级分类器进行检测。这样，由于采用两级分类器，可以提高准确率，同时提高了检测速度。在视频中将人的头肩检测出来，可以用于人体跟踪、事件检测，对于实现自动监控具有重要意义。The method for detecting the head and shoulders of a human body in a video of a complex scene proposed by the present invention is used for detecting the head and shoulders of a person in a surveillance video in the real world. The gradient direction histogram is chosen as the feature representation of the head and shoulders. The histogram of gradient direction has been used in the field of target detection in recent years. It is selected to represent the head and shoulders image, which can preserve the edge direction characteristics of the target. Experiments have proved that it is a robust feature that can improve the performance of detection. At the same time, two support vector machines are trained using different samples, and they form a cascade classifier. In the detection process, the first-level classifier removes the area that obviously does not contain the head and shoulders, and then the second-level classifier detects it through the window of the first-level classifier. In this way, due to the use of two-stage classifiers, the accuracy rate can be improved, and the detection speed is also improved. Detecting the head and shoulders of people in the video can be used for human body tracking and event detection, which is of great significance for realizing automatic monitoring.

具体实施方式Detailed ways

本发明提出的一种在复杂场景的视频中进行人体头肩检测的方法，结合实施例详细说明如下：A method for detecting human head and shoulders in a video of a complex scene proposed by the present invention is described in detail in conjunction with the embodiments as follows:

在视频中进行人体的头肩检测，实质方法仍是在图片上进行检测，本实施例采用直接在图片上搜索的方法，具体包括以下步骤：The head and shoulder detection of the human body in the video, the essential method is still to detect on the picture, this embodiment adopts the method of directly searching on the picture, specifically includes the following steps:

(1)从待检测的一类视频中选择一个视频；从该视频的各帧中人工标定一定数量(至少1000张)的头肩图片、一定数量(至少1000张)背景图片和一定数量(至少1000张)的身体其他部分的图片，其中要求这些图片的边长至少1厘米；以头肩图片作为正样本图片，以背景图片作为负样本图片；(1) Select a video from a class of videos to be detected; manually mark a certain number (at least 1000) of head and shoulder pictures, a certain number (at least 1000) of background pictures and a certain number (at least 1000 pictures) of other parts of the body, where the side length of these pictures is required to be at least 1 cm; the head and shoulders picture is used as a positive sample picture, and the background picture is used as a negative sample picture;

(2)将得到的正负样本图片进行左右镜像，增加样本的数量；(2) Mirror the obtained positive and negative sample pictures left and right to increase the number of samples;

(3)提取得到的正负样本图片的梯度方向直方图，并将梯度方向直方图转化为向量的形式，作为样本图片的梯度向量；具体包括：(3) Extract the gradient direction histogram of the obtained positive and negative sample pictures, and convert the gradient direction histogram into a vector form as the gradient vector of the sample picture; specifically include:

(31)将一个每个图片看作一个窗口，将该窗口划分为3×3的小块，小块之间有50％的重叠；(31) each picture is regarded as a window, and the window is divided into 3 × 3 small blocks, with 50% overlap between the small blocks;

(32)该每个小块又被平分为四个小单元；(32) Each small block is equally divided into four small units;

(33)对该每个单元内的像素计算梯度方向和大小，本实施例的梯度值采用模板[-1，0，1]的方法来计算，方向为：${\mathrm{grad}}_{(x,y)}=\arctan \frac{I_{(x,y+1)}-I_{(x,y-1)}}{I_{(x+1,y)}-I_{(x-1,y)}}$(33) Calculate the gradient direction and size of the pixels in each unit. The gradient value of this embodiment is calculated by the method of template [-1, 0, 1], and the direction is:${\mathrm{grad}}_{(x,\mathrm{the\; y})}=\arctan \frac{I_{(x,\mathrm{the\; y}+1)}-I_{(x,\mathrm{the\; y}-1)}}{I_{(x+1,\mathrm{the\; y})}-I_{(x-1,\mathrm{the\; y})}}$

大小为：The size is:

${\mathrm{<span><span>value</span>value</span>}}_{<span><span>(</span>(</span>\mathrm{<span><span>x</span>x</span>}<span><span>,</span>,</span>\mathrm{<span><span>y</span>the\; y</span>}<span><span>)</span>)</span>}<span><span>=</span>=</span>\sqrt{{<span><span>[</span>[</span>{\mathrm{<span><span>I</span>I</span>}}_{<span><span>(</span>(</span>\mathrm{<span><span>x</span>x</span>}<span><span>+</span>+</span>\mathrm{<span><span>1</span>1</span>}<span><span>,</span>,</span>\mathrm{<span><span>y</span>the\; y</span>}<span><span>)</span>)</span>}<span><span>-</span>-</span>{\mathrm{<span><span>I</span>I</span>}}_{<span><span>(</span>(</span>\mathrm{<span><span>x</span>x</span>}<span><span>-</span>-</span>\mathrm{<span><span>1</span>1</span>}<span><span>,</span>,</span>\mathrm{<span><span>y</span>the\; y</span>}<span><span>)</span>)</span>}<span><span>]</span>]</span>}^{\mathrm{<span><span>2</span>2</span>}}<span><span>+</span>+</span>{<span><span>[</span>[</span>{\mathrm{<span><span>I</span>I</span>}}_{<span><span>(</span>(</span>\mathrm{<span><span>x</span>x</span>}<span><span>,</span>,</span>\mathrm{<span><span>y</span>the\; y</span>}<span><span>+</span>+</span>\mathrm{<span><span>1</span>1</span>}<span><span>)</span>)</span>}<span><span>-</span>-</span>{\mathrm{<span><span>I</span>I</span>}}_{<span><span>(</span>(</span>\mathrm{<span><span>x</span>x</span>}<span><span>,</span>,</span>\mathrm{<span><span>y</span>the\; y</span>}<span><span>-</span>-</span>\mathrm{<span><span>1</span>1</span>}<span><span>)</span>)</span>}<span><span>]</span>]</span>}^{\mathrm{<span><span>2</span>2</span>}}}$

其中，grad_(x，y)是该像素点的梯度方向。I_(x，y)是该像素点的亮度值。Value_(x，y)是该像素点的梯度的值；Among them, grad_{(x, y)} is the gradient direction of the pixel. I_{(x, y)} is the brightness value of the pixel. Value_{(x, y)} is the value of the gradient of the pixel;

(34)将单元内各像素的梯度值按照梯度方向统计成颜色直方图，以向量的形式保存，将该向量作为该单元的梯度向量；(34) Statize the gradient value of each pixel in the unit into a color histogram according to the gradient direction, save it in the form of a vector, and use the vector as the gradient vector of the unit;

(35)将各单元的梯度向量连接起来，每个小块用一个归一化的4×9＝36维的向量来表示，作为该块的梯度向量；(35) The gradient vectors of each unit are connected, and each small block is represented by a normalized 4 * 9 = 36-dimensional vector, as the gradient vector of the block;

(36)将各块的梯度向量连接起来，每个窗口的梯度方向直方图就可以用一个9×36＝324维的向量来表示，作为该窗口的梯度向量；(36) The gradient vectors of each block are connected, and the gradient direction histogram of each window can be represented by a vector of 9 * 36=324 dimensions, as the gradient vector of this window;

(4)用从正负样本中提取出的梯度向量对第一级支持向量机(LibSVM)进行训练，生成一个用于分类的第一级模型；(4) Use the gradient vector extracted from the positive and negative samples to train the first-level support vector machine (LibSVM) to generate a first-level model for classification;

(5)用所述头肩图片作为新正样本，用所述身体其他部分的图片代替所述背景图片作为新负样本；(5) Use the head and shoulders picture as a new positive sample, and replace the background picture with pictures of other parts of the body as a new negative sample;

(6)提取得到的新正负样本图片的梯度方向直方图，并将梯度方向直方图转化为1乘324的向量的形式，作为新样本的梯度向量；(6) Extract the gradient direction histogram of the new positive and negative sample pictures obtained, and convert the gradient direction histogram into a vector form of 1 by 324, as the gradient vector of the new sample;

(7)用从新的正负样本中提取出的梯度向量对第二级支持向量机(LibSVM)进行训练，生成一个用于分类的第二级模型；(7) Use the gradient vector extracted from the new positive and negative samples to train the second-level support vector machine (LibSVM) to generate a second-level model for classification;

(8)调用openCV库读入一个待检测视频，解析出该视频上的一帧；(8) Call the openCV library to read a video to be detected, and parse out a frame on the video;

(9)确定该帧图像上某一个待检测窗口的位置和大小，采用步骤(31)到步骤(37)的方法提取该窗口的梯度方向直方图，并得到该窗口的梯度向量；(9) determine the position and the size of a certain window to be detected on the frame image, adopt the method of step (31) to step (37) to extract the gradient direction histogram of the window, and obtain the gradient vector of the window;

(10)将该梯度向量通过第一级LibSVM分类器进行分类检测，如果分类结果为负(即此窗口不包含头肩头像)，则结束该窗口的检测，则转步骤(13)；如果第一级分类结果为正(即第一级分类器判定此窗口包含头肩头像)，则将该梯度向量通过第二级LibSVM分类器进行分类检测；如果第二级分类结果为负，转步骤(13)，如果分类结果为正，则确认该窗口包含头肩，将窗口的坐标保存下来，作为该窗口的检测结果；(10) Classify and detect the gradient vector through the first-level LibSVM classifier, if the classification result is negative (that is, this window does not contain the head and shoulder head portrait), then end the detection of this window, then go to step (13); if the first The first-level classification result is positive (that is, the first-level classifier determines that this window contains the head and shoulder head portrait), then the gradient vector is classified and detected by the second-level LibSVM classifier; if the second-level classification result is negative, go to step ( 13), if the classification result is positive, it is confirmed that the window contains the head and shoulders, and the coordinates of the window are saved as the detection result of the window;

(11)移动该窗口的坐标，并保持30％-80％重叠度，对该待检测窗口进行检测；(11) Move the coordinates of the window, and keep 30%-80% overlap, and detect the window to be detected;

(12)改变窗口的尺寸大小(根据视频中头肩的大小确定取值范围)，并且依次移动窗口的位置，进行该窗口的分类检测，最终得到各窗口的检测结果。(12) Change the size of the window (determine the value range according to the size of the head and shoulders in the video), and move the position of the window in turn, carry out the classification detection of the window, and finally obtain the detection results of each window.

Claims (3)

1. carry out the method that human head and shoulder detects in the video of a complex scene, it is characterized in that, mainly may further comprise the steps:

(1) from a class video to be detected, selects a video; The picture of other parts of health of artificial head shoulder picture, some background pictures and some of demarcating some from each frame of this video wherein requires at least 1 centimetre of the length of side of these pictures; With head shoulder picture as samples pictures just, with background picture as the negative sample picture;

(2) the positive and negative samples pictures that will obtain is carried out left and right sides mirror image, increases the quantity of sample;

(3) extract the gradient orientation histogram of the positive and negative samples pictures obtain, and gradient orientation histogram is converted into the form of vector, as the gradient vector of samples pictures;

(4) use the gradient vector that from positive negative sample, extracts that first order support vector machine is trained, generate a first order model that is used to classify;

(5) with described head shoulder picture as the sample first month of the lunar year, replace described background picture as new negative sample with the picture of described other parts of health;

(6) extract the gradient orientation histogram of described first month of the lunar year of negative sample picture, and gradient orientation histogram is converted into 1 form of taking advantage of the vector of N, N is a positive integer, as the gradient vector of new samples;

(7) use the gradient vector that from new positive negative sample, extracts that second level support vector machine is trained, generate a second level model that is used to classify;

(8) read in a video to be detected, extract a two field picture of this video;

(9) determine some position of window to be detected and size on this two field picture, adopt the method for step (3) to extract the gradient orientation histogram of this window, and obtain the gradient vector of this window;

(10) this gradient vector is carried out classification and Detection by first order sorter,, change step (11) if classification results then finishes the detection of this window for negative; If first order classification results then carries out classification and Detection with this gradient vector by second level sorter for just; If second level classification results changes step (11) for negative, if classification results confirms then that for just this window comprises the head shoulder, preserves the coordinate of window get off, as the testing result of this window;

(11) change position of window and size, adopt the method for step (3) to extract the gradient orientation histogram of this window, and obtain the gradient vector of this window, change the classification and Detection that step (10) is carried out this window, finally obtain the testing result of each window.

2. method according to claim 1 is characterized in that the step of described method (3) specifically comprises the steps:

(31) with each samples pictures as a window, window is divided into the piece of M x N, the overlapping of 30%-50% arranged between the piece, M, N are positive integer;

(32) this each piece is divided equally into a plurality of unit;

(33) to pixel compute gradient direction and size in this each unit;

(34) gradient with the pixel in each unit becomes a histogram by directional statistics, and gradient orientation histogram is converted into vectorial form, as the gradient vector of this unit;

(35) gradient vector with each unit connects into a long vector, as the gradient vector of this piece;

(36) gradient vector with each piece in the window connects into a long vector, as the gradient vector of this window;

(37) gradient vector of this window is carried out normalization, as the gradient vector of this samples pictures.

3. method according to claim 1 is characterized in that, the step of described method (11) changes position of window and size, carries out the classification and Detection of this window, specifically comprises the steps:

(111) move the coordinate of this window, and keep the 30%-80% degree of overlapping, carry out the classification and Detection of this window;

(112) change the size of window, and the position of moving window successively, carry out the classification and Detection of this window.