CN103177248B - A kind of rapid pedestrian detection method of view-based access control model - Google Patents

Abstract

The invention discloses a kind of rapid pedestrian detection method of view-based access control model, first the method obtains the video image in vehicle forward march by the camera be arranged on vehicle, then class Lis Hartel is adopted to levy as pedestrian's Expressive Features, build multiple dimensioned cascade classifier as pedestrian detector, series connection concatenation tactic is adopted to realize the Classification and Identification of pedestrian and non-pedestrian in real time fast, finally determine and the moving window that pedestrian's feature is mated most with non-maxima suppression algorithm, determine the position of pedestrian.If the moving window do not matched with pedestrian's feature after above-mentioned steps judges, then judge in the image inputted without pedestrian.Pedestrian detection technology is pushed to practical by the inventive method, enables pedestrian detection technology be applicable in actual engineer applied, as all having huge applicable prospect in security protection video monitoring field, automobile Initiative Defense security fields.

Description

A Fast Pedestrian Detection Method Based on Vision

technical field

The invention relates to computer vision technology and the field of advanced assisted driving of automobiles, in particular to a vision-based rapid pedestrian detection method.

technical background

With the rapid increase in the number of cars in the last decade, road traffic safety has become an important issue worldwide. A World Health Organization report shows that traffic accidents are one of the main causes of casualties, and traffic accidents cause casualties worldwide every year There were about 10 million people, of which 2-3 million were serious casualties. Vulnerable road users (such as pedestrians, cyclists and other small vehicle occupants) account for the vast majority of victims in traffic accidents. According to statistics on traffic accidents reported in the United States in 2003, among the 35,000 road traffic casualties in the United States, 5,000 cases involved collisions between pedestrians and vehicles; in the European Union, 150,000 people were injured and 7,000 died due to collisions between vehicles and pedestrians . Therefore, in the face of frequent and increasingly serious road traffic accidents, domestic and foreign research institutions have proposed Advanced Driver Assistance Systems (Advanced Driver Assistance Systems, ADAS) from the perspective of vehicle autonomous defense to avoid traffic accidents or reduce the severity of traffic accidents. , Improve road traffic safety. And one of the important components is the pedestrian detection system (Pedestrian Detection Systems, PDS), which is an active defense system for automobiles from the perspective of protecting road pedestrians.

The pedestrian detection system refers to obtaining the road information of the vehicle's forward direction through the sensors installed on the vehicle (optical camera, infrared camera, radar, etc.), and then using a certain intelligent detection algorithm to judge the pedestrians appearing in the driving environment of the vehicle, and to judge the distance between pedestrians and vehicles. The spatial relationship of the vehicle can be used to warn the driver of possible dangerous situations or perform automatic braking on the vehicle. The vehicle-mounted pedestrian detection system based on vision uses an optical camera as the main sensor. On the one hand, it can assist in expanding the driver's field of vision, reduce the driver's visual blind spot caused by the vehicle structure, and can give early warning of pedestrians appearing in the blind spot to avoid vehicles. A collision with a pedestrian or vehicle emerging from the blind spot. Especially for large engineering vehicles with large visual blind spots, the vision-based pedestrian detection system has very important engineering significance; on the other hand, the vision-based on-board pedestrian detection system can assist inexperienced drivers to judge the distance relationship between vehicles and pedestrians , improve the driver's driving safety and reduce the occurrence of road traffic accidents. .

At present, the vision-based pedestrian detection system generally has poor adaptability to the road, and the detection speed is relatively slow, and the processing speed is generally lower than 1 second per frame (frame per second, fps). At the same time, the accuracy of pedestrian detection is generally not high. Therefore, aiming at the problems of low calculation rate and poor road adaptability of the vision-based pedestrian detection system, the present invention proposes a fast pedestrian detection method based on vision scheme, which can realize real-time pedestrian detection rate and ensure certain pedestrian detection At the same time, this method has a strong adaptability to road and pedestrian diversity, and has the prospect of engineering application.

Contents of the invention

The purpose of the present invention is to overcome the shortcomings of the existing pedestrian detection technology based on vision, and provide a fast pedestrian detection method based on vision.

The purpose of the present invention is achieved by the following technical solutions: a visual-based fast pedestrian detection method, the method includes the following:

(1) Obtain video images of the vehicle on the road ahead through the camera installed on the vehicle;

(2) Process the video image obtained in step 1 frame by frame: calculate the color invariant parameter feature channel image, HOG feature channel image, and gradient amplitude feature channel image for the input image respectively, and obtain , , and feature channel image, where, , , is the color-invariant parameter feature channel image, is the HOG feature channel image, gradient magnitude feature channel image;

(3) Calculate respectively in step 2 , , and The integer image representation method corresponding to the feature channel image, and the integer feature channel image corresponding to each feature channel image is obtained: , , , ;

(4) Use sliding windows of different scales to traverse the integer feature channel images obtained in step 3, and calculate the Haar-like features in each sliding window as pedestrian description features;

(5) Use the pedestrian detector to detect the pedestrian description features calculated in step 4, and judge whether the input features are related to pedestrians;

(6) Adopt a series cascading strategy to improve the rate and efficiency of the pedestrian detector detecting input features in step 5;

(7) Use the non-maximum value suppression algorithm to determine the sliding window that best matches the characteristics of the pedestrian, and determine the position of the pedestrian; if there is no sliding window that matches the characteristics of the pedestrian after the above steps, it is judged that there is no pedestrian in the input image.

The beneficial effect of the present invention is that, on the premise of ensuring the accuracy of pedestrian detection, the present invention improves the detection rate of pedestrians in the road by the pedestrian detection method, and makes the detection rate reach the level of real-time detection. At the same time, the fast pedestrian detection method has strong adaptability to road and pedestrian diversity. The above-mentioned technical improvement of the pedestrian detection method further pushes the pedestrian detection method to practicality, which makes the pedestrian detection method have engineering application value.

Description of drawings

Fig. 1 is a schematic diagram of image acquisition;

Figure 2 is a schematic diagram of a gradient histogram based on a 4*4 neighborhood;

Fig. 3 is a schematic diagram of the Laplacian operator;

Fig. 4 is a schematic diagram of Haar-like features.

detailed description

The purpose and effects of the present invention will become more apparent by describing the present invention in detail below in conjunction with the accompanying drawings. The background subtraction method based on the color invariant parameter of the present invention comprises the following steps:

Step 1: Obtain video images on the road ahead of the vehicle through the camera installed on the vehicle.

The specific method of image acquisition by the method of the present invention is shown in Figure 1, the camera adopted is PAL system, and the resolution of each frame of image is 352*288.

Step 2: Calculate the color-invariant parameter feature channel image, HOG feature channel image, and gradient magnitude feature channel image for the input image, and obtain , , , feature channel image, where , , is the color-invariant parameter feature channel image, is the HOG feature channel image, Gradient magnitude feature channel image.

The feature channel image refers to the image obtained by performing feature calculation on the input image, and the color invariant parameter feature channel image, HOG feature channel image, and gradient amplitude feature channel image refer to the input image to calculate the color invariant parameter feature, HOG feature , The feature channel image obtained by the gradient magnitude feature. The calculation process of the color invariant parameter feature, HOG feature, and gradient amplitude feature channel image is as follows:

2.1. Color invariant parameter feature channel image

The color invariant parameter is a characteristic parameter calculated by combining the spectral information and spatial structure information of the color in the image. This parameter has the characteristics of translation invariance, scale invariance and color invariance in the local neighborhood of the image. It has strong color discrimination ability and good adaptability to light changes. Calculating the color invariant parameters first requires physical modeling of the image according to the following formula:

;(1)

in, is the physical model of the image, represents the position in the image, is the wavelength of light, represents the spectrum of light, expressed in The Fresnel reflection of the position, Indicates the emissivity of the substance;

In the above physical model, the characteristic parameters H, , It has the property of color invariance, which are defined as follows:

(2)

(3)

(4)

in, for right The first-order partial derivative of for right The second partial derivative of , is the first-order partial derivative in the x direction of formula (1), is the first-order partial derivative of formula (1) in the y direction.

According to the formulas (1), (2), (3), and (4), the color invariant parameter feature channel image is calculated for the input image, and the color invariant parameter is obtained , , Corresponding respectively , , feature channel image.

2.2) HOG feature channel image

Calculate the gradient image of the input image, and then calculate the gradient histogram distribution of each pixel in the 8*8 neighborhood centered on each pixel in turn. The statistical rules of the histogram are as follows: the gradient magnitude of each pixel in the 8*8 neighborhood is the weight of the pixel, and the histogram is divided into 6 intervals based on the direction of the gradient (0-180°). Each pixel falls into the corresponding interval according to the direction of its own gradient, and then the corresponding gradient magnitudes of the pixels existing in each interval are added, and finally the gradient histogram is obtained, as shown in Figure 2.

2.3) Gradient magnitude feature channel image

The second order differential operator - Laplacian operator is used to calculate the gradient magnitude of the image. The definition of the second order partial differential of the image is as follows:

(5)

(6)

in represents the input image, , Indicates the position of the pixel in the image. Then the gradient of the 2D image get as follows:

which is,

(8)

Then the gradient magnitude of the image is .

In the actual calculation, the Laplacian operator shown in Figure 3 is used to filter each pixel of the image and then take the film to obtain the gradient magnitude feature channel image of the image.

Step 3: Calculate the integer image representation methods corresponding to the color invariant parameter feature channel image, HOG feature channel image and gradient magnitude feature channel image in step 2 respectively, and obtain the integer feature channel images corresponding to each feature channel image.

The calculation method of integer image representation is as follows:

(9)

In the formula, is the plastic representation of the image, is the pixel value in the original feature channel image, Indicates the position of a pixel in the image. For the feature channel images obtained in step 3), the reshaping images are calculated sequentially, and these reshaping images are denoted as , , , .

Step 4: Use sliding windows of different scales to traverse the integer feature channels obtained in step 3, and calculate the Haar-like features in each sliding window as pedestrian description features.

The method of the present invention detects pedestrians of different sizes in the input image by using sliding windows of different scales. In the embodiment, a sliding window with a size of 100*160 can be used as a standard scale window, and then The scale step size scales the sliding window.

Using a 4*4 size matrix in the sliding window, three Haar-like features are mainly calculated, which are the Haar-like features based on 2 adjacent rectangles, the Haar-like features based on 3 adjacent rectangles, and the Haar-like features based on 4 adjacent rectangles. Haar-like features of adjacent rectangles. As shown in Figure 4, the Haar-like features shown in A and B are based on two adjacent rectangles, and this feature is the difference between the sum of the median values of the two adjacent rectangles, namely:

(10)

in, sum of pixel values in a rectangle, Represents the sum of pixel values in another rectangle. Similar Haar-like features based on 3 adjacent rectangles and Haar-like features based on 4 adjacent rectangles are expressed as follows:

(11)

(12)

Step 5: Use the pedestrian detector to detect the pedestrian description features calculated in step 4, and judge whether the input features are related to pedestrians.

The pedestrian detector involved in this step refers to a pre-trained pedestrian detection classifier. The specific steps to train the pedestrian detection classifier are as follows:

5.1) The INRIA pedestrian image database is used as the image set for computing the training classifier sample data.

5.2) Calculate the pedestrian description feature set of images in the INRIA pedestrian image database according to steps 2-4, and the feature set is expressed in the form of a set as follows:

(13)

(14)

(15)

in, Represents the training samples of the classifier, Represents the set of pedestrian features, Represents the set of non-pedestrian features, Represents an element in the pedestrian feature set, Indicates the feature value corresponding to the element, 1 indicates that the classification attribute of the element is a pedestrian feature, Represents an element in the set of non-pedestrian features, Indicates the feature value corresponding to the element, and -1 indicates that the classification attribute of the element is a non-pedestrian feature.

5.3) A set of classifiers with a cascade structure composed of a two-level decision tree is used, and the cascade classifier is expressed as:

(16)

in, represents the learned classifier, Represent the weak classifier that constitutes the classifier, i.e. 2-level decision tree, i=1,...K, represent the subscript of the decision tree, K represent the number of decision trees in the classifier, get K=12 in the method of the present invention, express the corresponding weight.

5.4) Use the training sample data calculated in step 5.2) to train the classifier defined in step 5.3). Adaboost algorithm is used to train the classifier, and the parameters of each decision tree and their weights are determined.

5.5) Use the method of step 5.4) to train 5 standard scale classifiers. The classifier scale depends on the size of the sliding window corresponding to the sample data used to train the classifier. The method of the present invention uses windows of 25*15, 50*30, 100*60, 200*120, and 250*1505 sizes as sliding windows of standard sizes, and uses the pedestrian description features generated by traversing images of these five sizes as training sample data, 5 standard scale classifiers are obtained.

5.6) Based on the five standard scale classifiers trained in step 5.5), a set of classifiers with complete scales is constructed using the method of scale estimation. The build method is as follows:

(17)

(18)

, (19)

in , is the parameter of the standard scale classifier, , is the classifier parameter of the scale to be estimated, for eigenvalues at scale 1 and scale ratio. is the scale value, , , , The upsampling and downsampling parameters, respectively, require extensive experiments to determine their values. In the method of the present invention, HOG and gradient amplitude features are adopted , , , ; for color invariant parameter features use , , , .

5.7) Based on the 5 standard scale classifiers trained in step 5.5), use the method in step 5.6) to construct a set of 50 scale-complete classifiers, namely pedestrian detectors.

Step 6: Use the CrosstalkCascade (serial cascade) strategy to improve the speed and efficiency of the pedestrian detector in step 5) to detect input features.

The CrosstalkCascade strategy is the key to complete fast pedestrian detection. The specific execution steps are as follows:

6.1) According to the softcascade (loose cascade) rule, the pedestrian detection classifier is used to filter the pedestrian description feature parameters in the sliding window to screen potential feature parameters that belong to pedestrian features. The softcascade rules are:

(20)

(twenty one)

in, Pedestrian description feature parameters for classifying input pedestrian detections, Denotes the number of decision trees that make up the pedestrian detector, =1,... , =1,... , , Both represent the subscript of the decision tree, Indicates the first a decision tree, for the corresponding weight, Indicates the sum of the output values of the 1st to i-th decision trees, is the judgment threshold. If the formula (20) is established, the judgment process ends, and the judgment is a non-pedestrian feature, i.e. judging Pedestrians are not included in the sliding window.

6.2) If the pedestrian describes the characteristics After step 6.1), it is judged as a potential pedestrian description feature, and the sliding window is used as the unit, and the feature The sliding window where it is located is the center, and the pedestrian description feature parameters in 7*7*3 sliding windows are selected to input to the pedestrian detector. Among them, 7*7*3 corresponds to w*h*d, w represents the number of sliding windows in the horizontal direction, h represents the number of sliding windows in the vertical direction, and d represents the d adjacent scales of a certain position in the image sliding window. Denote the features in 7*7*3 sliding windows as:

(twenty two)

6.3) Use the excitationcascade (incentive cascade) rule to get in step 6.2) The pedestrian description feature parameters in the filter are screened. The excitationcascade rules are as follows:

(twenty three)

(twenty four)

in, Denotes the pedestrian description features obtained in step 6.2), is the decision threshold, Denotes the number of decision trees that make up the pedestrian detector, =1,... , =1,... , , Both represent the subscript of the decision tree, Indicates the first a decision tree, for the corresponding weight, Indicates the sum of the output values of the 1st to i-th decision trees, is the judgment threshold. When the formula (23) is established, the judgment is a non-pedestrian feature, i.e. judging Pedestrians are not included in the sliding window.

6.4) Use the inhibitory cascade (cut-off cascade) rule to filter the feature parameter sets obtained in steps 6.3) and 6.1). The inhibitory cascade rules are as follows:

(25)

in , Defined by equations (21)(24), is the judgment threshold. When the formula (25) is established, the characteristic parameter is judged is a non-pedestrian feature.

6.5) Determine the pedestrian description feature parameters that have not been screened out after the above steps are determined as pedestrian features, that is, feature The corresponding windows contain pedestrians.

Step 7: Use the non-maximum value suppression algorithm to determine the sliding window that best matches the characteristics of the pedestrian, and determine the position of the pedestrian. If there is no sliding window matching the characteristics of pedestrians after the above steps are judged, it is judged that there is no pedestrian in the input image.

After the detection of the pedestrian detector, there may be multiple windows that match the pedestrian profile. Therefore, it is necessary to select the best matching window from these windows. The non-maximum suppression algorithm can effectively and quickly select the window.

The method of the invention aims at the problem of time-consuming calculation in the traditional pedestrian detection method, and proposes a fast pedestrian detection method based on vision. This method not only improves the detection rate, but also ensures the accuracy of pedestrian detection. The method of the present invention mainly improves the execution speed of the pedestrian detection method from the following three aspects: 1) By defining effective and easy-to-calculate pedestrian description feature channels: color invariant parameter feature channel, gradient histogram feature channel (Histogram of Gradient, HOG), gradient Amplitude feature channel; 2) By building a multi-scale classifier, the time-consuming link in the detection process is completed in the training stage of the pedestrian detector in advance, and this method also improves the accuracy of detection; 3) Using the CrosstalkCascade strategy to make training A good classifier can quickly and efficiently detect pedestrian features when implementing real-time detection. The method of the invention promotes the practicality of the pedestrian detection technology, so that the pedestrian detection technology can be applied to practical engineering applications, such as in the field of security video monitoring and the field of automobile active defense safety, which has great application prospects.

Claims (5)

1. A vision-based fast pedestrian detection method is characterized in that the method comprises the following: (1) Obtain video images on the road ahead of the vehicle through a camera installed on the vehicle; (2) Process the video image obtained in step (1) frame by frame: calculate the color invariant parameter feature channel image, HOG feature channel image, and gradient amplitude feature channel image for the input image respectively, and obtain _SH , S _x , S _y , S _Histo and S _M feature channel images, where _SH , S _x , S _y are color-invariant parameter feature channel images, S _Histo is HOG feature channel images, and S _M is gradient magnitude feature channel images; (3) Calculate the integer image representation methods corresponding to the S _H , S _x , S _y , S _Histo and S _M feature channel images in step (2) respectively, and obtain the integer feature channel images corresponding to each feature channel image: IS _H , IS _x , IS _y , IS _Histo , IS _M ; (4) Use sliding windows of different scales to traverse the integer feature channel images obtained in step (3), and calculate the Haar-like features in each sliding window as pedestrian description features; (5) Use the pedestrian description feature calculated in the pedestrian detector detection step (4) to judge whether the input feature is a feature related to the pedestrian; (6) Adopting a series cascading strategy to improve the speed and efficiency of pedestrian detector detection input features in step (5); (7) Determine the sliding window most matching with the pedestrian feature with the non-maximum value suppression algorithm, and determine the position of the pedestrian; if there is no sliding window matching the pedestrian feature after the above-mentioned steps are judged, then judge that there is no pedestrian in the input image; In the step (2), the HOG feature channel is defined as follows: calculate its gradient image for the input image, and then calculate each pixel in the 8*8 neighborhood centered on each pixel in turn in the 8*8 neighborhood centered on the pixel The gradient histogram distribution of the gradient; the statistical rules of the histogram are as follows: the gradient amplitude of each pixel in the 8*8 neighborhood is the weight of the pixel, and the histogram is divided into 6 intervals based on the direction of the gradient (0-180°). Each pixel falls into the corresponding interval according to the direction of its own gradient, and then adds the corresponding gradient amplitudes of the pixels existing in each interval to finally obtain the gradient histogram; In the step (4), the Haar-like feature is defined as follows: a matrix of 4*4 size is used in the sliding window, and 3 Haar-like features are calculated, which are respectively based on 2 adjacent rectangle-like Haar-like features, based on 3 The Haar-like feature of two adjacent rectangles and the Haar-like feature based on 4 adjacent rectangles; the Haar-like feature based on 2 adjacent rectangles, which is the difference between the sum of the median values of two adjacent rectangles, which is: H ₂ ＝|∑rec1 _i -∑rec2 _j |(10) Among them, ∑rec1 _i represents the sum of pixel values in one rectangle, and ∑rec2 _j represents the sum of pixel values in another rectangle; Haar-like features based on 3 adjacent rectangles and Haar-like features based on 4 adjacent rectangles They are expressed as follows: H ₃ ＝|∑rec1 _i +∑rec3 _j -∑rec2 _j |(11) H ₄ ＝|∑rec1 _i -∑rec2 _j +∑rec3 _i -∑rec4 _j |(12); In the step (6), the series cascading strategy is defined as follows: (6.1) According to the loose cascade rule, the pedestrian detection classifier is used to filter and screen the potential characteristic parameters belonging to the pedestrian feature to the pedestrian description feature parameters in the sliding window; the loose cascade rule is: ${\mathrm{<span\; class="notranslate">\; h</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; <</span>{\mathrm{<span\; class="notranslate">\; \&theta;</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}^{\mathrm{<span\; class="notranslate">\; R</span>}}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; <</span>\mathrm{<span\; class="notranslate">\; K</span>}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 20</span>}<span\; class="notranslate">\; )</span>$ ${\mathrm{<span\; class="notranslate">\; h</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>{<span\; class="notranslate">\; \&Sigma;</span>}_{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}^{\mathrm{<span\; class="notranslate">\; k</span>}}{\mathrm{<span\; class="notranslate">\; a</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}{\mathrm{<span\; class="notranslate">\; h</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; twenty\; one</span>}<span\; class="notranslate">\; )</span>$ Among them, x is the pedestrian description feature parameter for input pedestrian detection classification, K represents the number of decision trees constituting the pedestrian detector, i=1,...k, k=1,...K, i, k all represent The subscript of the decision tree, h _i (x) represents the i-th decision tree, a _i is the weight corresponding to h _i (x), H _k (x) represents the sum of the output values of the 1st to i-th decision trees, is the threshold of judgment; if the formula (20) is established, the judgment process ends, and it is judged that x is a non-pedestrian feature, that is, the sliding window where x is judged does not contain pedestrians; (6.2) If the pedestrian description feature x is judged to be a potential pedestrian description feature after step (6.1), then take the sliding window as the unit and take the sliding window where the feature x is located as the center, select 7*7*3 pedestrians in the sliding window The description feature parameters are input into the pedestrian detector; where 7*7*3 corresponds to w*h*d, w represents the number of sliding windows in the horizontal direction, h represents the number of sliding windows in the vertical direction, and d represents a certain The sliding window of d adjacent scales for the position; the features in the 7*7*3 sliding windows are recorded as: N(x)＝{ _xi | _xi ∈N(x)}(22) (6.3) Use the incentive cascading rules to filter the pedestrian description feature parameters in N(x) obtained in step (6.2); the incentive cascading rules are as follows: Among them, x' represents the pedestrian description feature obtained in step (6.2), is the decision threshold, K represents the number of decision trees that constitute the pedestrian detector, i=1,...k, k=1,...K, i, k both represent the subscripts of the decision tree, h _i (x) Represents the i-th decision tree, a _i is the weight corresponding to h _i (x), H _k (x) represents the sum of the output values of the 1st to i-th decision trees, is the judgment threshold; when the formula (23) is established, it is judged that x is a non-pedestrian feature, that is, the sliding window where x is judged does not contain pedestrians; (6.4) adopt cut-off cascade rule to filter the feature parameter set obtained in step (6.3), step (6.1); cut-off cascade rule is as follows: $\frac{{\mathrm{<span\; class="notranslate">\; h</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; )</span>}{{\mathrm{<span\; class="notranslate">\; h</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; x</span>}}^{<span\; class="notranslate">\; \&prime;</span>}<span\; class="notranslate">\; )</span>}<span\; class="notranslate">\; <</span>{\mathrm{<span\; class="notranslate">\; \&theta;</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}^{\mathrm{<span\; class="notranslate">\; I</span>}}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; <</span>\mathrm{<span\; class="notranslate">\; K</span>}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 25</span>}<span\; class="notranslate">\; )</span>$ Among them, H _k (x), H _k (x') are defined by formula (21) (24), is the threshold of judgment; when the formula (25) is established, then the judgment feature parameter x is a non-pedestrian feature; (6.5) Determine the pedestrian description feature parameters that have not been screened out after the above steps are determined as pedestrian features, that is, the window corresponding to feature x contains pedestrians. 2. the vision-based fast pedestrian detection method according to claim 1, is characterized in that, in described step (2), color invariant parameter feature channel is defined as follows: The color invariant parameter is a characteristic parameter calculated by combining the spectral information and spatial structure information of the color in the image; this parameter has translation invariance, scale invariance and color invariance in the local neighborhood of the image, and has a strong Color discrimination ability, good adaptability to light changes; calculation of color invariant parameters first requires physical modeling of the image according to the following formula, in, is the physical model of the image, Indicates the position in the image, λ is the wavelength of the light, represents the spectrum of light, expressed in The Fresnel reflection of the position, Indicates the emissivity of the substance; In the above physical model, the characteristic parameters H, W _x , W _y have the property of color invariance, which are defined as follows: $\mathrm{<span\; class="notranslate">\; h</span>}<span\; class="notranslate">\; =</span>\frac{{\mathrm{<span\; class="notranslate">\; E.</span>}}_{\mathrm{<span\; class="notranslate">\; \&lambda;</span>}}}{{\mathrm{<span\; class="notranslate">\; E.</span>}}_{\mathrm{<span\; class="notranslate">\; \&lambda;</span>}\mathrm{<span\; class="notranslate">\; \&lambda;</span>}}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 2</span>}<span\; class="notranslate">\; )</span>$ ${\mathrm{<span\; class="notranslate">\; W</span>}}_{\mathrm{<span\; class="notranslate">\; x</span>}}<span\; class="notranslate">\; =</span>\frac{{\mathrm{<span\; class="notranslate">\; E.</span>}}_{\mathrm{<span\; class="notranslate">\; x</span>}}}{\mathrm{<span\; class="notranslate">\; E.</span>}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 3</span>}<span\; class="notranslate">\; )</span>$ ${\mathrm{<span\; class="notranslate">\; W</span>}}_{\mathrm{<span\; class="notranslate">\; the\; y</span>}}<span\; class="notranslate">\; =</span>\frac{{\mathrm{<span\; class="notranslate">\; E.</span>}}_{\mathrm{<span\; class="notranslate">\; the\; y</span>}}}{\mathrm{<span\; class="notranslate">\; E.</span>}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 4</span>}<span\; class="notranslate">\; )</span>$ where E _λ is The first-order partial derivative of λ, E _λλ is To the second-order partial derivative of λ, E _x is the first-order partial derivative to the x direction of formula (1), and E _y is the first-order partial derivative to the formula (1) y direction; According to the formulas (1), (2), (3), and (4), calculate the color invariant parameter feature channel image for the input image, and obtain the SH and S _x corresponding to the color invariant parameters _H , W _x , and W _y respectively , S _y feature channel image. 3. the vision-based fast pedestrian detection method according to claim 1, is characterized in that, in described step (2), the gradient magnitude feature channel is defined as follows: The second-order differential operator - Laplacian operator is used to calculate the gradient magnitude of the image; the definition of the second-order partial differential of the image is as follows: $\frac{{<span\; class="notranslate">\; \&part;</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}\mathrm{<span\; class="notranslate">\; f</span>}}{<span\; class="notranslate">\; \&part;</span>{\mathrm{<span\; class="notranslate">\; x</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 2</span>}\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 5</span>}<span\; class="notranslate">\; )</span>$ $\frac{{<span\; class="notranslate">\; \&part;</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}\mathrm{<span\; class="notranslate">\; f</span>}}{<span\; class="notranslate">\; \&part;</span>{\mathrm{<span\; class="notranslate">\; the\; y</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 2</span>}\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 6</span>}<span\; class="notranslate">\; )</span>$ Where f represents the input image, x, y represent the position in the image; then the gradient of the 2D image get as follows: ${<span\; class="notranslate">\; \&dtri;</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; =</span>\frac{{<span\; class="notranslate">\; \&part;</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}\mathrm{<span\; class="notranslate">\; f</span>}}{<span\; class="notranslate">\; \&part;</span>{\mathrm{<span\; class="notranslate">\; x</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}<span\; class="notranslate">\; +</span>\frac{{<span\; class="notranslate">\; \&part;</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}\mathrm{<span\; class="notranslate">\; f</span>}}{<span\; class="notranslate">\; \&part;</span>{\mathrm{<span\; class="notranslate">\; the\; y</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 7</span>}<span\; class="notranslate">\; )</span>$ which is, ${<span\; class="notranslate">\; \&dtri;</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; =</span><span\; class="notranslate">\; \&lsqb;</span>\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; \&rsqb;</span><span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 4</span>}\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 8</span>}<span\; class="notranslate">\; )</span>$ Then the gradient magnitude of the image is 4. the vision-based fast pedestrian detection method according to claim 1, is characterized in that, in the described step (3), the integer image representation calculation method is as follows: ii(x, y) = ∑ _{x'≤x, y'≤y} i(x', y') (9) In the formula, ii(x, y) is the plastic representation of the image, i(x', y') is the pixel value in the original feature channel image, x, y represents the position of the pixel in the image; for the obtained in step (3) The characteristic channel images of are calculated in turn for their reshaping images, and these reshaping images are denoted as IS _H , IS _x , IS _y , IS _Histo , and _ISM . 5. the vision-based fast pedestrian detection method according to claim 1, is characterized in that, in described step (5), the construction and training method of pedestrian detector are as follows: (5.1) Use the INRIA pedestrian image database as the image set for calculating the training classifier sample data; (5.2) According to step (2)-step (4), calculate the pedestrian description feature collection of image in INRIA pedestrian image database, represent feature collection with the form of collection below: S=P∪N(13) P = {p _i ∈ P: p _i = (v _P , 1)} (14) N = {n _i ∈ N: n _i = (v _N , -1)} (15) Among them, S represents the training sample of the classifier, P represents the pedestrian feature set, N represents the non-pedestrian feature set, p _i represents an element in the pedestrian feature set, v _P represents the feature value corresponding to the element, and 1 represents the classification of the element The attribute is a pedestrian feature, n _i represents an element in the non-pedestrian feature set, v _N represents the feature value corresponding to the element, and -1 indicates that the classification attribute of the element is a non-pedestrian feature; (5.3) A group of classifiers with a cascade (cascade) structure composed of a 2-level decision tree is adopted, and the cascade classifier is expressed as: $\mathrm{<span\; class="notranslate">\; h</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; h</span>}}_{\mathrm{<span\; class="notranslate">\; K</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>{<span\; class="notranslate">\; \&Sigma;</span>}_{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}^{\mathrm{<span\; class="notranslate">\; K</span>}}{\mathrm{<span\; class="notranslate">\; a</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}{\mathrm{<span\; class="notranslate">\; h</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 16</span>}<span\; class="notranslate">\; )</span>$ Among them, H(x) represents the learned classifier, h _i (x) represents the weak classifier constituting the classifier, that is, a 2-level decision tree, i=1,...K, represents the subscript of the decision tree, K Represent the number of decision trees in the classifier, take K=12, a _i represents the weight corresponding to h _i (x); (5.4) use the classifier defined in the training sample data training step (5.3) defined in the step (5.2); Adopt the Adaboost algorithm training classifier to determine the parameters of each decision tree and its corresponding weight; (5.5) Train 5 standard scale classifiers with the method of step (5.4), the classifier scale depends on the size of the sliding window corresponding to the sample data used to train the classifier; the method is 25*15, 50*30, 100* 60, 200*120, 250*1505 size windows are used as standard size sliding windows, and the pedestrian description features generated by traversing images of these five sizes are used as training sample data to obtain 5 standard scale classifiers; (5.6) Based on the five standard scale classifiers trained in step (5.5), use the method of scale estimation to construct a set of classifiers with complete scales; the construction method is as follows: v _rec '=v _rec *r(s)(17) τ'=τ*r(s)(18) $\mathrm{<span\; class="notranslate">\; r</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; the\; s</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\left\{\begin{array}{cc}{\mathrm{<span\; class="notranslate">\; a</span>}}_{\mathrm{<span\; class="notranslate">\; u</span>}}<span\; class="notranslate">\; *</span>{\mathrm{<span\; class="notranslate">\; the\; s</span>}}^{{\mathrm{<span\; class="notranslate">\; b</span>}}_{\mathrm{<span\; class="notranslate">\; u</span>}}}& \mathrm{<span\; class="notranslate">\; i</span>}\mathrm{<span\; class="notranslate">\; f</span>}\mathrm{<span\; class="notranslate">\; the\; s</span>}<span\; class="notranslate">\; ></span>\mathrm{<span\; class="notranslate">\; 1</span>}\\ {\mathrm{<span\; class="notranslate">\; a</span>}}_{\mathrm{<span\; class="notranslate">\; d</span>}}<span\; class="notranslate">\; *</span>{\mathrm{<span\; class="notranslate">\; the\; s</span>}}^{{\mathrm{<span\; class="notranslate">\; b</span>}}_{\mathrm{<span\; class="notranslate">\; d</span>}}}& \mathrm{<span\; class="notranslate">\; o</span>}\mathrm{<span\; class="notranslate">\; t</span>}\mathrm{<span\; class="notranslate">\; h</span>}\mathrm{<span\; class="notranslate">\; e</span>}\mathrm{<span\; class="notranslate">\; r</span>}\mathrm{<span\; class="notranslate">\; w</span>}\mathrm{<span\; class="notranslate">\; i</span>}\mathrm{<span\; class="notranslate">\; the\; s</span>}\mathrm{<span\; class="notranslate">\; e</span>}\end{array}\right.<span\; class="notranslate">\; ,</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 19</span>}<span\; class="notranslate">\; )</span>$ Among them, v _rec and τ are the parameters of the standard scale classifier, v _rec ', τ' are the classifier parameters of the scale to be estimated, r(s) is the ratio of the feature value at scale 1 to scale s; s is the scale value, a _u , b _u , a _d , b _d are the up-sampling and down-sampling parameters, and their values need to be determined through a lot of experiments; for HOG and gradient amplitude features, a _u = 1, b _u = 0, a _d = 0.89, b _d = 1.586; a _u = 1, b _u = 2, a _d = 1, b _d = 2 are used for color invariant parameter features; (5.7) Based on the 5 standard scale classifiers trained in step (5.5), use the method in step (5.6) to construct a set of 50 scale-complete classifiers, namely pedestrian detectors.