KR20160081190A - Method and recording medium for pedestrian recognition using camera - Google Patents

Abstract

The present invention relates to a method for recognizing a pedestrian by using a camera and a recording medium thereof, capable of determining whether or not an object positioned in front of a vehicle is a person or other obstacle by using the camera. The method for recognizing the pedestrian by using the camera comprises: a first step for collecting a model image; a second step for extracting an HOG feature for each model image by using HOG; a third step for performing model learning through a machine learning algorithm and generating a pedestrian model based on the HOG feature of each model image extracted through the second step; a fourth step for receiving an input image; a fifth step for performing scoring for calculating a score value in proportion to a similarity between the HOG feature of the pedestrian model and the HOG feature extracted in the input image; a sixth step for comparing the score value calculated through the fifth step with a predetermined threshold; and a seventh step for recognizing the input image as a pedestrian image when the score value is higher than the threshold based on a comparison result of the sixth step.

Description

BACKGROUND OF THE INVENTION 1. Field of the Invention [0001] The present invention relates to a pedestrian recognition method using a camera,

The present invention relates to a pedestrian recognition method and a recording medium thereof, and more particularly, to a pedestrian recognition method and a pedestrian recognition method using a camera that reads an image photographed through a camera installed in a vehicle and recognizes when a pedestrian appears in the vicinity of the vehicle And relates to the recording medium.

A smart car is a vehicle that can be operated automatically by using advanced computer, communication and measurement technology. It is informed of accurate latitude and longitude by a GPS (Global Positioning System) receiver installed in a car. And refers to the most efficient and autonomous moving vehicle from its current location to its destination.

Smart automotive technology can be largely classified into three categories: automobile safety technology, automobile comfort technology, and autonomous driving. Automobile safety technology is a technology to dramatically reduce traffic accidents and damage by protecting drivers and passengers from dangerous situations such as vehicle flaws, accident prevention and avoidance, collision, and automobile comfort technology is an interest in automobiles and maximizes driver convenience It is a technology that utilizes automobiles as the third residential space after home and office. Finally, autonomous driving refers to the technology of collecting information about traffic signals, lanes, obstacles, etc. by sensors and cameras installed in the vehicle, and collecting the information by the computer in the vehicle and operating the vehicle according to the surrounding situation.

The active safety technology of the smart car is the sensor technology to judge the external situation. Approximately 160 sensors are used in automobiles, and the usage is continuously increasing. There are RADAR, LIDAR (Light Detection And Ranging), and camera as main sensors used in smart cars except sensors used in existing vehicles.

RADAR and LIDAR sensors are used to find distance information between object and vehicle. Since the purpose of vehicle safety technology is to minimize damage to human life, two sensors that return only simple distance information are used to detect objects There is a limit that can not distinguish between obstacles. Therefore, it is necessary to recognize pedestrians using a camera in order for an autonomous vehicle to operate with the safety of the pedestrian as a top priority.

In the case of the pedestrian recognition method using the conventional camera, the pedestrian image (model image) is simply determined to determine whether it is similar to the input image, and then it is configured to distinguish whether it is a pedestrian or not.

However, since the characteristic of the pedestrian can be changed by clothes, hair style, and the like, if the pedestrian is detected only by the features of the single model image, there is a possibility that an error may be detected that detects only a specific pedestrian.

SUMMARY OF THE INVENTION The present invention has been made in order to solve the above problems, and it is an object of the present invention to provide a pedestrian detection apparatus and a pedestrian detection method capable of distinguishing whether an object existing in front of a vehicle is a person or other obstacle by using a camera, A pedestrian recognition method using a camera which can be recognized and a recording medium therefor.

According to another aspect of the present invention, there is provided a method for recognizing a pedestrian using a camera, comprising: a first step of collecting a plurality of images (hereinafter referred to as 'model images') through a camera; A second step of extracting HOG features for each model image using a histogram of gradient (HOG); A third step of generating a pedestrian model by performing model learning through a machine learning algorithm based on the HOG feature of each of the model images extracted through the second step; A fourth step of inputting an image (hereinafter, referred to as 'input image') requiring identification of the presence or absence of a pedestrian; A fifth step of performing scoring in which a score value is calculated in proportion to the degree of similarity between the HOG feature of the pedestrian model and the HOG feature extracted from the input image; A sixth step of comparing a score value calculated through the fifth step with a preset threshold value; And a seventh step of recognizing the input image as a pedestrian image when the score value is greater than the threshold value as a result of the comparison in the sixth step.

According to the pedestrian recognition method and the recording medium using the camera according to the present invention, it is possible to discriminate whether the object included in the image input through the camera is a pedestrian, and in particular, to generalize the characteristics of the pedestrian, So that pedestrians can be recognized accurately.

1 is a block flow diagram illustrating a flow of a learning step of a pedestrian recognition method using a camera according to the present invention.
BACKGROUND OF THE INVENTION 1. Field of the Invention [0002] The present invention relates to a method for recognizing a pedestrian using a camera.
3 is an example showing HOG feature transformation of an image collected through a camera of the present invention.
4 is an example showing various model images for constructing a support vector machine learning set according to the present invention.
5 is a graphical representation of HOG features for one of a plurality of model parameters that may be generated from the "004.jpg model image"
6 is a graphical representation of HOG features for one of a plurality of model parameters that may be generated from the "girin 3.jpg model image" of Fig.
7 is an example of a learned model image (pedestrian model) generated based on the HOG feature of the model image of the present invention.
8 is a process flow chart showing the flow of scoring according to the present invention.
9 (a) and 9 (b) illustrate an example of a score function according to the present invention.
10 is an example of a weight filter according to the present invention.

The terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The singular expressions include plural expressions unless the context clearly dictates otherwise. In the present application, the terms "comprises" or "having" and the like are used to specify that there is a feature, a number, a step, an operation, an element, a component or a combination thereof described in the specification, But do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, or combinations thereof.

Furthermore, although the terms first, second, etc. may be used herein to describe various components, the components should not be limited by the terms. The terms are used only for the purpose of distinguishing one component from another.

In the following, preferred embodiments, advantages and features of the present invention will be described in detail with reference to the accompanying drawings.

FIG. 1 is a block flow chart showing a flow of learning steps of a pedestrian recognition method using a camera according to the present invention, and FIG. 2 is a block flow chart showing a flow of a recognition step of a pedestrian recognition method using a camera according to the present invention.

Referring to FIGS. 1 and 2, the pedestrian recognition method using the camera of the present invention comprises a learning step (steps 1 and S10) and a recognition step (steps 2 and S20).

The learning step (step 1) includes a step of collecting an image through a camera (step 1-1), a step of extracting features of an image collected through a camera (step 1-2) (Step 1-3) of removing noise, and generating a pedestrian model (step 1-5) by performing model learning using a machine learning algorithm (steps 1-4) .

The recognition step (step 2) comprises the steps of discriminating the presence or absence of a pedestrian of the input image based on the pedestrian model constructed in the learning step, when an image (hereinafter, referred to as 'input image' to be.

In detail, the recognition step (step 2) includes an image input step (step 2-1), a step of performing scoring using the characteristics of the pedestrian model and features extracted from the input image (step 2-2) , And determining whether the pedestrian is present based on the calculated score value (step 2-3).

Hereinafter, a preferred embodiment of each step will be described in more detail.

≪ Step 1-1 (S11)

Step 1-1 (S11) of the present invention is a step of collecting a plurality of images through a camera to generate a pedestrian model.

The images collected in step 1-1 are divided into an image in which a pedestrian (person) is expressed and an image in which a non-pedestrian such as an object or an animal is expressed.

The image in which the pedestrian is expressed is used as a positive image in the training set construction of the model learning step S13 and the image in which the non-pedestrian is expressed is used as a negative image .

≪ Step 1-2 (S12)

Step 1-2 of the present invention is a step of extracting features of a plurality of images collected through a camera, and the feature of each image is extracted using Histogram of gradient (HOG) .

HOG (Histogram of Gradient) refers to an algorithm for dividing an image into a rectangular grid shape and calculating a local histogram of orientation in each of its square grid (Grid). In particular, A feature vector of the transformed image is extracted using a block of cells.

That is, HOG is a vector obtained by finding the histogram of the direction of the edge pixels for each cell and connecting these histograms in a line, and the histogram is accumulated for each angle. This HOG can be viewed as an orientation histogram template of an edge and is suitable for expressing features of an object shape.

The HOG feature for each pixel of the image can be composed of a HOG feature vector having an angle (direction) and a magnitude component, and the angle and size of the HOG feature can be derived by the following equation (1).

(In the equation (1)

i, j: Pixel index of the image,

f (i, j) is the intensity of the image,

m (i, j): size of feature,

θ (i, j): angle of feature)

3 is an example showing the HOG feature transformation of an image collected through the camera of the present invention. Referring to FIG. 3, the image (FIG. 3 (a)) collected in step 1-1 (S11) can extract the HOG feature according to equation (1) Can be visualized together.

≪ Step 1-3 (S13)

In step 1-3 of the present invention, only a target object (e.g., a person, an object, or an animal) is displayed at the time of HOG feature conversion of the corresponding image by removing noise included in the collected image This is the step for acquiring the model image.

The model image is an image for finding a desired object in the collected image. Since the model image includes not only the target object but also the background component, a HOG weight filter is used to remove it.

Equation (2) is a formula of a HOG weighted filter for removing such background components. The HOG weighted filter is applied only when the model image is white as shown in FIG. 10, and a value of '0' is assigned if the model image is white.

≪ Step 1-4 (S14)

Step 1-4 of the present invention is a step of generating a pedestrian model by performing model learning using a machine learning algorithm. According to a preferred embodiment, the machine learning algorithm of steps 1-4 (S14) is configured to use a support vector machine (SVM).

A support vector machine (SVM) is a kind of machine learning, and is an algorithm for classifying a given data belonging to a classification. Here, machine learning is a process of inputting given data into a computer and building a discrimination criterion by performing learning based on a certain algorithm to predict what kind of data will be discriminated when new data is given.

For example, if a feature vector of a giraffe and a hippopotamus exist, the optimal straight line (an optimal hyper plane) to distinguish the two animals is drawn in the feature space, and the distance (margin) It is the purpose of SVM.

In SVM, the straight line passing through the outermost vector and support vector of each class is expressed as "ω · x - b = 1" and "ω · x - b = -1" Is present in the region above or below the straight lines, the following condition (3) must be satisfied.

Here, the distance between the two straight lines becomes "2 / ∥ω∥" and "∥ω∥" should be minimized in order to maximize the distance (margin) between the two straight lines. In general, the extreme value for the function f (x, y) with two variables is obtained by setting the partial differential value for each variable to be '0' and the solution of two simultaneous equations. If two variables are used to find the extremum of f (x, y) under the condition of the constraint function g (x, y), the maximum or minimum value should be obtained by using a Lagrange multiplier. In the SVM, the Lagrangian multiplier is used to obtain the minimum value of "? ||

The general formula of the Lagrange function is shown in Equation 4 below, and the Lagrangian function for minimizing "? ||

(In equation (5)

α: Lagrange constant,

y _i : the input data class, the expression is

)

)

Hereinafter, the step of generating the pedestrian model by performing the model learning using the support vector machine (SVM) will be described in detail.

≪ Step 1-5 (S15)

In order to learn the model through the support vector machine, the support vector machine training set (SVM training set) must be configured first.

FIG. 4 shows an example of various model images for constructing a support vector machine learning set according to the present invention. In order to construct a support vector machine training set, a plurality of model images, that is, a positive image and a set of negative images, are required as shown in FIG.

A positive image is a model image classified as true data, a model image of a pedestrian, a human being, and a nagative image is classified as false data. The same non-pedestrian is expressed.

Step 1-5 of the present invention classifies "true data" and "false data" into HOG feature values for each angle of the collected positive image and negative image set, training set.

More specifically, the angle ('?' In Equation 1) that can be included in the HOG characteristic of the positive image and the negative image is divided into a predetermined unit range, and the HOG And extracts features to perform model learning.

In one embodiment, by dividing the angle division into 20-degree unit ranges to generate a total 18-dimensional HOG feature vector, 18 model parameters can be obtained for one positive image or negative image.

In another embodiment, 12-dimensional HOG feature vectors are generated by dividing the angle division into 30-degree unit ranges, thereby obtaining 12 model parameters for one positive image or negative image.

As described above, it is preferable that the angle division for determining the number of model parameters is set within a unit range of preferably 10 to 30 degrees.

In other words, it is preferable that the range angles composed of a plurality of angles are constituted by "range angles divided by 18 angles in total by 20 angles"

5 is an example of a graphical representation of HOG features for one of a plurality of model parameters that may be generated from the "004.jpg model image" of FIG.

The model parameters in FIG. 5 are graphs of the HOG characteristics of the model parameters corresponding to the angles in the range of 0 ° to 20 ° among the 18 model parameters generated when the unit range is set to 20 °.

For reference, "Model X" and "Model Y" in the graph of FIG. 5 indicate cell indexes on the X and Y axes of the " 004.jpg model image " O, 10O, ..., 700) means the size of the HOG feature value in the range of 0 ° to 20 ° of the "004.jpg model image".

Here, the HOG characteristic of the model parameter of FIG. 5 is for a person, and since the present invention is an algorithm for detecting a pedestrian, the model parameter of FIG. 5 becomes a positive sample for support vector machine (SVM) learning.

Meanwhile, in FIG. 5, the number of dimensions of features for support vector machine (SVM) learning is the product of the width and height of the model expressed in the image.

Fig. 6 is an example of a graphical representation of HOG features for one of a plurality of model parameters that may be generated from the "giraffe 3.jpg model image" of Fig.

Specifically, the model parameters of FIG. 6 are graphs of the HOG characteristics of the model parameters corresponding to angles ranging from 0 ° to 20 ° among the 18 model parameters generated when the unit range is set to 20 °.

The HOG feature of the model parameter of Fig. 6 is for an animal (i.e., a non-pedestrian), and since the present invention is an algorithm for detecting a pedestrian, the model parameter of Fig. 6 becomes a negative sample.

In this manner, the HOG feature of the image is extracted for each of the plurality of images (i.e., the positive image and the negative image) collected in step 1-5 (step S15) false data ", and the model learning is performed using the support vector machine (SVM) on the basis of the training set, thereby generating a learned model image (ie, a pedestrian model) as shown in FIG. 7 .

As a result, the pedestrian model according to the present invention independently distinguishes each angle as a unit range form, extracts HOG features for each range angle, and is generated based on the extracted HOG features for each angle, And is configured to have information on the HOG characteristic by range angle.

For example, in the case of the pedestrian model of Fig. 7, the HOG feature vector of 18 model parameters generated when the unit range is set to 20 [deg.] Is used, and as in the cell index set in the model image of Fig. 5, Is composed of 20 cells, and the vertical axis can be composed of 30 cells, and each cell is configured to have information on the HOG characteristic by range angle.

Each cell constituting the pedestrian model can be represented by various brightness (brightness), which means that the brightness of the cell is larger as the value of the HOG characteristic for the angle is larger.

≪ Step 2-1 (S21) >

Step 2-1 (S21) of the present invention is a step of receiving images around the vehicle through a camera installed in the vehicle.

The vehicle surroundings image taken through the camera may include pedestrians or may include non-pedestrians. According to the recognition method of the present invention, it is necessary to determine whether the object included in the image is a pedestrian or not (Hereinafter referred to as " input image ").

≪ Step 2-2 (S22) >

Step 2-2 (S22) of the present invention is a step of performing scoring using features of the pedestrian model generated through the learning step (Step 1) described above and features extracted from the input image, Scoring is used as a measure to determine how similar the HOG feature of the learned model image (ie, the pedestrian model) is to the HOG feature of the input image.

The scoring of the present invention is configured such that the pedestrian model, which serves as a criterion for performing the scoring, has HOG feature information for each of a plurality of range angles, and the scoring is also configured to generate a plurality of score values according to the number of angles to be classified .

The scoring of the present invention is characterized in that it is configured to calculate a score value for an input image based on a Gaussian function.

Hereinafter, the scoring of the present invention and the pedestrian discriminating method according to the present invention will be described in detail.

Scoring has HOG feature information of the pedestrian model and matches it with the HOG feature of the input image to calculate a high score value when the similarity is high and a low score value when the similarity is low.

Specifically, the score value is configured to move the cell from the upper left corner to the lower right corner of the input image in units of cells of a predetermined size, and to calculate a score value for each unit point of the input image. For reference, 'unit point' of the input image means each unit point corresponding to the current position of the moving cell with respect to the input image.

8 is a process flow chart showing the flow of scoring according to the present invention. 8, the scoring of the present invention places the cell 10 at the first unit point of the input image 30 as shown in FIG. 8A, and calculates the score value of the corresponding unit point 8 (b), the cell 10 is moved to the second unit point immediately adjacent to the first unit point to calculate the score value of the second unit point. Subsequently, as shown in FIG. 8 (c) And the score of the third unit point is calculated by moving the cell 10 to the third unit point immediately adjacent to the second unit point as shown in FIG. Point, ..., N-th unit point).

When the cell 10 moves with respect to the input image 30, as shown in FIGS. 8A, 8B and 8C, the model image learned through the step 1 (that is, Model 20 is also configured to move in the same manner as the cell 10 so that matching and scoring of the pedestrian model 20 can be performed for each unit point of the input image 30.

In other words, the pedestrian model 20 is matched to the input image on the basis of each unit point of the input image 30, and the score value is calculated for each unit point.

For reference, the pedestrian model image shown in FIG. 8 is displayed as an actual model image instead of the learned model image expressed in the form shown in FIG. 7 for the sake of convenience of explanation. Actual scoring, however, .

On the other hand, a score value for each unit point (first unit point, second unit point, ..., N-th unit point) of the input image 30 can be calculated by the following equation (6).

(In Equation 6,

score (x, y): score of the (x, y) cell unit of the input image,

N: Number of range angles of the pedestrian model,

H: Height of pedestrian model

W: Width of the pedestrian model,

M: HOG feature value of pedestrian model,

I: HOG characteristic value of input image

c: constant)

As can be seen from Equation (6), the scoring of the present invention is configured to calculate the score value of a unit point corresponding to each cell ("10" in FIG. 8) of the input image 30 based on the Gaussian function .

The number N of the range angles of the pedestrian model means the number N of the range angles when the angle division is set to 20 ° as shown in FIG. 5, for example, Quot; 18 ", and when the angle division is set to a unit range of 30 DEG, the number N of the range angles becomes "12 ".

The height H and the width W of the pedestrian model are represented by cell indices constituting a model image for generating a pedestrian model and / or a pedestrian model (the learned model image in FIG. 7) generated thereby . For example, the highest value "30" of the Y axis cell index (Model Y) of the "004.jpg model image" is the height H of the pedestrian model and the highest value "20" (W) of the pedestrian model.

Meanwhile, Equation (6) is constructed based on the Gaussian function. Specifically, the variable corresponding to the average of the Gaussian functions is replaced by "M" and the input is replaced by "I".

In Equation (6), "? "Has to be changed according to the size of the pedestrian model HOG, and therefore, a constant C is added so that it can be changed in proportion to the magnitude of" M ".

9 (a) and 9 (b) are examples of the score function according to the present invention. Specifically, FIG. 9A is a graph showing the change of the score function according to the HOG characteristic of the input image when "M" is set to 500. FIG. 9B is a graph showing the change of the score function when "M" It is a graph showing the change of the score function.

The difference between Figs. 9 (a) and 9 (b) is that the score acquisition range varies depending on the model size. As a result, the larger the feature of the model image, the wider the acquisition range of the score, so as to give more weight than the smaller feature.

When the score value according to Equation (6) is calculated for each unit point (i.e., the fourth unit point, ..., the N-th unit point) of the input image as described above, It is determined whether or not the user is a pedestrian.

Specifically, if a plurality of score values calculated for each unit point are compared with predetermined threshold values as in the following conditional expression 1, and if the score value of any one of the plurality of score values is greater than the threshold value And recognizes the input image as a pedestrian image.

<Conditional Expression 1>

On the other hand, if all of the plurality of score values are less than the threshold value as a result of the comparison, the input image is recognized as a non-pedestrian image.

Meanwhile, the pedestrian recognition method using the camera according to the present invention can be provided in the form of a recording medium readable by an electric / electronic apparatus in which a program for executing the pedestrian recognition method is stored in an electric / electronic apparatus such as a computer.

For example, in the recording medium, information on the pedestrian model obtained through the learning step S10 is recorded, and a program for executing the detailed steps S21 to S25 constituting the recognition step S20 is recorded And may be provided as a medium readable by an electric / electronic device.

As another example, the recording medium may include a medium readable by the electric / electronic apparatus in which a program for executing the sub-steps S12 to S15 constituting the learning step S10 is recorded, A program for executing the steps S21 to S25 may be provided as a medium readable by the recorded electric / electronic device.

While the preferred embodiments of the present invention have been described and illustrated above using specific terms, such terms are used only for the purpose of clarifying the invention, and it is to be understood that the embodiment It will be obvious that various changes and modifications can be made without departing from the spirit and scope of the invention. Such modified embodiments should not be understood individually from the spirit and scope of the present invention, but should be regarded as being within the scope of the claims of the present invention.

10: Cell
20: learned model image (pedestrian model)
30: input image

Claims (13)

A first step of collecting a plurality of images (hereinafter referred to as 'model images') through a camera;
A second step of extracting HOG features for each model image using HOG (histogram of gradient);
A third step of generating a pedestrian model by performing model learning through a machine learning algorithm based on the HOG feature of each of the model images extracted through the second step;
A fourth step of inputting an image (hereinafter, referred to as an 'input image') requiring identification of the presence or absence of a pedestrian;
A fifth step of performing scoring in which a score value is calculated in proportion to the degree of similarity between the HOG feature of the pedestrian model and the HOG feature extracted from the input image;
A sixth step of comparing a score value calculated through the fifth step with a preset threshold value; And
And a seventh step of recognizing the input image as a pedestrian image when the score value is greater than the threshold value as a result of the comparison in the sixth step.
The method according to claim 1,
The second step comprises:
The HOG features are extracted for each range angle by dividing the model image into a plurality of range angles,
In the third step,
Model learning is performed through the machine learning algorithm based on HOG features of a plurality of range angles of the model image,
Wherein the pedestrian model is configured to have information on a HOG feature for each range angle.
3. The method of claim 2,
In the fifth step,
The pedestrian model is matched to the input image on the basis of each unit point of the input image, and the score value is calculated for each unit point,
In the sixth step,
A plurality of score values calculated for each unit point are compared with predetermined threshold values,
In the seventh step,
Wherein the input image is recognized as a pedestrian image when the score value of any one of the plurality of score values is greater than the threshold value as a result of the sixth step.
The method of claim 3,
Wherein the score value for each unit point in the fifth step is calculated using a Gaussian function.
The method of claim 3,
Wherein the score value for each unit point in the fifth step is calculated using the following equation.

(score (x, y)): score of the (x, y) cell unit of the input image,

, c: constant, N: number of range angles of the pedestrian model,
H: Height of pedestrian model, W: Width of pedestrian model, M: Pedestrian
HOG feature value of model, I: HOG feature value of input image)
3. The method of claim 2,
Wherein the plurality of range angles are "range angles divided by 18 in total by 18 degrees" to "range angles divided by 12 in total by 30 degrees &quot;.
The method according to claim 1,
Wherein the second step further comprises removing a noise including a background included in the model image using a HOG weight filter.
The method according to claim 1,
Wherein the machine learning algorithm is configured to use a support vector machine (SVM).
Collecting a plurality of images (hereinafter referred to as 'model images') through a camera; Extracting a HOG feature for each model image using a histogram of gradient (HOG); And performing model learning through a machine learning algorithm based on the extracted HOG features of the respective model images, based on the pedestrian model obtained by the step (a) , &Quot; input image &quot;) as a pedestrian,
A first step of receiving the input image;
A second step of performing scoring in which a score value is calculated in proportion to the degree of similarity between the HOG feature of the pedestrian model and the HOG feature extracted from the input image;
A third step of comparing the score value calculated through the second step with a preset threshold value; And a fourth step of recognizing the input image as a pedestrian image when the score value is greater than the threshold value as a result of the comparison in the third step and recognizing the input image as a non-pedestrian image when the score value is smaller than the threshold value, A Pedestrian Recognition Method Using.
10. The method of claim 9,
In the pedestrian model,
Extracting HOG features for each range angle by dividing the model image into a plurality of range angles and performing model learning through the machine learning algorithm based on HOG features for each of a plurality of range angles of the model image, Wherein the pedestrian model is configured to have information on a HOG characteristic for each of the range angles.
11. The method of claim 10,
The second step comprises:
The pedestrian model is matched to the input image on the basis of each unit point of the input image, and the score value is calculated for each unit point,
In the third step,
A plurality of score values calculated for each unit point are compared with predetermined threshold values,
In the fourth step,
Wherein the input image is recognized as a pedestrian image when the score value of any one of the plurality of score values is greater than the threshold value as a result of the third step.
12. The method of claim 11,
Wherein the score value for each unit point in the second step is calculated using the following equation.

(score (x, y)): score of the (x, y) cell unit of the input image,

, c: constant, N: number of range angles of the pedestrian model,
H: Height of pedestrian model, W: Width of pedestrian model, M: Pedestrian
HOG feature value of model, I: HOG feature value of input image)
A computer-readable medium having recorded thereon a program for executing any one of the methods of any one of claims 1 to 11 in an electric / electronic apparatus.

Images