KR101270604B1 - Face detecting method for smart airbag - Google Patents

Abstract

PURPOSE: A face detection method for a smart airbag is provided to detect a face position of a passenger through gray scale conversion and simple algorithms, thereby simply detecting the face position. CONSTITUTION: An airbag system preprocesses an image inputted through a camera of a vehicle and makes a binary code for the pixels of the preprocessed image(S10,S20). The airbag system assigns label values to the pixels and calculates a candidate area of a human shape(S30). The airbag system assigns a distance value to each pixel of the candidate area(S50). The airbag system extracts a central line through sessionization and labeling value assignment for the pixels and simplifies the central line(S60,S70). The airbag system calculates a central coordinate of a face(S80). [Reference numerals] (AA) Start; (BB) End; (S10) Preprocess an image; (S20) Binarize the image; (S30) Assign label values to each pixel and calculate a candidate area of a human shape; (S40) Fill an empty space filling the label values in an empty pixel; (S50) Assign a distance value to each pixel of the candidate area; (S60) Extract a central line through sessionization and labeling value assignment; (S70) Simplify the central line; (S80) Calculate a central coordinate of a face

Description

Face detection method for smart airbag {FACE DETECTING METHOD FOR SMART AIRBAG}

The present invention relates to a face detection method for a smart airbag, and more particularly, to a face detection method for a smart airbag that enables the passenger's face position to be detected through a simple algorithm without generating a complicated parallax image.

In general, a vehicle is equipped with an airbag system that is deployed between the structure of the vehicle and the passenger at the time of an accident, so as to prevent the passenger from being directly injured by the structure of the vehicle.

These airbag systems are to protect the passengers by controlling the deployment of the airbag cushion (ACU) according to the signal detected from the sensor for detecting the accident of the vehicle.

However, since airbag cushions are generally designed to be deployed at a deployment pressure that is adapted to a general adult in consideration of general adult physical conditions, they may be injured due to the deployment pressure of the airbag cushion when infants or children sit in the passenger seat. there is a problem.

In addition, even if an adult is sitting forward or rearward or sitting deep in the seat, the airbag cushion does not protect the passenger sitting in the passenger seat with the designed deployment pressure.

Accordingly, there is a need for an airbag apparatus that allows the deployment pressure and direction of the airbag cushion to vary depending on the type or location of the passenger.

In order to determine the type and location of such a passenger, a method using a weight sensor, a method using a stereo vision system and a method using a distance sensor are generally applied.

The method using the weight sensor is to determine whether or not the adult according to the weight of the passenger in the passenger seat to control the deployment pressure of the airbag cushion, there is a problem that can misjudge a heavy child as an adult or a light adult as a child.

In addition, the stereo vision system obtains disparity information through a CCD and an infrared light emitting diode, and uses the same to estimate a passenger's position information and a face position. However, such a stereo vision system needs to acquire a parallax image in which a difference image of several frames is accumulated, and process the correlation between the parallax images, which consumes a lot of computational time for image processing and a face position due to parallax distortion. There was a problem that the reliability of the calculation is lowered.

The present invention for solving the problems of the background technology converts the image input through the camera installed in the vehicle to grayscale, and binarize each pixel by comparing the grayscale converted image with a threshold value, each of the binarized pixels After labeling each pixel located in one closed curve with the same label value, extract each line of the labeled image into the centerline by using the distance conversion method, and then extract the line that is the most central, and then use the center value of these pixels. The present invention provides a face detection method for a smart airbag that can detect a face position.

The present invention relates to a method for detecting a face for a smart airbag, the method comprising: preprocessing an image input through a camera installed in a vehicle, comparing brightness values of each pixel in the preprocessed image with a threshold value, and the brightness value of the target pixel is critical A binarization processing step of assigning a value corresponding to white if the value is less than the threshold value, and a value corresponding to a black value if the threshold value is less than the threshold value, giving a label value for each binarized pixel, and calculating a candidate region having a human shape. Assigning a distance value to each pixel based on any one of the pixels in the candidate region using a distance conversion algorithm, and assigning a single pixel to each pixel to which the distance value is assigned by sessioning and labeling values. Extracting a center line consisting of a plurality of branch points continuously connected, the number of lines connected to each branch point among the center lines A step of removing the reference count or less lines simplify center line, calculating a coordinate average of the number of lines based on the number associated with each branch point of the simplified center line than to the branch point and a step of calculating the coordinates in the coordinates of the center of the face.

In this case, the preprocessing step may include an image size conversion step of reducing the size of the input image, converting the size-converted image to gray scale, and removing noise of the image converted to gray scale.

In addition, after the step of calculating the candidate region in the form of a person, the empty space filling step of copying and filling the label value of the candidate region into the empty pixel to which the label value is not given in the candidate region may be further performed.

In addition, the filling of the empty space may track an outline of the candidate area and copy and fill a label value of the candidate area in the empty space inside the outline.

The present invention can detect a passenger's face position through simple algorithms such as grayscale transformation, binarization, labeling, and distance transformation, and thus can easily detect face position by generating complex parallax images and using no algorithm. In addition, it is possible to prevent a decrease in the accuracy of calculating face position due to parallax distortion.

1 is a flowchart sequentially showing a face detection method for a smart airbag according to the present invention.
2 is a flowchart illustrating an image preprocessing process of FIG. 1.
Figure 3 is an example screen showing a step-by-step method for detecting a face for a smart airbag according to the present invention.
4 is an exemplary view for explaining a grayscale conversion process when detecting a face for a smart airbag according to the present invention.
FIG. 5 is a graph illustrating a histogram for explaining an Otsu algorithm applied when detecting a face for a smart airbag according to the present invention. FIG.

Hereinafter, with reference to the accompanying drawings will be described in detail a preferred embodiment of the present invention.

1 is a flowchart sequentially showing a face detection method for a smart airbag according to the present invention, FIG. 2 is a flow chart showing the image preprocessing process of FIG. 1, and FIG. 3 is an image showing the face detection method for a smart airbag according to the present invention step by step. 4 is an exemplary view illustrating a grayscale conversion process when detecting a face for a smart airbag according to the present invention, and FIG. 5 is an Otsu Algorithm applied to detecting a face for a smart airbag according to the present invention. A graph showing a histogram for explaining

First, the face detection method for a smart airbag according to the present invention includes an image preprocessing step (S10), a binarization processing step (S20), a human shape candidate area calculation step (S30), a blank space filling step (S40), and a pixel-specific distance value assignment step. S50, a centerline extraction step S60, a center point simplification step S70, and a face center coordinate calculation step S80.

First, the image preprocessing step S10 is performed for image processing optimization in an embedded system environment, and includes an image size converting step S11, a gray scale converting step S12, and a noise removing step S13.

The image size converting step (S11) is to reduce the size of the input image since the method for detecting a face for a smart airbag according to the present invention is performed in an embedded system which is a limited resource. .

In the gray scale conversion step 12, only the luminance signal from which the color signal is removed from the size-converted image is extracted and converted into the gray scale. That is, the input image is an IR (infrared) image, and the visible portion is a gray scale image, but in fact, a luminance signal Y having a YCbCr type and a chrominance signal Cb, Cr are summed. In general, YCbCr is a color expression method that is used as a standard when compressing a video. Each pixel is used to share four Yb Cb and one Cr as shown in FIG.

The noise removing step S13 is for removing noise from the gray scale converted image, and may use a linear smoothing filter. The method using a linear smoothing filter can smooth the edge portion by making a specific pixel value equal to the average value of neighboring pixels.

The binarization process step (S20) is a step of allocating a value corresponding to white or a value corresponding to black according to the brightness value of the target pixel. When the brightness value of the target pixel is less than or equal to the threshold value, a value (0) corresponding to white is assigned. If the value is less than the threshold value, a black value (255) is assigned. In this case, the threshold for the binarization process may be obtained using an Otsu algorithm.

The Otsu Algorithm finds a valley point and assumes a threshold value T when assuming that the histogram of the image is bimodal. In other words, the histogram is obtained for each pixel in the image as shown in FIG. 5, and the histogram 2 is divided by setting a threshold value so that the variance divided into two regions is maximized when the histogram is divided into two regions.

For example, when the density level is 1 to m and the frequency of the density level i is n ₁ , the number of pixels and the probability of each level are as follows.

All pixels

Probability of each level

의 2개의 영역으로 나누었다고 하면,

의 출현 확률과 평균값은 아래의 식과 같다.Split into two parts with level k

If we divide it into 2 areas of,

The probability of occurrence and the average value of are given by

의 출현 확률

Probability of appearance

의 출현 확률

Probability of appearance

의 평균값

Average of

의평균값

Average value of

는 전 샘플 값의 평균값 only,

Is the mean of all sample values

Therefore, f becomes as follows.

In addition, the dispersion between the two regions is as follows.

을 부여하는 k를 구해서 이것을 임계값으로 한다. 이 방식은 2개의 영역으로 분리하였을 경우 양쪽 영역의 정리가 쉽고 영역 사이의 분리가 우수하다.Using this formula max

Let k be given and set this as the threshold. This method is easy to organize both areas when the two areas are separated and excellent separation between the areas.

On the other hand, the step of calculating the candidate shape of the human shape (S30) is to calculate the candidate region for morphologically distinguishing the human shape by labeling each of the binarized pixels.

In this case, labeling is a process of considering areas connected to the current pixel or connected areas or areas having the same color range as one object, and the present invention uses a binary image with values corresponding to white and search tablets. As a result, the same label value is assigned to pixels having the same color value.

In the blank space filling step S40, when there is an empty pixel to which no label value is assigned in the candidate area after the above-mentioned labeling, the empty value is copied and filled with the label value of the candidate area.

In other words, the candidate area selected after labeling has a hole area therein. As a result of the distance transform, the candidate area has a significant influence in determining the candidate area in human form. It must be filled. In this case, the fill image algorithm has a large amount of calculation in terms of processing speed when using the region filling algorithm generally used in the existing image processing. Therefore, the outline of the candidate region is traced to track the outline of the candidate region. By eliminating holes, the amount of calculation can be reduced to shorten the empty space filling time.

The distance value assigning step (S50) is to assign a distance value for each pixel through a distance transform algorithm based on one pixel in the candidate area obtained through labeling to extract a human-like feature.

The distance transform algorithm calculates the distance from one pixel on the image to the boundary of the nearest object. Intuitive distance transformation is to find the foreground boundary and assign the minimum distance from each pixel. However, since this method requires too many calculations when the boundary is composed of many pixels, the present invention will be described in the case of applying an algorithm that is linearly proportional to the size of the image.

First, in order to be linear in the number of pixels, the algorithm should be able to be implemented by the raster scan method. The distance transformation considers a hypothetical secondary paraboloid at each point in the foreground, and eventually finds the one that gives the smallest value at any point. In other words, it is necessary to find the paraboloid located at the lowest point at any point, but it is difficult to find the paraboloid in the secondary region.

While scanning in the x-axis direction, the distance in the x-axis direction is stored, and scanning in the y-axis direction is used again to use information of the square of the distance in the x-direction. This is more easily solved by storing the square of the distance as the gray value of the image when scanning in the x-axis direction. Then, when scanning in the y-axis direction, the vertex is found at the bottom of the quadratic curves previously moved by the square of the distance stored in the x-axis scan.

Therefore, Euclidean distance transform can be obtained by performing one-dimensional distance transform twice. In the one-dimensional distance transform, one of the quadratic curves whose values are moved at each grid point is selected and taken from the bottom of the corresponding area.

The one-dimensional distance transformation is DT (p) = min_ {q} ((p-q) ^ 2 + F (q)).

Q is all grid points.

The initial value of F (q) can be set to 0 on the foreground grid and + INFINITY on the background. The quadratic curve formed at the boundary of the foreground is the lowest, which determines the distance. Given a general value of F (q), the problem is which one is the lowest quadratic curve seen at each grid point, and the problem is to find the area where it is established. The point 2, the point where the intersection of the quadratic curve is s = ((F [p] + p * p)-(F [q] + q * q)) / (2 * p-2 * q), is the lowest 2 Find the area of the difference curve.

The center line extraction step (S60: Peak Detection) is a step of extracting a center line composed of a plurality of branching points by successive single pixels through thinning and labeling the pixels of the candidate area to which the distance value is assigned.

In other words, the centerline extraction step (human shape feature extraction) is to find the floor portion to which the highest distance value is assigned from the distance map. In this case, when the threshold value is applied locally or globally, the extracted peaks are broken, making it difficult to recognize a human feature. Therefore, in order to solve this problem, the kernel mask value of the edge detection algorithm commonly used in image processing is modified to have a distance value of 90% of the peak pixel (distance value), and the peak point and its surrounding area are applied. Extract. Then, the area around the selected peak point is labeled. At this time, the labeling is for removing the small area far from the peak area. Subsequently, by performing sessionization on the extracted peak point and its surrounding area, a single pixel is continuously connected without any break and a center line composed of a plurality of branch points is extracted.

In the centerline simplifying step S70, the number of lines connected to each branch point among the centerlines is less than the reference number. That is, since the extracted human centerline has branching points (intersections of the lines), the centerline is simplified by eliminating branching points where the number of lines connected based on the branching point is three or less, for example.

In calculating the center coordinates of the face (S80), a coordinate average value of the branch points having the number of lines connected to each branch point among the simplified center lines is greater than or equal to the reference number is calculated to calculate the coordinates as the center coordinates of the face. That is, the coordinate average value of the points having three or more branch points satisfying the characteristics of the seated person (the form of the Korean vowel “ㅗ”) and the knowledge base (one-third of the vertical length to locate the face) are obtained. It is extracted from the coordinates of the center of the final face.

Although the present invention has been described in connection with the above-mentioned preferred embodiments, it is possible to make various modifications and variations without departing from the spirit and scope of the invention. Accordingly, the appended claims are intended to cover such modifications or changes as fall within the scope of the invention.

Claims (4)

Preprocessing an image input through a camera installed in the vehicle;
A binarization processing step of comparing a brightness value of each pixel in the pre-processed image with a threshold value and allocating a value corresponding to white when the brightness value of the target pixel is less than or equal to the threshold value and assigning a value corresponding to black when the threshold value is less than the threshold value;
Assigning a label value to each binarized pixel and calculating a candidate region in a human form;
Allocating a distance value for each pixel based on any one of each pixel in the candidate region using a distance conversion algorithm;
Extracting a center line consisting of a plurality of branching points with successive single pixels through sessionization and labeling value assignment for each pixel to which the distance value is assigned;
Simplifying the center line by removing lines whose number of lines connected to each branch point among the center lines is equal to or less than a reference number;
And calculating a mean value of branch coordinates whose number of lines connected to each branch point among the simplified center lines is equal to or greater than a reference number, and calculating the coordinates as the center coordinates of the face.
The method of claim 1,
The pre-
An image size conversion step of reducing the size of the input image;
Converting the size-converted image to gray scale;
And a noise removing step of the image converted into the gray scale.
The method of claim 1,
After the step of calculating the candidate region of the human shape, the smart airbag face characterized in that for performing the empty space filling step of copying and filling the label value of the candidate region to the empty pixel that is not given a label value in the candidate region Detection method.
The method of claim 3, wherein
The empty space filling step tracks an outline of the candidate area and copies and fills a label value of the candidate area into a blank space inside the outline to fill the blank space.

Images