KR100829904B1 - Vehicles detecting method using multiple resolution technique - Google Patents

Abstract

A vehicle detecting method using multiple resolution technique is provided to calculate correlation of coefficient between a vehicle area and a vehicle template, using a specific map for checking vehicles ahead and behind accurately through template matching and CCD camera sensors. Light/darkness information is extracted from input images(S100). Contour information is extracted from the light/darkness information and images of multiple resolution are created(S120). Symmetrical information is extracted from the light/darkness information(S150). The contour and symmetrical specific images are created using contour and symmetrical information(S140,S170). The specific map is created by adding the contour and symmetrical specific images(S180). A following vehicle area is calculated using entropy algorithm(S200). An actual vehicle area is selected from the following vehicle area(S270). The correlation of coefficient between the following vehicle area and the vehicle template is calculated by putting the specific map.

Description

Vehicle detection method using multiple resolution technique

1 is a control block diagram of a device to which a steering control method using a monovision disclosed in the conventional patent application No. 10-1996-0073015 "steering control method using monovision" is applied.

2 is a control flowchart of a steering control method using the monovision of FIG.

3 is a view showing a field driving apparatus using the color vision disclosed in the conventional patent application No. 10-1996-0079563 "field driving apparatus and method using color vision".

4 is a view showing a front vehicle detecting apparatus disclosed in the conventional patent application No. 10-2001-0033824 "front vehicle detecting apparatus".

5 illustrates a vehicle detection model of the present invention.

6 is a flow chart illustrating a vehicle detection algorithm of the present invention.

7 is a view showing an input image input from the outside of the present invention.

8 is a flowchart illustrating a process of extracting symmetry information of the present invention.

9 illustrates a method of implementing the fast NTGST algorithm of the present invention.

10 is a diagram illustrating a method for calculating a candidate vehicle region using the entropy algorithm of the present invention.

11 shows an example of the actual implementation of the algorithm of the present invention.

The present invention relates to a vehicle detection method using a multi-resolution technique that can efficiently detect a vehicle by comparing the outline information, contrast information, and symmetry information of the vehicle with the basic information of the vehicle through the multi-resolution image obtained from the input image.

Intelligent vehicle research is actively underway to provide driver safety and convenience. In particular, the vision field that will serve as the eye of intelligent cars is called the “Grand Challenge” and is attracting worldwide attention. In addition, efforts are being made to apply this vision technology to the front and rear vehicle detection of intelligent vehicles.

Hereinafter, with reference to the drawings will be described the problems of the prior art and the prior art.

1 is a control block diagram of a device to which a steering control method using a monovision disclosed in the conventional patent application No. 10-1996-0073015 "method of steering control using monovision" is applied, Figure 2 is a view using the monovision of FIG. Control flow chart of the steering control method.

Referring to FIG. 1, a device to which a conventional steering control method using a monovision is applied includes a monovision (1) mounted in a traveling direction of a vehicle to configure an environment in front of the vehicle as a two-dimensional screen and outputting the image as an electric signal (1). ), A lane detection device (2) for capturing an image signal in the monovision to perform image compression and edge detection, and outputting a control value for the driving lane of the vehicle, and the lane detection device A vehicle control device 3 for receiving an error signal of a vehicle traveling direction with respect to the lane at 2 and outputting a signal for adjusting the vehicle traveling direction, and receiving a control signal from the vehicle controller 3 for steering; It includes a steering control device 4 for driving.

Referring to FIG. 2, the lane detection apparatus 2 captures an image of an environment image in front of a vehicle using a mono vision (S20). The image is of high resolution and has a large amount of processing data and thus cannot be processed in real time. For this reason, the lane detection apparatus 2 compresses a high resolution image to a low resolution for real time image processing (S30). Determination of the slope from the three edges to determine whether the left edge or the right edge, and determines whether the slope of the left and right edge is correct to detect the lane (S40). The lane detecting apparatus 2 models the lane of the road using the detected edge and calculates an error between the lane direction and the traveling direction of the vehicle (S50). The vehicle controller 3 outputs the steering control value obtained above to the steering controller 4 and controls steering by the control value (S60).

3 is a view showing a field driving apparatus using the color vision disclosed in the conventional patent application No. 10-1996-0079563 "field driving apparatus and method using the color vision".

Referring to FIG. 3, a color vision 6 which is mounted in a traveling direction of a vehicle to configure an environment in front of the vehicle as a two-dimensional screen and outputs the image as an electric signal, and three central processing units are configured. An image processing apparatus 7 which captures an image signal in a vision, performs image compression and edge detection, and outputs a control value for a driving lane of the vehicle, and an error in the vehicle traveling direction with respect to the lane in the image processing apparatus 7. And a vehicle control device 8 for receiving a signal and outputting a signal for adjusting a traveling direction of the vehicle, and a steering control device 9 for driving the steering by receiving a control signal from the vehicle control device 8.

The prior art disclosed in Patent Application No. 10-1996-0073015 and Patent Application No. 10-1996-0079563 is for the purpose of detecting a lane, not for the purpose of detecting the vehicle, and automatically searches for a lane by processing image signals in parallel. It's just a technology that provides road driving capabilities. That is, there is a problem in that the vehicle can be operated using the conventional technology only when there is no vehicle on the road.

4 is a diagram illustrating a front vehicle detecting apparatus disclosed in the patent application No. 10-2001-0033824 "front vehicle detecting apparatus".

Referring to FIG. 4, the apparatus for detecting a front vehicle includes a transmitter 10 for transmitting a millimeter wave, that is, a measurement wave, to a front vehicle, a receiver 30 for receiving a measurement wave reflected from the front vehicle, and a received measurement wave. It includes a digital signal processing unit 20 for calculating the distance with the front vehicle based on.

The transmitter 10 includes a voltage controlled oscillator 12, a linearizer ramp generator 11, an amplifier 13, a frequency doubler 14, and a transmit antenna 15. The linear signal generator 11 supplies an input voltage to the voltage controlled oscillator 12, and the voltage controlled oscillator 12 outputs a variable oscillation frequency of the supplied input voltage. The amplifier 13 amplifies and outputs the voltage output from the voltage controlled oscillator 12, and the frequency doubler 14 doubles and outputs the frequency of the amplified voltage.

Receiving unit 30 includes receiving antennas 31 and 31 ', first and second harmonic mixers 32 and 32', first and second intermediate frequency amplifiers 33 and 33 ', first and second automatic It includes an auto gain controller 34, 34 ', first and second buffers 35, 35', and a signal converter 36. The first and second low frequency mixers 32 and 32 'mix the received signals received from the receiving antennas 31 and 31' and the signals output from the amplifier 13 of the transmitter 10, respectively, to generate an intermediate frequency signal. The first and second intermediate frequency amplifiers 33 and 33 'amplify and output the generated intermediate frequency signals, respectively. The first and second automatic gain adjusters 34 and 34 ′ are used to adjust the intermediate frequency signals amplified by the first and second frequency amplifiers 33 and 33 ′ according to the gain control signals output from the digital signal processor 20. The gain is adjusted and output, and the first and second buffers 35 and 35 'temporarily store the gain-adjusted intermediate frequency signal and output it to the signal converter 36. The signal converter 35 converts the signal output from the digital signal processor 20 into an analog signal and outputs the D / A converters 361 and 262, and converts an intermediate frequency signal into a digital signal to the digital signal processor 20. And output A / D converters 363 and 364.

The digital signal processor 20 controls the overall operations of the transmitter 10 and the receiver 30, calculates a distance from the vehicle ahead based on the signals received and processed by the receiver 30, and calculates the distance from the front vehicle. Display the distance to the vehicle or perform a warning.

The conventional technology is characterized by detecting the distance between the vehicle and the vehicle traveling in front by using a millimeter (mm) wave, but the exact position of the vehicle when changing the lane of the vehicle or when the vehicle in front of the vehicle is outside the detection range of the transmitter It is difficult to detect, and there is a problem that it is not easy to distinguish between the vehicle and other obstacles.

In order to solve this problem, the development of a vehicle detection technology using vision technology has recently been made. In order to develop a vision-based intelligent vehicle, it is essential to develop a cognitive mechanism and its real-time model. Nevertheless, the conventional vehicle detection technology remains at the extent of detecting a vehicle by detecting lanes on a highway and analyzing specific information such as color, contour information, texture, symmetry degree, and shadow, respectively, based on the detected lanes. These models are sensitive to noise caused by natural phenomena such as changes in illumination over time, shadows, lane conditions, and the effects of rain, making them difficult to use.

The present invention devised to solve the above problems is an object of the present invention to provide a vehicle detection method using a multi-resolution technique that can accurately detect the front and rear vehicles using a template matching less sensitive to disturbance.

The present invention for achieving the above object, (a) setting the reference position and extracting the contrast information by removing the image of the reference position or more; (b) extracting contour information from the contrast information and creating a multi-resolution image; (c) extracting symmetry information from the contrast information; (d) adding and normalizing all layers of the contour information and the symmetry information to form a contour, symmetric feature image; (e) constructing a feature map by adding and averaging the contour and symmetric feature images; (f) combining and normalizing Gaussian contour information and symmetry information of each layer to construct a normalized feature map; (g) extracting a car candidate region through a local maximum search for the normalized feature map; (h) obtaining a candidate vehicle region using an entropy algorithm; And (i) selecting an actual vehicle region from the candidate vehicle region.

Here, the step (b) includes generating an image having different multi resolutions through Gaussian filtering and enlarging the multi resolution image to the resolution size of the input image using Gaussian filtering.

In the step (c), the symmetry information is extracted using the NTGST algorithm.

Also, in the step (h), an area satisfying the entropy algorithm is selected and set as the candidate vehicle size while increasing the size by a predetermined size from the candidate vehicle center position in all directions with respect to the normalized feature map.

In addition, the step (i) may include obtaining an average value and standard deviation of the symmetry information of each candidate vehicle region, calculating a correlation coefficient between the candidate vehicle region and the vehicle template by inputting a feature map, and symmetry. And selecting an area in which the standard deviation of the information is larger than the reference standard deviation and the correlation coefficient of each candidate vehicle area is larger than the reference correlation coefficient as the final vehicle area.

In order to fully understand the present invention, the advantages of the operability of the present invention, and the objects achieved by the practice of the present invention, reference should be made to the accompanying drawings which illustrate preferred embodiments of the present invention and the contents described in the accompanying drawings.

Hereinafter, exemplary embodiments of the present invention will be described in detail with reference to the accompanying drawings. Like reference numerals in the drawings denote like elements.

5 is a diagram showing a vehicle detection model of the present invention, Figure 6 is a flow chart showing the vehicle detection algorithm of the present invention.

5 and 6, the motor vehicle generally has a rectangular-like shape. Therefore, the vehicle detection method of the present invention expresses the automobile area more prominently than the surroundings by using the multi-resolution image of the edge and symmetry information, and uses the characteristic information to search the local maximum searching. The candidate vehicle region is obtained by using the method, and the candidate vehicle region is calculated using an entropy maximization algorithm, which is one of information maximization algorithms. In addition, the actual vehicle region is finally selected among the candidate vehicle regions by using template matching and symmetry information.

5 and 6 will be described in more detail with reference to FIGS. 7 to 10, when an input image is input from the outside (S100), as shown in FIG. 7, unnecessary operations are reduced and complicated backgrounds are removed for vehicle detection. In order to set the reference position (y _m ) to remove the image above the reference position (S110). Next, contour information is extracted using a Sobel operator from the input image from which the background of the reference position or more is removed for vehicle detection (S120). When the extraction of the contour information is completed, a multi-resolution image is generated. In more detail, Gaussian pyramid contour information is extracted using a Gaussian filter (S130).

For example, if the resolution of the input image is 360 * 240, a multi-resolution image may generate 180 * 120 resolution, 90 * 60 resolution, 45 * 30 resolution, and 22 * 15 resolution images. Once created, we filter it using a Gaussian filter. Subsequently, the multi-resolution image is enlarged to the resolution size of the input image by using a Gaussian filter, and these images are added and normalized to construct a contour feature map (S140).

When generating the multi-resolution image, a Gaussian filter or a smoothing filter may be used to blur each image to remove high frequency noise and to express prominent information about the object.

On the other hand, the paper on the input image Context-free Maker-controlled watershed transform for efficient multi-object detecton and segmentation, K. S. Seo. et al, IEICE Trans, 384-A, no. 6, 1392-1400pp, 2001. The use of the noise-tolerant generalized symmetry (NTGST) algorithm disclosed in Figure 8 to extract the symmetry information through the procedure shown in Figure 8 (S150) Gaussian pyramid symmetry information is extracted (S160).

Referring to FIG. 8, a pyramid image of contrast information is extracted to generate pyramid symmetry information (S160a). After extracting the pyramid image of the contrast information, the direction information of the pyramid contour information is extracted (S160b), and the pyramid contour information is extracted (S160c). The phase weighting function is calculated on the basis of the extracted position information of the pyramid contour information and the symmetry calculation position sampling window of the vehicle template size (S160d), and the distance weighting function is calculated (S160e). Pyramid symmetry information is extracted using the pyramid contour information extracted in step S160c and the phase weighting function and the distance weighting function calculated in steps S160d and S160e (S160f). Thereafter, symmetry information is extracted from the sum of the respective pyramid images (S160g).

The Gaussian pyramid symmetry information extracted through the above step is enlarged to the size of the input image with Gaussian filtering. Then, these images are added and normalized to form a symmetric feature map (S170).

The NTGST algorithm is composed of a phase weight function, a distance weight function and a symmetry operator. Here, to implement the NTGST algorithm having a fast processing speed, as shown in FIG. 9, a predetermined window size is set, a window size is set according to the vehicle template size, and a position to be calculated is set. Create a look-up table with the distance weighting function and other parameters.

Meanwhile, all layers of Gaussian contour information and Gaussian symmetry information are added and normalized to form a contour image. In addition, a contour map and a symmetrical feature image are simply added and the average value is taken to form a feature map (S180). Further, the Gaussian contour information and the symmetry information of each layer are added and normalized to form a normalized feature map (S190).

Next, since most cars have the same shape, a car candidate area is obtained through local maximum searching for a normalized feature map by creating a window having a certain shape having the same size (S200). The candidate vehicle region is then calculated using an entropy algorithm based on the histogram disclosed in the paper Scale saliency and image description, Timor Kadir et al., International Journal of Computer Vision, 45 (2001), 83-105. The method proposed by Timor Kadir consists of a histogram-based entropy equation, a size weighting function, an entropy reflecting the size weighting function, and a condition for selecting an optimal size.

Referring to FIG. 10 as an example, when a normalized feature map is obtained for the input image of FIG. 10, first, the normalized feature map is obtained, and as shown in FIG. 10B for the obtained normalized feature map. Entropy and size weighting are obtained by increasing the size by a certain size from the center of the candidate car up, down, left, and right. When the condition for selecting the optimal size is satisfied by obtaining the entropy and the size weighting value, it is set as the size of the candidate car as shown in FIG. This can significantly reduce the amount of computation compared to the method obtained by calculating the size of the object for each pixel of the input image by Timor Kadir et all.

If the candidate vehicle region is selected through the above process, the actual vehicle region is selected from the candidate vehicle regions through the following process. In more detail, using the symmetry information extracted through the step S150 to obtain the average value (Sy _mean ) of the symmetry information of each candidate vehicle region (S210). Subsequently, the average symmetry value of the candidate vehicle area is compared with the reference _threshold (Sy _threshold ), and when the symmetry average value is less than or equal to the reference value, the area is excluded from the candidate vehicle area (S230). In general, cars have a strong symmetry component, so symmetry is taken into consideration in order to select candidate car regions.

If the average symmetry value is greater than or equal to the reference value, the standard deviation Sy _std of the symmetry information is obtained for the selected candidate vehicle region through the above process (S240). In the symmetry information on the automobile region, the symmetry information is not expressed locally but the symmetry information is displayed overall.

In operation S90, the preset rectangular car-shaped template is reduced or enlarged by the size of the selected car candidate region. In this case, the car-shaped template is configured with feature maps for several cars in advance, and the car region is directly divided by the person and stored in a predetermined size. The degree of similarity is calculated by calculating a correlation coefficient (Mat _corr ) between the selected candidate vehicle region and the vehicle template reduced or enlarged by the size of the candidate vehicle region by inputting the feature map obtained through the operation S180 (S250).

The standard deviation Sy _std of the symmetry information for each candidate vehicle region obtained through the step S240 for the selected candidate vehicle regions is greater than a standard _{std threshold} (Sy _std > Std _threshold ), and the step S250. In operation S270, candidate regions having a correlation coefficient value Matt _corr of each candidate vehicle region calculated through S 1 (Cat _corr > Corr _threshold ) larger than a reference correlation coefficient value Corr _threshold are selected (S270).

11 is a diagram illustrating an example in which the above described algorithm is actually implemented.

Although the present invention has been described with reference to one embodiment shown in the drawings, this is merely exemplary, and those skilled in the art will understand that various modifications and equivalent other embodiments are possible therefrom. . Therefore, the true technical protection scope of the present invention will be defined by the technical spirit of the appended claims.

The present invention provides a vehicle detection method using a multi-resolution technique capable of accurately detecting front and rear vehicles using template matching which is less sensitive to disturbances. In addition, by using the CCD camera sensor to efficiently detect the front and rear vehicles while driving, accidents with the front and rear vehicles can be reduced.

Claims (5)

(a) extracting contrast information from an input image; (b) extracting contour information from the contrast information and creating a multi-resolution image; (c) extracting symmetry information from the contrast information; (d) adding and normalizing all layers of the contour information and the symmetry information to form a contour, symmetric feature image; (e) constructing a feature map by adding and averaging the contour and symmetric feature images; (f) combining and normalizing Gaussian contour information and symmetry information of each layer to construct a normalized feature map; (g) extracting a car candidate region through a local maximum search for the normalized feature map; (h) obtaining a candidate vehicle region using an entropy algorithm; And (i) selecting a real vehicle region from the candidate vehicle region. delete The method of claim 1, In step (c), Vehicle detection method using multiple resolution technique that extracts symmetry information using NTGST algorithm. The method of claim 1, (H) step, A method for detecting a vehicle using a multi-resolution technique in which a region satisfying an entropy algorithm is selected and set to a candidate vehicle size while increasing a size by a predetermined size in all directions from a candidate vehicle center position with respect to a normalized feature map. The method of claim 1, In step (i), Obtaining an average value and standard deviation of symmetry information of each candidate vehicle region; Calculating a correlation coefficient between the candidate car region and the car template by inputting the feature map; and And selecting a region in which the standard deviation of the symmetry information is larger than the reference standard deviation and the correlation coefficient of each candidate vehicle region is larger than the reference correlation coefficient as the final vehicle region.

Images