KR100335753B1 - Vehicle Detection Method Based on Wavelet Transforms - Google Patents

Abstract

The present invention relates to a method for detecting a vehicle from information input from an image detector installed at an intersection of a city, and more specifically, to increase reliability in detecting a vehicle by using wavelet transforms for detection areas for each lane. The present invention relates to a vehicle detection method using wavelet transform.

The present invention comprises the steps of setting a detection area by dividing a predetermined distance after the stop line at equal intervals with respect to the image information input to the image detection means; Converting the image signal of the set detection area into a discrete wavelet transform and converting the image signal into a frequency signal; Calculating a wavelet coefficient by decomposing the frequency-converted signal through a high frequency biorthogonal wavelet filter; Restoring the calculated wavelet coefficient to a contrast value between 0 and 255 using an inverse high frequency biorthogonal wavelet filter; And determining whether or not the vehicle exists in the corresponding detection area by comparing the number of pixels of the restored intensity value with a threshold value.

Therefore, the present invention eliminates the additional processing for removing noise caused by the change of weather by processing the image processing in the existing space in the frequency domain, and the contrast value of the road as the brightness changes according to the time zone of the day. The detection error rate is minimized by excluding the operation of estimating.

Description

Vehicle Detection Method Based on Wavelet Transforms {Vehicle Detection Method Based on Wavelet Transforms}

The present invention relates to a method for detecting a vehicle from information input from an image detector installed at an intersection of a city, and more specifically, to increase reliability in detecting a vehicle by using wavelet transforms for detection areas for each lane. The present invention relates to a vehicle detection method using wavelet transform.

In general, a device that detects traffic information, both at home and abroad, is an inductive loop detector (ILD) that detects traffic information by embedding a coil on a road and then detecting a change in electromotive force generated by the passage of a vehicle. Is mainly used.

However, in the case of ILD, the construction method is difficult because it must be buried on the road where the vehicle passes, and ILD must be buried again according to the renovation of the road surface, and the maintenance of the ILD coil is shorted or broken due to frequent road construction. There are many problems with this.

In addition, in the case of traffic information detected through ILD, the types of information available are limited to collecting traffic volume and occupancy time and non-occupancy time that pass through the point, and thus provide various information such as travel time prediction. There is a disadvantage that can not.

Therefore, as an alternative to the ILD, various detectors such as an ultrasonic detector, a microwave detector, and an image detector have recently been installed. Time, Non-Occupancy Time, etc. can be measured, as well as measuring the waiting length, about 3 to 4 types of vehicles, the speed of the spot, etc. Research is also underway to collect global information such as forecasts.

As an example, a device called Wide Area Detection System (WADS) applied in the United States (Ref. RMInigo, Traffic Monitoring and Control using Machine Vision: A Survey, IEEE Trans.on Industrial Electronics , Vol. IE-32, No. 3, Aug) , 1985, pp. 177-185, and RM Inigo, Application of Machine Vision to Traffic monitoring and Control, IEEE Trans.on Vehicular Technology , Vol. 38, Aug, 1989, pp112-122). Is a comprehensive traffic information providing system for vehicle detection, vehicle velocity measurement, and vehicle tracking using image processing for the first time in the world. This function is essential for collecting and analyzing information.

In the vehicle detection algorithm in the WADS, 50 pixels were set as one line in the horizontal direction for each lane, and five lines were searched in units of ten frames using the vehicle detection area. In other words, search for the 1 + j line in the first frame and the 6 + j line in the 11th frame (where j = 0,1,2,3,4) to find the vehicle if the difference with the road exceeds a certain threshold. How to detect.

The system searched in 10-frame units assuming the average speed of the vehicle approaching 20Km / h or more, and searched in 5-pixel units assuming that the vehicle length was less than 15m.

If the algorithm for this is expressed as an equation, it is expressed as Equation 1 below.

here : A new road determined as i = 1,2,... , 50,

L _i : Contrast value in the entering lane i.

In Equation 1 above Is an estimate of the contrast of the road by time zone, and fixed at ω = 0.8, but can be adjusted slightly according to the movement of the cloud, and the threshold T is determined by B _i , the minimum change of road for each time zone. Generally, it is determined by three times value of B _i .

Therefore, in the case of V _D > T for one of the five lines, it is determined that the vehicle is detected.

In addition, as another example, the proposed method in Japan (Reference: Modeling Machine, Eui-Jung Jung, Development of Waiting Length Measurement Algorithm using Voting Technique Based on Image Processing, Journal of Korean Society of Transportation, Vol. 16, No. 3, 1998, pp. 113) -121 and RMInigo, Traffic Monitoring and Control using Machine Vision: A Survey, IEEE Trans.on Industrial Electronics , Vol.IE-32, No. 3, Aug, 1985, pp.177-185). The traffic information is collected in advance and the vehicle is detected when there is a change according to the change of time t. The maximum value L _u and the minimum value L _l of the contrast value on the road surface are determined, and then the n th frame of the t th frame. If the contrast value of the pixel is I (n, t), P (n, t) is determined based on this I value as shown in Equation 2 below.

When P (n, t) is determined in Equation 2 above, the vehicle can be detected by comparing with the current p value and whether or not the vehicle is detected at time t-1 as shown in Equation 3 below.

In Equation 3, τ means the number of frames, and in general, the search is performed in units of 3 to 5 frames. When P = 1, the vehicle is detected, and when P = 0, the vehicle is not detected.

In addition, the contrast value of the road surface was estimated by τ for each time zone, and the estimation method is expressed by Equation 4 below.

(n, t) = (1-ε) (n, t-1) + εI (n, t)

if│I (n, t-1) -I (n, t) │> α then I (n.t) = I (n, t-1)

Here, α is a coefficient determined empirically, and generally, α = 1.5, ε = 0.01, or 0.02.

In another example, Iwasaki in Japan sets a detection area by dividing a road regularly by 30m, extracts an outline using a Sobel mask in the detection area, and detects a vehicle through a histogram of contrast values. The histogram distribution of contrast values in the presence of a vehicle shows a large distribution of values of 255, whereas in the absence of vehicles, the histogram distribution of contrast values is between 0 and 50. It can be detected.

This is expressed as Equation 5 below.

Here, P _i is the number of pixels of the edge portion in the region i, T _i is the threshold value in the region i, P _i-255 means the number of pixels of the contrast value 255 in the region i.

In addition, M.Higashikubo, T.Hinenoya, K.Takeuchi, SPITS (Special Image Processing Traffic Flow Sensor) in Traffic Queue Length Measurement using an Image Processing Sensor, 3rd Annual World Congress on Intelligent Transport System , Orland USA., Oct, 1996. It is a system that collects comprehensive traffic information such as traffic volume, speed, vehicle type, and waiting length through image processing, and sets sampled points in the detection area to specify the waiting length. The vehicle is detected by processing the time difference, the background difference, and the spatial difference with respect to the sample point.

Here, the time difference is a method for detecting moving vehicles, the background difference is a method for extracting a vehicle, and the space difference is a method for extracting an outline of the vehicle.

Among them, in order to extract a vehicle by the background difference, modification is required for each time zone of the background information, that is, road information, and the contour extraction by the space difference is greatly influenced by noise such as the direction indication of the road.

As described above, the existing methods using the image detector obtain the intensity or brightness or gray-level information corresponding to the road in advance, and the vehicle exists when the change amount of the intensity is larger than the set threshold. Since it is detected as having a very sensitive to noise (snow, rain, shadow, etc.) due to the change of weather, there was a problem that the accuracy is low.

In addition, since the intensity of the road is changed by time according to the intensity of light that changes by time of day, it is necessary to estimate the intensity of the road using the 'Kalman Filter' method to correct this, but the estimated intensity is continuously updated. For sudden changes such as daytime with heavy clouds, there is a problem that the accuracy of the vehicle detection is lowered because the reliability of the estimated value is lowered.

In particular, there is a problem that the occurrence of a detection error is very high due to the headlight of the vehicle at night.

The present invention has been made in view of the above-described general problems, and its object is to perform a discrete wavelet transform of a spatial domain image input from an image detector into a frequency domain, and then perform filtering to perform a change in weather or a day. It removes the noise due to the change of brightness in each time zone to increase the detection rate and increase the accuracy.

1 is a schematic configuration diagram of an image detection apparatus using wavelet transform according to the present invention.

2 is an exemplary diagram in which a detection area is set for an image signal input to an image detector according to the present invention.

(A), (b), (c), and (d) of FIG. 3 are filter values for resolving and reconstructing a frequency of a wavelet transformed video signal according to the present invention.

4 is a procedure for calculating coefficients by decomposing a frequency of a wavelet transformed video signal according to the present invention.

5 is a procedure for reconstructing a wavelet coefficient decomposed into an image contrast value according to the present invention.

6 (a) and 6 (b) are characteristic extraction results of contrast values for a case where a vehicle is present and a case where a vehicle does not exist in the daytime time zone detection region according to the present invention.

Figure 7 (a) (b) is a feature extraction result of the contrast value for the case where the vehicle is present in the detection area and the vehicle does not exist in the night time zone in accordance with the present invention.

According to an aspect of the present invention, there is provided a method for detecting a vehicle using an image, the method comprising: setting a detection area by dividing a predetermined distance after a stop line at equal intervals for the image information input to the image detection means; Converting the image signal of the set detection area into a discrete wavelet transform and converting the image signal into a frequency signal; Calculating a wavelet coefficient by decomposing the frequency-converted signal through a high frequency biorthogonal wavelet filter; Restoring the calculated wavelet coefficient to a contrast value between 0 and 255 using an inverse high frequency biorthogonal wavelet filter; And determining whether or not the vehicle exists in the corresponding detection area by comparing the number of pixels of the restored intensity value with a threshold value.

Hereinafter, exemplary embodiments of the present invention will be described in detail with reference to the accompanying drawings.

As can be seen in Figure 1, the traffic information detection apparatus according to the present invention is an image detection unit 10, the control unit 20, wavelet conversion unit 30, the first filter unit 40 and the second filter unit ( 50), the image detector 10 is located on a local space such as an intersection for detecting traffic information, and inputs an image of a vehicle and outputs a signal thereof.

The controller 20 performs a function of performing overall preprocessing on the input image. When the input of the image is detected, the control unit 20 sets the detection regions at equal intervals up to a predetermined distance after the stop line.

The wavelet converting unit 30 performs a function of converting an image signal into a frequency signal by performing wavelet transform on the image detected at equal intervals through the preprocessing.

The first filter part 40 is composed of a low frequency filter having a band as shown in FIG. 3A and a high frequency filter having a band as shown in FIG. 3B, and each of which receives a frequency signal applied through a wavelet transform. Decomposes the components.

The second filter unit 50 is composed of a low frequency filter having a band as shown in FIG. 3 (c) and a high frequency filter having a band as shown in FIG. 3 (d), and decomposed by the first filter unit 40. It is responsible for restoring each component to the image.

First, when an image signal input from an image detector 10 located on a local space such as an intersection is input to the controller 20 for detecting traffic information, the controller 20 performs preprocessing on the input image signal. As shown in FIG. 2, the detection area is set at a predetermined distance, up to about 120 m, and the like after the stop line according to a preset algorithm.

Subsequently, the wavelet converting unit converts the input image signal to detect whether the vehicle exists in the set detection area to measure the length of the vehicle waiting when the vehicle stops after the stop line while the stop signal is input from the intersection traffic light. Is applied to (30).

The wavelet converter 30 converts an image signal into a frequency domain so that the image signal can be divided into approximation information and detail information, and then applied to the first filter unit 40. ) Filter the frequency signal of the input image into a low frequency band as shown in (a) of FIG. 3 and a high frequency band as shown in (b) of FIG. Is authorized.

The second filter unit 50 restores the decomposed signal by filtering the low frequency band as shown in (c) of FIG. 3 and the high frequency band as shown in (d) of FIG. Extract the input image.

The operation of decomposing the wavelet transformed frequency by the first filter unit 40 will be described in more detail with reference to FIG. 4 as follows.

When the frequency X (i, j) of the wavelet transformed image signal is input to the first filter unit 40, the intensity of the input image has a band as shown in FIG. 3A in the direction of row. Multiply the low-frequency value h (u) and then multiply the calculated result by the low-frequency value h (u) with the band as shown in Fig. 3 (a) in the direction of the column (Column). _LL is calculated, and the intensity value of the input image is multiplied by the low frequency value h (u) having the band as shown in FIG. 3 (a) in the direction of the row, and then the calculated result value is again a column. By multiplying the high frequency value g (u) as shown in FIG. 3 (b) in the direction of, the coefficient value d _LH for the vertical component of the image is calculated.

In addition, the contrast value of the input image is multiplied by the value of the high frequency g (u) having the band as shown in FIG. 3 (b) in the direction of the row, and then the calculated result is plotted again in the direction of the column. Multiplying the low frequency values h (u) with the band as in (a) of 3, the coefficient value d _HL for the horizontal component of the image is calculated, and the intensity value of the input image is shown in the direction of row in FIG. Multiply the value of the high frequency g (u) with the band as (b) of and then multiply the calculated result by the value of the high frequency g (u) as shown in (b) of FIG. 3 in the direction of the column. The odd value d _HH of the diagonal component of the image is calculated.

In order to restore the original image to each coefficient value of the extracted image signal as described above, the restoration is performed by the procedure as shown in FIG. 5. In the present invention, the diagonal line of the image is extracted from various values extracted during the decomposition of the frequency. to a component of constant value d taken by only the value _HH and performs the following operation for the restoration of an image through the second filter unit 50.

This As can be seen in Figure 5, the diagonal components of the image station having a band, such as (d) of Figure 3 in a direction to open the high-frequency value of the filter value d _HH Inverse high frequency filter value having the band as shown in (b) of FIG. If you multiply by, it will be restored to image with contrast value between 0 and 255.

The different biorthogonal wavelet filters used when decomposing and restoring the frequency of the wavelet transformed video signal are as follows.

The linear vector space is ..., V ₁ , V ₀ , V _-1 , ...

When f (x, y), (1≤x≤X, 1≤y≤Y) is an image of the size of coordinate X × Y, f (x, y) in the vector space V ₀ is expressed by Equation 6 below. Is expressed.

Where? (?) And? (?) Are the scaling factors.

<F (x, y) · Φ (x-n) Φ (y-p)> = 0, <f (x, y) · Φ (x-n) ψ (y-p)> = 0

<F (x, y) · ψ (x-n) Φ (y-p)> = 0, <f (x, y) · ψ (x-n) ψ (y-p)> = 0

〈·〉 Means the inner product of the vector.

Here, a ₀ (n, p) corresponding to the approximate information is a coefficient obtained by down-sampling the image f (x, y) by multiplying the low frequency filters in the x and y directions, respectively, and b ₀ (n, corresponding to the detailed information). p) is a coefficient calculated by multiplying a high frequency filter in the x direction and a low frequency filter in the y direction.

C ₀ (n, p) is a coefficient obtained by multiplying a low frequency filter in the x direction and a high frequency filter in the y direction, and d ₀ (n, p) is a coefficient obtained by multiplying a high frequency filter in both the x and y directions.

When the diagonal component of the detailed information as described above is restored to an image, when there is a vehicle in the detection area in which the restored value is set as shown in FIG. 2, the number of pixels having a contrast value of 200 or more as shown in FIG. If there are more than and there are no vehicles, the number of pixels of 200 or more appears as shown in Fig. 6B.

As can be seen from (a) and (b) of FIG. 7, the same result can be extracted even at night time.

Especially, at night, the effect of noise from the headlights of the vehicle is very large in the general space, so the detection rate is often lowered to 85% or less in the normal state, but the wavelet transform is 95% as in the daytime. The above detection rate can be obtained.

As described above, the present invention is not sensitive to the image noise generated by the change of the weather by processing the image processing in the existing space in the frequency domain, and thus, an additional processing for removing the noise is eliminated. As the brightness changes for each time zone, the contrast value of the road should be estimated by the conventional method, but the present invention excludes this, thereby minimizing the detection error rate due to the error according to the estimation.

Claims (4)

In the vehicle detection method using an image, Setting a detection area by dividing a predetermined distance after the stop line at equal intervals with respect to the image information input to the image detection means; Converting the image signal of the set detection area into a discrete wavelet transform and converting the image signal into a frequency signal; Calculating a wavelet coefficient by decomposing the frequency-converted signal through a high frequency biorthogonal wavelet filter; Restoring the calculated wavelet coefficient to a contrast value between 0 and 255 using an inverse high frequency biorthogonal wavelet filter; And determining whether the vehicle exists in a corresponding detection area by comparing the number of pixels of the restored intensity value with a threshold value. delete The method according to claim 1, The count value of the video signal used to restore the contrast between 0 and 255 uses only the diagonal component count value d _HH of the decomposed wavelet coefficients. The method according to claim 1, If the number of pixels with a contrast value of 200 or more is greater than or equal to a threshold value, the vehicle occupies the corresponding detection area. If the number of pixels with a contrast value of 200 or more is less than or equal to a threshold value, the vehicle is determined not to occupy the detection area. Vehicle detection method using a wavelet transform, characterized in that.