JP3355027B2 - Target detection method - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

FIELD OF THE INVENTION This invention relates to a blue sky and clouds, such as the land is, to accurately detect a target object from an image with complex background that exists at the same time, it relates to target detection how.

[0002]

2. Description of the Related Art FIG.
FIG. 1 is a configuration diagram showing a target detection device of a conventional example shown in Japanese Patent Application Laid-Open Publication No. H10-207, in which 9 is an isolated point extraction filter, 901 is a background detection unit, 902 is a target detection unit, and 903 is a target determination unit. FIG. 22 is a diagram showing a shape of an isolated point extraction filter used in the conventional example, and FIG. 23 is a diagram showing isolated points to be extracted.

[0003] The isolated point extraction filter 9 has a shape as shown in FIG. 22, and a hatched portion in a size of 9 × 9 pixels is a background detection region 904, and a target detection region 905 in the background detection region 904. Has a size of 3 × 3 pixels. The target point 906 to be detected has a size of 1 × 1 pixel located at the center of the target detection area.

Next, the operation will be described. Background detector 9
01 calculates the average luminance level Ib of the background detection area 904, compares the luminance levels of all the pixels in the area with the average luminance level Ib, and determines that the difference from the average luminance level is equal to or more than the reference value a × N. Only when a certain pixel does not exist, it is detected as a valid background area and a logic “0” (hereinafter, described as “0”) is output. Here, N indicates a noise level. a
Is a parameter, and the parameter a is set according to the condition of the target to be extracted.

[0005] The target detection section 902 includes a background detection area 904.
Is compared with the luminance level of each pixel in the target detection area 905, and the average luminance level Ib is compared with the reference value b ×
A pixel having a luminance level exceeding the value obtained by adding N is detected, and a logic “1” (hereinafter, referred to as “1”) is output. Here, N is a noise level, b is a parameter, and the parameter b is set according to a target condition to be extracted.

The output of the target detector 903 is "0" and the output of the target detector 902 is "1".
In the case of, the point of interest 906 is determined to be an isolated point, and "1"
Is output.

For example, at the isolated point 907 shown in FIG. 23, all the pixels in the background detection area 904 have a luminance difference from the average luminance level Ib of the background detection area 904 having a reference value a × N.
, The background detection area 904 has a predetermined uniform brightness level. In the target detection area 905, there is a pixel having a luminance level exceeding a value obtained by adding the reference value b × N to the average luminance level Ib, and this pixel is extracted as an isolated point.

As described above, the isolated point extraction filter 9 extracts an isolated point having a luminance level exceeding a predetermined reference of the target detection area from a background detection area having a predetermined uniform luminance level. 1 "is output. Also, 9 × 9
Output data is output while sequentially scanning from the upper left to the lower right for the pixel target area.

[0009]

Since the conventional target detecting device is configured as described above, the target can be detected only when the background region has a predetermined uniformity. In the case of detection, the target can be detected if the background is sunny, but if the cloud exists in the background, the target cannot be detected because the background area does not have a predetermined uniformity.

If the luminance level of the target (isolated point) is sufficiently higher than the luminance level of the background, the target can be detected. On the contrary, for example, the frequency distribution of the luminance of the background and the frequency distribution of the target luminance are one. When the luminance level of the target is not sufficiently higher than the luminance level of the background, as in the case of overlapping, there is a problem that the target cannot be detected or the background is erroneously detected.

Further, in determining a reference value for evaluating the uniformity of the background and a reference value for detecting the target, a priori knowledge such as a background state and a luminance difference between the target and the background is required. There were also problems.

The present invention solves the above problems.
Was made in order, the target luminance for the case or background near the target pixel and the target pixel around the brightness variation is large clouds and land by examining the luminance information in more detail in the background background is sufficiently large It is an object of the present invention to obtain a target detection method capable of detecting a target with higher accuracy even when it cannot be said.

Further , the present invention uses the standard deviation of the luminance of the background as a measure of the luminance variation, thereby more accurately evaluating the variation of the background luminance. An object is to obtain a target detection method capable of detecting a target.

Furthermore, this invention aims to obtain a target detection method do not require a priori knowledge, such as the luminance difference between the state and the target and the background in the background.

[0015]

According to the present invention, there is provided a method for detecting a target, comprising the steps of:
Second, third, and fourth local pixel regions are set, and the first local pixel region surrounds the target pixel region around the target pixel region and is equal to or smaller than the expected size of the target to be detected. A target local pixel area having a size of
The second and third local pixel regions are located around the target pixel region around the target pixel region and are target local pixel regions having a size equal to or smaller than the expected size of the target to be detected. The fourth local pixel region is a background local pixel region surrounding the target pixel region around the target pixel region and having a size larger than the expected size of the target to be detected. A plurality of primary primary features are calculated by calculation using the luminance of the pixel area, the luminance of the target local pixel area, and the luminance of the background local pixel area.

According to the target detecting method of the present invention, in a background local pixel region, a sensor for obtaining at least any two luminances among a maximum luminance, an average luminance, and a minimum luminance and obtaining a digital image from these luminances. The standard deviation of the original luminance of the background local pixel region from which the noise component has been removed is estimated, and the primary primary feature is calculated by using at least the estimation result, the luminance of the pixel region of interest, and the luminance of one of the background local pixel regions. Calculate the amount.

In the target detection method according to the present invention, at the time of determining a target, at least one primary characteristic amount is calculated for each pixel of at least one digital image known in advance as a background, and at least one primary characteristic amount is calculated. The primary feature distribution is generated by mapping the primary feature in at least a one-dimensional space having as one axis, and the statistic of the primary feature distribution is calculated. Calculating a distance between a position in a space of at least one primary feature amount calculated for a target pixel area including a certain or unknown pixel and a distribution center position of a primary feature amount distribution determined by a statistic; When the distance is larger than a predetermined distance threshold, it is determined that the target pixel area includes the target.

[0018]

In the target detection method according to the present invention, a plurality of primary primary features are calculated by calculation using at least the luminance of the target pixel area, the luminance of the target local pixel area, and the luminance of the background local pixel area. Specifically, the luminance data of the target local pixel region, the luminance data of the background local pixel region, and the luminance data of the target pixel region are input, and the amount representing the difference in the sharpness and the shape of the luminance peak around the target pixel is expanded. Then, an amount representing the variation in the luminance of the background is calculated, and the difference in luminance between the target pixel region and the background and the difference in the sharpness and spread of the luminance peak are normalized based on the variation. Therefore, by combining the target detection method and the primary feature value according to Claims 2 and 3 and Claim 4, it is considered that the target luminance is sufficiently large in the case where a cloud or land having a large luminance variation is a background or the background is large. The target can be detected even when it cannot be said.

According to the target detecting method of the present invention, in a background local pixel area, at least any two of the maximum luminance, the average luminance, and the minimum luminance are obtained, and a sensor for obtaining a digital image from these luminances is obtained. A standard deviation of the original luminance of the background local pixel region from which the noise component has been removed is estimated, and a primary primary feature amount is calculated by using at least the estimation result, the luminance of the pixel region of interest, and the luminance of one of the background local pixel regions. Is calculated. That is, the luminance data of the background local pixel region and the luminance data of the pixel region of interest are input, the standard deviation of the luminance of the background is calculated, and the luminance difference between the pixel region of interest and the background is normalized by the standard deviation. The luminance information with a small amount of calculation such as the maximum value and the average value of the background is extracted from the background local pixel area, and the standard deviation of the luminance of the background is estimated based on the luminance information, and this is used as an amount representing the variation of the luminance of the background. By evaluating the primary feature amount obtained by normalizing the luminance difference between the pixel area of interest and the background with the variation in the luminance of the background, the target can be detected even when the background is a cloud or land having a large variation in the luminance.

In the target detecting method according to the present invention, at the time of determining a target, at least one primary characteristic amount is calculated for each pixel of at least one digital image known in advance as a background, and at least one primary characteristic amount is calculated. A primary feature distribution is generated by mapping a primary feature in at least a one-dimensional space defined as one axis, and a statistic of the primary feature distribution is calculated to be a background or a target. Or the distance between the position in the space of at least one primary feature amount calculated for the pixel region of interest including the unknown pixel and the distribution center position of the primary feature amount distribution determined by the statistic is calculated. If the distance is larger than the predetermined distance threshold, it is determined that the target pixel area includes the target. That is, after mapping the primary features as calculated in claims 2 to 7 into a one-dimensional or more multidimensional space,
A threshold for target determination is automatically determined using the statistics of the distribution. Therefore, it is not necessary to have a priori knowledge of the state of the background or the luminance difference between the target and the background.

[0021]

EXAMPLES Reference Example 1 FIG. Hereinafter, a reference example of the present invention will be described with reference to the drawings. FIG.
FIG. 5 is a flowchart showing the flow of processing of a target detection method according to Embodiment 1 of the present invention. In the figure, step ST1 is an image input processing, step ST2 is a primary feature amount calculation processing, and step ST3 is a target determination processing.

Further, FIG. 2 is a diagram showing a luminance distribution of the target in this Example 1, 10 is a luminance distribution of the target in FIG.

[0023] Next, the operation will be described. First, in step ST1, image data of Ix pixels in the horizontal direction and Iy pixels in the vertical direction captured by a visible or infrared TV camera are input.

At this time, it is assumed that the background of the input image includes a blue sky having a relatively small variation in luminance and a flat luminance distribution, and a cloud or land having a relatively large variation in luminance. However, it is assumed that the swell of the brightness of the clouds and the land is relatively gentle (the brightness peak is gentle). FIG.
As shown in the figure, the size of the target to be extracted
x pixels or less than Ty pixels in the vertical direction, and the luminance distribution is a two-dimensional Gaussian distribution (the luminance of the central pixel is the largest,
As the distance from the center increases, the luminance decreases while drawing a curve represented by a two-dimensional Gaussian distribution equation). Also, consider a case where the target sizes Tx and Ty are small and the central luminance has a sharp peak.

In the primary feature amount calculation processing of step ST2, the vicinity of a point of interest having arbitrary coordinates (x, y) of the image data (target pixel area: including at least one target pixel)
And pixel data of a target local pixel area (a local area set within a target size range to be detected centering on the target pixel is referred to as a target local pixel area). Then, whether or not the target point has a local luminance peak with respect to the periphery is evaluated (step ST21), and at the same time, a primary primary feature value representing the shape of the target point is calculated (step ST22). To step ST24). In the reference example 1 , as a primary feature amount representing a shape, an S / N ratio (S / N ratio) calculated with reference to a target point and a background local pixel region at a distance that is equal to or larger than a target size assumed therefrom is calculated. Step ST22), the sharpness of the luminance peak calculated with reference to the target point and the range of the target size assumed thereabout (step ST23), the spread as the shape to be calculated with reference to the target point and its surroundings Estimation amount of whether one is two-dimensional or one-dimensional (step ST24)
Ask for.

In response to the result, in the target determination process in step ST3, it is determined whether or not a plurality of primary feature values, which are the result of the primary feature value calculation process in step ST2, satisfy a predetermined target likelihood. . Then, when it is determined that the attention point satisfies the target likelihood, the determination result is output in the form of a signal or the like.

[0027] Thus, the luminance variation and the background by the way the primary feature quantity calculation processing, sharpness and brightness peak around target point, by calculating a quantity representing the difference in spread how the shape, by the target determination process Since it is determined whether or not these primary feature values are likely to be a target, a target can be detected even when a cloud or land having a large variation in luminance is a background or when the luminance of the target is not sufficiently large with respect to the background.

[0028] In this reference example 1, but running all the steps ST24 from step ST21, performing any one or more steps of these steps as described in the following Reference Example <br/> You may do so.

[0029] Reference Example 2. FIG. 3 is a flowchart showing the flow of a process of a target detection method according to Embodiment 2 of the present invention. In the figure, step ST1 is an image input process, step ST2 is a primary feature amount calculation process, and step ST3 is a target determination process. is there. Step ST201 is a point-of-interest extraction process.
T202: target local pixel area A extraction processing, step ST
203 is target feature information A extraction processing, step ST204
Is a feature information calculation process. In the second embodiment, a plurality of processes (steps ST201 to ST203) are described in parallel as shown in FIG. 3, but the present invention is not limited to this, and the processes may be performed in series.

Further, FIG. 4 is a diagram showing an example of a local pixel region near the point of interest in the reference example 2, in the drawing,
11 is an assumed target size, 12 is a target pixel area, 1
3 is a target local pixel area A.

[0031] Next, the operation will be described. The assumed target luminance, shape, size, and background state are the same as in the first embodiment . In step ST202, as shown in FIG. 4, the position of the target local pixel area A13 is set so that its center coincides with the attention point coordinates (x, y), and its size M
Atx and MAty are almost the same or slightly smaller than the target sizes Tx and Ty, and their widths WAtx and WAty are 1 or more, and the sum of the sizes MAtx and MAty (MAtx + 2)
WAtx, MAty + 2WAty) is the target size T
Set so as not to exceed x and Ty.

On the other hand, the image data input process in step ST1, and target point extracting process of extracting the pixel data of the point of interest from the input image data in step ST 201 (x, y) is performed in the same manner as in Reference Example 1 . Step ST20
The pixel data of the target local pixel area A read in the target local pixel area A extraction processing of Step 2 is the next step ST203.
In the target characteristic information A extraction processing, for example, an average value of the pixel data of the target local pixel area A is calculated and extracted as the characteristic information.

In the characteristic information calculation processing in step ST204, the characteristic information of the target local pixel area A, which is the result of the target characteristic information A extraction processing in step ST203, and step ST204.
Using the pixel data of the target point, which is the result of the target point extraction processing of 201, a primary feature amount is calculated according to, for example, the following equation (1). Here, G is the pixel data of the target point, and Bamean is the characteristic information (average value) obtained from the target local pixel area A.

[0034] The primary feature quantity = (G-BAmean) ÷ G (1)

In general, targets have sharp luminance peaks. On the other hand, a background such as a cloud often has a gentle peak. Considering the difference in the sharpness of the peak with respect to the primary feature amount calculated according to the above equation (1), the difference between the pixel data of the point of interest and the average value of the target local pixel area A in the numerator of the primary feature amount If the center has a sharp luminance peak like the target, the luminance value drops sharply as the distance from the center increases, so it becomes a large positive value, and if it has a gentle peak like a cloud, the point of interest This means that the peripheral brightness decreases gradually and becomes a small positive value. By the way, when the center has no luminance peak, it takes a negative value. Note that G in the denominator of the primary feature value is a term for normalizing the sharpness of the luminance peak expressed by the numerator in order to obtain the same primary feature value even when the contrast of the input image changes. Therefore, the sharpness of the luminance peak can be evaluated based on the magnitude of the primary feature amount. The primary feature amount (sharpness of the luminance peak) calculated in the feature information calculation process in step ST204 is passed to the next target determination process in step ST3.

In the target determination process in step ST3, the threshold value set in advance is compared with the primary feature amount (sharpness of the luminance peak) calculated in step ST2. Is determined, and the coordinates (x, y) of the target point are output together with a signal or symbol indicating the target.

Further, the series of processing is performed for all the pixels or pre given limited area of the pixels of the image while sequentially changing point of interest coordinates from the upper left of the image to the bottom right.

In the second embodiment , only one pixel is used as the pixel data of the target point. However, depending on the size of the target, data obtained by averaging several pixels near the target point may be used. In the target determination processing in step ST3, a signal or symbol indicating the target and the coordinates of the target point are output as a determination result. However, a signal indicating whether the target or the background is present, for example, logic "1" for the target. "," 0 "may be output for the background.

Reference Example 3 FIG. 5 is a flowchart showing the flow of processing of a target detection method according to Embodiment 3 of the present invention. In the figure, step ST1 is an image input processing, step ST2 is a primary feature amount calculation processing, and step ST3 is a target determination processing. is there. Step ST201 is a point-of-interest extraction process.
T205: target local pixel area B extraction processing, step ST
207, a target local pixel area C extraction process, step ST2
06 is target characteristic information B extraction processing, step ST208 is target characteristic information C extraction processing, and step ST204 is characteristic information calculation processing. Although a plurality of processes are described in parallel in the reference example 3 shown in this figure, the present invention is not limited to this, and these processes may be performed in series.

Further, FIG. 6 is a diagram showing an example of a local pixel region in this reference example 3, reference numeral 11 is the magnitude of the target envisaged, 12 noticed pixel region, 14 targeted local pixel region B, Reference numeral 15 denotes a target local pixel area C.

[0041] Next, the operation will be described. The assumed target luminance, shape, size, and background state are the same as in the first embodiment . The position of the target local pixel area B is determined by the attention point coordinates (x,
y), that is, the coordinates (x-1, y) and the coordinates (x +
1, y), each of which is one pixel.
The position of the target local pixel area C is above and below the coordinates (x, y) of the target point, that is, coordinates (x, y-1) and coordinates (x, y + 1).
And the size is one pixel.

The image data input process in step ST1, and the point of interest from the input image data (x, y) of interest point extraction process of step ST201 for extracting pixel data of
This is performed in the same manner as in Reference Example 1. The pixel data of the target local pixel area B read out in the target local pixel area B extraction processing in step ST205 is used as the characteristic information in the next target characteristic pixel B extraction processing in step ST206, for example, the pixel of the target local pixel area B. The average of the data is calculated and extracted. Similarly, the target local pixel area C read in the target local pixel area C extraction processing in step ST207
In the target characteristic information C extraction processing in the next step ST208, the pixel data of is extracted by calculating, for example, the average value of the pixel data of the target local pixel region C as the characteristic information.

In the characteristic information calculation processing in step ST204, the target local pixel area B obtained in step ST206
Using the characteristic information of the target local pixel area C obtained in step ST208 and the pixel data of the target point obtained in step ST201, for example, the following equation (2) is used.
Is calculated according to the following. Here, G is pixel data of a point of interest, BBmean is feature information (average value) obtained from the target local pixel area B, and BCmean is feature information (average value) obtained from the target local pixel area C. It is.

[0044] The primary feature quantity = (BBmean-BCmean) ÷ G (2)

The target handled in this reference example has a two-dimensional spread. On the other hand, when considering noise, an input image is input to a system as a one-dimensional signal after being obtained by a two-dimensional sensor, and noise is often added to the one-dimensional signal. That is, the target has a two-dimensional spread, and the noise has a one-dimensional spread. Considering this difference in spread with respect to the primary feature value calculated from the above equation (2), in the numerator of the primary feature value, the difference between the average value on the left and right of the attention point and the average value on the top and bottom of the attention point is However, when the target has a two-dimensional spread, the numerator takes a small value because the brightness difference between the left and right and the top and bottom is small. Means that the numerator takes a large value because the luminance difference is large. The denominator G of the primary feature is a term for normalizing the difference in the spread expressed by the numerator. Therefore, it is possible to evaluate the difference in the spread around the point of interest based on the magnitude of the primary feature amount. The primary feature value (difference in spread) calculated in the feature information calculation process in step ST204 is passed to the target determination process in the next step ST3.

In the target determination process of step ST3, the threshold value set in advance is compared with the primary feature value (difference in spread) calculated in step ST204. If the primary feature value is smaller than the threshold value, the point of interest is the target. It is determined that there is, and the coordinates (x, y) of this point of interest are output together with a signal or symbol indicating that it is the target.

Further , these series of processes are performed on all pixels of the image or pixels of a limited area given in advance while sequentially changing the coordinates of the point of interest from the upper left to the lower right of the image.

In the third embodiment , only one pixel is used as the pixel data of the target point. However, depending on the size of the target, data obtained by averaging several pixels near the target point may be used. In the target determination process of step ST3, a signal indicating the target and the coordinates of the target point are output as the determination result. However, a signal indicating whether the target is the background or the background, for example, a logic "1" for the target, Logic "0" for background
May be output.

Reference Example 4 FIG. 7 is a flowchart showing the flow of processing of a target detection method according to Embodiment 4 of the present invention. In the figure, step ST1 is an image input processing, step ST2 is a primary feature amount calculation processing, and step ST3 is a target determination processing. is there. Step ST201 is a point-of-interest extraction process.
T209: background local pixel area D extraction processing, step ST
210: background feature information D extraction processing, step ST204
Is a feature information calculation process. In the fourth embodiment, a plurality of processes are described in parallel as shown in the figure. However, the present invention is not limited to this, and these processes may be executed in series.

[0050] FIG. 8 is a diagram showing an example of a local pixel region in this reference example 4, reference numeral 11 is the magnitude of the target envisaged, 12 noticed pixel region, 16 in the background local pixel region D is there.

[0051] Next, the operation will be described. The assumed target luminance, shape, size, and background state are the same as in the first embodiment . In step ST209, as shown in FIG. 8, the position of the background local pixel region D is set so that its center matches the attention point coordinates (x, y), and its size MDb
x and MDby are sufficiently larger than the target sizes Tx and Ty, and their widths WDbx and WDby are set to one or more appropriate sizes.

[0052] Image data input process in step ST1, and the point of interest from the input image data (x, y) of interest point extraction process of step ST201 for extracting pixel data of
This is performed in the same manner as in Reference Example 1. The pixel data of the background local pixel area D read out in the background local pixel area D extraction processing in step ST209 is used as the characteristic information in the background characteristic information D extraction processing in the next step ST210. Calculate and extract the average and maximum values of the data.

In the characteristic information calculation processing in step ST204, the background local pixel area D obtained in step ST210
, And the pixel data of the target point obtained in step ST201, only when the pixel data of the target point is larger than the maximum value of the background local pixel region D, for example, according to the following equation (3). Calculate the amount. If the pixel data of the target point is smaller than the maximum value of the background local pixel area D, the primary feature amount is set to a very small value (for example, 0). Here, G is pixel data of a point of interest, and BDmax
Is one of feature information obtained from the background local pixel region D (maximum value), and BDmean is the other feature information (average value) obtained from the background local pixel region D.

[0054] The primary feature quantity = (G-BDmax) ÷ ( BDmax -BDmean) (3)

The difference between the pixel data at the point of interest and the maximum value of the background local pixel region D in the numerator of the primary feature value is a positive value when the center has a luminance peak as in the target. The magnitude depends on the relative difference between the center luminance and the background luminance. Conversely, if the center does not have a luminance peak, it means a negative value. Further, in the denominator of the primary feature amount, the maximum value and the average value of the background local pixel region D are different, and the magnitude of the difference indicates the magnitude of the swell of the luminance of the background. Therefore, the primary feature amount can be compared to the magnitude of the signal in the numerator and the magnitude of the noise in the denominator, and the result indicates the S / N ratio. Step ST2
04 primary feature amount (S /
N ratio) is passed to the target determination processing in the next step ST3.

In the target determination process in step ST3, the threshold value set in advance is compared with the primary characteristic amount (S / N ratio) calculated in step ST204. , And the coordinates (x,
y) is output.

Further , these series of processes are performed on all pixels of the image or pixels of a predetermined limited area while sequentially changing the point-of-interest coordinates from the upper left to the lower right of the image.

In the fourth embodiment , the maximum value and the average value are used as the characteristic information of the background local pixel region D. However, the same effect can be expected by using the maximum value and the minimum value.
If it is considered that the maximum value or the minimum value is vulnerable to noise and shows a sudden singular value, when the pixel data of the background local pixel area D are arranged in descending order, the second largest and the third smallest are the two. The third or third may be used, or the average of several from the larger and the average of several from the smaller may be used. In the fourth embodiment , only one pixel is used as the pixel data of the target point. However, depending on the size of the target, data obtained by averaging several pixels near the target point may be used. In the target determination process in step ST3, a signal indicating the target and the coordinates of the target point are output as the determination result.
For example, a logic “1” may be output for a target, and a logic “0” may be output for a background.

Reference Example 5 FIG. 9 is a flowchart showing the flow of processing of a target detection method according to Embodiment 5 of the present invention. In the figure, step ST1 is an image input processing, step ST2 is a primary feature amount calculation processing, and step ST3 is a target determination processing. is there. Step ST201 is a point-of-interest extraction process.
T209: background local pixel area D extraction processing, step ST
210: background feature information D extraction processing, step ST202
Is a target local pixel region A extraction process, step ST203 is a target feature information A extraction process, and step ST204 is a feature information calculation process. In the fifth embodiment , as shown in the drawing, a plurality of processes are described in parallel. However, the present invention is not limited to this, and these processes may be executed in series.

[0060] Further, FIG. 10 is a diagram showing an example of a local pixel region in this Reference Example 5, in FIG, 11 is the magnitude of the target envisaged, 12 noticed pixel region, 16 background local pixel region D, 13 is a target local pixel area A.

[0061] Next, the operation will be described. The assumed target luminance, shape, size, and background state are the same as in the first embodiment . In step ST209, the position of the background local pixel region D is set so that the center thereof coincides with the attention point coordinates (x, y), and the sizes MDbx and MDby are sufficiently larger than the target sizes Tx and Ty. Width WDbx, W
Dby is set to one or more appropriate sizes. In step ST202, the position of the target local pixel area A is set so that its center coincides with the attention point coordinates (x, y), and its size MAtx and MAty are set to the target size T.
x, Ty are almost the same or slightly smaller, and their widths WAtx, W
Aty is set to 1 or more so that the sum of the sizes MAtx and MAty does not exceed the target sizes Tx and Ty.

[0062] Image data input process in step ST1, and the point of interest from the input image data (x, y) of interest point extraction process of step ST201 for extracting pixel data of
This is performed in the same manner as in Reference Example 1. The pixel data of the background local pixel area D read out in the background local pixel area D extraction processing in step ST209 is used as the characteristic information in the background characteristic information D extraction processing in the next step ST210. Calculate and extract the average and maximum values of the data. Similarly, the pixel data of the target local pixel area A read in the target local pixel area A extraction processing in step ST202 is used as the target characteristic information A extraction processing in the next step ST203, for example, as the target local pixel area A. Calculate the average value of the pixel data,
Extract.

In the characteristic information calculation processing of step ST204, the background local pixel area D obtained in step ST209 is obtained.
Using the characteristic information of the target local pixel area A obtained in step ST202 and the pixel data of the target point obtained in step ST201, the pixel data of the target point is calculated from the maximum value of the background local pixel area D. Only when it is large, for example, two primary feature amounts are calculated according to the following equations (4) and (5). If the pixel data of the target point is smaller than the maximum value of the background local pixel area D, the two primary feature amounts are both set to very small values (for example, 0). Here, G is pixel data of a point of interest, BDmax is one (maximum value) of feature information obtained from the background local pixel area D,
Dmean is the other characteristic information (average value) obtained from the background local pixel region D, and Bmean is the characteristic information (average value) obtained from the target local pixel region.

[0064] The primary feature quantity 1 = (G-BDmax) ÷ (BDmax -BDmean) (4)

[0065] The primary feature quantity 2 = (G-BAmean) ÷ (BDmax -BDmean) (5)

The primary characteristic amount 1 is the S / N ratio described in the reference example 4, and the primary characteristic amount 2 is obtained by modifying the sharpness of the luminance peak described in the reference example 2 (the normalization term is changed from the luminance of the target point to the background undulation). Has been changed to the size of). Primary feature amount 1 (S / N ratio) and primary feature amount 2 (sharpness of luminance peak) calculated in the feature information calculation processing in step ST204.
Is passed to the target determination process in the next step ST3.

In the target determination process in step ST3, the primary feature 1 (S / N ratio) and primary feature 2 (sharpness of the luminance peak) are compared with respective preset thresholds, and the primary feature 1 and the primary feature 1 are compared. If both the primary feature values 2 are larger than the respective thresholds, the point of interest is determined to be the target, and the coordinates (x,
y) is output.

[0068] Also, the series of processing is performed for all the pixels or pre given limited area of the pixels of the image while sequentially changing point of interest coordinates from the upper left of the image to the bottom right.

Although the maximum value and the average value are used as the characteristic information of the background local pixel region in the reference example 5 , almost the same effect can be expected by using the maximum value and the minimum value. Considering that the value or the minimum value is vulnerable to noise and shows a sudden singular value, the pixel data of the background local pixel area is arranged in the descending order, the second from the largest and the second from the third and the second from the smallest. The third may be used, or the average of several pieces from the larger one and the average of several pieces from the smaller one may be used. Further, in this reference example 5 , only one pixel is used as the pixel data of the target point. However, depending on the size of the target, data obtained by averaging several pixels near the target point may be used.

In the fifth embodiment , independent threshold values are set for the two primary feature values. However, for example, the two feature values are associated with each other and a function given by the following equation (6) is used. A threshold may be set. Here, a and b are constants.

[0071] The primary feature quantity 2 threshold = a × (primary characteristic amount 1) + b (6)

In the fifth embodiment , the two primary feature values are described. For example, the three primary feature values are evaluated using the primary feature value for evaluating the difference in the spread around the point of interest shown in the third embodiment. The determination may be made in combination. In the target determination process of step ST3, a signal indicating the target and the coordinates of the target point are output as the determination result. However, a signal indicating whether the target is the background or the background, for example, a logic "1" for the target, For a background, a logic “0” may be output.

Reference Example 6 FIG. 11 is a flowchart showing the flow of processing of a target detection method according to Embodiment 6 of the present invention. In the figure, step ST1 is an image input processing, step ST2 is a primary feature amount calculation processing, and step ST3 is a target determination processing. is there.
Step ST201 is a point-of-interest extraction process, step ST209 is a background local pixel region D extraction process, step ST209.
T210 is background characteristic information D extraction processing, step ST21
1 is a standard deviation calculation process, and step ST204 is a feature information calculation process. In the sixth embodiment , as shown in the drawing, a plurality of processes are described in parallel. However, the present invention is not limited to this, and these processes may be executed in series.

[0074] FIG. 8 is a diagram showing an example of a local pixel region in this Example 6, reference numeral 11 is the magnitude of the target envisaged, 12 noticed pixel region, 16 in the background local pixel region D is there.

[0075] Next, the operation will be described. As shown in FIG. 8, the target brightness, shape, size, and background state assumed in step ST209 are the same as those in the first embodiment . As for the position of the background local pixel area D, the center thereof is the attention point coordinate (x,
y) and the size MDbx, M
Dby is sufficiently larger than the target sizes Tx and Ty, and the widths WDbx and WDby are set to one or more appropriate sizes.

The image data input processing in step ST1 and the attention point extraction processing in step ST201 for extracting the pixel data of the attention point (x, y) from the input image data are as follows.
This is performed in the same manner as in Reference Example 1. The pixel data of the background local pixel area D read out in the background local pixel area D extraction processing in step ST209 is used as the characteristic information in the background characteristic information D extraction processing in the next step ST210, for example, the pixel data of the background local pixel area D. The average value and the maximum value of are calculated and extracted. In the standard deviation calculation process in step ST211, the standard deviation is calculated using the average value and the maximum value obtained in the background feature information D extraction process in step ST210, for example, according to the following equation (7). Here, BDmax is the maximum value of the luminance obtained from the background local pixel area D, and BDmean is the average value. M is a constant determined by the number of pixels included in the background local pixel region, and σn is a fixed value known in advance as the standard deviation of sensor system noise.

[0077] The standard deviation σc = √ [{(BDmax -BDmean ) ÷ M} 2 -σn 2] (7)

In the characteristic information calculation processing in step ST204, the background local pixel area D obtained in step ST210
Using the standard deviation σc obtained in step ST211 and the pixel data of the target point obtained in step ST201, only when the pixel data of the target point is larger than the maximum value of the background local pixel area D. For example, the primary feature amount is calculated according to the following equation (8). If the pixel data of the target point is smaller than the maximum value of the background local pixel area D, the primary feature amount is set to a very small value (for example, 0).
Here, G is the pixel data of the point of interest, BDmax is the maximum value which is one of the characteristic information obtained from the background local pixel area D,
σc is a standard deviation calculated from the background local pixel area D.

[0079] The primary feature quantity = (G-BDmax) ÷ σc (8)

[0080] The primary feature quantity is obtained by variation of the standard deviation normalization section of the primary feature quantity (S / N ratio) shown in Reference Example 4 from simple brightness difference, meaning of Reference Example 4 in which this has Same as the ones. The primary feature amount (S / N ratio) calculated in the feature information calculation process in step ST204 is passed to the next target determination process in step ST3.

In the target determination process in step ST3, a preset threshold value is compared with a primary characteristic amount (S / N ratio),
If the primary feature value is larger than the threshold value, the point of interest is determined to be the target, and the coordinates (x, y) of the point of interest are output together with a signal or symbol indicating that the point is the target.

Further , these series of processes are performed on all the pixels of the image or the pixels of a predetermined limited area while sequentially changing the coordinates of the point of interest from the upper left to the lower right of the image.

In the sixth embodiment , the maximum value and the average value are used as the characteristic information of the background local pixel region D. However, when the maximum value and the minimum value are used, the expression shown in the equation (8) is used. By setting BDmax-BDmean to BDmax-BDmin and setting the constant M to 2 x M, almost the same effect can be expected. Also, consider that the maximum value and the minimum value are vulnerable to noise and exhibit sudden singular values. Then, when the pixel data of the background local pixel region D are arranged in the descending order, the second or third from the largest and the third from the smallest may be used, or the average of several from the largest and the small Even if several averages are used, the constant M in the equation (8) may be changed accordingly. Although only one pixel is used as the pixel data of the point of interest in the sixth embodiment , data obtained by averaging several pixels near the point of interest may be used depending on the size of the target. In the target determination process of step ST3, a signal indicating the target and the coordinates of the target point are output as the determination result. However, a signal indicating whether the target is the background or the background, for example, a logic "1" for the target, Logic if background ”
0 "may be output.

[0084] Example 1. FIG. 12 is a flowchart showing the flow of the process of the target detection method according to the first embodiment of the present invention. In the figure, step ST1 is an image input process, step ST2 is a primary feature amount calculation process, and step ST3 is a target determination process. is there.
Step ST201 is a point-of-interest extraction process, step ST209 is a background local pixel region D extraction process, step ST209.
T202: target local pixel area A extraction processing, step ST
205 is a target local pixel area B extraction process, step ST2
07 is a target local pixel area C extraction process, step ST21.
0 is background characteristic information D extraction processing, step ST203 is target characteristic information A extraction processing, step ST206 is target characteristic information B extraction processing, and step ST208 is target characteristic information C.
Extraction process, step ST211 is standard deviation calculation process, step ST212 is feature information calculation process α, step ST212.
213 is a feature information calculation process β, and step ST4 is a region determination process. In the first embodiment, as shown in the drawing, a plurality of processes are described in parallel. However, the present invention is not limited to this, and these processes may be executed in series. Although a plurality of processes are described in parallel in the figure, this does not specify parallel processing.

[0085] Further, FIGS. 4, 6, 8 is a diagram showing an example of a local pixel region in the first embodiment, in FIG,
11 is an assumed target size, 12 is a target pixel area, 1
6 is a background local pixel area D, 13 is a target local pixel area A,
14 is a target local pixel area B, 15 is a target local pixel area C
It is.

[0086] Next, the operation will be described. The assumed target luminance, shape, size, and background state are the same as in the first embodiment . The image data input process of step ST1 and the attention point extraction process of step ST201 for extracting pixel data of the attention point (x, y) from the input image data
This is performed in the same manner as in Reference Example 1. In the setting process of the background local pixel region D, the feature information is extracted from the background local pixel region D in the background feature information D extraction process in step ST210, and the feature information is extracted from the background local pixel region D in step ST211.
The processing is passed to feature information calculation processing in T212, and step ST21 is performed.
The calculation of the standard deviation of the luminance of the background in the standard deviation calculation process of 1 is the same as that of the sixth embodiment . The setting of the target local pixel areas A to C and the extraction of the characteristic information of the target extraction processing A to C in the target characteristic information A to C extraction processing are the same as in Reference Example 2 and Reference Example 3.

[0087] characteristic information calculation processing in step ST212 alpha
Then, the feature information of the background local pixel region D and the step ST2
Using the standard deviation σc of the background luminance calculated in the standard deviation calculation processing of No. 11, the characteristic information of the target local pixel area A obtained in step ST203, and the pixel data of the target point obtained in step ST201, Only when the pixel data is larger than the maximum value of the background local pixel area D, two primary feature amounts are calculated according to, for example, the following equations (9) and (10). If the pixel data of the target point is smaller than the maximum value of the background local pixel area D, the two primary feature amounts are both set to very small values (for example, 0). Here, G is the pixel data of the point of interest, BDmax is one of the feature information (maximum value) obtained from the background local pixel region D, and Bamean is one of the feature information (average value) obtained from the target local pixel region A. value),
σc is a standard deviation calculated from the background local pixel area D.

[0088] The primary feature quantity 1 = (G-BDmax) ÷ σc (9)

[0089] The primary feature quantity 2 = (G-BAmean) ÷ σc (10)

[0090] normal primary feature amount shown primary characteristic quantity 1 in Reference Example 6 (S / N ratio), the primary feature quantity 2 is primary feature amount shown in Reference Example 5 and Reference Example 2 (sharpness of the luminance peak) This is obtained by changing the chemical term into the standard deviation of the luminance of the background, and has the same meaning as in the above embodiment. Step ST21
The primary feature amount (S /
N ratio and sharpness of the luminance peak) are passed to the target determination processing in the next step ST3.

[0091] The feature information calculation processing of step ST213 β
Then, the characteristic information of the background local pixel area D obtained in step ST210, the standard deviation σc of the background luminance calculated in the standard deviation calculation processing in step ST211 and the step ST2
06 and the characteristic information of the target local pixel areas B and C obtained in ST208 and the pixel data of the target point obtained in step ST201, the pixel data of the target point is larger than the maximum value of the background local pixel area D. Only in this case, for example, two primary feature amounts are calculated according to the following equations (11) and (12). If the pixel data of the target point is smaller than the maximum value of the background local pixel area D, the two primary feature amounts are both set to very small values (for example, 0). Here, G is pixel data of a point of interest, BDmax is one (maximum value) of feature information obtained from the background local pixel area D, BBmean, BC
mean is one (average value) of feature information obtained from the target local pixel regions B and C, and σc is a standard deviation calculated from the background local pixel region D.

[0092] The primary feature quantity 1 = (G-BDmax) ÷ σc (11)

[0093] The primary feature quantity 2 = (BBmean-BCmean) ÷ σc (12)

[0094] The primary feature quantity indicated primary characteristic quantity 1 in Reference Example 6 (S / N ratio), the primary feature quantity 2 is normalized primary feature amount shown in Reference Example 5 and Reference Example 3 (the difference of the spread) The term is changed to the standard deviation of the background luminance, and it has the same meaning. Feature information calculation processing β in step ST213
The primary feature amounts (differences in S / N ratio and spread) calculated in step (1) are passed to the target determination processing in the next step ST3.

In the area determination processing in step ST4, the standard deviation σc of the background luminance is compared with a fixed value σn which is known in advance from the standard deviation of the intrinsic noise of the sensor system.
If it is smaller than n, for example, a logic "0" is output. Conversely, if σc is larger than σn, a logic "1" is output.
This is because, when noise is added to an area having an inherently constant luminance, such as a blue sky, σn is larger and a logical “0” is output. In addition to the undulation and the addition of noise, the logic "1" is output because .sigma.c is larger.

In the target determination process of step ST3, if the area determination process of step ST4 outputs a logic "0", the primary feature value output by the feature information calculation process β of step ST213 is used in advance for this primary feature value. The coordinates (x, y) of this point of interest are output together with a signal or symbol indicating that the target is attained by comparing with the set threshold value to determine whether the target is the target. If the area determination processing in step ST4 is logic "
If “1” is output, the primary feature value output by the feature information calculation processing α in step ST212 is compared with a threshold value preset for the primary feature value, and it is determined whether or not the target is a target. And the coordinates (x, y) of this point of interest are output together with a signal or symbol indicating the step ST21 when the influence of noise is large (logic "0" for area determination).
The feature information calculation process β of step 3 evaluates the difference between the S / N ratio and the spread output by the feature information calculation process β. If the effect of the cloud is large (logic “1” for area determination), the feature information calculation process α of step ST212 is output. This means that the S / N ratio and the sharpness of the luminance peak are evaluated.

Further , these series of processes are performed on all the pixels of the image or the pixels in a predetermined limited area while sequentially changing the coordinates of the point of interest from the upper left to the lower right of the image.

In this embodiment, the feature information calculation process is performed for each region. However, one feature information calculation process calculates all three primary feature amounts, and uses the region determination result in the target determination process in step ST3. This is the same even if the primary feature is selected. As a result of the target determination processing in step ST3, a signal indicating the target and the coordinates of the target point are output. However, a signal indicating whether the target is the background or the background is obtained. For a background, a logic “0” may be output.

Embodiment 2 FIG . FIG. 13 is a flowchart showing the flow of a process of a target detection method according to the second embodiment of the present invention. In the figure, step ST2 is a primary feature amount calculation process, step ST3 is a target determination process, step ST301 is a mode switching process, Step ST302 is a background class generation process, step ST304 is a multidimensional feature space mapping process, step ST305 is a class statistic calculation process, step ST305.
306 is a threshold value calculation process, step ST303 is a class determination process, step ST307 is a multidimensional distance calculation process, and step ST308 is a determination process.

[0100] Next, the operation will be described. Step ST
In second primary feature quantity calculation process, for example to connect 2 described in Reference Example 5 and Reference Example 7 calculates the three primary feature quantity.

In the target determination process of step ST3, a determination process is performed using a plurality of primary feature amounts as input. Specifically, the input primary feature amount is passed to the background class generation processing in step ST302 or the class determination processing in step ST303 in the mode switching processing in step ST301 according to the mode instructed from the outside. At this time, the mode is set to step ST at a stage where it is known that the target is not included in the input image such as an introduction stage.
It is set so as to shift to the background class generation processing at 302, and it is not known whether the target is included in the input image, such as the operation stage thereafter, and at the stage where the automatic detection is actually desired, the class determination processing at step ST303 is performed. It is set to migrate. Such a mode switching instruction is given by, for example, a user or an operator via an external interface.

When the primary feature values are passed to the background class generation process in step ST302 by the mode switching process in step ST301 according to the above judgment, all of the primary feature values are background primary feature values. All can be unconditionally treated as a background class. In addition, a plurality of primary feature amounts (for example, three primary feature amounts described in Reference Example 5) are given to all pixels in the primary feature amount image. Therefore, in the multidimensional feature quantity space mapping processing in step ST304,
The three primary feature amounts for arbitrary coordinates (x, y) in the image are mapped to a three-dimensional feature amount space around each feature amount. Hereinafter, the primary feature amounts for other coordinates in the image are similarly mapped in the feature amount space, and the primary feature amounts for each coordinate calculated from another image after an arbitrary time are similarly mapped in the feature amount space. . By this process, the feature space is mapped as a distribution having a large amount of primary features calculated from the background (this distribution is a distribution of only the primary features calculated from the background pixels. Call).

[0103] Next, the class statistic calculation processing in step ST 305, from the background class distribution generated by the multi-dimensional feature space mapping process, an average value expressed as a vector of statistic as for example 3-dimensional distribution thereof dispersed Calculate the covariance value. The average value and the variance-covariance value calculated here are used in the next threshold calculation process in step ST306, and are also used in the multidimensional distance calculation process in step ST307 in the class determination process in step ST303. Finally, in the threshold value calculation process of step ST306, using the average value and the variance-covariance value calculated in the class statistic calculation process of step ST305, for example, a distance away from the average value by (3 × variance-covariance value) And hold it as the threshold.

[0104] On the other hand, when the primary feature quantity according to the judgment as described above is passed to the class determination process in step ST303 by the mode switching process in step ST301, the primary feature quantity from one target pixel that is calculated from the background pixel It is not known whether the calculated value is obtained or not, and it is determined that the target belongs to the detection of the target. Therefore, first, the primary feature amount input in the multi-dimensional distance calculation processing of step ST307 and step S307
Step ST3 in the background class generation processing of T302
Using the average value and the variance-covariance value of the background class calculated in the class statistic calculation process of 05, the distance at which the primary feature value is away from the average value of the background class is calculated. Next, in the determination processing in step ST308, the distance calculated in the multidimensional distance calculation processing in step ST307 is compared with the distance calculated in step ST3.
Step ST306 in the 02 background class generation process
Is compared with the distance calculated as the threshold in the threshold calculation process, and when the distance of the primary feature is larger than the distance as the threshold, it is determined that the primary feature is calculated from the target pixel, and the primary feature is determined. Is smaller than the threshold distance, it is determined that the primary feature amount has been calculated from the background.

Reference Example 7 FIG. 14 is a flowchart showing the flow of processing of the target detection method according to Embodiment 7 of the present invention. In the figure, step ST2 is a primary feature amount calculation processing, step ST3 is a target determination processing, and step ST5 is a reproducibility determination processing. It is.

[0106] Next, the operation will be described. Step ST
In the second primary feature value calculation process, the primary feature value of each point of interest is calculated for an input image at a certain time T, and the primary feature value is passed to the next step ST3. Step ST3
Ginseng processing until detection of a target in the target determination process
This is the same as that described in Examples 1 to 6 and Examples 1 and 2 . In the target determination processing in step ST3, the pixel coordinates of the target point having the primary feature amount determined as the target are output, and are passed to the reproducibility determination processing in step ST5. In this embodiment, the target determination result passed to step ST5 is treated as a target candidate.

In the reproducibility determination process of step ST5, the result of extracting the target candidate for the input image before S unit time before time T (time TS) is stored, and the newly input image at time T is input. Is compared with the target candidate extraction result. The comparison is made to determine the spatial proximity of the coordinates. That is, an arbitrary I-th candidate among a plurality of target candidates at the time (TS) is represented by coordinates (x [TS,
I], y [TS, I]), and among a plurality of target candidates at time T, an arbitrary J-th candidate has coordinates (x [T,
J], y [T, J]), the distance is calculated by the following equation (13). If the distance is small, the J-th target candidate at time T is determined to be a true target. If the distance is large, the distance is calculated in the same manner as the coordinates of another target candidate at time (TS), and the size is checked. If all the target candidates and the distance at time (TS) are calculated and the distance is not small, the time T
Is determined as a non-target (= background).

[0108] distance = √ {{x [T- S, I] -x [T, J]} 2 + {y [T-S, I] -y [T, J]} 2} (13)

In the seventh embodiment , the time T and the time (T
−S), the time T and the time (T−
S) and time (T−2 × S) may be compared for three or more images. Further, in this reference example 7 , the equation shown in equation (13) was used as the distance, but the distance near the coordinates (x [T, J], y [T, J]) of the J-th candidate at time T was simply obtained. A condition of N pixels may be used.

Reference Example 8 FIG. 15 is a flowchart showing the flow of processing of a target detection method according to Embodiment 8 of the present invention. In the figure, step ST1 is an image input processing, step ST2 is a primary feature amount calculation processing, and step ST3 is a target determination processing. is there.
Step ST214 is a pixel-of-interest extraction process, step ST215 is a characteristic-of-interest information extraction process, and step S215 is
T209: background local pixel area D extraction processing, step ST
210: background feature information D extraction processing, step ST204
Is a feature information calculation process. In this reference example 8 ,
As shown in the figure, multiple processes are described in parallel,
The present invention is not limited to this, and these processes may be performed in series.

[0111] FIG. 16 is a diagram showing an example of a local pixel region in this reference example 8, in the figure, the spatial noticed pixel region time 17, space background local pixel region D (time-space local pixel region time 18 ).

[0112] Next, the operation will be described. The assumed target luminance, shape, size, and background state are the same as in the first embodiment . However, regarding the target moving speed, between a certain image input time T and the next image input time (T + 1),
Assume that at most one pixel is moved.

As shown in FIG . 16, the position of the spatiotemporal pixel region of interest 17 is set so that the center thereof coincides with the coordinate of the point of interest (x, y), and its horizontal and vertical dimensions Mcx,
Mcy is about 3 pixels each, and its depth Mcz in the time direction is about 2 pixels. The position of the spatiotemporal background local pixel area 18 is set so that the center thereof coincides with the attention point coordinates (x, y), and its horizontal and vertical dimensions MDb
x and MDby are sufficiently larger than the target sizes Tx and Ty, the widths WDbx and WDby are set to one or more appropriate sizes, and the depth MDbz in the time direction is set to about 2.

[0114] pixel data of the spatial noticed pixel region when read in the target pixel region extraction process in step ST214 is
In the attention feature information extraction processing in the next step ST215, as the feature information, for example, the maximum value of the attention pixel area at time T in the spatiotemporal attention pixel area and the attention at time (T + 1) in the spatiotemporal attention pixel area The average value with the maximum value of the pixel area is calculated and extracted. Step ST
The pixel data of the spatiotemporal background local pixel area D read out in the background local pixel area D extraction processing in 209 is used as background information, for example, in the spatiotemporal background local pixel area in the background feature information D extraction processing in step ST210. The average value and the maximum value of the D pixel data are calculated and extracted.

In the characteristic information calculation processing in step ST204, the spatio-temporal characteristic information of the spatiotemporal background local pixel area D obtained in step ST210 and the characteristic information of the spatiotemporal target pixel area obtained in step ST215 are used. Only when the feature information of the target pixel region is larger than the maximum value of the spatiotemporal background local pixel region D, the primary feature amount is calculated according to, for example, the following equation (14). If the feature information of the spatiotemporal pixel region of interest is smaller than the maximum value of the spatiotemporal background local pixel region D, the primary feature amount is set to a very small value (for example, 0). Here, G is the feature information of the pixel area of interest, BDmax
Is one of the feature information obtained from the spatiotemporal background local pixel region D (maximum value), and BDmean is the other feature information (average value) obtained from the spatiotemporal background local pixel region D.

[0116] The primary feature quantity = (G-BDmax) ÷ ( BDmax -BDmean) (14)

The difference between the feature information of the spatio-temporal pixel region of interest and the maximum value of the spatio-temporal background local pixel region D in the numerator of the primary characteristic amount given by the equation (14) can be obtained by calculating the luminance peak near the center like the target. Has a positive value, which means that the magnitude depends on the relative difference between the luminance near the center and the luminance of the background. Here, assuming that BDmax is constant, if the target exists at both the time T and the time (T + 1) and the moving distance is about 1 pixel (within the range covered by the spatio-temporal pixel area of interest). Both the maximum value of the attention point area at time T and the maximum value of the attention point area at time (T + 1) become large values, and as a result, the feature information G of the spatio-temporal attention pixel area, which is the average, also increases.
However, when the luminance peak existing at a certain time T, such as random noise, is generated at a position completely separated at a time (T + 1) (outside the range covered by the spatio-temporal pixel region of interest), Although the maximum value of the point area is large, the maximum value of the attention point area at time (T + 1) becomes small, and as a result, the feature information G of the spatiotemporal attention pixel area, which is the average, becomes small. Therefore, the target and the noise can be distinguished from each other by this difference. The meaning of the denominator and the entire meaning are the same as in Reference Example 4.

In the target determination process of step ST3, the primary feature value is compared with a preset threshold value, and when the primary feature value is larger than the threshold value, the target is determined, and the target is determined together with a signal or symbol indicating the target. Output the coordinates (x, y) of the point.

Further , these series of processes are performed on all the pixels of the image or the pixels in a predetermined limited area while sequentially changing the coordinates of the point of interest from the upper left to the lower right of the image.

Reference Example 9 FIG. 17 is a block diagram showing a configuration of a target detection apparatus according to Embodiment 9 of the present invention. In FIG. 17, reference numeral 5 denotes a primary feature amount calculation unit, reference numeral 6 denotes a target determination unit (target determination means), and reference numeral 501 denotes a local data extraction circuit. (Local data extraction means), 502 is a feature information extraction unit (feature information extraction means), 503 is an arithmetic operation circuit, 504 is a comparison and selection circuit, 505 is a feature amount operation circuit, 601 is a threshold value holding circuit, 602 is a comparison determination circuit It is. In this reference example , the local data extraction circuit 501
Is included in the primary feature value calculation unit 5, but is not limited thereto, and may be provided independently of the primary feature value calculation unit 5.

[0121] Next, the operation will be described. The image to be processed is captured by an ITV camera, infrared camera, etc.
In the input stage to the target detection device, the image is quantized into horizontal Ix pixels × vertical Iy pixels × Ib gradations, and pixel data is input to the primary feature amount calculation unit 5 in order from the upper left to the lower right of the image. All subsequent processing is performed on the pixels of the entire image.

[0122] local data extraction circuit 501, for example, reference
Row delay circuit number corresponding to the magnitude of the background local pixel area shown in FIG. 8 of Example 4 (for example, a line buffer or FIFO)
Then, the pixel data of the local pixel region is extracted using the register and the register, and the image data is sent to the feature information extracting unit 502. The feature information extraction unit 502 includes a local data extraction circuit 50
The pixel data at the point of interest (the center pixel of the local pixel region) is selectively extracted from the pixel data of the local pixel region sent from No. 1 and the pixel data at the periphery of the local pixel region (the hatched portion in FIG. 8) is extracted. The average value is calculated by the arithmetic operation circuit 503, and the maximum value is extracted by the comparison and selection circuit 504. Feature calculation circuit 505
Is, for example, a memory (look-up table: in this case, a combination of a combination of a multiplier / divider and an adder / subtractor or a combination of a target pixel data, an average value, and a maximum value). And feature value calculation results for all combinations of values that can take the maximum value are stored in advance at a location where the combination value is used as an address). For example , the primary feature amount shown in Reference Example 4 is calculated from the average value and the maximum value.

[0123] comparative determination circuit 602 of the target determination unit 6 (e.g., comparator), the feature amount calculating circuit primary characteristic amount calculated compared to the pre-given threshold by 505, when greater than the threshold, for example a logic "1 Is output, and if it is smaller than the threshold value, for example, logic “0” is output. The threshold value is stored in a threshold value holding circuit 601 (for example, a memory), read out in synchronization with the input of the primary feature amount to the comparison / determination circuit 602, and sent to the comparison / determination circuit 602.

In the ninth embodiment , the local data extraction circuit 501, the arithmetic operation circuit 503, the comparison / selection circuit 50
4 and the feature value calculation circuit 505 are each described as one circuit. However, when processing a plurality of local pixel regions and primary feature values as in Reference Example 5, the number of circuits corresponding to the number should be provided. It goes without saying that processing can be performed at high speed.

Reference Example 10 FIG. 18 is a block diagram showing a configuration of a target detecting apparatus according to Embodiment 10 of the present invention. In FIG. 18, reference numeral 5 denotes a primary feature amount calculating unit, 6 denotes a target determining unit, and 7 denotes a reproducibility determining unit (reproducibility determining unit). ), 501 is a local data extraction circuit, 502 is a feature information extraction unit, 503 is an arithmetic operation circuit, 504 is a comparison / selection circuit, 505 is a feature amount operation circuit, 601 is a threshold value holding circuit, 602 is a comparison and judgment circuit, and 701 is a candidate. Holding circuit, 7
02 is a neighborhood comparison circuit.

[0126] Next, the operation will be described. After the image to be processed is captured and quantized, the primary feature amount is calculated by the primary feature amount calculation unit 5 and executed in the same manner as in the reference example 9 until it is compared with the threshold value in advance by the target determination unit 6. Is done. this
The target detection device according to the ninth embodiment handles the determination result output from the target determination unit 6 as a candidate detection result.

The result of the candidate detection by the target determination unit 6 (the logic “1” is output when the primary feature value is larger than the threshold value, and the logic “0” is output when the primary feature value is smaller than the threshold value) is reproduced. It is sent to the candidate holding circuit 701 and the neighborhood comparison circuit 702 of the gender determination unit 7.

The candidate holding circuit 701 stores all the outputs of the target judgment section 6. The candidate holding circuit 701 is composed of a memory having the same size (Ix × Iy pixels) as the input image and a depth of 1 bit, for example, and outputs of the target determination unit 6 are stored in order from address 0 of the memory. In this case, the result of processing the coordinates (x, y) of the input image as the point of interest is stored in the (x, y) coordinates of the memory of the candidate holding circuit 701. However, considering the data of the (x, y) coordinates in the memory of the candidate holding circuit 701, the result of processing the coordinates (x, y) of the input image at time T as the point of interest is represented by the coordinates ( Immediately before being written to (x, y), the result of processing using the coordinates (x, y) of the input image at time (T−1) as a point of interest is read from the coordinates (x, y) of the memory of the candidate holding circuit 701. It is. This is because two data (one is input directly from the target determination unit 6 and one is input to the candidate holding circuit 7
(Input from 01) means that the data given from the candidate holding circuit 701 is delayed by one time.

The neighborhood comparison circuit 702 includes a candidate holding circuit 70
It is composed of, for example, a line buffer and a register for extracting neighboring data from the determination result at the time (T-1) sequentially given from 1 and a comparator for comparing the neighboring data with the output of the target determining unit 6, and The determination result of the coordinates (x, y) at time T output by the unit 6 is logic “1”, and the vicinity of the coordinates (x, y) at time (T−1) (centering on the coordinates (x, y)) (The (n × n) region to be determined) (the determination result at time (T−1) is stored in the candidate holding circuit 701. The neighborhood comparing circuit 702 is given by the reading operation of the candidate holding circuit 701). If there is at least one logic “1”, the determination result of the coordinates (x, y) at time T output by the target determination unit 6 is determined to be correct, and the logic “1” is output as it is, and the coordinates ( x, y) If there is no logic "0" in the determination result near the coordinates (x, y) at time T-1 even though the logic is "1", the coordinates at time T output by the target determination unit 6 It is determined that the determination result of (x, y) is incorrect, and is changed to logic "0" and output. Note that the determination result of the coordinates (x, y) at time T is a logical "
If it is "0", the logic "0" is output unconditionally.

Reference Example 11 FIG. 19 is a block diagram showing a configuration of a target detecting apparatus according to Embodiment 11 of the present invention. In FIG. 19, reference numeral 5 denotes a primary feature amount calculating unit, 6 is a target determining unit, 506 is a spatiotemporal local data extracting unit, and 507 is An input image holding circuit, 501 is a local data extraction circuit, 502 is a feature information extraction unit, 503 is an arithmetic operation circuit, 504 is a comparison / selection circuit, 505 is a feature amount operation circuit, 601 is a threshold value holding circuit, and 602 is a comparison determination circuit. is there.

[0131] Next, the operation will be described. After an image to be processed is captured and quantized, the process is performed in the same manner as in the ninth embodiment until the image is input to the primary feature amount calculation unit 5.

[0132] Image data of time T which is input to the space-time local data extracting section 506 is sent to the input image holding circuit 507 and the local data extraction circuit 501. Input image holding circuit 5
Reference numeral 07 includes, for example, a two-port memory, and stores image data at time (T-1) at the start of input of time T. When the image at time T is sequentially input from the upper left to the lower right, the image at the time (T-1) is sequentially read from the upper left to the lower right in synchronization with the input, and
Are sequentially stored. That is, when the pixel data of the coordinates (x, y) of the image data at the time T is transmitted at each address (x, y) of the memory, the input image holding circuit 507 stores the data at the address (x, y). The pixel data at the coordinates (x, y) of the image data at the time (T-1) is read out, and immediately thereafter, the pixel data at the coordinates (x, y) of the image data at the time T is newly stored. By this series of operations, the input image is delayed by one time in the input image holding circuit 507. On the other hand, the local data extraction circuit 501 extracts, for example , pixel data of the spatiotemporal background local pixel area shown in FIG. A delay circuit (for example, a line buffer or FIFO) and a time T
And two sets of registers for the input image at time (T-1) and the input image at time (T-1) (sent from the input image holding circuit 507). The pixel data of the background local pixel region is extracted from the input image and sent to the feature information extraction unit 502.

[0133] Arithmetic circuits 503, comparison and selection circuit 504
The feature value calculation circuit 505 receives the pixel data of the local pixel region at each time (T and (T−1)) extracted by the spatiotemporal local data extraction unit 506, and receives the pixel data of the target pixel region (FIG. 1).
6) and pixel data in the peripheral portion of the local pixel region (the hatched portion in FIG. 16) are selectively extracted, for example, as a reference example.
The primary characteristic amount as shown in FIG. 8 is calculated, and the result is sent to the target determination unit 6.

The target judging section 6 operates in the same manner as the target judging section 6 shown in Embodiment 9 to detect a target.

Reference Example 12 FIG. 20 is a block diagram showing a configuration of a target detection apparatus according to Embodiment 12 of the present invention. In the figure, reference numeral 5 denotes a primary feature amount calculation unit, 6 is a target determination unit, 501 is a local data extraction circuit, and 502 is feature information. An extraction unit, 503 is an arithmetic operation circuit,
Reference numeral 504 denotes a comparison / selection circuit, reference numeral 505 denotes a feature value calculation circuit, reference numeral 8 denotes a primary feature value holding circuit (primary feature value holding circuit), reference numeral 601 denotes a threshold value holding circuit, and reference numeral 603 denotes a calculation comparison / determination circuit.

[0136] Next, the operation will be described. After an image to be processed is captured and quantized, the image is input to the primary feature amount calculation unit 5, and the primary feature amount is calculated by the feature information extraction unit 502 via the local data extraction circuit 501. The steps up to this point are performed in the same manner as in the ninth embodiment . The primary feature amount calculated for the input image at time T is sent to the target determination unit 6 and at the same time to the primary feature amount holding circuit 8.

The primary characteristic amount holding circuit 8 is formed of, for example, a two-port memory, and stores the primary characteristic amount at time (T-1) at the start of input of time T. When the image at the time T is sequentially input from the upper left to the lower right, and the primary feature amount is calculated with each pixel as a point of interest, the primary feature amount at the time (T-1) is sequentially read out in synchronization with the primary feature amount. The primary feature values at time T are sequentially stored. That is, the primary feature value holding circuit 8 calculates the address (x, y) of the memory
When the primary feature amount having the coordinates (x, y) of the image data at the time T as the point of interest is transmitted, the coordinates (the coordinates of the image data at the time (T-1) stored at the address (x, y) ( x,
y) is read out and sent to the target determination unit 6, and immediately after that, the coordinates (x,
The primary feature amount with y) as the point of interest is newly stored. By this series of operations, the primary feature amount is delayed by one time in the primary feature amount holding circuit 8.

The primary feature quantity and time (T-1) using the coordinates (x, y) at time T input to the target determination unit 6 as a point of interest
The primary feature amount having the coordinates (x, y) of the target point as an attention point is subjected to an average value processing by the operation comparison and determination circuit 603, and the result is compared with the contents of the threshold value holding circuit 601. As a result of the comparison, if the value is larger than the threshold value, logic "1" is output as a target, and if it is smaller, logic "0" is output.

[0139]

As described above, according to the first aspect of the present invention, many local pixel areas are set, and at least
Calculation using the brightness of the area and the brightness of the target local pixel area
It is configured to calculate more primary primary features.
Therefore, there is an effect that the target can be detected with higher accuracy even in the case where the background of the cloud or land has a large variation in luminance or the case where the luminance of the target is not sufficiently large with respect to the background. Further, the processing for calculating the primary feature amount is very simple, and has an effect that it can be realized at a relatively low cost.

[0140] According to the second aspect of the invention, in the background the local pixel area, the maximum luminance, the average luminance, and obtains at least any two of the luminance of the lowest luminance, the sensor for obtaining digital images from these brightness The standard deviation of the original luminance of the background local pixel region from which the noise component has been removed is estimated, and the primary primary feature is calculated by using at least the estimation result, the luminance of the pixel region of interest, and the luminance of one of the background local pixel regions. Because it is configured to calculate the amount, by evaluating the primary feature amount obtained by normalizing the luminance difference between the target pixel area and the background with the variation in the background luminance, even if the cloud or land with large luminance variation is the background There is an effect that the target can be detected. Further, the processing for calculating the primary feature amount is very simple, and has an effect that it can be realized at a relatively low cost.

[0141] According to the invention of claim 3, in determining the target, to calculate at least one primary feature amount for each pixel of the at least one digital image is known in advance the background, at least one primary feature quantity The primary feature distribution is generated by mapping the primary feature in at least a one-dimensional space having as one axis, and the statistic of the primary feature distribution is calculated. Calculating a distance between a position in a space of at least one primary feature amount calculated for a target pixel area including a certain or unknown pixel and a distribution center position of a primary feature amount distribution determined by a statistic; Since the target pixel area is determined to include the target when the distance is larger than the predetermined distance threshold, a priori knowledge such as the background state and the luminance difference between the target and the background is required. There is no effect with. Further, the processing for calculating the primary feature amount is very simple, and has an effect that it can be realized at a relatively low cost.

[Brief description of the drawings]

FIG. 1 is a flowchart illustrating a target detection process according to a first embodiment of the present invention;

2 is a diagram showing a luminance distribution of <br/> Ru target put to an Example of the invention of Example 1 of the present invention.

FIG. 3 is a flowchart illustrating a target detection process according to a second embodiment of the present invention;

FIG. 4 is a diagram showing an example of a local pixel region according to a second embodiment of the present invention;

FIG. 5 is a flowchart illustrating a target detection process according to a third embodiment of the present invention;

FIG. 6 is a diagram showing an example of a local pixel region in a third embodiment of the present invention.

FIG. 7 is a flowchart illustrating a target detection process according to a fourth embodiment of the present invention;

FIG. 8 is a diagram showing an example of a local pixel region in a fourth embodiment of the present invention.

FIG. 9 is a flowchart showing a target detection process according to Embodiment 5 of the present invention.

FIG. 10 is a diagram showing an example of a local pixel region in a fifth embodiment of the present invention.

FIG. 11 is a flowchart showing a target detection process according to Embodiment 6 of the present invention.

FIG. 12 is a flowchart illustrating a target detection process according to the first embodiment of the present invention.

FIG. 13 is a flowchart illustrating a target detection process according to the second embodiment of the present invention.

FIG. 14 is a flowchart illustrating a target detection process according to Embodiment 7 of the present invention;

FIG. 15 is a flowchart showing a target detection process according to Embodiment 8 of the present invention.

FIG. 16 is a diagram showing an example of a local pixel area in Embodiment 8 of the present invention.

FIG. 17 is a block diagram illustrating a configuration of a target detection device according to Embodiment 9 of the present invention;

FIG. 18 is a block diagram showing a configuration of a target detection device according to Embodiment 10 of the present invention.

FIG. 19 is a block diagram showing a configuration of a target detection device according to Embodiment 11 of the present invention.

FIG. 20 is a block diagram illustrating a configuration of a target detection apparatus according to Embodiment 12 of the present invention;

FIG. 21 is a block diagram showing a configuration of a conventional target detection device.

FIG. 22 is a diagram showing the shape of an isolated point extraction filter used in a conventional target detection device.

FIG. 23 is a diagram showing isolated points extracted by conventional target detection.

[Explanation of symbols]

6 target determination unit (target determination unit), 7 reproducibility determination unit (reproducibility determination unit), 8 primary feature amount holding circuit (primary feature amount holding unit), 12 pixel area of interest, 13, 14, 15
Target local pixel area, 16 Background local pixel area, 17
Spatio-temporal attention pixel area, 18 spatio-temporal background local pixel area (spatio-temporal local pixel area), 501 local data extraction circuit (local data extraction means), 502 feature information extraction unit (feature information extraction means), ST2 primary feature amount calculation Processing, ST
3 Target judgment processing.

──────────────────────────────────────────────────続 き Continued from the front page (56) References JP-A-2-196381 (JP, A) JP-A-2-46264 (JP, U) JP-A-3-59784 (JP, U) JP-A-5 83859 (JP, U) Japanese Utility Model Showa 58-125883 (JP, U) Japanese Utility Model Utility Model No. 1-116853 (JP, U) (58) Fields investigated (Int. Cl. ⁷ , DB name) G06T 7/00-7 / 60 G06T 1/00 H04N 7/18

Claims (3)

(57) [Claims] 1. A target detection method for setting a whole or a part of a digital image as a detection target area and sequentially moving a pixel of interest to detect a target, wherein the target area includes at least one pixel in the detection target area. A pixel area is set, and at least first,
Setting second, third and fourth local pixel regions,
Of the local pixel region is centered on the pixel region of interest.
Surrounds the pixel area and predicts the target to be detected.
A target local pixel area having a size less than or equal to the expected size
And the second and third local pixel regions are
Pixel region around the pixel region of interest with the pixel region as the center.
Of the target to be detected
A target local pixel region having a size,
The pixel area is centered on the target pixel area,
Surrounding the area and the expected
Background local pixel area with a size larger than
At least the brightness of the target pixel area, the target local
The brightness of the pixel area and the brightness of the background local pixel area are used.
And calculating a plurality of primary primary features by calculation, and evaluating the primary features to determine whether the target pixel region includes a target. 2. In the background local pixel area, at least any two luminances of a maximum luminance, an average luminance, and a minimum luminance are obtained, and a noise component of a sensor for obtaining a digital image is removed from these luminances. A standard deviation of the original luminance of the background local pixel area is estimated, and a primary primary feature amount is calculated by calculation using at least the estimation result, the luminance of the pixel area of interest, and one luminance of the background local pixel area. The target detection method according to claim 1 , wherein 3. The method of claim 1, further comprising: calculating at least one primary feature value for each pixel of at least one digital image known in advance as a background; and determining at least one primary feature value on one axis. By generating the primary feature distribution of the background pixel by mapping the primary feature in at least a one-dimensional space, and calculating the statistic of the primary feature distribution, whether the background is a background or a target is unknown. A position in the space of at least one primary feature amount calculated for the pixel region of interest including
Calculating a distance from the distribution center position of the primary feature amount distribution determined by the statistic, and determining that the target pixel region includes a target when the distance is larger than a predetermined distance threshold value; The target detection method according to claim 1 or 2 .