JP5215955B2 - Object detection device - Google Patents

Description

  The present invention relates to an object detection device that detects an object from an image obtained by imaging a monitoring space, and more particularly to an object detection device that detects an object based on edge information extracted from the image.

  In order to detect an object from an image, a discriminator (discriminator) that learns the object and features other than the object in advance is used. At this time, edge information representing the outline of the object may be used as a feature for identifying the object. In order to obtain sufficient edge information for identification, sufficient contrast is required between the object and the background.

  Japanese Patent Application Laid-Open No. H10-228561 describes that a correction process for normalizing a luminance value is performed on a partial image in a detection target image in order to improve a problem of failing to distinguish an object due to a change in contrast. If there is a person wearing dark clothes in the shade, or there is a person wearing light clothes in the sun, an improvement in detection accuracy can be expected when the contrast is insufficient. .

JP 2007-25766 Gazette

  However, an extremely bright sunlight portion and an extremely dark shade portion are adjacent to each other in an image captured in a monitoring space into which strong light such as direct sunlight enters, such as an entrance of an apartment. In such an environment, it has been difficult to extract the edge of the contour of the object that straddles the boundary between the sun and the shade to the extent that it is detected by the discriminator.

  Further, the edge of the contour of the object that straddles the boundary between the sun and the shade is extracted by being connected to the edge having a strong boundary. At this boundary, since the contrast is particularly strong, a strong edge is extracted, and the strength of the edge greatly affects detection of an object. Therefore, there has been a problem that the detection accuracy of the object is lowered due to the influence of the boundary edge.

  FIG. 9 exemplifies an image 900 captured when a person in dark clothes straddles the boundary between the sun and the shade, and an edge image 910 extracted from the image 900. The contour edge 911 of the person in the sun is strong, but the boundary edges 912 and 913 between the sun and the shade are also strong. On the other hand, the contour edge 914 of the person in the shaded part is very weak. In detecting a person, the contribution of the weak contour edge 914 is smaller than that of the strong contour edge 911 and the strong boundary edges 912 and 913, and the convex shape portion composed of the strong contour edge 911 and the strong boundary edges 912 and 913 is mainly evaluated. End up. As a result, the person is missed from the edge image 910.

  The present invention has been made to solve the above-described problem, and an object of the present invention is to provide an object detection device that can detect an object straddling a boundary between light and shadow with high accuracy.

  An object detection device according to the present invention is an object detection device that detects an object from an image in which a monitoring space is imaged, and is imaged when there is no object in advance and an imaging unit that images the monitoring space. A storage unit for storing the background image, a luminance distribution analyzing means for analyzing the background image and calculating the background side luminance distribution, and a bimodal distribution in which the background side luminance distribution is unevenly distributed in the high luminance region and the low luminance region The bimodality judging means for judging whether or not the image is inputted, and when the bimodal distribution is judged, luminance conversion for expanding the high luminance side distribution range in which the pixels in the high luminance region are distributed is inputted from the imaging unit. A low-luminance correction image is generated by applying luminance conversion to each pixel of the input image to generate a high-luminance correction image on each input image and expanding the low-luminance distribution range in which pixels in the low luminance region are distributed. Generates the high brightness side corrected image and the low brightness side corrected image Image correction means for generating an average image; edge extraction means for extracting edge information from the input image; and object detection means for detecting an object based on the edge information. The edge extraction means generates an average image. If so, edge information is extracted from the average image.

According to such a configuration, the bimodal distribution is determined if the background includes the boundary between light (high luminance region) and shadow (low luminance region), and the contrast between the background and the object is corrected in the high luminance region. The high brightness side corrected image and the low brightness side corrected image in which the contrast between the background and the object is corrected in the low brightness area are generated. Further, in the high luminance side corrected image and the low luminance side corrected image, the contrast between light and shadow is weakened by the above correction.
The edge information extracted from the average image having the properties of the high-luminance correction image and the low-luminance correction image is emphasized on the contour edge of the object in the high luminance region and the contour edge of the object in the low luminance region, In addition, since the boundary edge between light and shadow is suppressed, it is possible to detect an object that straddles the boundary between light and shadow with high accuracy.

  In a preferred aspect of the present invention, the luminance distribution analyzing means analyzes the pixels of the input image corresponding to the pixels in the high luminance region, calculates the high luminance side distribution range, and corresponds to the pixels in the low luminance region. The low luminance side distribution range is calculated by analyzing the pixels of the input image.

When the object spans light and shadow, there is a situation in which sufficient contrast is obtained on the other side even if the contrast is insufficient on one side of the high luminance range and the low luminance range. If the brightness distribution range with sufficient contrast is greatly expanded, even the fine edges that cause disturbances in object detection are extracted, resulting in an excessive edge state and the accuracy of object detection decreases. To do. However, if attention is paid to the input image, the luminance is distributed over a wide range in a case where sufficient contrast is obtained.
Therefore, as described above, by calculating the high luminance side distribution range and low luminance side distribution range expanded by luminance conversion from the input image, the luminance distribution range is greatly expanded in the luminance region where the contrast is insufficient. In the luminance range where the contour edge is extracted and sufficient contrast is already obtained, the expansion of the luminance distribution range is suppressed and excessive edges can be prevented, so that the object across the boundary between light and shadow can be detected with high accuracy. .

In a preferred aspect of the present invention, the object detection device further includes candidate area setting means for setting a candidate area having a size set in advance according to the object at a predetermined position in the input image, and the luminance The distribution analysis means analyzes the pixels in the candidate area.
According to such a configuration, the luminance distribution is analyzed in an area having a minimum size necessary for detecting an object. Therefore, it is possible to perform an accurate analysis that is not affected by an excessive background portion and an effective luminance conversion that maximizes the distribution range.

In a preferred aspect of the present invention, the candidate area setting means compares the input image with the background image and sets a candidate area at a position where a difference is detected.
When a corrected image is generated from an input image of only the background, even a fine edge of the background is extracted, resulting in excessive edges, causing erroneous detection of the object. According to the above configuration, since only an area with a high probability of existence of an object is set as a candidate area, generation of a corrected image from an input image with only a background is prevented, and erroneous detection of the object can be prevented.

In a preferred aspect of the present invention, the candidate area setting means detects a human head area having an elliptical shape from the input image, and sets the candidate area at a position where the human head area is detected.
According to such a configuration, since only a region having a high probability that the human body that is the target exists is set as a candidate region, it is possible to prevent a corrected image from being generated from an input image of only the background, and to prevent erroneous detection of the target. it can.

  According to the present invention, it is possible to suppress the boundary edge between light and shadow while enhancing the contour edge of the object, so that it is possible to detect the object across the boundary between light and shadow with high accuracy. Become.

1 is an overall configuration diagram of an image monitoring apparatus according to the present embodiment. It is a figure which shows an example of the process by the luminance distribution analysis means. It is a figure which shows an example of the process by the image correction means. It is a figure which shows the process by the image correction means 43 in another example. It is a figure which shows the flowchart of an image monitoring process. It is a figure which shows the one part flowchart of a passer-by detection process. It is a figure which shows the flowchart of the remaining part of a passer-by detection process. It is the figure which showed typically the edge information extracted by this invention. It is the figure which showed typically the edge information extracted when not applying this invention.

  As an example of a preferred embodiment of the present invention, an image monitoring apparatus will be described in which an entrance of a building is used as a monitoring space, a passerby passing through the monitoring space is detected as an object, and unauthorized traffic is monitored.

[Configuration of Image Monitoring Apparatus 1]
A functional block diagram of the image monitoring apparatus 1 is shown in FIG. The image monitoring apparatus 1 includes an imaging unit 2, a storage unit 3, and an output unit 5 connected to a signal processing unit 4.

  The imaging unit 2 is a so-called surveillance camera. The imaging unit 2 sequentially outputs images obtained by imaging the monitoring space at predetermined time intervals to the signal processing unit 4. Hereinafter, the unit of time that is recorded at the above time interval is referred to as time. In the present embodiment, the imaging unit 2 captures and outputs a color image in which each RGB color has 256 gradations.

  The imaging unit 2 is installed on the ceiling of the entrance of a building with the optical axis directed vertically downward. Direct sunlight from the outside enters the entrance, and the shadows of the building structures are reflected. Therefore, the image picked up by the image pickup unit 2 includes a high-luminance sun part and a low-brightness shade part.

  The storage unit 3 is a memory device such as a ROM or a RAM. The storage unit 3 stores various programs and various data, and inputs and outputs such information to and from the signal processing unit 4. Various data includes a background image 30 and object identification information 31.

  The background image 30 is an image that includes only the background image of the monitoring space and does not include the passerby image. The background image 30 is generated and stored prior to the passer-by detection process.

  The object identification information 31 is reference information for identifying whether or not a passerby exists in the image. Prior to the passer-by detection process, the object identification information 31 is generated and stored based on edge information (luminance gradient information) extracted from at least a sample image in which the passer-by is captured.

As the edge information, a feature amount called HOG (Histograms of Oriented Gradients) can be adopted. The HOG feature amount is a parameter series including an edge angle distribution and an edge strength calculated in each of a plurality of small local areas set in the image. In the HOG feature amount, the edge strength is used as a weighting coefficient of the edge angle distribution. As a reference for HOG features, N. Dalal, B. Triggs: Histograms of oriented gradients for human detection: Proc. Of IEEE Conference on Computer Vision and Pattern Recognition
(CVPR): 886-893 (2005).

  The object identification information 31 is a parameter representing an identification surface that discriminates passers-by from others in a feature space spanned by a multi-dimensional HOG feature. This parameter is used for supporting vector machines and boosting methods for HOG feature values extracted from a large number of sample images taken by passersby and HOG feature values extracted from a number of sample images not taken by passersby. It is learned by applying.

  In another embodiment, the object identification information 31 is generated as a passer's edge pattern that is obtained by averaging the edge intensities extracted from a plurality of sample images in which the passer is imaged. In yet another embodiment, the edge angle extracted from a plurality of sample images in which a passerby is imaged is generated as an edge pattern averaged for each pixel.

  The object identification information 31 thus generated mainly represents the characteristics of the passer's edge common to a plurality of sample images, that is, the outline (outer shape) characteristics of the passer.

The signal processing unit 4 includes a DSP (Digital Signal Processor), an MCU (Micro Control
Unit). The signal processing unit 4 describes the operations of the background image generation unit 40, the candidate area setting unit 41, the luminance distribution analysis unit 42, the image correction unit 43, the edge extraction unit 44, the object detection unit 45, the abnormality detection unit 46, and the like. It functions as each means by reading the program from the storage unit 3 and executing it. The signal processing unit 4 appropriately corrects an image input from the imaging unit 2 (hereinafter referred to as an input image) to detect a passer-by, and when detecting unauthorized traffic by tracking the detected passer-by, the signal processing unit 4 outputs to the output unit 5. An abnormal signal is output.

  The background image generation unit 40 stores the entire input image captured when no passerby exists in the monitoring space in the storage unit 3 as the background image 30. In addition, the background image generation unit 40 updates the background image 30 by synthesizing the background image 30 with an image of a portion of the input image where no passer-by is detected, in order to adapt to illumination fluctuations.

  The candidate area setting unit 41 sets a partial area having the size of the passer's image in the input image as a candidate area, and sets information on the candidate area as a luminance distribution analysis unit 42, an image correction unit 43, an edge extraction unit 44, Output to the object detection means 45.

  The size of the passerby image changes depending on the positional relationship between the passerby and the imaging unit 2 or the like. Therefore, the candidate area setting means 41 sets the size of the passer's image using the position on the image as a variable. Specifically, a camera parameter such as the installation height of the image pickup unit 2 and the angle of view, and a three-dimensional model that approximates a standing human figure with a spheroid is applied to the pinhole camera model for monitoring. An image in which the three-dimensional model virtually arranged at each position in the space is projected on the image is calculated, and a rectangular frame surrounding the projected image is associated with the position of the projected image on the image and stored in the storage unit 3 in advance. Keep it. The candidate area setting means 41 reads out and sets a rectangular frame corresponding to the position where the candidate area is to be set as a candidate area.

  Here, the candidate area setting means 41 can also set candidate areas sequentially over the entire area of the image, but it is better to set candidate areas by narrowing down to positions where there is a high possibility that a passerby exists. The probability of false detection is reduced, and the processing cost is also reduced.

  Therefore, the candidate area setting unit 41 extracts a change area having a difference from the background image 30 in the input image, sets the candidate area at a position based on the change area, and an elliptical image is captured in the input image. The human head region is extracted and a candidate region is set at a position based on the human head region. In particular, by limiting the extraction range of the human head region to the change region, the accuracy of narrowing down can be increased, and the processing cost can be reduced.

  The change area can be extracted by a difference process between the input image and the background image 30 or a correlation calculation between the input image and the background image 30. The head region can be extracted by generating an edge image from the input image and performing a Hough transform on the generated edge image, or by matching the generated edge image with a preset elliptic pattern. The edge image can be generated by applying filtering such as a Sobel filter or a Canny filter to the input image.

  When a predetermined image region is designated by the image correction unit 43, the luminance distribution analyzing unit 42 calculates the characteristic amount of the luminance distribution of the background image 30 and the input image in the region, and calculates the calculated luminance distribution feature amount. It outputs to the image correction means 43. As a feature value of the luminance distribution, the luminance distribution analyzing means 42 calculates a luminance histogram of each image, detects a peak from the calculated luminance histogram, and calculates a luminance range of the peak.

  The luminance histogram is the frequency of pixels for each luminance. The luminance histogram can be calculated by counting the pixels constituting each image for each luminance. In this embodiment, each frequency is divided by the total number of pixels and normalized to a range of 0-1. Note that the input image and the background image 30 are RGB color images, but the luminance distribution analysis means 42 converts these into 256-tone monochrome images, and then calculates the characteristic amount of the luminance distribution.

  The mountain is a part where high-frequency luminance is dense. The luminance distribution analyzing means 42 applies the two peak detection thresholds Tg1, Tg2 (Tg1> Tg2) to the luminance histogram to detect the peak luminance distribution. That is, the luminance distribution analyzing unit 42 specifies the luminance whose frequency exceeds Tg1, and sets the luminance range in which the frequency continuously exceeds Tg2 from the specified luminance to the low luminance side and the high luminance side as the peak distribution range. To detect. If there are a plurality of peaks, the same number of distribution ranges are detected. Each of Tg1 and Tg2 may be set to a value corresponding to about 20% or about 5% of the total number of pixels in the designated image area, for example.

  Hereinafter, the feature amount of the luminance distribution calculated for the background image 30 is also referred to as a background side luminance distribution, and the feature amount of the luminance distribution calculated for the input image is also referred to as an input side luminance distribution.

Here, in order to sufficiently obtain the effect of the contrast correction process, it is preferable that the distribution range of the peaks is sufficiently narrow. That is, it is better that the mountain portion is unevenly distributed.
Therefore, the luminance distribution analysis means 42 determines the uneven distribution of the peaks and detects only the uneven peaks. Specifically, the luminance distribution analyzing means 42 compares the width of the peak distribution range (luminance width) with the uneven distribution determination threshold value Tw, and detects only the peak portions whose distribution range is equal to or smaller than Tw. Output. Specifically, if the peak satisfies (Hb−Lb) ≧ Tw, the peak is detected. If (Hb−Lb) <Tw, the peak is not detected. For example, Tw may be set to a value of about one third of the gradation. In this example, 256 ÷ 3≈85 is set.

FIG. 2 shows an example in which the luminance distribution analyzing means 42 calculates the background side luminance distribution from the background image 100.
In the background image 100, the sunlit portion is imaged on the upper side and the shaded portion is imaged on the lower side. The luminance distribution analyzing means 42 analyzes the background image 100 to calculate the luminance histogram 101, and detects two peaks that exceed the peak detection threshold Tg1. Further, the luminance distribution analyzing means 42 detects the luminance exceeding the peak detection threshold Tg2, and sets the distribution range of the peak portion on the low luminance region side as Lb1 to Hb1, and the distribution range of the peak portion on the high luminance region side as Lb2. Calculate as Hb2.

  The image correction unit 43 performs a contrast correction process on the input image in the candidate area, and outputs the processed image to the object detection unit 45. The processing in the image correction unit 43 is largely divided into a case where the background side luminance distribution is a bimodal distribution unevenly distributed in a high luminance region and a low luminance region, and a case where the background correction is not, and the image correction unit 43 is a bimodal distribution. Features in processing. Here, the processing in the case of the bimodal distribution will be described in detail, and the processing in the case of the non-bimodal distribution will be outlined in the description of the operation.

In order to determine the bimodal distribution, the image correcting unit 43 includes a bimodal determining unit 431. The bimodal determination means 431 determines that the background side luminance distribution is a bimodal distribution if the number of pieces of peak information included in the background side luminance distribution is two, and if the number is not two, It is determined that the distribution is not peak.
The image correcting unit 43 causes the luminance distribution analyzing unit 42 to calculate the background side luminance distribution for each candidate region, and the determination by the bimodal determination unit 431 is performed for each candidate region.

  When the bimodal distribution is determined, the image correcting unit 43 detects pixels distributed in each of the high luminance region and the low luminance region of the background side luminance distribution, and the luminance of the pixel group detected in the high luminance region. The luminance conversion for expanding the distribution range is performed on each pixel of the input image, and the luminance conversion for expanding the luminance distribution range of the pixel group detected in the low luminance region is performed on each pixel of the input image. Contrast correction processing is performed by averaging the results of the types of luminance conversion.

  With reference to FIG. 3, a processing example when the bimodal distribution is determined in the image correction unit 43 will be described.

  The background image 100 and its luminance histogram 101 in FIG. 3 are the same as those in FIG. The image correcting unit 43 sets the intermediate luminance between the lowest luminance value Lb2 of the high luminance side peak and the highest luminance value Hb1 of the low luminance side of the two peak portions to the division threshold Td, and the background image Among the 100 pixels, a pixel in a high luminance region whose luminance is higher than Td and a pixel in a low luminance region whose luminance is Td or less are detected. Td can be set to {(Lb2-Hb1) / 2} or the like. The image area formed by the detected pixel group in the high luminance area is the bright area 102, and the image area formed by the detected pixel group in the low luminance area is the dark area 103. They correspond to the sun and shade respectively.

  An image 110 is an input image corresponding to the background image 100. In the image 110, a passerby in light clothes is captured. The image correcting unit 43 causes the luminance distribution analyzing unit 42 to specify the bright region 102 and calculate the input side luminance distribution. The luminance distribution analyzing means 42 calculates a luminance histogram 113 and peak distribution ranges Li to Hi of the input image 112 in the bright region 102.

The image correcting unit 43 converts the luminance I of each pixel of the input image 110 into {(I−Li) × 255 ÷ (Hi−Li)}. The distribution range Li to Hi is expanded to 0 to 255. At this time, the luminance of the pixel whose luminance is less than Li in the input image 110 is 0, and the luminance of the pixel whose luminance is higher than Hi in the input image 110 is 255.
Note that the entire input image 110 is subjected to luminance conversion based on the luminance distribution of the input image 112 in the bright region 102.

  Luminance conversion that shifts the luminance at a constant rate is called histogram expansion (Histogram Spreading or Histogram Stretching). The expansion of the histogram is one of contrast correction processes for correcting the contrast by expanding the distribution biased to a narrow luminance range.

  Another contrast correction process is a process called histogram equalization. In the flattening of the histogram, the luminance I of each pixel of the input image 110 is converted into {255 ÷ N × ΣP (I)}. However, N represents the total number of pixels having a luminance of Li to Hi, and ΣP (I) represents the sum of the frequencies from the luminance Li to I.

When the luminance conversion is performed in this way, a high luminance side corrected image 114 corrected based on the luminance distribution of the bright region 102 on the high luminance side is generated. Incidentally, the luminance histogram 111 of the input image 110 becomes like the luminance histogram 115.
The pixels of the input image 110 distributed in the high luminance region, that is, the pixels related to the passersby who are dressed in bright colors with the sunny portion, have a large luminance difference from the surrounding pixels due to the expansion of the luminance distribution range. Therefore, the contrast between the object and the background becomes strong, and the contour edge of the object is emphasized.
On the other hand, the pixels of the input image 110 that have been distributed in the low luminance region, that is, the pixels in the shaded portion, have a smaller luminance difference from the pixels in the sunny portion due to the expansion of the luminance distribution range. Therefore, the contrast between the sun and the shade is weakened, and the boundary edge between light and shadow is suppressed.

  Further, the image correction unit 43 causes the luminance distribution analysis unit 42 to specify the dark region 103 and calculate the input side luminance distribution, and obtains the luminance histogram 123 of the input image 122 and the peak distribution ranges Li to Hi in the dark region 103. . In this example in which there are two peaks, the minimum luminance value of the low luminance side peak is Li, and the maximum luminance value of the high luminance side peak is Hi. Then, the image correction unit 43 performs luminance conversion on the input image 110 to generate a low luminance side corrected image 124 that is corrected based on the luminance distribution of the dark region 103 on the low luminance side.

When the low luminance side corrected image 124 is generated, in the luminance histogram 123, Li is 0 and Hi is a value close to 255, so the luminance distribution range is hardly expanded. This result is best from the point of view of preventing excessive edges. The reason will be described.
What is imaged in the dark region 103 of the input image 110 is a passerby who is dressed in a shaded background and light colors. In such a portion, sufficient contrast is already obtained between the background and the object, and an appropriate amount of edge information can be obtained, so that it is not necessary to expand the luminance distribution range.
If the luminance distribution range is greatly enlarged despite being unnecessary, a weak luminance difference such as a shadow or a pattern in the background region or the passer-by region is emphasized, and extra edge information that becomes a disturbance for the object detection processing increases. That is, if the luminance distribution range is expanded in a portion where sufficient contrast has already been obtained, excessive edges occur and the detection accuracy is lowered.
For such a reason, it is preferable to prevent the excessive edge by suppressing the expansion of the luminance distribution range in a portion where sufficient contrast has already been obtained.
The image correction means 43 performs luminance conversion that expands the luminance distribution range in the input image 110, thereby substantially suppressing enlargement in a portion where sufficient contrast has already been obtained. Can be prevented.

When the high-luminance correction image 114 and the low-luminance correction image 124 are generated as described above, the image correction unit 43 replaces the corresponding pixels of the high-luminance correction image 114 and the low-luminance correction image 124 with each other. An average image 130 is generated on average. By this averaging, the high luminance side corrected image 114 and the low luminance side corrected image 124 are smoothly synthesized. For this reason, it is possible to prevent the brightness difference between the sun and the shade from being excessively emphasized.
The average image 130 has the properties of both the high-luminance correction image 114 and the low-luminance correction image 124. In the bright region 102 and the dark region 103, the contrast between the background and the object is insufficient. The contrast is enhanced, the contrast between the shade and the sun is suppressed, and the contour edge of the object is easily extracted accurately.

  With reference to FIG. 4, another processing example when the bimodal distribution is determined in the image correction unit 43 will be described.

  In the example of FIG. 4, the candidate areas are the same as in the example of FIG. 3, and the background image 100, the luminance histogram 101, Hb1, Lb2, the bright area 102, and the dark area 103 are the same as in the example of FIG. Unlike the example of FIG. 4, the input image 110 captures a passerby in dark clothes.

  The image correcting unit 43 obtains luminance histograms 213, Li, and Hi from the input image 212 in the bright area 102, and performs luminance conversion on the input image 210 to generate a high luminance side corrected image 214. In the example of FIG. 3, the low luminance side corrected image 124 in which the luminance distribution range is not substantially enlarged is generated, but in the example of FIG. 4, the high luminance side corrected image 214 in which the luminance distribution range is not substantially enlarged. Is generated. In the bright region 102 where sufficient contrast has already been obtained, the expansion of the luminance distribution range is suppressed, and excessive edges are prevented.

  The image correcting unit 43 obtains luminance histograms 223, Li, and Hi from the input image 222 in the dark region 103, performs luminance conversion on the input image 210, and generates a low luminance side corrected image 224. In the input image 222 of the dark region 103, a passerby of dark clothes is imaged in the shade, and the luminance distribution ranges Li to Hi are unevenly distributed in the low luminance region. In the low-luminance correction image 224, the unevenly distributed luminance distribution range Li to Hi is expanded to 0 to 255, the contrast between the background and the object is enhanced, and the contrast between the shade and the sun is suppressed.

  The image correction means 43 averages the high-luminance correction image 214 and the low-luminance correction image 224, emphasizes the contrast that was lacking in the bright region 102, and suppresses the contrast between the shade and the sunny boundary. 230 is generated.

As described above, when the bimodal distribution is determined, the image correction unit 43 inputs luminance conversion for expanding the luminance distribution range of the pixel group (bright region) distributed in the high luminance region of the background side luminance distribution. Apply to each pixel of the input image to generate a high-luminance correction image for each pixel of the image, and to perform luminance conversion that expands the luminance distribution range of the pixel group (dark region) distributed in the low-luminance region of the background-side luminance distribution To generate a low-luminance correction image, and generate an average image of the high-luminance correction image and the low-luminance correction image.
As a result, the boundary edge between light and shadow can be suppressed while enhancing the contour edge of the input image where the contrast between the background and the object is insufficient, and the object detection accuracy can be improved.

The image correcting unit 43 sets the original luminance distribution range to be enlarged as the luminance distribution range in the input-side luminance distribution calculated from the input images in the bright region and the dark region.
Thereby, the expansion of the luminance distribution range is suppressed in the luminance conversion with respect to the portion with sufficient contrast, the occurrence of excessive edges can be prevented, and the object detection accuracy can be improved.

  Here, if only the background area in which no passerby exists is set as a processing target, an excessive edge tends to occur. Therefore, the image correcting unit 43 processes the candidate region set by the candidate region setting unit 41, that is, the candidate region set in the position where the change area and the head region are extracted and the passerby exists with high probability. By doing so, it is avoided that contrast correction is performed on only the background region, and the occurrence of excessive edges is prevented.

  Further, in order to perform accurate luminance distribution analysis and maximum luminance conversion without being affected by the extra background portion, it is desirable that the area consisting only of the passerby and the background in the vicinity thereof is a processing target. For this reason, the luminance distribution analyzing unit 42 and the image correcting unit 43 use the candidate area set by the candidate area setting unit 41, that is, the candidate area set in accordance with the size of the passer's image as a processing target. An accurate luminance distribution analysis and maximum luminance conversion based on necessary and minimum image information consisting only of a person's image and a background image in the vicinity thereof are performed.

The edge extraction unit 44 extracts edge information from the average image when the average image is generated by the image correction unit 43. If the average image is not generated, the edge extraction unit 44 performs an input image or a contrast correction process if necessary. Edge information is extracted from the applied input image, and the extracted edge information is output to the object detection means 45.
The edge information extracted by the object detection means 45 is the same type of edge information as the object identification information 31, and the above-described HOG feature value or edge pattern can be adopted.

  FIG. 8 illustrates edge information (edge strength) extracted from the average image 130 of FIG. 3 or the average image 230 of FIG. Both the contour edge 301 of the sunlit portion and the contour edge 302 of the shaded portion are extracted with moderate strength, and the boundary edges 303 and 304 of the sun and shade are extracted weaker than the contour edges.

  The object detection unit 45 detects the object by comparing the edge information input from the edge extraction unit 44 with the object identification information 31, and information on the input image in the candidate area where the object is detected, that is, the traffic Information about the human image is output to the abnormality detection means 46.

When the parameter of the identification plane in the HOG feature space is learned as the object identification information 31, the object detection unit 45 reads the parameter, sets it in a preset classifier, and uses the extracted HOG feature value. The comparison is performed by inputting to the discriminator. The discriminator discriminates whether or not the input HOG feature amount is in the space on the object side of the discrimination plane boundary in the feature space, and outputs the discrimination result.
When an edge pattern is created as the object identification information 31, the object detection means 45 performs the above comparison by reading the edge pattern and performing pattern matching with the extracted edge information.

  When the information on the passerby image is input from the object detection unit 45, the abnormality detection unit 46 extracts a feature amount such as a color histogram, a position, and a shape from the passerby image and stores it in the storage unit 3 and moves back and forth. By identifying the feature value extracted at the time, the passerby is tracked to detect unauthorized traffic. That is, the location of the building entrance and the location of the authentication device are set in the storage unit 3 in advance, the movement trajectory obtained as a result of tracking is compared with these positions, and the movement trajectory reaching the location of the building entrance is identified by the authentication device. If it does not go through the position of, it is determined that the traffic is illegal. When unauthorized traffic is detected, the abnormality detection means 46 outputs an abnormality signal to the output unit 5.

  When an abnormality signal is input from the abnormality detection means 46 of the signal processing unit 4, the output unit 5 is a security company through a sound device such as a buzzer or a speaker that performs sound output for warning of unauthorized traffic, and a communication line. Including a communication device that transmits an abnormal signal to a host computer installed in a security center in the building or a security room in the building, and outputs to the outside that unauthorized traffic has occurred.

[Operation of the image monitoring apparatus 1]
Hereinafter, the operation of the image monitoring apparatus 1 will be described with reference to FIG.

When an administrator who confirms that the monitoring space is unmanned turns on the apparatus, each unit and each means are initialized and start operating (S1). The background image generation means 40 stores the input image being initialized in the storage unit 3 as the background image 30.
After initialization, every time a new image is input from the imaging unit 2 to the signal processing unit 4, the processing of S2 to S8 is repeated.

  When a new image is input (S2), processing for detecting passers-by from the input image is performed (S3). Details of the passer-by detection process will be described below with reference to the flowchart of FIG.

  In steps S30 to S32, candidate areas are set by the candidate area setting means 41 of the signal processing unit 4.

  First, the candidate area setting means 41 performs a difference process between the background image 30 and the input image, and extracts a group of pixels having a difference equal to or greater than a preset difference threshold value as a change area (S30). As an extraction result, a mask image in which the pixel value in the change area is set to 1 and the pixel value outside the change area is set to 0 is generated.

  Next, the candidate area setting unit 41 generates an edge image from the input image, masks the edge image with the input mask image, removes the edge outside the change area, and performs the Hough transform on the masked edge image. Then, an ellipse is detected, and an area where the ellipse is detected is extracted as a head area (S31). As a result, a region where an elliptical image exists within the change region is extracted. The candidate area setting means 41 calculates the barycentric position of the head area, and reads out from the storage unit 3 a rectangular frame stored in association with the position as a candidate area (S32).

  If no change area is extracted or if a head area is not extracted, no candidate area is set. If the candidate area is not set, the signal processing unit 4 ends the passer detection process on the assumption that no passer is detected (NO in S33 to S4).

  On the other hand, when one or more candidate areas are set (YES in S33), signal processing unit 4 sequentially sets the individual candidate areas as processing targets and executes the loop process of steps S34 to S51.

  In the loop processing, first, the image correction unit 43 of the signal processing unit 4 designates a candidate region to be processed and causes the luminance distribution analysis unit 42 of the signal processing unit 4 to calculate the background side luminance distribution (S35). The luminance distribution analyzing means 42 outputs the luminance histogram of the background image 30 in the candidate region to be processed and the information on the peak portion determined to be unevenly distributed as the background side luminance distribution.

  If the information on the peaks detected in step S35 is not two, the bimodal determination unit 431 of the image correction unit 43 determines that the background-side luminance distribution is not a bimodal distribution (NO in S36), and processing Advances to steps S44 to S49 in FIG. This process will be described later.

  On the other hand, if there are two pieces of peak information detected in step S35, the bimodal determination means 431 determines that the background side luminance distribution is a bimodal distribution (YES in S36), and steps S37 to S43. Is performed.

  The image correcting means 43 sets the intermediate luminance between the two peaks to the division threshold Td, and consists of a bright region composed of pixels with a luminance higher than Td and a pixel with a lower luminance than Td in the background image 30 of the candidate region. A dark region is detected (S37).

  When a bright / dark region is detected, the image correcting unit 43 designates the bright region and causes the luminance distribution analyzing unit 42 to calculate the input-side luminance distribution, and the width of the peak portion of the luminance distribution in the input image of the bright region. Is acquired as a high luminance side distribution range (S38), and luminance conversion for expanding the high luminance side distribution range to all gradations is performed on all pixels constituting the input image of the candidate region to generate a high luminance side corrected image ( S39).

  Further, the image correction unit 43 designates a dark region and causes the luminance distribution analysis unit 42 to calculate the input side luminance distribution, and acquires the width of the peak portion of the luminance distribution in the input image of the dark region as the low luminance side distribution range. (S40) A luminance conversion that expands the low luminance side distribution range to all gradations is performed on all the pixels constituting the input image of the candidate area to generate a low luminance side corrected image (S41).

  When the high luminance side corrected image and the low luminance side corrected image are generated, the image correcting unit 43 generates an average image obtained by averaging these corresponding pixels (S42), and the edge extracting unit 44 of the signal processing unit 4 is used. Extracts HOG feature values from the generated average image (S43).

  With reference to FIG. 7, the process of step S44-S49 performed when background side luminance distribution is not bimodal distribution is demonstrated.

  If there is only one peak information included in the background-side luminance distribution, the image correcting unit 43 assumes that the background-side luminance distribution is a unimodal distribution unevenly distributed in one place (YES in S44), and selects candidate areas. The luminance distribution analyzing means 42 is designated to calculate the input-side luminance distribution (S45), and if the calculated input-side luminance distribution is similar to the background-side luminance distribution, contrast correction is performed on the input image in the candidate area (S46). YES → S47). The edge extraction unit 44 extracts the HOG feature amount from the corrected image generated as a result of the contrast correction (S48). For example, if there are passers in dark clothes in the shade, passers in light clothes in the sun, etc., if the contrast is insufficient, similarity with the unimodal distribution is determined, Edge information emphasized by contrast correction is extracted.

  On the other hand, when the background-side luminance distribution is not determined to be a unimodal distribution (NO in S44) and when similarity is not determined even for a unimodal distribution (NO in S46), edge extraction is performed. The means 44 extracts the HOG feature value from the input image of the candidate area (S49). If there is already sufficient contrast between the background and the passerby, a unimodal distribution or a difference is determined, and the occurrence of excessive edges is prevented by not performing contrast correction.

  When the HOG feature value is extracted as described above, the object detection unit 45 compares the extracted HOG feature value with the object identification information 31 in the storage unit 3 and a passerby exists in the candidate area. And the determination result is temporarily stored in the storage unit 3 in association with the candidate area information (S50).

  When the processing is completed for all candidate areas (YES in S51), the object detection unit 45 passes the input image in the candidate area and the position and size of the candidate area where the passer-by is detected. Output as information.

  Returning to FIG. 6 again, the continuation of the image monitoring process will be described.

  When a passer-by is detected (YES in S4), the abnormality detection means 46 of the signal processing unit 4 associates the detected passer-by information with the passer-by information detected before the current time. By detecting the movement trajectory of each passerby, the presence of an abnormal passerby who has not stopped at the authentication device is determined based on the movement trajectory (S5).

  When an abnormality is determined (YES in S5), the abnormality detection means 46 outputs an abnormality signal to the output unit 5 (S7), and the output unit 5 to which the abnormality signal is input performs an output operation such as an alarm or a report. . If no passerby is detected (NO in S4), steps S5 to S7 are skipped, and if no abnormality is determined (NO in S6), step S7 is skipped.

  When the processing up to step S7 is completed, the background image generation means 40 of the signal processing unit 4 updates the background image 30 by synthesizing the background image 30 with an image of an area in which no passerby is detected in the input image ( S8).

[Modification]
In the above embodiment, the luminance distribution analyzing unit 42 calculates a luminance histogram, and the image correcting unit 43 determines bimodality based on the luminance histogram and detects a bright region (high luminance region) / dark region (low luminance region). The brightness distribution range to be expanded was specified.
In another embodiment, the luminance distribution analyzing means 42 counts the number N1 of pixels having a luminance equal to or higher than the upper limit luminance threshold TH1 and the number N2 of pixels having a luminance value equal to or lower than the lower limit luminance threshold TH2. The image correction means 43 determines bimodality when N1 ≧ TN and N2 ≧ TN of the background side luminance distribution, detects a light / dark region by the division threshold Td = (TH2−TH1) / 2, 0 to TH1 or TH2 to 255 is specified as the luminance distribution range to be enlarged.
TH1 can be set to about 85, TH2 can be set to about 170, and TN can be set to a value corresponding to 30% or more of the total number of pixels.

  In yet another embodiment, the luminance distribution analyzing means 42 calculates the average A of luminance and the variance σ. The image correcting unit 43 determines bimodality when the background side luminance distribution TH1 <A <TH2 and σ ≧ Tσ2, and detects a bright / dark region by the division threshold Td = A, and A ± α × σ Is specified as the luminance distribution range to be enlarged.

In the above embodiment, the image correction unit 43 includes the bimodal determination unit 431. In another embodiment, the luminance distribution analysis unit 42 includes a bimodal determination unit 431. In this case, the luminance distribution analyzing unit 42 outputs the determination result of the bimodality together with the feature amount of the luminance distribution, and the image correcting unit 43 switches the processing with reference to the input determination result.

DESCRIPTION OF SYMBOLS 1 ... Image monitoring apparatus 2 ... Imaging part 3 ... Memory | storage part 4 ... Signal processing part 5 ... Output part 30 ... Background image 31 ... Object identification information 40 ... Background image generation means 41 ... Candidate area setting means 42 ... Luminance distribution analysis means 43 ... Image correction means 44 ... Edge extraction means 45 ... Object detection means 46 ... Anomaly detection means 431 ... Measuring means for bimodality

Claims (5)

An object detection device that detects an object from an image obtained by imaging a monitoring space,
An imaging unit for imaging the monitoring space;
A storage unit that stores a background image captured when the object does not exist in advance;
A luminance distribution analysis means for analyzing the background image and calculating a background side luminance distribution;
Bimodal determination means for determining whether the background side luminance distribution is a bimodal distribution unevenly distributed in a high luminance region and a low luminance region;
When the bimodal distribution is determined, each pixel of the input image input from the imaging unit is subjected to luminance conversion that expands the high luminance side distribution range in which the pixels in the high luminance region are distributed. A correction image is generated and a luminance conversion that expands a low luminance side distribution range in which pixels in the low luminance region are distributed is applied to each pixel of the input image to generate a low luminance side correction image, and the high luminance side correction image And an image correction means for generating an average image of the low luminance side corrected image,
Edge extraction means for extracting edge information from the input image;
Object detection means for detecting an object based on the edge information;
With
The object detection unit is characterized in that the edge information is extracted from the average image when the average image is generated.   The luminance distribution analyzing unit further analyzes the pixels of the input image corresponding to the pixels in the high luminance region to calculate the high luminance side distribution range, and calculates the distribution of the input image corresponding to the pixels in the low luminance region. The object detection apparatus according to claim 1, wherein the low-luminance distribution range is calculated by analyzing a pixel. Candidate area setting means for setting a candidate area of a size set in advance according to the object at a predetermined position in the input image;
The object detection apparatus according to claim 1, wherein the luminance distribution analysis unit performs the analysis on pixels in the candidate area.   The object detection device according to claim 3, wherein the candidate area setting unit sets the candidate area at a position where a difference is detected by comparing the input image with the background image. 5. The object according to claim 3, wherein the candidate area setting unit detects a human head area having an elliptical shape from the input image, and sets the candidate area at a position where the human head area is detected. Detection device.