JP6266461B2 - Object detection device - Google Patents

Description

  The present invention relates to an object detection device that detects a predetermined object from an input image.

  Search processing by a classifier or the like is performed in order to detect an image of an object such as a person from an input image taken by a surveillance camera or the like. In this search process, a window area is set at each position in the input image, and an image in the window area is input to a discriminator or the like. Then, a window region whose score output from the discriminator or the like exceeds a threshold is extracted as a candidate region for the target object. Here, since there is a tendency that candidate areas are also extracted around the true object, the candidate areas having overlap are grouped, and the candidate area having the maximum score for each group is selected as the object area.

  Conventionally, a predetermined common value has been used for any input image as a threshold value of a score used for extraction of candidate areas.

JP 2010-160640 A

  However, although the score tends to vary depending on the shooting environment of the input image (background complexity, lighting conditions, resolution, etc.), a common threshold is used for all input images. Depending on the shooting environment, a candidate area in which a part of the background is erroneously extracted frequently appears, which causes a decrease in detection accuracy of the target object.

  Specifically, depending on the shooting environment of the input image, a group consisting of only candidate areas with only the background appearing, and eventually the background becomes the object area, or the background is erroneously extracted between multiple objects. As a result, the number of target objects is often detected as a single target object, and the detection accuracy is reduced.

  However, if the threshold value is tightened in advance in accordance with a shooting environment in which a candidate region is likely to be erroneously extracted, the failure to detect an object in another shooting environment increases.

  This problem will be described with reference to examples shown in FIGS. FIG. 8 is a schematic diagram of an image 900 photographed in a relatively simple environment, and FIG. 9 is a schematic diagram of an image 910 photographed in a complicated environment. A rectangle in each image indicates a candidate area. FIGS. 8A and 9A show a case where candidate regions are extracted with a relatively low threshold. In the image 900 of FIG. 8A, candidate areas 901a, 901b, around a small (far) person image on the left side, around a central desk / chair, and around a large (near) person image on the right side, respectively. 901c is extracted. Like image 900, candidate areas 911a, 911b, and 911c are extracted from the left small person image, the center desk / chair, and the right large person image, respectively, in the image 910 of FIG. 9A. Has been. Here, the candidate areas 901b and 911b are obtained by erroneously extracting the vicinity of the chair, which is the background, as a person. In addition, the left and right human images of the image 910 in a complicated environment are partially concealed by a desk or the like constituting the background, and the score is lowered accordingly.

  In this example, it is conceivable to increase the extraction threshold value of the candidate region in order to delete the erroneously extracted candidate regions 901b and 911b. FIGS. 8B and 9B show how candidate areas are extracted by setting a common threshold higher than those in FIGS. 8A and 9A. Since the erroneous extraction candidate area basically has a low score, it is possible to delete the erroneous extraction candidate areas 901b and 911b by increasing the threshold. As a result, candidate areas 901a and 901c including a person image remain in the image 900 shown in FIG. 8B. On the other hand, in the image 910 shown in FIG. 9B, the candidate area 911a having a high degree of concealment is also deleted from the candidate areas corresponding to the person, and the detection of the person is omitted.

  As described above, it has been difficult to accurately detect an object from input images taken in various environments using a common threshold.

  The present invention has been made in view of the above-mentioned problems, and found that if there are misextractions in candidate areas, the score of the extracted candidate areas varies, and images were taken in various environments based on this knowledge. An object of the present invention is to provide an object detection device capable of detecting an object from an input image with high accuracy.

  An object detection apparatus according to the present invention detects an object region in which a predetermined object appears in an input image, and is an index representing the likelihood that the object exists in an attention region set in the input image A storage unit that stores in advance an index value calculation function for calculating a value using a feature amount extracted at various points in the input image, and sets the attention area at a plurality of positions in the input image Then, an index value calculation unit that calculates the index value in the attention area by the index value calculation function, and a target area that extracts the attention value that exceeds the predetermined first threshold among the attention areas A candidate region extraction unit; and a target region determination unit that determines the target region using the target candidate region extracted by the candidate region extraction unit, the candidate region extraction unit for each region of interest When the degree of variation of the index value exceeding the first threshold is equal to or greater than a predetermined erroneous extraction estimation threshold, a second threshold larger than the first threshold is set, and the index value is determined from the target candidate area. That are less than or equal to the second threshold are deleted.

  In another object detection apparatus according to the present invention, the candidate region extraction unit is a value smaller than a representative value of the index value that exceeds the first threshold value as the second threshold value, and the first threshold value Is set to a value that increases the difference with the representative value.

  In still another object detection apparatus according to the present invention, the candidate area extraction unit sets the second threshold value by adding a value corresponding to the variation degree to the first threshold value.

  In the target detection device according to another aspect of the present invention, the candidate area extraction unit may perform the first step when the variation degree of the index value is lower than the erroneous extraction estimation threshold and is equal to or less than a preset extraction omission estimation threshold. One threshold is decreased by a predetermined amount, and the target candidate region is re-extracted.

  According to the present invention, it is possible to adaptively set the extraction threshold value of the candidate area for each input image, so that the target can be detected with high accuracy from the input images taken in various environments.

1 is a schematic block configuration diagram of a person detection device according to an embodiment of the present invention. It is a schematic diagram which shows the example of an input image and a reduction image. It is the typical graph which plotted the score of the candidate area | region with the horizontal axis | shaft and the frequency on the vertical axis | shaft. It is a flowchart which shows operation | movement of the outline of the person detection apparatus which concerns on embodiment of this invention. It is a typical image explaining the subsequent process with respect to the candidate area | region extracted by the score calculation part. FIG. 10 is a schematic process flow diagram of a candidate area deletion unit. It is a schematic diagram explaining the downward correction of a 1st threshold value. It is a schematic diagram of the input image image | photographed in the comparatively simple environment. It is a schematic diagram of the input image image | photographed in the complicated environment.

  Hereinafter, embodiments of the present invention (hereinafter referred to as embodiments) will be described with reference to the drawings. The target detection apparatus according to the present embodiment is a person detection apparatus 1 that targets a person shown in an image as a detection target.

[Configuration example]
FIG. 1 is a schematic block diagram of a person detection device 1 according to the embodiment. The person detection device 1 includes an image input unit 2, a control unit 3, a storage unit 4, and an output unit 5. The image input unit 2, the storage unit 4, and the output unit 5 are connected to the control unit 3.

  The image input unit 2 is, for example, an imaging device such as a surveillance camera or a recording device such as a digital video recorder that records video, and outputs an image to the control unit 3. Hereinafter, an image input from the image input unit 2 to the control unit 3 is referred to as an input image. As will be described later, the control unit 3 performs processing for detecting a target in each input image. Therefore, it is not necessary to sequentially input images of a plurality of frames from the image input unit 2 to the control unit 3. Even when images are sequentially input, the input image does not need to be an image captured by a camera installed in a fixed place, and may be an image captured by a camera attached to a moving body. Further, the input image may be an image taken by a camera capable of panning, tilting, and zooming, such as a PTZ camera.

  The control unit 3 is configured using an arithmetic device such as a CPU (Central Processing Unit) or a DSP (Digital Signal Processor). The control unit 3 processes the input image from the image input unit 2 to determine the presence / absence of a person and outputs the determination result to the output unit 5. For this purpose, the control unit 3 reads out and executes a program from the storage unit 4, and executes an image reduction unit 30, a feature amount calculation unit 31, a score calculation unit 32, a candidate area deletion unit 33, an area group generation unit 34, and an object area. It functions as the calculation unit 35.

  The image reduction unit 30 reduces the input image at a plurality of preset magnifications in response to the various sizes of the person captured in the input image. As a result, it is possible to detect images of persons of various sizes without changing the size of the window region set for detecting the person in the image. For example, the image reduction unit 30 sequentially reduces the input image at a predetermined interval until it reaches a predetermined minimum width or height, and generates a reduced image. For example, the reduction ratio is set to 10 levels until the vertical and horizontal sizes are halved. For example, the image 100 shown in FIG. 2A is an input image of the original size, and the images 110 and 120 shown in FIGS. 2B and 2C are examples of input images obtained by reducing the image 100.

  The feature amount calculation unit 31 divides each of the original size input image and the reduced input image into predetermined block sizes, and calculates the feature amount for each block image. Conventionally known features such as Histograms of Oriented Gradients (HOG), Local Binary Pattern (LBP) features, Haar-like features, etc. It can be used alone or in combination.

  The score calculation unit 32 sets a window area (attention area) of a predetermined person size as a frame for detecting a person at each position in the original size input image and the reduced input image, and the window This is an index value calculation unit that calculates a score, which is a multi-valued index value representing the likelihood that an object exists in a region, by using an index value calculation function that has been learned in advance from an input image. For example, the score calculation unit 32 calculates the score for the window region by inputting the feature amount in each window region to the index value calculation function, or considers the posture variation that the person's arm or the like protrudes from the window region. Then, feature values inside and outside the window area including a predetermined range around the window area are input to the index value calculation function to calculate a score for the window area.

  In FIG. 2, an example of the rectangular window region 101 set in the images 100, 110, and 120 is indicated by a dotted line. The score calculation unit 32 repeatedly sets the window area 101 while gradually shifting it, and scans the entire image. For example, scanning of the window region 101 starts in the horizontal direction from the upper left of the image. The horizontal scanning is repeated while shifting the vertical position little by little.

  In this embodiment, the index value calculation function is a discriminator that discriminates between “persons” to be detected and those other than “persons”. The discriminator is learned in advance using a large number of images in which “people” are reflected and a large number of images in which “people” are not reflected, and is stored in a discriminator storage unit 40 described later. The score calculation unit 32 calculates a score by giving a feature amount to the classifier according to the position of the window region.

The score calculation unit 32 also has a function as a candidate region extraction unit for obtaining a candidate region (target candidate region) where a person exists in the input image together with a candidate region deletion unit 33 described later. Specifically, the score calculation unit 32 uses the rectangular information (position, width, height, and score in the input image) of the window region where the score exceeds a predetermined first threshold T ₁ as a candidate region, which will be described later. Store in the storage unit 41.

The candidate area deletion unit 33 sets a second threshold value T ₂ larger than the first threshold value T ₁ as a threshold value for the score based on the information on the candidate area stored in the candidate area storage unit 41, and the second threshold value T _The candidate area is deleted and narrowed down by ₂ .

The score of the candidate area tends to vary depending on the shooting environment such as background complexity, illumination state, and resolution, and this variation may increase false extraction. Then, there is a tendency that a significant difference occurs in the variation in scores between the case where there are many erroneous extractions and the case where there are not many erroneous extractions. Therefore, the candidate area deleting unit 33 dynamically sets the second threshold value T ₂ of the scores according to the distribution of the scores of the candidate region obtained for each input image.

  Here, how the score distribution of the candidate area differs depending on the environment will be described with reference to FIG. FIG. 3 is a schematic graph plotted with the score of the candidate region on the horizontal axis and the frequency on the vertical axis. FIG. 3 (a) shows the distribution of the scores of candidate areas in an input image with a person on a simple background, while FIGS. 3 (b) and 3 (c) show the candidate areas in an input image with a person on a complex background. The distribution of scores. For example, FIG. 3B shows a case where a part of the person is concealed, and FIG. 3C shows a case where an area having characteristics similar to the person exists in a complicated background. .

  In a simple background, the score is concentrated at a high frequency distribution (mountain 300a in FIG. 3A). On the other hand, in a state where the background is erroneously extracted as a person with a complicated background (FIG. 3C), the distribution varies to the side where the score is lower than the mountain 300a in FIG. 3A. For example, FIG. It has been experimentally confirmed that a mountain 301c having a lower frequency appears at a position having a lower score than the mountain 300c corresponding to the mountain 300a of a).

That is, if the number of erroneous extractions increases depending on the shooting environment, the score distribution varies greatly. Therefore, the candidate area deletion unit 33 determines whether or not many erroneous extractions are included from the score distribution of the candidate areas, and then uses the second threshold value T ₂ higher than the first threshold value T ₁ for erroneous extraction. The candidate area is deleted.

The candidate area deleting unit 33 to calculates the stored candidate regions in the candidate area storage unit 41, i.e. the degree of dispersion of the distribution of scores for the window area where the score exceeds the first thresholds T ₁ by the score calculation unit 32, The calculated variation degree is compared with an erroneous extraction estimation threshold. Candidate area deleting unit 33, as false if the degree of variation is erroneous extraction estimated threshold above extraction contained many sets a second threshold value T ₂ of the the score, the score is the second threshold value T ₂ or less A candidate area is deleted from the candidate area storage unit 41. On the other hand, when the degree of variation is less than the erroneous extraction estimation threshold, the deletion is not performed.

The candidate area deleting unit 33 calculates, for example, the variance value σ ² as the degree of variation. For example, the erroneous extraction estimation threshold value V _H for the variance value can be set to 0.04.

Here, in a state where a part of the person is concealed in a complicated background, the score of the person is lowered, and a low mountain can be generated as indicated by 301b in FIG. 3B. Therefore even no more erroneous extraction may equal to or greater than the extraction estimated threshold erroneous degree of variation, to control the second threshold value T ₂ so that accidental deletion low mountain 301b is desirable. Here, the mountain 301b of the person whose score has decreased appears in the vicinity of the peak of the score distribution, and the peak of the score distribution can be approximated by an average value of the score distribution. That is, when the variance σ ² is large, although the number of erroneous extractions included in the candidate region is relatively large, the candidate region has a score that exceeds at least the first threshold value T ₁ and is a correctly extracted candidate The territory is dominant. Therefore, the average score value m ₁ of all candidate areas indicates the vicinity of the peak of the peak (the peak 300 in FIG. 3) of the frequency distribution formed by the correctly extracted candidate area scores. Therefore, the candidate area deleting unit 33 sets a second threshold T ₂ corresponding to the average value m ₁ of the score distribution. For example, T ₂ = κ ₀ m ₁ can be set using a preset coefficient κ ₀ . κ ₀ is a value greater than 0 and less than 1, and can be set to 0.4, for example.

Furthermore, the second threshold value T ₂ corresponding to the average value m ₁ can be set based on the value of the first threshold value T ₁ . Candidate area deleting unit 33 Specifically, an average value m ₁ is less than value of in excess of first thresholds T ₁ of the score calculated for each window area, the difference between the first threshold value T ₁ is the average A value that increases according to the value m ₁ is set as the second threshold value T ₂ . For example, let m ₂ be the average score value of the candidate regions whose score is lower ξ%. In a shooting environment where there are many erroneous extractions, the variation of the score of the candidate area becomes large and m ₂ becomes low. Therefore, T ₂ can be determined by the following equation using the difference between m ₁ and m _{2 that} varies depending on the shooting environment.
T ₂ = T ₁ + κ ₁ (m ₁ −m ₂ ) (1)

κ ₁ is a positive coefficient and can be set to 2/3, for example. Incidentally, the coefficient kappa, through preliminary experiments, at least T ₂ may be set to a low enough to not to delete the highest score of the region group including the true person area. For example, ξ = 10 (%) can be set. (1) According to the equation, when it is erroneously extracted is large shooting environment, the second threshold value T ₂ is set high, the detected candidate region erroneous background is easily removed. On the other hand, if extraction errors is less shooting environment, it is possible to the second threshold value T ₂ is set low, promote the removal of erroneously extracted candidate region background while suppressing the detected impair the person area.

Or, the candidate area deleting unit 33 may also set the second threshold value T ₂ by adding a value corresponding to the degree of variation in the first threshold value T _1. In this case, for example, the candidate area deleting unit 33 determines T ₂ by the following expression using the variance value σ ² .
T ₂ = T ₁ + κ ₁ · SQRT (σ ² ) (2)

However, SQRT () represents a square root. Further, κ ₂ is a positive coefficient, and it may be set to a low value that does not delete the highest score of the region group including at least T _{2 that} is a true human region through experiments.

As the degree of score variation, the standard deviation of the score, the difference between the maximum value and the minimum value of the score, or the like can be used instead of the distribution of the score of the candidate area described above. Further, instead of the average values m ₁ and m ₂ described above, a median value or a mode value which is another representative value can be used.

  The region group generation unit 34 and the target region calculation unit 35 constitute a target region determination unit that determines a person region (target region) from the candidate regions left by the candidate region deletion unit 33.

  The area group generation unit 34 performs processing for grouping candidate areas stored in the candidate area storage unit 41 that are attributed to the same person. Specifically, the region group generation unit 34 generates region group information by assigning the same label to candidate regions having a predetermined overlap or more. The label information of each candidate area defined by the area group generation unit 34 is stored in the candidate area storage unit 41 together with the rectangle information and the score.

  The object area calculation unit 35 obtains a final person area from the candidate areas stored in the candidate area storage unit 41. The object area calculation unit 35 determines one person area for each group calculated by the area group generation unit 34, and stores area information of the person area in the object area storage unit 42 together with the score. For example, the object region calculation unit 35 selects one candidate region having the maximum score in each region group as the final person region. Alternatively, the target object area calculation unit 35 calculates the final person area by averaging the candidate areas constituting the area group for each area group.

  When at least one final person area is detected from the input image, the control unit 3 outputs the information to the output unit 5.

  The storage unit 4 is a storage device such as a ROM (Read Only Memory), a RAM (Random Access Memory), and a hard disk, and stores programs and data used by the control unit 3. The storage unit 4 inputs and outputs these programs and data to and from the control unit 3. The storage unit 4 has functions as a discriminator storage unit 40, a candidate region storage unit 41, and an object region storage unit 42.

The discriminator storage section 40 calculates an index value for calculating a score representing the likelihood that the target exists in the window area set in the input image using the feature amount extracted in each block in the input image. calculation function, and stores in advance a first threshold value T _1. As described above, the index value calculation function is a discriminator. Specifically, the index value calculation function is obtained by applying a support vector machine (SVM) to learning images of human beings and learning images other than human beings. The discriminator parameters are stored in the discriminator storage unit 40. When linear SVM is used as the learning algorithm, the parameter of the discriminator is a weight vector generated from the learning image. This weight vector is a weight for each element of the feature amount. The weight vector is adjusted in learning so that the inner product of the weight vector and the feature amount extracted from the learning image is identified as a person when it is greater than 0, and when it is equal to or less than 0, it is identified as a person other than the person. The value of the inner product of the quantity and the weight vector represents the score. Therefore, the threshold value for identifying a score of non-human and human is zero in principle, usually, the first thresholds T ₁ may be set to 0. However, in order to reduce the error identified as other than human to human, the first thresholds T ₁ may be set to a value smaller than 0.

  As the learning algorithm of the discriminator, various conventionally known ones such as the AdaBoost method can be used in addition to the SVM.

  In addition, a pattern matching device can be used instead of the discriminator. In this case, the score is the reciprocal of the distance between the average pattern of feature values extracted from the human learning image and the feature value of the input image, and the index value. The calculation function can be a function having the score as an output value and the feature quantity of the input image as an input value.

  The candidate area storage unit 41 stores information on the candidate area of the person obtained by the score calculation unit 32 and label information indicating the area group obtained by the area group generation unit 34. The candidate area information includes the position / size and score of the window area as described above. For example, the upper left coordinates of the rectangle forming the window area are stored as the position of the window area in the input image.

  The object area storage unit 42 stores information on a person area that is finally determined by the object area calculation unit 35 to have a person. The information on the person area is the rectangular information (the coordinates and size of the upper left corner of the rectangle) and the score of the person area in the input image, similarly to the information on the candidate area.

  The output unit 5 receives the result of the object area calculation unit 35 and displays an input image on an external display device such as a display or sends an abnormal signal to the center device.

[Example of operation]
Next, the operation of the person detection device 1 will be described. FIG. 4 is a flowchart showing a schematic operation of the person detection apparatus 1. When an image is input from the image input unit 2 (step S10), the control unit 3 creates a reduced image by reducing the input image at a plurality of magnifications by the image reduction unit 30 (step S20). For example, as illustrated in FIG. 2, reduced images 110 and 120 are generated from the input image 100.

  The feature quantity calculation unit 31 calculates the feature quantity at each location in the image for each of the input image and the plurality of reduced images (step S30).

The score calculation unit 32 calculates a score corresponding to the window region set in each place in the image by using the feature amount calculated by the feature amount calculation unit 31 and the classifier stored in the classifier storage unit 40. score from the the window region is stored in the candidate area storage unit 41 in excess of the first thresholds T ₁ as a candidate region of a person (step S40).

  In FIG. 2, the window area 101 is indicated by a dotted rectangle, and an example of a candidate area is indicated by a solid rectangle having a size corresponding to the window area. In the image 100, candidate regions 102a and 102b are extracted around a small (far) person image on the left side. In the image 110, a candidate area 112 is detected around a central desk / chair, and in the image 120, candidate areas 122a and 122b are extracted around a large (near) person image on the right side. As shown in FIG. 2, a plurality of overlapping candidate regions can be extracted for one image such as a person.

  FIG. 5 is a schematic image for explaining the subsequent processing for the candidate area extracted by the score calculation unit 32. The image of FIG. 5 shows the same content as that shown in FIG. 2, and the image 130 of FIG. 5A displays the candidate areas of the images 100, 110, and 120 together on one image. Is. The image 130 is the same size as the input image 100, and the candidate areas 102a and 102b of the image 100 become candidate areas 131a and 131b on the image 130 at the same magnification. On the other hand, the candidate areas 112, 122a, and 122b in the reduced image become candidate areas 132, 133a, and 133b normalized to the magnification of the input image 100, respectively.

Candidate area deleting unit 33, from the score stored in the candidate area storage unit 41, performs for determining whether or not performing the process of deleting the candidate area, the case of performing the deletion processing sets the second threshold value T ₂ deletes the _{T 2} following candidate region from the candidate region storing unit 41 (step S50).

FIG. 6 is a schematic process flow diagram of the candidate area deletion unit 33. The operation of the candidate area deletion unit 33 will be described with reference to FIG. The candidate area deletion unit 33 calculates the variance value σ ² as the degree of variation of the score of the candidate area stored in the candidate area storage unit 41 (step S501). If the variance value σ ² is larger than the predetermined extraction omission estimation threshold v _L (in the case of “NO” in step S502), the candidate area deletion unit 33 further sets the variance value σ ² to the predetermined erroneous extraction estimation threshold v _H. (Step S503).

If the variance value σ ² is greater than or equal to the erroneous extraction estimation threshold v _H (“YES” in step S503), a second threshold T ₂ greater than the first threshold T ₁ is set (step S504).

On the other hand, when the variance value σ ² is less than the erroneous extraction estimation threshold value V _H (in the case of “NO” in step S503), the process by the candidate area deleting unit 33 is terminated without setting the second threshold value T ₂ , Proceed to step S60 of FIG. That is, in this case, the control unit 3 does not delete the candidate area.

If the variance value σ ² is equal to or higher than V _H , the second threshold value T ₂ is set as described above, and the candidate area deletion unit 33 performs processes S505 to S508 using this T ₂ to delete the candidate area. This process is performed as a loop process that is repeated for each candidate area stored in the candidate area storage unit 41 one by one. Specifically, selecting the candidate region loop processing has not been processed (step S505), the score of the selected candidate region as a processing target is determined whether the second threshold value T ₂ or less (step S506), If the second threshold value _{T 2} or less (if at step S506 is "YES"), deletes the candidate area from the candidate area storage unit 41 (step S507). If the score is greater than the second threshold value _{T 2} (when step S506 is "NO"), the candidate region is not removed, (returns from step S507 to S505) and determines the next candidate area. If processing has been completed for all candidate areas, that is, if no unprocessed candidate areas exist and cannot be selected in step S505 (“NO” in step S508), the process proceeds to step S60 in FIG.

Now, ( "YES" in step S502) if the variance value sigma ² is the following extraction failure estimated threshold _{v L,} the candidate area deleting unit 33 downward adjustment of the first threshold value _{T 1.} Specifically, the candidate area deleting unit 33 decreases by a predetermined amount a first threshold value T _1. And the control part 3 starts a process again from step S40. That is, performs the processing of step S40 by using the first thresholds T ₁ to the score calculation unit 32 is adjusted downward, and stored in the candidate area storage unit 41 re-extracts the candidate region, the candidate area deleting unit 33 is re-extracted The above-described processing of step S50 is performed for the candidate area.

The extraction omission estimation threshold v _L is set based on the variation of the score assumed for the human image, that is, the variance value in the present embodiment. That is, as shown in the distribution shown in FIG. 3A, the score obtained by scanning a human image in which no concealment or the like is generated in the window region inherently varies to some extent without erroneous extraction. Have. As the extraction omission estimation threshold v _L , a value that is less than the original minimum variation degree of the score distribution of the candidate region is set in advance based on experiments. Specifically, if the first threshold value T ₁ is set to the position of the peaks 300 of the distribution of scores resulting from the figures as shown in FIG. 7, the first thresholds T ₁ following part of the mountain 300 candidates will be extracted as a region, the degree of variation of a portion exceeding the first threshold value T ₁ of the the mountains is smaller than the degree of variation in overall mountain. Downgrades first thresholds T ₁ modifies the T ₁ set in this way the wrong position, to be able to extract the entire mountain 300.

Incidentally, if the score exceeds a first threshold value T ₁ is not more than a few possible estimate the presence of a mountain 300 due to figures, the determination and the first threshold value T ₁ of the downward revision of step S510 in step S502 is omitted Is preferred. For example, if score exceeds a first threshold value T ₁ is such that a few may be omitted. Although not shown in FIG. 3, there is a mountain of distribution due to the background in the low score area. If this distribution of mountain score range exceeding the background is low, the first threshold value T ₁ of the configuration be a suitable small score beyond that dispersion can be reduced. Thus case, T ₁ to T ₁ to a position including As you downgraded mountain background can occur may be reduced. Therefore, if there is the possibility downward revision T ₁ is omitted. In this case, basically, the processes in steps S503 to S508 are also omitted, and the process proceeds to step S60.

  An image 140 in FIG. 5B shows the processing result of the candidate area deletion unit 33 for the image 130, and the candidate area 132 caused by the background is deleted.

  When the processing of the candidate area deletion unit 33 described above is completed, the area group generation unit 34 calculates the degree of overlap between candidate areas stored in the candidate area storage unit 41, and the degree of overlap between candidate areas is determined in advance. A group (region group) composed of candidate regions equal to or greater than the group determination threshold is generated, and a label number indicating the group is added to the candidate region information and stored in the candidate region storage unit 41 (step S60 in FIG. 4).

The degree of overlap is calculated by, for example, (the area of the common area between candidate area A and candidate area B in the input image) / (the smaller of the areas of candidate area A and candidate area B in the input image). The Also,
The degree of overlap can also be calculated by (area of common area between candidate area A and candidate area B in input image) / (area of sum area of candidate area A and candidate area B in input image). . The group determination threshold can be set to 0.5, for example.

  In addition, in order to avoid that candidate areas relating to a plurality of people in the vicinity are combined into one group, a candidate area having a high score is set as the core of the group with priority, and the candidate area and the candidate area overlap. It is desirable to group other candidate areas whose degrees are greater than or equal to the group determination threshold.

  For this purpose, for example, the area group generation unit 34 first arranges the candidate areas stored in the candidate area storage unit 41 in descending order of the scores to initialize the unallocated list, and sets the candidate area at the head of the unallocated list as the core of the group To do. Next, the area group generation unit 34 calculates the degree of duplication of each of the second and subsequent candidate areas in the unallocated list with respect to the candidate area at the top of the unallocated list, and the candidate areas having the degree of duplication equal to or greater than the group determination threshold and unallocated The same label number is assigned to the candidate area at the head of the list and stored in the candidate area storage unit 41, and these candidate areas are deleted from the unallocated list. Thereafter, the candidate area at the head of the updated unallocated list is sequentially set as a nucleus, and label number allocation is repeated until there are no candidate areas in the unallocated list.

  When the processing of the area group generation unit 34 described above is completed, the object area calculation unit 35 obtains a final person area and stores it in the object area storage unit 42 (step S70 in FIG. 4).

  An image 150 in FIG. 5C shows the processing result of the object area calculation unit 35 for the image 140, and the candidate area having the maximum score among the candidate areas 133a and 133b constituting the group with the label number “0”. 133a is selected as the person area, and the candidate area 131a having the maximum score is selected as the person area among the candidate areas 131a and 131b constituting the group with the label number “1”.

  If there is even one person in the image after calculating the person area in step S70 (in the case of “YES” in step S80), for example, the output unit 5 detects the information of the detected person area and the person area is detected. An abnormal signal including the input image thus transmitted is sent to the center device (step S90).

  As described above, in the present invention described using the embodiment, the final detection result is obtained by setting the second threshold value of the score for deleting the candidate area according to the environment estimated from the score of the candidate area for each input image. Accuracy can be improved.

  DESCRIPTION OF SYMBOLS 1 Person detection apparatus, 2 Image input part, 3 Control part, 4 Storage part, 5 Output part, 30 Image reduction part, 31 Feature-value calculation part, 32 Score calculation part, 33 Candidate area deletion part, 34 Area group generation part, 35 object region calculation unit, 40 classifier storage unit, 41 candidate region storage unit, 42 object region storage unit.

Claims (4)

A target detection device for detecting a target region where a predetermined target appears in an input image,
An index value calculation function for calculating an index value representing the likelihood that the target exists in a region of interest set in the input image using feature amounts extracted at various points in the input image is stored in advance. Storage unit
An index value calculation unit that sets the attention area at a plurality of positions in the input image and calculates the index value in the attention area by the index value calculation function;
A candidate area extraction unit that extracts, as a target candidate area, an area in which the index value exceeds a predetermined first threshold among the attention areas;
A target region determination unit that determines the target region using the target candidate region extracted by the candidate region extraction unit;
With
The candidate region extraction unit, when the degree of variation of the index value for each region of interest exceeding the first threshold is equal to or greater than a predetermined erroneous extraction estimation threshold, a second threshold greater than the first threshold And deleting the index value equal to or less than the second threshold value from the target candidate area,
An object detection device characterized by the above.   The candidate area extraction unit is a value that is smaller than a representative value of the index value that exceeds the first threshold value as the second threshold value, and a difference from the first threshold value increases according to the representative value. The object detection device according to claim 1, wherein a value is set.   The target detection apparatus according to claim 1, wherein the candidate area extraction unit sets the second threshold value by adding a value corresponding to the degree of variation to the first threshold value.   The candidate area extraction unit reduces the first threshold by a predetermined amount when the degree of variation of the index value is lower than the erroneous extraction estimation threshold and equal to or less than a preset extraction omission estimation threshold, The object detection apparatus according to claim 1, wherein the target candidate area is re-extracted.

Images