JP5106356B2 - Image monitoring device - Google Patents

Description

  The present invention relates to an image monitoring apparatus that detects an object that has entered a monitoring area from an image obtained by photographing the monitoring area.

  Conventionally, an object detection device has been proposed in which an image is input to a classifier trained in advance using an image in which the target object exists and an image in which the target object does not exist, and the target object in the image is detected by the classifier. ing.

For example, Patent Document 1 discloses a target object detection device having a discriminator including a plurality of weak discriminators subjected to collective learning by boosting and an adder that performs weighted majority determination of these outputs.
Patent Document 2 discloses a subject identification device having an integrated classifier having a plurality of classifiers and an identification result integration unit. Patent Document 2 discloses a learning discriminating type discriminator in which each discriminator discriminates by learning learning data based on a predetermined subject image or both a predetermined subject image and an image other than the subject. It is described that.
Furthermore, Patent Document 3 discloses a person tracking device. The person tracking apparatus estimates the state transition of the person using the particle filter, and generates N hypotheses indicating the state of the person at time t. The person tracking device selects a classifier to be applied to the generated hypothesis among a plurality of classifiers, integrates the N hypotheses based on the identification result of the classifier, and determines the combined hypothesis. Use to track a person. Patent Document 3 describes that each classifier is a cascade type AdaBoost-based classifier that uses Haar-like features for hypothesis evaluation.

JP 2005-157679 A JP 2006-18707 A JP 2008-26974 A

  In the object detection apparatus as described above, an object to be detected is learned in advance by a classifier. Therefore, when it is not assumed that the object detection device is installed at a specific location at the time of learning the discriminator, the learning result of the prior learning may match the installation location depending on the installation location of the object detection device. There was a risk that it would not be. For example, when an object (for example, clothes hung on a mannequin or a hanger) similar to a detection target object (for example, a person) included in the image used for learning exists in the monitoring area, the object detection device There is a risk of erroneous determination that a detection target object exists in the monitoring image obtained by photographing the monitoring area.

  Therefore, an object of the present invention is to provide an image monitoring apparatus capable of preventing erroneous detection of a similar object as a detection target object even when an object similar to the detection target object exists in the monitoring region. .

  This invention for solving this subject provides the image monitoring apparatus which detects a detection target object from the image which image | photographed the specific monitoring area | region. The image monitoring apparatus includes an image input unit that acquires a monitoring image obtained by photographing a monitoring region, a storage unit that stores a background image that is an image obtained by photographing the monitoring region in a situation where a detection target does not exist in the monitoring region, and a monitoring It has a discriminator that calculates the degree of likelihood of a detection object in a predetermined area of the image and the background image, and the discriminator calculates the degree of the predetermined area of the monitoring image acquired by the image input unit as the reliability. And an identification unit that calculates the degree of the predetermined area of the background image corresponding to the predetermined area of the monitoring image as a similarity by the identifier, and the relative relationship between the reliability and the similarity, And a determination unit that determines whether or not a detection target exists.

  In the image monitoring apparatus according to the present invention, it is preferable that the determination unit determines that the detection target exists in the monitoring area when the difference obtained by subtracting the similarity from the reliability is larger than a predetermined threshold.

  Furthermore, in the image monitoring apparatus according to the present invention, the classifier included in the classifier is a cascade type classifier in which a plurality of classifiers are connected in series, and each of the plurality of classifiers connected in series is a predetermined classifier. The identification result indicating whether or not the detection target exists in the region of the output is output at the same time, and the cascade type discriminator continues until the discriminator at any stage outputs the identification result that the detection target does not exist. The discriminator at the stage outputs the discrimination result, and the reliability and similarity are the number of stages of the last discriminator that output the discrimination result that the detection target exists in the predetermined area in the cascade discriminator. It is preferable.

  Furthermore, in the image monitoring apparatus according to the present invention, the classifier included in the classifier is a parallel classifier in which a plurality of classifiers are connected in parallel, and each of the plurality of classifiers connected in parallel is a predetermined classifier. The identification result indicating whether or not the detection target exists in the region of the output is output, and the parallel type discriminator outputs the identification result that the detection target exists among the plurality of classifiers connected in parallel. It is preferable that the number of discriminators be reliability and similarity.

  Furthermore, in the image monitoring apparatus according to the present invention, the discriminator included in the discriminating unit has a degree of coincidence between the feature information of the detection target extracted from the predetermined area and the feature information of the detection target stored in the storage unit. It is preferable to calculate reliability and similarity by inputting.

  The image monitoring apparatus according to the present invention further includes a change area extracting unit that performs a difference process between the monitor image and the background image to extract a change area having a luminance difference between the monitor image and the background image, and It is preferable that the unit calculates the reliability and the similarity with the change area as a predetermined area.

  The image monitoring apparatus according to the present invention has an effect that even when an object similar to the detection target object exists in the monitoring area, it is possible to prevent erroneous detection of the similar object as the detection target object.

Hereinafter, an image monitoring apparatus according to an embodiment of the present invention will be described with reference to the drawings.
This image monitoring apparatus detects a detection target object by inputting a monitoring image obtained by capturing a monitoring area to a discriminator previously learned about the detection target object. In particular, this image monitoring apparatus stores an image obtained by photographing a monitoring area as a background image when it is known that an object to be detected does not exist in the monitoring area. Then, the image monitoring apparatus calculates a similarity that is a degree of confusingness with the detection target obtained by inputting the background image to the discriminator. When determining the presence of the detection target object by inputting the monitoring image to the discriminator, it is difficult to detect the detection target object for a region in the monitoring image corresponding to a region having a high similarity in the background image. To do. As a result, the image monitoring apparatus prevents erroneous detection of an object that exists in the monitoring area and is confused with the detection target object as the detection target object. In the present embodiment, the image monitoring apparatus detects an intruder that has entered the monitoring area as a detection target object. However, the detection target object in the present invention is not limited to a person, and may be any object that can be preliminarily learned as a detection target by assuming that it enters the monitoring area in advance.

  FIG. 1 is a diagram showing a schematic configuration of an image monitoring apparatus 1 according to an embodiment of the present invention. As illustrated in FIG. 1, the image monitoring apparatus 1 includes an image input unit 2, a storage unit 3, an output unit 4, and a signal processing unit 5.

The image input unit 2 is a two-dimensional detector composed of a photoelectric converter having sensitivity to visible light or near-infrared light, such as a CCD, and an image for forming an image of a monitoring region on the two-dimensional detector. It has an optical system. Further, the image input unit 2 performs imaging at regular time intervals (for example, 1/5 second). Then, the image input unit 2 generates a monitoring image obtained by capturing the monitoring area as a grayscale image having, for example, horizontal 340 pixels × vertical 240 pixels, and the luminance of each pixel expressed by 256 gradations. The image input unit 2 may have an illumination light source that illuminates the monitoring area, such as a plurality of infrared LEDs, in order to keep the monitoring area at a certain level of illuminance.
Note that the camera that captures the monitoring area may be provided separately from the image monitoring apparatus 1. In this case, the camera is connected to the image monitoring apparatus 1 via a communication network, and the image input unit 2 can be a communication interface for the communication network connecting the camera and the image monitoring apparatus 1 and its control circuit.
The monitoring image acquired by the image input unit 2 is sent to the signal processing unit 5.

  The storage unit 3 includes a storage device such as a nonvolatile semiconductor memory such as a flash memory, a volatile semiconductor memory, or a magnetic disk (HDD). The storage unit 3 stores various programs and data used in the image monitoring apparatus 1. In addition, the storage unit 3 stores, as a background image, an image of a monitoring area in which an intruder is not obtained, which is acquired from the image input unit 2 when the image monitoring device 1 is activated, for example.

The output unit 4 includes a light emitting device such as a speaker or an LED. When an intruder into the monitoring area is detected by the signal processing unit 5, the output unit 4 emits a sound according to an alarm signal from the signal processing unit 5, or emits an alarm by emitting or blinking.
The output unit 4 may include a communication interface connected to a communication network such as a local area LAN or a public line network and a control circuit thereof. The output unit 4 receives from the signal processing unit 5 an abnormality detection signal indicating that an intruder into the monitoring area has been detected by the signal processing unit 5 or a monitoring image showing the intruder, and the abnormality detection signal. Alternatively, the monitoring image may be output to a security device or a monitoring center device connected to the image monitoring device 1 via a communication network.

  The signal processing unit 5 has an embedded microprocessor unit and its peripheral circuits. The signal processing unit 5 controls the entire image monitoring apparatus 1. The signal processing unit 5 detects an intruder in the monitoring area based on the monitoring image received from the image input unit 2 and the background image stored in the storage unit 3. For this purpose, the signal processing unit 5 includes a change area extraction unit 51, an identification unit 52, a determination unit 53, and a background image update unit 54. Each of these units included in the signal processing unit 5 is mounted as a functional module of a program that operates on the microprocessor unit, for example.

The change area extraction unit 51 extracts an area where an intruder may appear in the latest monitoring image as a person candidate area. Therefore, the change area extraction unit 51 obtains a luminance difference between corresponding pixels between the latest monitoring image and the background image, and calculates a difference image in which the pixel value of each pixel is represented by the absolute value of the luminance difference. create. Then, for each pixel of the difference image, the change area extraction unit 51 sets a pixel whose absolute value of the luminance difference is equal to or greater than a predetermined threshold as a variable pixel, while selecting a pixel whose absolute value of the luminance difference is less than the predetermined threshold. A difference binarized image binarized so as to be a background pixel is created. The predetermined threshold value can be, for example, an average value of the pixel values of the difference image. The change area extraction unit 51 detects, as a change area, an area in which change pixels are connected together by using a known labeling method or the like. Then, the change area extraction unit 51 sets a circumscribed rectangular area of the change area as a person candidate area.
The change area extraction unit 51 passes information indicating the person candidate area, for example, the coordinates of the upper left end point of the person candidate area and the sizes in the vertical and horizontal directions to the identification unit 52.

The process of the change area extraction unit 51 will be described with reference to FIGS. 2 (a) to 2 (c). FIG. 2A shows a background image 200. The planting 201 is reflected in the background image 200. FIG. 2B shows the latest monitoring image 205. The monitoring image 205 shows a planting 207 corresponding to the planting 201 shown in the background image 200. Further, the intruder 206 is shown in the monitoring image 205.
FIG. 2C shows a difference binarized image 210 created by the change region extraction unit 51 performing binarization by performing difference processing between corresponding pixels between the monitoring image 205 and the background image 200. Show. In the difference binarized image 210, a rectangular area corresponding to the intruder 206 shown in the monitoring image 205 is extracted as the person candidate area 211. Furthermore, if the planting existing in the monitoring area is shaken by the influence of wind or the like, the planting 201 in the background image 200 and the planting 207 in the monitoring image 205 do not completely match. Therefore, in the difference binarized image 210, an area 212 whose luminance has changed due to planting shake is also extracted as a person candidate area. Therefore, in order to prevent erroneous detection of an intruder based on a person candidate area in which an object that is not a detection target is shown, the signal processing unit 5 determines whether or not each extracted person candidate area is an intruder. It is necessary to identify.

  Therefore, the identification unit 52 obtains a reliability indicating the degree of intruder-likeness of the person candidate area by inputting information on any person candidate area in the monitoring image to the classifier. Further, the identification unit 52 inputs information related to the region in the background image corresponding to any one of the candidate human regions to the candidate human region by inputting the information to the same classifier used for calculating the reliability. Find the degree of similarity that represents the degree of confusion with the intruder in the area.

  FIG. 3 shows a schematic configuration of a discriminator used in the present embodiment. As shown in FIG. 3, the cascade type discriminator 300 included in the discriminating unit 52 has a configuration in which four discriminators 310, 320, 330, and 340 are connected in series. When each classifier 310, 320, 330, and 340 receives information on a person candidate area in the monitoring image or an area in the background image corresponding to the person candidate area, an intruder appears in that area. The identification result of whether or not there is output. Then, the cascade type discriminator 300 executes discrimination in order from the first stage discriminator 310 until one of the discriminators outputs an identification result indicating that no intruder is shown in the area. , Repeat identification. Cascade type discriminator 300 then outputs the number of stages of the last discriminator that has output the discrimination result that an intruder appears in the area. The discriminator 52 sets the number of discriminators output from the cascade discriminator 300 as the reliability Ri or the similarity Rb. The identification unit 52 obtains the reliability and the similarity by using the cascade type classifier 300 as described above, so that the number of classifiers that perform the classification process is reduced in a region that is not like an intruder. Therefore, the signal processing unit 5 can perform the detection process in a short time by reducing the calculation amount as the entire intruding object detection process.

  Hereinafter, an example of the operation of the identification unit 52 when the reliability is obtained by inputting information related to the person candidate region in the monitoring image to the cascade type identification device 300 will be described. Note that the identification unit 52 obtains the similarity by inputting information about the area in the background image corresponding to the person candidate area to the cascade classifier 300 instead of the information about the person candidate area in the monitoring image. be able to.

  First, the identification unit 52 inputs information related to the person candidate area in the monitoring image to the first-stage classifier 310. When the discriminator 310 outputs a discrimination result (False) that no intruder is shown in the area, the cascade discriminator 300 outputs 0 as the reliability. On the other hand, when the discriminator 310 outputs a discrimination result (True) that an intruder is captured in the area, information regarding the area in the monitoring image corresponding to the person candidate area is stored in the second-stage discriminator 320. Entered. When the discriminator 320 outputs a discrimination result (False) indicating that no intruder is shown in the area, the cascade discriminator 300 outputs 1 as the reliability. On the other hand, when the discriminator 320 outputs a discrimination result (True) indicating that an intruder appears in the area, information on the person candidate area in the monitoring image is input to the third-stage discriminator 330. Thereafter, the same processing is repeated until identification by the fourth-stage classifier 340 is performed. When the discriminator 340 outputs an identification result (True) indicating that an intruder appears in the area, the cascade discriminator 300 outputs 4 as the reliability.

  In the present embodiment, the classifiers 310, 320, 330, and 340 at each stage are Adaboost classifiers using Haar-like features. The Haar-like feature is a luminance difference between a plurality of adjacent rectangular areas arbitrarily set in the input image area. The Adaboost classifier includes a plurality of weak classifiers and a strong classifier that is determined by integrating the determination results of the weak classifiers. Therefore, in the present embodiment, the information related to the person candidate area in the monitoring image that is input to each of the classifiers 310 to 340 is the luminance value of each pixel included in the person candidate area. Similarly, the information related to the background image area corresponding to the person candidate area is the luminance value of each pixel included in the background image area corresponding to the person candidate area.

  Each weak classifier calculates different Haar-like features from the input image area, and whether or not an intruder appears in the input image area based on the calculated Haar-like feature Is output. For example, if the weak classifier determines that an intruder appears in the input image area, it outputs 1, while if it determines that no intruder appears in the input image area, -1 Is output. On the other hand, the strong classifier weights the output result from each weak classifier, and obtains the weighted sum. Then, when the obtained weighted sum is higher than a predetermined threshold (for example, 0), the strong classifier determines that an intruder appears in the input image area.

In addition, the weak classifier that uses Haar-like features for identification in which position in the image area is used, and the weight for the weak classifier includes a plurality of sample images that do not show people and people It is determined by prior learning using a plurality of sample images. The outline of the learning procedure is as follows.
(1) For all the weak classifiers that can be used, a threshold value is set for identifying whether or not an intruder appears in the input image area based on the value of the Haar-like feature.
(2) The weight for each sample image is determined. The initial value of the weight is the same value for each sample image.
(3) Each sample image is input to all weak classifiers, and the weights assigned to the sample images that have failed to be identified for each weak classifier are summed. The sum of the weights is used as the evaluation value.
(4) The weak classifier having the smallest evaluation value is selected as the weak classifier used in the Adaboost classifier. Then, the weight assigned to the output of the weak classifier selected from the evaluation value is determined.
(5) The weight of the sample image for which the selected weak classifier has failed to be identified is increased.
(6) Repeat steps (3) to (5).
For details of the Haar-like feature and the Adaboost discriminator, see, for example, Paul Viola and Michael Jones, "Rapid Object Detection using a Boosted Cascade of Simple Features", IEEE CVPR, vol.1, pp.511-518, 2001. Is disclosed.
In addition, the classifiers at each stage can construct different classifiers by learning in advance using different sample images, for example.

  The identification unit 52 passes the reliability obtained for each candidate person region in the monitoring image and the similarity obtained for the region in the background image corresponding to each person candidate region to the determination unit 53.

The determination unit 53 determines whether or not an intruder appears in the corresponding person candidate area based on the relative relationship between the reliability and the similarity received from the identification unit 52. Here, the determination unit 53 increases the criteria for determining that an intruder is shown in the person candidate region corresponding to the similarity as the obtained similarity is larger. For example, the determination unit 53 has a degree of reliability obtained for the target person candidate area that is greater than a predetermined threshold Th1 and a degree of similarity obtained for the area in the background image corresponding to the person candidate area. If it is larger, it is determined that an intruder is shown in the person candidate area.
The predetermined threshold Th1 is set in accordance with the accuracy with which the intruder is allowed to be erroneously detected with respect to the image monitoring apparatus. In the present embodiment, the predetermined threshold Th1 is set to 2. That is, when at least the third-stage discriminator 330 outputs an identification result (True) indicating that an intruder appears in the target human candidate area, the reliability is greater than the predetermined threshold value Th1. In addition, the predetermined threshold Th1 may be set to 0 if the accuracy with which the intruder is allowed to be erroneously detected is not strict. That is, the comparison between the reliability and the predetermined threshold value Th1 may be omitted.

  In addition, in the monitoring image, when a person candidate area of interest is confused with an intruder, the reliability value required for the person candidate area by the identification unit 52 is increased. Therefore, if the determination unit 53 determines the presence or absence of an intruder based only on the reliability value, there is a possibility that an object that is confused with a person existing in the monitoring area is erroneously detected as an intruder. However, if the object shown in the person candidate area of interest is an object that originally exists in the monitoring area, the same object is shown in the area corresponding to the person candidate area in the background image. For this reason, the similarity value obtained by the identification unit 52 for the region in the background image corresponding to the target human candidate region also increases in the same manner as the reliability. Therefore, as described above, the determination unit 53 determines that only when the reliability obtained for the target candidate candidate region is higher than the similarity obtained for the region in the background image corresponding to the person candidate region. By determining that an intruder appears in the person candidate area, it is possible to prevent erroneous detection of the intruder.

  An outline of the process of the determination unit 53 will be described by taking the person candidate areas 211 and 212 shown in FIG. In this case, regarding the person candidate area 211, the background image 200 shown in FIG. 2A only shows a background scene, and there is nothing that is confusing with an intruder. Therefore, the similarity obtained by the identification unit 52 for the area in the background image 200 corresponding to the person candidate area 211 is a low value. On the other hand, in the monitoring image 205 shown in FIG. 2B, an intruder is shown in the person candidate area 211. Therefore, the reliability calculated for the candidate person area 211 by the identification unit 52 is a high value. Therefore, the determination unit 53 can determine that an intruder is captured because the reliability of the person candidate region 211 is higher than the similarity.

  On the other hand, regarding the person candidate region 212, planting is shown in both the background image 200 shown in FIG. 2A and the monitoring image 205 shown in FIG. Therefore, there is a high possibility that the reliability and the similarity obtained for the person candidate area 212 by the identification unit 52 have the same value. For this reason, the determination unit 53 does not determine that an intruder is captured even when the reliability obtained for the person candidate region 212 is high. As described above, the determination unit 53 can prevent erroneous determination that an intruder is captured even if an area where an object that is confused with the intruder is extracted as a person candidate region.

  In addition, the determination part 53 shows the intruder in the focused candidate candidate area when the difference obtained by subtracting the similarity from the reliability calculated for the focused candidate candidate area is higher than a predetermined threshold Th2. May be determined. However, the predetermined threshold Th2 is a real number greater than or equal to 0, and when Th2 is 0, it is the same as the above embodiment in which the reliability and the similarity are simply compared. Alternatively, when the value obtained by multiplying the reliability obtained for the target human candidate region by a predetermined coefficient α (where α is a positive value) is greater than the similarity, the determination unit 53 determines the target human candidate to be focused on. It may be determined that an intruder appears in the area. In this way, the determination unit 53 intrudes the target person candidate area in the background image and the target person candidate area in the monitoring image, rather than determining whether or not the intruder is captured based on the mere magnitude of the reliability and the similarity. Since it is determined that the intruder is reflected in the target person candidate area when the difference in personiness is large, erroneous detection of the intruder can be further prevented.

Further, the determination unit 53 determines the average value of the reliability obtained from the person candidate areas extracted over a plurality of frames of the continuously acquired monitoring image and the background image corresponding to the person candidate area of each frame. If the difference between the average value of the similarities extracted from the area or the maximum value of the similarities is larger than a predetermined threshold value Th2, it may be determined that an intruder appears in the series of person candidate areas. In this way, by using the reliability and similarity obtained for a plurality of frames as determination criteria, the determination unit 53 erroneously detects an area that has been confused with an intruder for some reason as an intruder. Can be prevented.
In this case, the signal processing unit 5 may use any one of various known tracking processes in order to associate areas related to the same object among the candidate human areas extracted in each frame. it can.

  Further, the determination unit 53 uses a discriminator such as a discriminant function, a perceptron type neural network, or a support vector machine that outputs whether or not an intruder is shown in the person candidate area, with the reliability and similarity as inputs. Then, it may be determined whether or not an intruder appears in the person candidate area. Even in this case, the discriminant function and the discriminator are pre-learned using the reliability and similarity obtained for each of a plurality of images in which a person is photographed and a plurality of images in which a person is not photographed. The discriminant function and the discriminator are also learned so that the higher the similarity, the stricter the criterion for determining that an intruder is in the person candidate area corresponding to the similarity.

  If the determination unit 53 determines that an intruder is present in any person candidate area, the determination unit 53 determines that an intruder is present in the monitoring area, while no intruder is present in all person candidate areas. If it is determined that there is no intruder in the monitoring area, it is determined.

  When the determination unit 53 determines that there is an intruder into the monitoring area, the signal processing unit 5 transmits an alarm signal to the output unit 4 and causes the output unit 4 to issue an alarm.

  The background image update unit 54 updates the background image stored in the storage unit 3 by replacing it with a monitoring image when the determination unit 53 determines that there is no intruder. In this way, the background processing image always represents the latest state of the monitoring region, so that the signal processing unit 5 can perform the change in weather, the sun in the person candidate region extraction process based on the background difference of the variation region extraction unit 51. It is possible to reduce false detection of a human candidate region caused by the daily fluctuation of the person. Also, with respect to the similarity calculation processing in the identification unit 52, the similarity to the background image representing the latest state of the monitoring area is obtained. Therefore, since the signal processing unit 5 can compare the similarity and the reliability with respect to the latest state in the determination unit 53, the reliability of the determination result can be improved. Note that the background image update unit 54 may perform the background image update process at regular intervals (for example, every 10 minutes), not every time a monitoring image is acquired.

The operation of the intruder detection process by the image monitoring apparatus will be described with reference to the flowchart shown in FIG. This operation is controlled by the signal processing unit 5. The following operations are repeated at the monitoring image capturing interval.
When the detection process is started, the signal processing unit 5 acquires a monitoring image of the monitoring area captured by the image input unit 2 (step S101). Next, the change area extraction unit 51 of the signal processing unit 5 performs background difference processing between the acquired monitoring image and the background image stored in the storage unit 3 to extract a person candidate area (step S102). The change area extraction unit 51 passes information indicating all extracted person candidate areas to the identification unit 52 of the signal processing unit 5.

Next, the identification unit 52 selects a person candidate area of interest from each person candidate area. And the identification part 52 calculates | requires the reliability Ri showing the degree of intruder-likeness of the person candidate area | region by inputting the information of the person candidate area | region to which attention is paid to a discriminator in the monitoring image (step S103). Then, the identification unit 52 passes the obtained reliability Ri to the determination unit 53 of the signal processing unit 5. The determination unit 53 determines whether or not the reliability Ri is higher than a predetermined threshold Th1 (step S104). In step S104, when the reliability Ri is equal to or less than the predetermined threshold Th1, the signal processing unit 5 shifts the control to step S108.
On the other hand, when the reliability Ri is higher than the predetermined threshold Th1 in step S104, the determination unit 53 notifies the identification unit 52 to that effect. Then, the identification unit 52 obtains the similarity Rb indicating the degree of confusingness with the intruder by inputting information on the region in the background image corresponding to the target human candidate region to the classifier (step S105). Then, the identification unit 52 passes the obtained similarity Rb to the determination unit 53.

  The determination unit 53 determines whether or not the reliability Ri is higher than the similarity Rb (step S106). In step S106, when the reliability Ri is equal to or less than the similarity Rb, the signal processing unit 5 shifts the control to step S108. On the other hand, in step S106, when the reliability Ri is higher than the similarity Rb, the determination unit 53 determines that an intruder appears in the target person candidate area. And the detection flag showing the determination result that the intruder is reflected is temporarily memorize | stored in the memory | storage part 3 (step S107).

  After step S107 or when the reliability Ri is equal to or less than the similarity Rb in step S106, or when the reliability Ri is equal to or less than the predetermined threshold Th1 in step S104, the signal processing unit 5 It is determined whether or not the determination process for determining whether or not an intruder has been captured for all person candidate areas has been completed (step S108). In step S108, when the process for determining whether or not an intruder is shown in any person candidate area has not been completed, the signal processing unit 5 focuses on the person candidate area for which the determination process has not been completed. Let it be a candidate area. And the signal processing part 5 performs the process of step S103-S108.

  On the other hand, if it is determined in step S108 whether or not the intruder has been captured for all person candidate areas, the determination unit 53 refers to the storage unit 3 and stores a detection flag. By checking whether or not there is an intruder in the monitoring area, it is determined (step S109). In step S109, if the detection flag is stored in the storage unit 3 and the determination unit 53 determines that an intruder exists in the monitoring area, the signal processing unit 5 issues an alarm via the output unit 4 (step S110). ). In addition, the signal processing unit 5 may output an abnormality detection signal indicating that an intruder exists in the monitoring area to a security device or a monitoring center device connected to the image monitoring device 1 via a communication network. . Thereafter, the signal processing unit 5 erases the detection flag stored in the storage unit 3.

On the other hand, in step S109, if the detection flag is not stored in the storage unit 3 and the determination unit 53 determines that there is no intruder in the monitoring area, the background image update unit 54 of the signal processing unit 5 The background image is updated by replacing the background image stored in 3 with the latest monitoring image (step S111).
After step S110 or S111, the signal processing unit 5 ends a series of controls.

  As described above, the image monitoring apparatus according to an embodiment of the present invention provides information on a candidate human area corresponding to an area having a luminance difference from a background image in a monitoring image obtained by capturing the monitoring area. Input to the classifier to obtain the reliability indicating the degree of intruder-likeness of the area, and input the information about the corresponding area in the background image to the same classifier to disguise the intruder in the area. Find the degree of similarity that represents the degree of height. The image monitoring apparatus makes it difficult to detect a person candidate area having a higher degree of similarity as an intruder, so that even if an object confusing with the intruder appears in the monitoring area, the confusing object is mistakenly identified as an intruder. Detection can be prevented.

  The preferred embodiments of the present invention have been described above, but the present invention is not limited to these embodiments. For example, if the discriminator used by the discriminating unit 52 of the signal processing unit 5 is constructed so as to obtain the reliability for the partial region in the monitoring image by scanning the entire monitoring image, the change region extracting unit 51 is omitted. May be. In this case, the identification unit 52 inputs the monitoring image itself obtained by the image input unit 2 or an image obtained by performing preprocessing on the monitoring image to the discriminator, so that the reliability for each partial region in the monitoring image is obtained. Find the degree. The pre-processing includes, for example, a luminance correction process for making the average value of the luminance of the entire image constant, and a contrast correction process for making the difference between the maximum value and the minimum value of the luminance in the image constant. Similarly, the identification unit 52 obtains the similarity to each partial region in the background image by inputting the background image to the classifier. Then, the determination unit 53 determines whether or not the detection target object is captured in the partial area based on the reliability and similarity obtained for the corresponding partial areas of the monitoring image and the background image.

Further, the discriminator used in the discriminating unit 52 of the signal processing unit 5 is not limited to the cascade type discriminator.
FIG. 5 shows another example of a classifier that can be used in the classifier 52. The parallel type discriminator 500 shown in FIG. 5 has a configuration in which four Adaboost discriminators 510 to 540 are connected in parallel. This parallel type discriminator 500 inputs information related to the person candidate area in the monitoring image or information related to the area in the background image corresponding to the person candidate area to each of the discriminators 510 to 540, so that an intruder enters the area. The identification result of whether or not is shown is output. The parallel type discriminator 500 sets the number of discriminators that output the discrimination result that the intruder appears in the input area among the discriminators 510 to 540 as the reliability Ri or the similarity Rb.
Further, the number of classifiers included in the cascade classifier 300 illustrated in FIG. 3 is not limited to four. Similarly, the number of classifiers included in the parallel classifier 500 shown in FIG. 5 is not limited to four.

  In addition, the feature used in the classifier is not limited to the Haar-like feature. For example, the classifiers 310 to 340 of each stage of the cascade type classifier 300 shown in FIG. 3 or the classifiers 510 to 540 of the parallel type classifier 500 shown in FIG. 5 have an EOH (Edge of Orientation Histograms), HOG. Features such as (Histograms of Oriented Gradients), Edgelet, Joint-Haar-like, Joint HOG, and Shapelet may be calculated from the input image area and used for identification. Alternatively, the classifiers 310 to 340 at each stage of the cascade type classifier 300 or the classifiers 510 to 540 of the parallel type classifier 500 calculate different types of feature quantities from among any of the above feature quantities. May be used for identification.

Each discriminator is used to identify the degree of coincidence representing the degree of coincidence between various feature amounts extracted from a predetermined area in the monitoring image or the background image and the corresponding feature amount of the person shown on the image. You may use for. For example, the shape of a person shown on the image or the feature amount such as EOH or HOG is stored in advance in the storage unit 3 as a template. The identification unit 52 extracts feature amounts of the same type as those templates from a predetermined region of the input monitoring image or background image, and obtains the degree of coincidence between the extracted feature amounts and the template. Each discriminator receives the matching degree as an input and outputs a discrimination result as to whether or not an intruder is shown in the predetermined area. For example, if the template has a predetermined value, the difference between the feature value extracted from the predetermined area of the input image and the value of the template is δ, and the maximum value that δ can take. As δ _max , (δ _max −δ) / δ _max can be set. If the template is an image block of m × n pixels (where m and n are natural numbers), the maximum correlation value calculated by pattern matching between the template and a predetermined area of the input image is calculated as the degree of coincidence. The

Furthermore, each classifier of the cascade classifier shown in FIG. 3 or each of the parallel classifiers shown in FIG. 5 is not limited to the Adaboost classifier. Each of these classifiers may be a support vector machine, a perceptron type neural network, or the like.
Furthermore, if the discriminator used by the discriminating unit 52 is a discriminator that outputs a reliability or similarity in multiple levels of three or more levels by inputting a monitoring image or a background image, or a partial region thereof, A single discriminator may be used. As such a discriminator, for example, a stochastic model such as a Bayesian network can be used. For example, when a Bayesian network is used as a discriminator, the discriminator inputs each feature amount as described above obtained from a predetermined region in the monitoring image to at least one input node of the Bayesian network. The reliability is output as a posterior probability for the combination of feature value values input to the input node. Similarly, the discriminator inputs each feature amount as described above, which is obtained from an area in the background image corresponding to a predetermined area in the monitoring image, to at least one input node of the Bayesian network. Similarity is output as a posteriori probability for a combination of feature value values input to a node. And the determination part 53 can determine with the same reference | standard as the determination criterion mentioned above, for example, when a reliability is higher than a similarity degree, that the intruder is reflected in the predetermined area | region.
As described above, those skilled in the art can make various modifications in accordance with the embodiment to be implemented within the scope of the present invention.

It is a schematic block diagram of the image monitoring apparatus which is one Embodiment of this invention. (A) is a figure which shows an example of a background image, (b) is a figure which shows an example of a monitoring image, (c) is a figure which shows an example of the difference binarized image of a background image and a monitoring image. . It is a schematic block diagram of the cascade type discriminator which is an example of the discriminator used with an image monitoring apparatus. It is a flowchart which shows the operation | movement of the intruder detection process by an image monitoring apparatus. It is a schematic block diagram of the parallel classifier which is another example of the classifier used with an image monitoring apparatus.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 Image monitoring apparatus 2 Image input part 3 Memory | storage part 4 Output part 5 Signal processing part 51 Change area extraction part 52 Discriminator 53 Judgment part 54 Background image update part 300 Cascade type discriminator 500 Parallel type discriminator

Claims (6)

An image monitoring device for detecting a detection target from an image obtained by photographing a specific monitoring area,
An image input unit for acquiring a monitoring image obtained by photographing the monitoring area;
A storage unit that stores a background image that is an image of the monitoring area in a situation where there is no detection target in the monitoring area;
A discriminator that calculates the degree of likelihood of a detection object in a predetermined area of the monitoring image and the background image, and the reliability of the degree of the predetermined area of the monitoring image acquired by the image input unit by the discriminator; And an identification unit that calculates the degree of the predetermined region of the background image corresponding to the predetermined region of the monitoring image as the similarity by the identifier.
A determination unit that determines that the detection target exists in the monitoring area when the reliability satisfies a determination criterion that becomes more severe as the similarity increases ;
An image monitoring apparatus comprising: The determination criterion is that a difference obtained by subtracting the similarity from the reliability is greater than a predetermined threshold, and the determination unit determines that a difference obtained by subtracting the similarity from the reliability is greater than a predetermined threshold. The image monitoring apparatus according to claim 1, wherein it is determined that a detection target exists in the monitoring area.   The discriminator is a cascade type discriminator in which a plurality of discriminators are connected in series, and each of the plurality of discriminators connected in series has a detection object in the predetermined region. The cascade type discriminator outputs the discrimination result until the discriminator of the next stage outputs the discrimination result that the detection target does not exist. The reliability and the similarity are the number of stages of the last discriminator that outputs a discrimination result that the detection target exists in the predetermined region in the cascade discriminator. The image monitoring apparatus described.   The discriminator is a parallel discriminator in which a plurality of discriminators are connected in parallel, and each of the plurality of discriminators connected in parallel has a detection object in the predetermined area. And the parallel type classifier outputs the number of classifiers that output the identification result that the detection target exists among the plurality of classifiers connected in parallel with the reliability and the The image monitoring apparatus according to claim 1, wherein the image monitoring apparatus has similarity.   The discriminator inputs the degree of coincidence between the feature information of the detection target extracted from the predetermined area and the feature information of the detection target stored in the storage unit, so that the reliability and the similarity The image monitoring apparatus according to claim 1, wherein the degree is calculated. It further includes a change area extraction unit that performs a difference process between the monitor image and the background image, and extracts a change area having a luminance difference between the monitor image and the background image,
The image monitoring apparatus according to claim 1, wherein the identification unit calculates the reliability and the similarity using the change area as the predetermined area.

Images