JP2011186719A - Apparatus, method and program for detecting object - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To count the accurate number of objects by controlling adaptation of an object to a background model even when the object (person) rests for a comparatively long period. <P>SOLUTION: An object detection processing part 3 divides an input image photographed by a fixed camera 1 and a background model into blocks each of which is composed of a plurality of pixels and extracts an object area corresponding to a contour of the object in each block. Further, the object detection processing part 3 performs expansion processing to fill an area corresponding to the inside of the contour part and performs contraction processing for controlling excess detection of the object area due to the expansion processing. A load value table generation part 5 obtains a load value expressing a contribution ratio of each pixel to the number of objects to generate a load value table. An object counting part 6 integrates the load values of pixels in the object area to count the number of objects. <P>COPYRIGHT: (C)2011,JPO&INPIT

Description

  The present invention relates to an object detection device, an object detection method, and an object detection program for detecting a region of an object present in an input image taken by an image input device such as a fixed camera and counting the number of objects in the input image. In particular, the present invention relates to a technique suitable for counting objects such as limbs, toes, and heads that slightly move due to wobbling or the like, such as a person who is busy.

  As a technique for counting (measuring) the number of objects from an input image (video) from a fixed camera or the like, the one described in Patent Document 1 is known. In this case, the load corresponding to the contribution ratio of each pixel to the number of objects when there is an object at each pixel in the input image based on geometric conditions including the posture and position of the camera used for shooting. While obtaining values for all the pixels, a background model excluding the object is constructed from the input image, and an object region in which the object exists is obtained from the difference between the input image and the background model. Then, the number of objects existing in the input image is measured by integrating the load values of the pixels constituting the object region.

JP 2009-294755 A

  The background model is generated using, for example, an average of a plurality of input images taken in time series, and further excludes temporal changes in illumination outdoors (differences in morning and night brightness). Updated regularly. Therefore, when a person as an object stays on the spot for a long time, the movement of the object cannot be detected, and the object is likely to be assimilated to the background, and there is a possibility that the detection of the object may be omitted. Similarly, when there are many similar objects, since the difference between the objects is small, a similar object that has existed in the same place in the past is likely to be taken in as a background, and it is easy to cause an object detection failure.

  When a person or the like stays and stays for a long period of time, in general, the person will not be in a completely stationary state, and a slight change in movement appears in the contour parts such as the hands and feet due to wobbling. In the present invention, paying attention to this point, by extracting the outline of the object and then performing the expansion process and the contraction process, there are objects that remain stationary or staying in the long term, or there are many similar objects as well. Even in this case, the detection omission due to the assimilation of the object with the background is reduced, and the number of objects is accurately counted.

  That is, the object detection device according to the present invention constructs a background model based on an input image photographed by an image input device installed at a predetermined depression angle with respect to a target range of object detection, and the input image and Object detection processing means for detecting an object region in which an object is present in the input image based on a background model, and an object exists in each pixel in the input image based on a geometric condition of the image input device A load value table generating means for determining a load value representing a contribution ratio of each pixel to the number of objects in the case, and generating a load value table; and integrating the load values of the pixels constituting the object region, the input image And object counting means for counting the number of objects present therein.

  In the present invention, the object detection processing unit divides the input image and the background model into blocks composed of a plurality of pixels, and for each block, based on the similarity between the input image and the background model. The object region corresponding to the contour of the object is extracted, and the object region corresponding to the contour of the object is expanded so that the region inside the contour of the object is corrected to the object region. The object region after the expansion processing is contracted so as to suppress excessive expansion of the object region due to the above.

  In this way, by detecting the object region based on the similarity between the input image and the background model quantitatively in units of blocks including a plurality of pixels set to an appropriate size for the object, for example, individual Compared to the case where the object area is detected finely in units of pixels, the followability to the shape of the object is reduced, but the influence of erroneous detection of individual pixels can be offset and absorbed. It is possible to reliably extract the contour of an object that causes Then, the region inside the contour of the object is corrected as the object region by the expansion process, and the object region can be appropriately detected by suppressing the excessive expansion of the object region due to the expansion process by the contraction process. . As a result, it is possible to suppress object detection omission due to assimilation of the object with the background model, and to accurately count the number of objects.

  However, through the expansion process and the contraction process, a non-object area between adjacent objects is inevitably detected as an object area, and the number of objects is excessive due to this combined area. A new problem arises of being counted. Therefore, preferably, the probability that the combined region occurs is obtained, and based on this probability, the number of objects counted by the object counting means is corrected to the lower side. This makes it possible to suppress the number of objects that are excessively counted by the amount of the combined region.

  As described above, the detection accuracy of the object region and the extraction accuracy of the contour of the object greatly depend on the correlation between the size of the object and the size of the block, and if the block (grid) is too small for the object, As shown in FIG. 1A, the number of undetected blocks existing inside the contour of the object increases, and it becomes difficult to correct the filling of the missing portion inside the contour by the expansion processing described above.

  Therefore, preferably, the average size of the object in the input image is calculated based on the geometric condition of the image input device, and the shorter length of the average width and width of the average object is calculated. The size of the block is set based on Specifically, the length of one side of the block is within a range of 1/2 to 1/3 with respect to the shorter length of the average vertical and horizontal width of the average object. Adjust and set the size automatically.

  Alternatively, the input image and the background model may be enlarged or reduced with the block having a certain size. That is, based on the geometric condition of the image input device, the average size of the object in the input image is calculated, and based on the shorter length of the average width and width of the average object. Then, the size of the input image and the background model is enlarged or reduced so that the size of the object for each block becomes a predetermined level, and then the object is detected for each block. In this case, the object counting means counts the number of objects after returning the detection result to the original size.

  Each processing content described above is executed by a computer in the form of a program, for example, and this program can be recorded on an appropriate recording medium.

  As described above, according to the present invention, the contour of an object is extracted by detecting the object region in units of blocks including a plurality of pixels, and the expansion processing for filling the inside of the contour of the object and excessive expansion are suppressed. Even if an object such as a person stays or stays in place for a relatively long period of time, or if the background model is updated to cope with long-term lighting fluctuations, It is possible to extract the contour of an object that inevitably causes a fine movement due to wobbling of the object, to separate the object to be detected from the background model, to detect an appropriate object region, and to count the number of objects.

Explanatory drawing which shows the concept of the area | region detected as an object when (A) has a small block size and (B) has a large block size. Explanatory drawing which shows an example of the expansion process (A) and shrinkage | contraction process (B) by morphological calculation. Explanatory drawing which shows an example of suppression of the excessive detection pixel of an object periphery by shrinkage | contraction processing. Explanatory drawing which shows the coupling | bonding area | region of the object which remains by an expansion process and a contraction process. Explanatory drawing which shows the area | region where the coupling | bonding of an object may generate | occur | produce by an expansion process. The block diagram which shows simply the basic composition in case the image storage part which concerns on this invention is included. The block diagram which shows simply the basic composition in case the image storage part which concerns on this invention is not included. The flowchart which shows the flow of the control processing which concerns on 1st Example of this invention. The flowchart which shows the flow of the control processing which concerns on 2nd Example of this invention. The flowchart which shows the flow of the control processing which concerns on 3rd Example of this invention.

  Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the drawings. First, basic principles and concepts will be described with reference to FIGS.

  When a person or the like stays or stays on the spot for a long period of time, it is generally not completely stationary, and slight fluctuations in the contours of the hands, feet, head, etc. due to wobbling etc. Appears. Focusing on this point, an input image (also simply referred to as “image”) and a background model (hereinafter also simply referred to as “background”) captured by an image input device such as a fixed camera are formed in a grid pattern composed of a plurality of pixels. By comparing each block (also referred to as “lattice”), the contour of the object with such a slight change is extracted, and by performing expansion processing and contraction processing described later on the contour of the object, It is possible to detect an appropriate object region according to an actual object. In other words, by detecting the object area in units of blocks consisting of a plurality of pixels, even when background update is performed to cope with long-term illumination fluctuations, the contour of the object due to object fluctuations in the image, etc. It is possible to detect a minute change in the portion and extract the outline of the object to be detected separately from the background. Therefore, even when there are a plurality of similar objects, it is possible to detect a minute difference between similar objects.

  FIG. 1 shows a conceptual diagram of the first stage of object detection according to an embodiment of the present invention. An arbitrary image and a background model are divided into blocks 10 each having an appropriate size in a square lattice shape, and each block 10, that is, the block 10 is compared by comparing the input image and the background model for each block 10. It is detected whether or not all the pixels included in the object region are object regions. In the case of a moving object 11 such as a person, the background and the object area fluctuate in the contour / peripheral area of the object 11 due to slight changes such as wobbling. 1 and the background model, the contour portion of the object 11 can be extracted as the object region as shown in the white outline portion 12 in FIG.

  Here, the size of the block 10 can be set to an arbitrary size. However, when the block 10 is too small with respect to the size of the object 11, as shown in FIG. The number of blocks that become the undetected area 13 that is not detected as an object located inside the portion 12 increases, and correction by an expansion process, which will be described later, becomes difficult. Accordingly, the block 10 has a certain size or more, and has a uniform size in the image in order to correct an undetected area by an expansion process described later.

  When the change in the surface of the object is poor, as shown in FIG. 1, the contour portion 12 of the object is detected as the object region, but the region 13 inside the object hardly moves / changes the object, and therefore includes a plurality of pixels. Even in the block-by-block comparison, the object cannot be detected assimilated into the background model. Therefore, in order to correct and detect the undetected region 13 inside the contour portion 12 of the object that cannot be detected as an object region, as shown in FIG. 2, for example, an expansion process using a known morphological operation. Further, a contraction process is performed to reduce the excessive expansion of the object region due to the expansion process. The undetected area 13 inside the area 12 corresponding to the contour of the object is corrected as an object area by the expansion process of the morphological operation, and the area 12A around the object that is excessively detected by the expansion process is suppressed by the subsequent contraction process. . At this time, the block as a structural element used for the expansion process and the contraction process of the morphological operation is, for example, a 3 × 3 pixel square lattice, and the expansion process and the contraction process are half the number of pixels on one side of the square lattice. By repeatedly executing each time (5 times in the case of 10 pixels per side), the undetected area 13 is detected as an object area. As shown in FIG. 3, by comparing each block including a plurality of pixels divided in a grid shape, the actual object size is also detected by the morphological operation shrinkage process for the pixels around the object that are over-detected. It is possible to suppress the object area to an appropriate size according to the above.

  FIG. 4 shows an example in which a plurality of objects 11 in the image are close to each other within a certain range. Here, “within a certain range” corresponds to the distance between the objects that can generate the coupling region 14 in which the objects are coupled by the expansion process of the morphological operation. Since the pixels combined by the expansion process of the morphological operation do not return to the original state even in the subsequent contraction process, they are detected as an object region and cause excessive detection of the number of objects and objects. Therefore, it is necessary to calculate the probability of the object coupling due to such expansion processing and to correct the error due to excessive detection. The size of the object in the image varies depending on the actual size of the object, the focal length of the camera, and the distance from the camera to the object, but measures the size of the object and the distance from the camera to the moving object in advance. In this case, we simply assume that an object of average size exists uniformly in the image, and use the average size of the object that exists in the image instead of the real space. Calculate the probability of combining.

Due to the nature of the morphological operation, the combination causes the expansion only in the regions β1 to β4 in the vertical and horizontal directions of the object shown in FIG. 5 except for the oblique direction of the object. For example, when the distance between the region β2 on the right side of a certain object 11 and the region β4 on the left side of another adjacent object (not shown) is closer than a predetermined level, coupling due to expansion processing occurs, , The left side area β4 of the object 11 and the right side area β2 of the other object, the upper area β1 of the object 11 and the lower area β3 of the other object, the lower area β3 of the object 11 and the other area Coupling occurs when the distance between the regions β1 on the upper side of the object is short. From this, the probability that an error due to combining occurs with respect to a certain pixel (x, y) is in the case of a combination in which the corresponding pixel is not an object and is combined by the above-described expansion processing. The probability P1 ′ (x, y) that there is only one object and a pixel (x, y) is not an object is that the average object area (number of pixels) in the image is α and the area of the entire image (pixel When the number is A, it can be expressed by the following formula (1).
P1 ′ (x, y) = 1−α / A (1)
In addition, there are two probabilities P1 ″ (x, y) that there are two objects and a pixel (x, y) is not an object, and two in a region other than the first object with respect to the probability of P1 ′. Since this is the probability that there is no eye object, if the average value of the floor area (number of pixels) of the object in the image is a, it can be expressed by the following equation (2).
P1 ″ (x, y) = P1 ′ (x, y) × {1−α / (A−a)} (2)
From the above formulas (1) and (2), when there are N objects, the probability P1 (x, y) that a certain pixel (x, y) is not an object is generalized and expressed by the following formula: be able to.

  The number N of objects is calculated based on, for example, the object area α and the load value obtained from the geometric relationship between the camera and the floor surface obtained from Patent Document 1 described above.

  Note that when the object detection target range is performed not on the entire image but only on a part of the image, the value of the object detection area A is not the area of the entire image but the area of the target range (number of pixels). Can be calculated as

  If the geometric conditions of the image input device, that is, internal parameters such as focal length, image center, lens distortion, external parameters including posture and position, and object detection target range are not changed, α, A , A is a fixed value, and P1 (x, y) is a function dependent on N. Since N is the number of objects in the object detection target range, the maximum number is not greatly changed if the floor area or the size of the object to be measured is fixed. Therefore, it is easy to determine the upper limit value of N, and it is possible to speed up the processing by calculating P1 (x, y) in which N is changed from 1 to the upper limit value and making a table in advance. It becomes possible.

Next, assuming that the regions to be combined by the expansion processing are β1, β2, β3, and β4 shown in FIG. 5, the probability of combining by the expansion processing is that the region of β1 of the object A is different from that of the object A. It is considered that two or more regions overlap the β3 region of the object (hereinafter, object X). Here, using the assumption that the size of the object A and the object X is the average size of the object in the image, as shown in the following equations (4) and (5), the object A and the object Β1 and β3, and β2 and β4 of X have the same area.
β1 = β3 = β ′ (4)
β2 = β4 = β ″ (5)
From the above, the probability that a certain pixel (x, y) is not combined by the expansion processing is the probability that none of β ′ or β ″ exists or only one exists. Probabilities Pb01 (x, y) and Pb02 (x, y) that no β ′ or β ″ exists in a certain pixel (x, y) can be expressed by the following equations (6) and (7), respectively. it can.
Pb01 (x, y) = 2 × (1−β ′ / A) ^ N (6)
Pb02 (x, y) = 2 × (1-β ″ / A) ^ N (7)
Further, the probabilities Pb11 (x, y) and Pb12 (x, y) that there is only one β ′ or β ″ in a certain pixel (x, y) are expressed by the following equations (8) and (9). be able to.
Pb11 (x, y) = 2 × N × β ′ / A × (1-β ′ / A) ^ (N−1) (8)
Pb12 (x, y) = 2 × N × β ″ / A × (1-β ″ / A) ^ (N−1) (9)
Using the above equations (6) to (9), the probability P2 (x, y) that two or more β ′ or β ″ exists in a certain pixel (x, y) is expressed by the following equation (10). be able to.
P2 (x, y) = 1- {Pb01 (x, y) + Pb02 (x, y) + Pb11 (x, y) + Pb12 (x, y)} (10)
From Expression (3) and Expression (10), the probability P (x, y) that a certain pixel (x, y) is combined by the expansion process can be expressed as the following Expression (11).
P (x, y) = P1 (x, y) × P2 (x, y) (11)
Here, of P1 (x, y) and P2 (x, y), P1 (x, y) has a characteristic that decreases as the number N of objects increases, and P2 (x, y) The number N becomes larger as the number N increases. Therefore, by multiplying these P1 (x, y) and P2 (x, y), the final probability P (x, y) is the object until the number N of objects reaches a certain predetermined level. As the number N of the object increases, the characteristic increases, and when the value exceeds a predetermined level, the characteristic decreases as the number N of objects increases.

  Of the number of pixels detected as the object region, the number of pixels derived from the probability P (x, y) is considered to be a pixel that is excessively detected without being contracted by the contraction process due to the combination generated by the expansion process. Can do. Here, since the size of the object in the image is assumed as the average size of the object, a proportional relationship is established between the number of pixels and the number of objects, and the probability of overdetection represented by Expression (11) Can be viewed as the number of objects rather than as a pixel. Therefore, as shown in the following equation (12), by correcting the number of objects N using the probability P (x, y) calculated by the equation (11), the number of objects N ′ in consideration of excessive detection. Can be derived.

  FIG. 6 is a diagram showing an example of a basic configuration according to the present invention. In FIG. 6, 1 is a fixed camera as an image input device, 2 is an image storage unit that stores images acquired by the camera 1, 3 is a background model generated from the images stored in the image storage unit 2, and a plurality of pixels are By comparing the background model with the input image for each block included, the contour part of the object is detected as a region including the object, and the detected region is expanded by filling the missing portion inside the contour, or by expansion processing An object detection processing unit that performs an error reduction process such as a contraction process that reduces an excessively detected portion.

  For the generation of the background model, a known method can be used. For example, an image of a plurality of frames can be created on average. It is not always necessary to generate the background model at the time of inputting all the images. By temporarily holding the background model, the background model can be updated at an arbitrary interval using the background model generated in the past.

  A camera information input unit 4 inputs camera information (geometric conditions) such as internal parameters (focal length, image center, lens distortion, etc.) and external parameters (posture and position) of the camera 1. 5 is a view projected on each pixel on the observed image on the premise of a perspective projection relationship in which internal parameters and external parameters input from the camera information input unit 4 are associated with three-dimensional points in the real space. Calculate the volume of the volume that contributes to the volume of the object, and calculate the load value as a contribution ratio that quantitatively represents how much the pixel contributes to the number of objects, and calculate the load value table for all the pixels. It is a load value table production | generation part as a load value table calculation means to produce | generate.

  Reference numeral 6 denotes an object counting unit that measures the number of objects from each pixel in the region including the object detected by the object detection processing unit 3 and the load value of each pixel obtained by the load value table generating unit 5. That is, the object counting unit 6 counts the number of objects by integrating the load values of the pixels in the object region.

  7 is an error reduction process (expansion) of the object detection processing unit 3 based on each pixel of the region including the object detected by the object detection processing unit 3 and the number of objects measured by the object counting unit 6.・ Calculate the occurrence probability of the excessive detection of the number of objects due to the joint area between the objects inevitably caused by the contraction process), and based on this probability, correct the error again for the measured number of objects, It is an error correction unit as an object number correcting means for recalculating the number of objects.

  Reference numeral 8 denotes an object number output unit that outputs the number of objects recalculated by the error correction unit. Each process in the object detection processing unit 3, the load value table generation unit 5, the object counting unit 6, and the error correction unit 7 can be executed by, for example, a known computer system. The image storage unit 2 may use a recording medium such as a hard disk, a RAID device, or a CD-ROM, or may use a remote data resource via a network.

  FIG. 7 shows another example of the basic configuration according to the present invention. As in this example, it is also possible to adopt a configuration in which processing is performed in real time without having a storage device such as an accumulation unit or a recording medium. In FIG. 7, the same parts as those in FIG. 6 are denoted by the same reference numerals, and redundant description will be omitted as appropriate. In the example of FIG. 7, the image storage unit 2 is omitted, and instead, the object detection processing unit 3 holds the background model and updates the background model for the image input from the camera 1. At this time, it is not always necessary to update the background model when inputting all the images, and the background model can be set at an arbitrary interval of one or more sheets.

  Next, the processing flow of each part of the apparatus of FIGS. 6 and 7 will be described with reference to the flowchart of FIG. First, the overall flow will be described. An image taken by the camera 1 in step S1 is captured. Here, in the case of FIG. 6, the captured image is stored in the image storage unit 2. In step S2, a background model is generated or updated by the object detection processing unit 3 from the captured image. Next, in step S3, it is determined whether or not there is an unprocessed image. If there is an unprocessed image, the process proceeds to step S4, and a plurality of pixels are set between the background model and the input unprocessed image. A region corresponding to the contour of the object is detected as a region including the object by comparing each block including the object. That is, for each block including a plurality of pixels, the similarity based on the normalized correlation is calculated between the background model and the input image, and when the similarity does not reach a predetermined level, the input image of the block Is detected and detected as an object region, and all pixels in the block are detected as pixels in which an object exists.

  Next, in step S5, as described above, with respect to the region including the object, as described above, the expansion processing for detecting the missing portion inside the object contour portion as the object region, and the object region over-detected by the expansion processing are reduced. -Reduce contraction processing to reduce errors in the object area. The processing from step S3 to step S5 is performed in the object detection processing unit 3 shown in FIGS.

  Next, in step S10, based on the load value table generated in step S9, the load values of the pixels existing in the area including the object detected in step S5 are added, that is, integrated, and this integrated value is calculated as the object. Count and measure as a number. These processes are performed by the object counting unit 6 shown in FIGS. At this time, the load value table generated in step S9 is calculated for each pixel in step S8 based on the camera information input in step S6 and the geometric conditions such as camera parameters input and set in step S7. Is done. Since the load value table depends on the installation position of the camera, if it is after the installation of the camera, it can be calculated in advance before measuring the number of objects.

  Next, in step S11, by performing the above-described expansion processing and contraction processing, the number of excessively counted objects is suppressed by the amount of the combined region in which the non-object region between adjacent objects is excessively detected as the object region. To correct the number of objects. That is, based on the number of objects measured in step S10 and the number of pixels in the region including the object calculated in step S5, an error in the measured number of objects is calculated. Perform error correction. Specifically, the number of objects N ′ taking into account excessive detection is obtained using the above equation (12). The processing in step S11 is performed by the error correction unit 7 in FIGS. In step S12, the number of objects N ′ corrected in step S11 is output. The processing in step S12 is performed by the object number output unit 8 in FIGS.

  Details of each process will be described below. Images acquired via the camera 1 are stored in the image storage unit 2, and the object detection processing unit 3 updates the background model based on the images extracted from the image storage unit 2. At this time, the background model is updated at an arbitrary interval of one or more. For example, when creating a background model by averaging images of multiple frames, the time until the background model is completely updated by multiple updates corresponds to the above "arbitrary interval" 5 to 10 minutes, and about 30 minutes if indoor.

  Next, the extracted image and the updated background model are divided into grid-like blocks (lattices), and each pixel in the block corresponding to the position of the input image and the background model for each block including a plurality of pixels. Is used to calculate the similarity based on the normalized cross-correlation. This similarity is a value between 0 and 1 and is determined according to the camera installation environment (geometric conditions) such as indoors or outdoors. A comparison process is performed between the similarity and a threshold value set in advance by an experiment in accordance with the installation environment of the camera, and a block having a certain similarity or less is detected as an object region.

  After finishing the calculation of the similarity and the threshold processing for all the blocks, the defect in the undetected area inside the object contour portion detected as the object area is corrected as the object area by the expansion process of the morphological operation shown in FIG. Next, correction is performed to reduce / reduce the area that has been excessively detected by the expansion process or the area that has been excessively detected by calculation in units of blocks by the contraction process. Here, after the morphological calculation, not the calculation for each block but the calculation of the expansion process and the contraction process for each pixel is performed.

  Next, from the object area detected by the object detection processing unit 3 and the load value table generated by the load value table generating unit 5, the object counting unit 6 applies the load to the pixels included in the object area. By integrating the values, the number of objects included in the image is counted.

  Here, assuming a fixed camera, all of the three-dimensional position of the camera input by the camera information input unit 4, the rotation angle representing the posture of the camera, and the focal length of the camera (hereinafter referred to as camera information) are all It shall be known. The load value table generation unit 5 associates the image space with the three-dimensional real space from the camera information input by the camera information input unit 4, and all the pixels on the object surface when an object exists in a certain pixel The ratio to the number is calculated as a load value. The object counting unit 6 calculates the number of objects included in the entire image by integrating the values in the load value table corresponding to the object area detected by the object detection processing unit 3. Here, in the object detection processing unit 3, as shown in FIG. 4, there is a possibility that a combined area that is not an original object area is also detected as an object area by the expansion process of the morphological operation. For this reason, the error correction unit 7 corrects the number of objects so as to suppress the excessive detection of the number of objects due to the combined region. In the error correction unit 7, the number of objects calculated by the object counting unit 6, the number of pixels of the entire image, the average size of an object present in the image, and the pixel of the object region detected by the object detection processing unit 3 Based on the number, the coupling probability P (x, y) is calculated as shown in the above-described equations (1) to (11). Since the calculated P (x, y) corresponds to the occurrence probability of the error due to the combination, the error correction unit 7 uses the above equation using P (x, y) and the number of objects N calculated by the object counting unit 6. From (12), the corrected number of objects N ′ is calculated. The object number output unit 8 acquires the corrected object number N ′ from the error correction unit 7 and outputs it as the final object number.

  With reference to FIG. 8 again, a first embodiment of the present invention will be described. The processing contents of the present embodiment are roughly divided into a routine for calculating a load value table based on camera information set in advance and a flow for counting the number of objects from an input image based on the routine.

  First, in a routine for calculating a load value table, information on cameras installed in advance and camera parameters are input in step S6 and step S7, and a load value is calculated for each pixel in step S8 based on the respective values. A load value table is generated in step S9 based on the calculated load value and camera parameters. In the load value table, the ratio of the object surface area to each pixel of the input image is recorded as a load value, and the camera parameters used for the calculation are also recorded.

  On the other hand, in the object detection process in the flow for counting the number of objects, first, an input image from the camera 1 is acquired in step S1, and a background model is constructed in step S2 based on the acquired input image. At this time, the background model is not generated for all input images, and the background model can also be updated based on a preset processing interval. In parallel, it is determined in step S3 whether there is an unprocessed input image. If there is an unprocessed input image, the process proceeds to step S4. If there is no unprocessed input image, the process proceeds to step S4. The process proceeds to step S1.

  In step S4, the background model generated or updated in step S2 and the input image are each divided into grid blocks, and normalized cross-correlation is used for all the divided blocks based on the number of pixels in the block. The background model is compared with the input image. Here, when the value of the normalized cross-correlation is equal to or less than a predetermined value set in advance, the difference from the background is large, so it is determined as an object region, and when the value exceeds a certain value, the difference from the background is Since there are few, it determines as area | regions other than an object. Subsequently, in step S5, processing for suppressing defects and excess detection that occur due to the division and comparison in a lattice shape is performed by expansion and contraction processing of morphological operations. Next, the number of objects is counted in step S10 based on the object region corrected in step S5 and the load value table created in advance in step S9. Next, in step S11, the occurrence probability of the excessive detection region due to the combination of the objects generated by the processing in step S5 shown in FIG. 4 is calculated from the object region corrected in step S5 and the number of objects calculated in step S10. Finally, in step S12, the number of objects calculated in step S10 is corrected based on the error occurrence probability calculated in step S11, and the number of objects is counted.

  In the first embodiment, normalized cross-correlation is used for comparison between the input image and the background model, and the block size at that time is arbitrary. However, as shown in FIG. 1, if the size of the block 10 is too small relative to the size of the object, the defect detection process (expansion process) in the object detection processing unit 3 is performed inside the object contour portion. The deficit cannot be corrected. Further, if camera information is input by the camera information input unit 4 using a fixed camera, the average size of an object present in the input image can be calculated.

  Therefore, in the second embodiment shown in FIG. 9, the object region detection step S4A receives the camera information from the camera information input step S6, calculates the average object size based on the received camera information, and inputs the input image. The size of the block used for comparison with the background model is automatically determined. For example, assuming that the target object to be detected is projected in a state closest to the center of the image, the size of the object is used as the average size of the object. Here, since it is possible to detect in the object peripheral area due to a slight fluctuation of the object, the size of the block is short among the vertical width and the horizontal width in the image of the target object as shown in FIG. By setting the size within the range of one-third to one-half of the length dimension, a defect region that is not detected as an object region in the center portion of the object excluding the periphery of the contour of the object that is detected by wobbling is generally The size is one block, and good correction is possible by the defect correction processing in the object detection processing unit 3.

  In the first embodiment and the second embodiment, the block size using normalized cross-correlation is arbitrary. As the block size is larger, the amount of information (number of pixels) in one block increases, so that it is easy to catch a minute change, but there is a problem that the amount of calculation per block increases. Therefore, in the third embodiment, as shown in FIG. 10, image size changing steps S13 and S14 are newly added. In the image size changing step S13, the information from the camera information input step S6 is received, the average size of the object is calculated based on the received camera information, and the size of the input image acquired in step S1 is changed. The relative block size is changed without changing the block size. At this time, as in the second embodiment, the size of the block is one third to one half of the longer length of the vertical and horizontal widths in the image of the target object. As described above, by changing the size of the input image, the defect at the center of the object is also approximately one block, and good correction by the defect correction processing in the object detection processing unit 3 is possible. Since the load value table is created based on the input image of the original size, the object region detected by the object detection processing unit is subjected to the image size changing step S14 after the correction processing of the defect / overdetection in step S5. After returning to the size of the original input image, the process proceeds to step S10, and the object count unit 6 counts the number of objects. In this image size changing step S14, the information from the camera information input step S6 may be received and the image size may be restored based on this information, as in the image size changing step S13.

  In the present invention, a part or all of the processing contents of each part of the above-described object detection apparatus can be configured as a program that causes a computer to function.

1 ... Camera (image input device)
2 ... image storage unit 3 ... object detection processing unit (object detection processing means)
4 ... Camera information input unit 5 ... Load value table generation unit (load value table generation means)
6 ... Object counting section (object counting means)
7. Error correction unit (object number correction means)
8 ... Number of objects output section

Claims (6)

A background model is constructed based on an input image photographed by an image input device installed at a predetermined depression angle with respect to a target range for object detection, and the input image is determined based on the input image and the background model. An object detection processing means for detecting an object region in which an object is present;
A load value for generating a load value table by obtaining a load value indicating a contribution ratio of each pixel to the number of objects when an object exists in each pixel in the input image based on a geometric condition of the image input device Table generating means;
Object counting means for accumulating the load values of the pixels in the object region and counting the number of objects present in the input image;
In an object detection apparatus having
The object detection processing means is
The input image and the background model are divided into blocks composed of a plurality of pixels, and for each block, an object region corresponding to the contour of the object is extracted based on the similarity between the input image and the background model,
In order to correct the region inside the contour of the object to the object region, the object region corresponding to the contour of the object is expanded.
And, in order to suppress the excessive expansion of the object region due to the expansion processing, to perform the contraction processing of the object region after the expansion processing,
An object detection apparatus characterized by that.   Through the expansion process and the contraction process, the probability that a non-object area between adjacent objects is excessively detected as an object area is generated, and the object counting means counts based on this probability. The object detection apparatus according to claim 1, further comprising an object number correcting unit that corrects the number of objects to be detected.   The object detection processing unit calculates an average size of an object in the input image based on a geometric condition of the image input device, and based on a shorter length of the vertical width and the horizontal width of the object. The object detection device according to claim 1, wherein a size of the block is set. The object detection processing means includes
An average object size in the input image is calculated based on the geometric condition of the image input device, and the object size for each block is calculated based on the shorter one of the vertical and horizontal widths of the object. Detecting an object for each block after enlarging or reducing the size of the input image and the background model so that the size becomes a predetermined level,
3. The object detection device according to claim 1, wherein the object counting unit counts the number of objects after returning the detection result to the original size.   An object detection program characterized by being a program that causes a computer to execute each means according to any one of claims 1 to 4. A background model construction step of constructing a background model based on an input image captured by an image input device installed at a predetermined depression angle with respect to the object detection target range;
An object detection processing step for detecting an object region where an object is present in the input image based on the input image and a background model;
A load value for generating a load value table by obtaining a load value indicating a contribution ratio of each pixel to the number of objects when an object exists in each pixel in the input image based on a geometric condition of the image input device A table generation step;
An object counting step of accumulating the load values of the pixels in the object region and counting the number of objects present in the input image;
In an object detection method comprising:
The object detection processing step includes
The input image and the background model are divided into blocks composed of a plurality of pixels, and for each block, an object region corresponding to the contour of the object is extracted based on the similarity between the input image and the background model,
In order to correct the region inside the contour of the object to the object region, the object region corresponding to the contour of the object is expanded.
And, in order to suppress the excessive expansion of the object region due to the expansion processing, to perform the contraction processing of the object region after the expansion processing,
An object detection method characterized by the above.