JP6326622B2 - Human detection device - Google Patents

Description

  The present invention relates to a person detection apparatus, and more particularly to a person detection apparatus that analyzes an image taken by a surveillance camera and detects a person from the image.

  Conventional human detection devices divide the entire feature quantity into parts and perform part-by-part matching to synthesize the part-by-part matching results as a countermeasure against face recognizability due to face orientation, hiding, and illumination fluctuations In this way, an image with improved accuracy of identifying an object included in an image is known (for example, see Patent Document 1 below).

  In addition, in a target detection apparatus that detects a pedestrian using an in-vehicle camera and a laser radar, an apparatus that detects only candidates close to a distance measured by a laser radar as a detection target in order to suppress erroneous detection of a pedestrian far away. Is also known (see, for example, Patent Document 2 below).

  In Non-Patent Document 1, a rough shape of a person is expressed by a HOG (Histograms of Oriented Gradients) feature value, which is a feature value obtained by histogramating luminance gradient directions in a local region, and a person is identified from an image by a classifier. A person detection method for detection is disclosed.

JP 2012-234497 A JP 2011-48485 A JP 2012-118854 A

Fujikichi, “Gradient-based feature extraction -SIFT and HOG-”, 2007-CVIM-160, Information Processing Society of Japan Research Report, September 4, 2007, pp211-224

  Here, when a person is detected from an image obtained by photographing a wide range of monitoring locations with a monitoring camera, the size of the image differs greatly between a person existing in front and a person existing far away as viewed from the monitoring camera. In particular, a person who is far away appears small on the image and is difficult to detect because there is little information to be obtained.

For this reason, the method of Patent Document 1 described above has a problem that it can cope with a change in the orientation of the face or the like, but it is difficult to cope with a case where the size of the target image becomes small.
In addition, the method of Patent Document 2 has a problem that although a false detection can be suppressed, a distant person cannot be detected.
Further, in the method of Non-Patent Document 1, when the size on the image becomes small like a distant person, a luminance gradient histogram in nine directions is created as in the case where the size of the object on the image is large. This makes it difficult to detect a distant person.

  In view of the above, an object of the present invention is to provide a person detection apparatus capable of analyzing a wide range of monitoring place images taken by a monitoring camera and detecting a person with high accuracy from the image.

A person detection device according to a first invention for solving the above-described problem is
An image input means for inputting data of an image acquired by an imaging means for photographing the monitoring target area;
At least a processing setting unit for setting a plurality of object sizes corresponding to the size of a person appearing in the image and a person determination threshold value for determining the presence or absence of a person;
Object size setting means for sequentially selecting the plurality of object sizes and setting them as object sizes to be processed at present;
HOG feature quantity extraction means for extracting a HOG feature quantity, which is a feature quantity obtained by histogramating the gradient direction of luminance in a local area, based on the object size to be processed at the present time, for the image;
Identification means for judging the presence or absence of a person based on the HOG feature quantity extracted by the HOG feature quantity extraction means and the person judgment threshold;
A detection result aggregating unit that outputs the determination results of the presence / absence of a person based on all the object sizes together as a person detection result by collecting the determination results of the identification unit;
A plurality of the HOG feature amount extraction means are provided corresponding to a preset range of the object size to be processed, and a plurality of the identification means are provided corresponding to each of the HOG feature amount extraction means. A human detection device,
Set the number of divisions to divide the object size into a grid by the processing setting unit,
The plurality of HOG feature amount extraction means include
When the size of the small area obtained by dividing the object size into a grid shape based on the number of divisions is equal to or larger than a preset size, the object size is configured from a luminance gradient direction histogram in eight directions with respect to the object size. An 8-way HOG feature quantity extraction unit for extracting 8-way HOG feature quantities;
A four-way HOG feature quantity extraction unit that extracts a four-way HOG feature quantity configured from a luminance gradient direction histogram in four directions for the object size when the size of the small area is smaller than a preset size ; ,
The plurality of identification means are:
An 8-way discriminator that determines the presence or absence of a person based on the 8-way HOG feature value extracted by the 8-way HOG feature value extraction unit;
A four-way discriminator that determines the presence or absence of a person based on the four-way HOG feature value extracted by the four-way HOG feature value extraction unit.

A person detection device according to the second invention is the person detection device according to the first invention.
A luminance gradient direction image creating means for obtaining a luminance gradient direction image by obtaining eight luminance gradient directions and luminance gradient intensities in each pixel of the input image is provided.

A person detection device according to a third aspect of the invention is the person detection device according to the second aspect of the invention,
While the 8-direction HOG feature amount extraction unit extracts the 8-direction HOG feature amount based on the luminance gradient direction image,
Based on the luminance gradient direction image, the four-direction HOG feature amount extraction unit divides the luminance gradient strength in the intermediate direction with respect to the four directions out of the luminance gradient strengths in the eight directions into two adjacent directions to each component in the four directions. The four-way HOG feature value is extracted by synthesis.
A human detection device according to a fourth invention is
An image input means for inputting data of an image acquired by an imaging means for photographing the monitoring target area;
At least a processing setting unit for setting a plurality of object sizes corresponding to the size of a person appearing in the image and a person determination threshold value for determining the presence or absence of a person;
Object size setting means for sequentially selecting the plurality of object sizes and setting them as object sizes to be processed at present;
HOG feature quantity extraction means for extracting a HOG feature quantity, which is a feature quantity obtained by histogramating the gradient direction of luminance in a local area, based on the object size to be processed at the present time, for the image;
Identification means for judging the presence or absence of a person based on the HOG feature quantity extracted by the HOG feature quantity extraction means and the person judgment threshold;
A detection result aggregating unit that outputs the determination results of the presence / absence of a person based on all the object sizes together as a person detection result by collecting the determination results of the identification unit;
A plurality of the HOG feature amount extraction means are provided corresponding to a preset range of the object size to be processed, and a plurality of the identification means are provided corresponding to each of the HOG feature amount extraction means. A human detection device,
A luminance gradient direction image creating means for obtaining a luminance gradient direction image by obtaining eight luminance gradient directions and luminance gradient intensities in each pixel of the input image;
While the 8-direction HOG feature amount extraction unit extracts the 8-direction HOG feature amount based on the luminance gradient direction image,
Based on the luminance gradient direction image, the four-direction HOG feature amount extraction unit divides the luminance gradient strength in the intermediate direction with respect to the four directions out of the luminance gradient strengths in the eight directions into two adjacent directions to each component in the four directions. Combine and extract the 4-way HOG feature
It is characterized by that.

  According to the above-described person detection device according to the present invention, dense histogram data can always be used for collation of the discriminator in accordance with the size of the object on the image, and at a position away from the installation location of the surveillance camera. It is possible to detect a person who exists and appears small in the image with high accuracy.

It is a block diagram which shows schematic structure of the person detection apparatus which concerns on the Example of this invention. It is explanatory drawing which shows an example of the image data input into the person detection apparatus which concerns on the Example of this invention. FIGS. 3A to 3C are explanatory diagrams illustrating an example in which raster scanning is performed according to the object size, and FIG. 3D is an explanatory diagram illustrating an example of a person detection result. It is explanatory drawing which shows an example of the 8-direction luminance gradient direction histogram corresponding to the cell in the person detection apparatus which concerns on the Example of this invention, and the said cell. It is explanatory drawing which shows the learning example of the 8-way discriminator in the person detection apparatus which concerns on the Example of this invention. It is explanatory drawing which shows an example of the cell in the person detection apparatus which concerns on the Example of this invention, and a 4-way luminance gradient direction histogram corresponding to the said cell. It is explanatory drawing which shows the learning example of the 4-way discriminator in the person detection apparatus which concerns on the Example of this invention. FIG. 8A is an explanatory diagram illustrating an example of a luminance gradient direction histogram in eight directions, and FIG. 8B is an explanatory diagram illustrating an example of a luminance gradient direction histogram in four directions. It is a flowchart which shows the flow of the person detection process by the person detection apparatus based on the Example of this invention. It is explanatory drawing which shows an example of the 8-way luminance gradient direction histogram corresponding to the conventional cell and the said cell.

  Hereinafter, the details of the person detecting apparatus according to the present invention will be described with reference to the drawings.

A person detection apparatus according to an embodiment of the present invention will be described with reference to FIGS.
As shown in FIG. 1, a person detection apparatus 1 according to the present embodiment analyzes an image obtained by photographing a monitoring target area by a monitoring camera 2 and extracts a person from the image. 101, a processing setting unit 102, a storage unit 103, a luminance gradient direction image creation unit 104, an object size setting unit 105, an eight-direction HOG feature quantity extraction unit 106, an eight-direction discriminator 107, and four directions A HOG feature amount extraction unit 108, a four-direction discriminator 109, a detection result totaling unit 110, and a result output unit 111 are provided.

  The image input unit 101 inputs data of an image 3 (hereinafter referred to as image data) as shown in FIG. 2 taken by the monitoring camera 2 from the monitoring camera 2. Image data is stored in the storage unit 103.

The process setting unit 102 sets a plurality of target object sizes S1 to Sn, a target object size division number a, a person determination threshold value b, and the like as various process parameters necessary for person detection. These processing parameters are stored in the storage unit 103.
The object sizes S1 to Sn are sizes set in advance in a plurality of stages corresponding to the size of the person shown in the image 3 taken by the surveillance camera 2, and in this embodiment, as shown in FIG. N sizes. Further, the object size division number a is a number by which the object sizes S1 to Sn are divided in a lattice shape as shown in FIG. The person determination threshold value b is a threshold value for determining whether or not the probability obtained by the classifier is a person.
The storage unit 103 stores processing parameters and various data.

  The luminance gradient direction image creation unit 104 receives the image data from the storage unit 103 and creates a luminance gradient direction image obtained by obtaining the luminance gradient direction and the luminance gradient intensity in eight directions for each pixel. The created brightness gradient direction image data (hereinafter referred to as brightness gradient direction image data) is stored in the storage unit 103.

  The object size setting unit 105 sets the object size Si (i = 1, 2,..., N) to be currently processed among the object sizes S1 to Sn stored in the storage unit 103 (for example, FIG. 3). S2, S3,..., Sn and the object size Si to be processed are set in order from the largest object size S1 shown in FIG. The set object size Si to be processed at present is stored in the storage unit 103.

  The 8-direction HOG feature amount extraction unit 106 inputs the object size Si to be processed currently, the object size division number a, and the luminance gradient direction image data from the storage unit 103, and based on these, the 8-direction luminance gradient direction histogram HOG feature values (hereinafter, referred to as 8-way HOG feature values) are extracted. The extracted 8-way HOG feature value is sent to the 8-way discriminator 107.

  Specifically, as shown in FIG. 4, a square-shaped small region (this is referred to as a “cell” obtained by dividing the object size Si to be processed on the image in a grid shape based on the object size division number a. The luminance gradient direction histogram is created for each cell Cj for Cj (j = 1, 2,..., M), and a high-dimensional vector in which the luminance gradient direction histograms are arranged is extracted as an 8-way HOG feature quantity. To do. For example, when an eight-direction luminance gradient direction histogram is created for each cell Cj having a size of 5 × 5 pixels, the eight-direction HOG feature amount is obtained by data on luminance gradient directions obtained from 25 points for one cell Cj. It will be expressed as a histogram.

  Here, in this embodiment, when the size of the cell Cj is 5 × 5 pixels or more, the 8-way HOG feature quantity extraction unit 106 extracts the 8-way HOG feature quantity. The calculation of the luminance gradient direction can be performed by using a method of obtaining from the tan of the luminance gradient intensity in the vertical direction and the horizontal direction with respect to the target pixel or a method disclosed in Patent Document 3, and detailed description thereof is omitted here.

  The 8-direction discriminator 107 is a discriminator corresponding to the 8-way HOG feature quantity, and the person determination threshold value b inputted from the storage unit 103, the object size Si to be processed at present, and the 8-way HOG feature quantity extraction unit 106. Person detection (determination of whether or not the object is a person) is performed based on the eight-way HOG feature value sent from. The result of person detection is stored in the storage unit 103 as detection result data.

Here, the 8-direction discriminator 107 first performs learning as preparation before performing person detection. Specifically, as shown in FIG. 5, a large number of human images normalized to a certain size and non-human images (hereinafter collectively referred to as “learning image samples”) are prepared, and each learning image sample is prepared. The 8-way HOG feature value is extracted from the image and the 8-way HOG feature value component effective for human detection is automatically selected, and the probability that the selected component exists within the object size S1 to Sn and other than the person Is obtained, and a person judgment threshold value b is derived.
Then, the eight-way HOG feature value corresponding to the person selected by learning using the learning image sample in advance is compared with the eight-way HOG feature value input from the storage unit 103 to detect the person based on the person determination threshold value b. I do.

  The person detection using the above-described 8-direction HOG feature amount extraction unit 106 and the 8-direction discriminator 107 has a cell size of 5 × 5 pixels or more among the object sizes S1 to Sn with respect to the luminance gradient direction image. For example, as shown in FIG. 3A, the 8-direction HOG feature quantity extraction unit 106 extracts the 8-way HOG feature quantity and the 8-direction discriminator 107 detects the person by raster scanning. Among the object sizes S1 to Sn, person detection is performed for a cell size of 5 × 5 pixels or more.

  The 4-direction HOG feature amount extraction unit 108 inputs the object size Si to be processed currently, the object size division number a, and the luminance gradient direction image data from the storage unit 103, and based on these, the four-direction luminance gradient direction histogram HOG feature values (hereinafter referred to as 4-way HOG feature values) are extracted. The extracted 4-way HOG feature value is sent to the 4-way classifier 109.

  Here, in the human detection device 1 according to the present embodiment, the HOG feature amount calculated from the luminance gradient direction by the four-direction HOG feature amount extraction unit 108 is used, similarly to the above-described 8-direction HOG feature amount extraction unit 106. A rough shape of the person is expressed, and the person is detected by a four-way HOG discriminator 109 described later.

However, in the present embodiment, the configuration of the HOG feature amount corresponding to the object sizes S1 to Sn on the image is configured in the above-described eight directions so that a person far from the installation location of the monitoring camera 2 can be detected. It is assumed that switching is performed between a luminance gradient direction histogram and a luminance gradient direction histogram in four directions described later.
For example, as shown in FIG. 4, when a histogram is created based on information obtained from a large number of pixels with respect to the number of luminance gradient directions, a moderately dense histogram can be obtained. However, when the size of the object on the image is reduced, and the size of the cell Cj is reduced accordingly, for example, when a luminance gradient direction histogram in 8 directions is created for the cell Cj having a size of 3 × 3 pixels. Represents the data of the luminance gradient direction obtained from the nine points as a histogram. In this case, even the luminance gradient direction histogram obtained from the cell Cj at the same location relative to the object is a coarse histogram as shown in FIG. This is the main reason why it is difficult to detect the object when the object appears on the image 3 in a small size.
Therefore, in this embodiment, it is assumed that the number of directions at the time of creating the luminance gradient direction histogram is switched according to the size of the object on the image 3. Specifically, when the size of the object on the image 3 is equal to or larger than a preset size, the HOG feature amount is configured from the luminance gradient direction histograms in eight directions. That is, for example, when the cell size is 5 × 5 pixels or more, the luminance gradient direction histogram of 8 directions is used. On the other hand, when the cell size is smaller than 5 × 5 pixels, a luminance gradient direction histogram in four directions is used.

It is troublesome to obtain the luminance gradient direction in the object sizes S1 to Sn every time the HOG feature value is obtained by raster scanning as shown in FIGS. Therefore, in the person detection device 1 according to the present embodiment, as described above, the luminance gradient direction image creation unit 104 creates luminance gradient direction images in which eight luminance gradient directions and luminance gradient intensities are previously obtained from the image data. Keep it.
As a result, when the brightness gradient direction histogram of 8 directions is created, the histogram can be created as it is from the data of the brightness gradient direction image.
On the other hand, when a luminance gradient direction histogram in four directions is created, the luminance gradient direction histogram in eight directions (D1 to D8) shown in FIG. 8A obtained from the data of the luminance gradient direction image is shown in FIG. divided into two directions histogram values adjacent intermediate direction (d _1-2 ~d _4-1) for four directions (d1 to d4) as shown in the synthesis to the components in the four directions (d1 to d4) It shall be. As a result, as shown in FIG. 6, among the object sizes S1 to Sn on the image, a square shape obtained by dividing a cell size smaller than 5 × 5 pixels into a lattice shape based on the object size division number a. For each cell Cj, a luminance gradient direction histogram in four directions is created for each cell Cj, and a high-dimensional vector in which the luminance gradient direction histograms are arranged is extracted as a four-way HOG feature quantity.

  The four-direction discriminator 109 is a discriminator corresponding to the four-way HOG feature quantity, and the person determination threshold value b inputted from the storage unit 103, the object size Si to be processed at present and the four-way HOG feature quantity extraction unit Person detection (determination as to whether or not the object is a person) is performed based on the four-way HOG feature value sent from. The result of person detection is stored in the storage unit 103 as detection result data.

Here, in the four-way discriminator 109, first, as in the eight-way discriminator 107, learning is performed as preparation before performing person detection. Specifically, as shown in FIG. 7, a large number of learning image samples are prepared, and four-way HOG feature values are extracted from each learning image sample, and components of the four-way HOG feature values effective for human detection are automatically detected. A distribution of probability of being a person and a probability of being a person other than the person when the selected components exist within the object sizes S1 to Sn is obtained, and a person determination threshold value b is derived.
Then, the four-way HOG feature value corresponding to the person selected by learning using the learning image sample in advance and the four-way HOG feature value input from the storage unit 103 are compared, and the person is detected based on the person determination threshold value b. I do.
In the person detection using the above-described four-direction HOG feature amount extraction unit 108 and the four-direction discriminator 109, the cell size is less than 5 × 5 pixels among the object sizes S1 to Sn with respect to the luminance gradient direction image. For example, as shown in FIG. 3C, the four-way HOG feature quantity extraction unit 108 extracts the four-way HOG feature quantity and the four-way discriminator 109 detects the person by raster scanning. Among the object sizes S1 to Sn, person detection is performed for a cell size of less than 5 × 5 pixels.

  The detection result totaling unit 110 collects the results of human detection (detection result data) performed based on the object sizes S1 to Sn in a plurality of stages from the storage unit 103 as the human detection results for all sizes, and this is the detection result. Stored in the storage unit 103 as aggregated data. The detection result data is the number of people detected and the person detection position on the image.

  The result output unit 111 outputs the detection result total data input from the storage unit 103 as a person detection result to an external device (for example, a display device) not shown.

  The flow of person detection by the person detection apparatus according to this embodiment configured as described above will be described with reference to FIG.

As shown in FIG. 9, in the human detection apparatus according to the present embodiment, first, image data is input from the monitoring camera 2 by the image input unit 101 (step P1).
Subsequently, the above-described luminance gradient direction image is generated using the image data input from the storage unit 103 by the luminance gradient direction image generation unit 104 (step P2).
Subsequently, the target object size Si to be processed among the target object sizes S1 to Sn stored in the storage unit 103 is set by the target object size setting unit 105 (step P3).

  If the size of the cell obtained by dividing the object size Si into a lattice shape based on the object size Si to be processed and the object size division number a is 5 × 5 pixels or more, the process proceeds to step P5. If the cell size is less than 5 × 5 pixels, the process proceeds to step P7 (step P4).

When the process proceeds from step P4 to step P5, the object size Si to be processed, the object size division number a, and the luminance gradient direction image input from the storage unit 103 by the 8-direction HOG feature amount extraction unit 106 in step P5. An 8-way HOG feature value is extracted using the data.
Following step P5, the presence / absence of a person is determined from the eight-way HOG feature value input from the eight-way HOG feature value extraction unit 106 by the eight-way discriminator 107 and the person determination threshold value b input from the storage unit 103. Judge (step P6) and move to step P9.

When the process proceeds from step P4 to step P7, the object size Si to be processed currently, the object size division number a, and the luminance gradient input from the storage unit 103 by the four-way HOG feature quantity extraction unit 108 in step P7. The four-way HOG feature value is extracted using the directional image data.
Following step P7, the presence / absence of a person is determined from the four-way HOG feature value input from the four-way HOG feature value extraction unit 108 by the four-way classifier 109 and the person determination threshold value b input from the storage unit 103. Judge (Step P8) and proceed to Step P9.

In Step P9, it is determined whether or not the search using all the object sizes S1 to Sn is completed. If the search using all the object sizes S1 to Sn is not completed (NO), the process of Step P3 is performed. Returning, when the search by all the object sizes S1 to Sn is completed (YES), the detection result totaling unit 110 stores the person detection results based on the respective object sizes S1 to Sn from the storage unit 103. Enter and total.
Subsequently, the detection result summary data input from the storage unit 103 by the result output unit 111 is output to an external device (for example, a display device).

  By performing the above processing, an 8-way HOG feature amount extraction unit 106 and an 8-way classifier 107 are used for a person who is relatively large in the input image (in this embodiment, the cell size is 5 × 5 pixels or more). While human detection is performed, the human image is detected by the four-way HOG feature amount extraction unit 108 and the four-way classifier 109 for a person appearing relatively small in the input image (in this embodiment, the cell size is less than 5 × 5 pixels). When the search is completed for all the object sizes S1 to Sn, the detection results are collectively output.

According to the human detection apparatus according to the present embodiment described above, dense histogram data can always be used for collation of the discriminator according to the size on the image 3 of the object, and the installation location of the monitoring camera 2 It is possible to detect with high accuracy even a person who is present at a position away from the image 3 and is reflected in the image 3 small, and analyzes the image 3 of a wide range of monitoring locations photographed by the monitoring camera 2 and extends from the vicinity to the distance Person detection can be performed in a wide range.
Even when the surveillance camera 2 swings or zooms, the photographed image 3 of a wide range of surveillance locations can be analyzed to detect people over a wide range from near to far.

  The present invention relates to a person detection apparatus, and is particularly suitable for application to a person detection apparatus that detects a person in an image using HOG feature values.

DESCRIPTION OF SYMBOLS 1 Person detection apparatus 2 Surveillance camera 3 Image 101 Image input part 102 Process setting part 103 Memory | storage part 104 Brightness gradient direction image creation part 105 Object size setting part 106 8-way HOG feature-value extraction part 107 8-way identifier 108 Four-way HOG Feature extraction unit 109 Four-direction discriminator 110 Detection result totaling unit 111 Result output unit S1 to Sn Object size Cj cell

Claims (4)

An image input means for inputting data of an image acquired by an imaging means for photographing the monitoring target area;
At least a processing setting unit for setting a plurality of object sizes corresponding to the size of a person appearing in the image and a person determination threshold value for determining the presence or absence of a person;
Object size setting means for sequentially selecting the plurality of object sizes and setting them as object sizes to be processed at present;
HOG feature quantity extraction means for extracting a HOG feature quantity, which is a feature quantity obtained by histogramating the gradient direction of luminance in a local area, based on the object size to be processed at the present time, for the image;
Identification means for judging the presence or absence of a person based on the HOG feature quantity extracted by the HOG feature quantity extraction means and the person judgment threshold;
A detection result aggregating unit that outputs the determination results of the presence / absence of a person based on all the object sizes together as a person detection result by collecting the determination results of the identification unit;
A plurality of the HOG feature amount extraction means are provided corresponding to a preset range of the object size to be processed, and a plurality of the identification means are provided corresponding to each of the HOG feature amount extraction means. A human detection device,
Set the number of divisions to divide the object size into a grid by the processing setting unit,
The plurality of HOG feature amount extraction means include
When the size of the small area obtained by dividing the object size into a grid shape based on the number of divisions is equal to or larger than a preset size, the object size is configured from a luminance gradient direction histogram in eight directions with respect to the object size. An 8-way HOG feature quantity extraction unit for extracting 8-way HOG feature quantities;
A four-way HOG feature quantity extraction unit that extracts a four-way HOG feature quantity configured from a luminance gradient direction histogram in four directions for the object size when the size of the small area is smaller than a preset size ; ,
The plurality of identification means are:
An 8-way discriminator that determines the presence or absence of a person based on the 8-way HOG feature value extracted by the 8-way HOG feature value extraction unit;
Human object detection apparatus you wherein there in the four directions classifier to determine the presence or absence of a person based on the four directions HOG features extracted by the four directions HOG features extraction section. Human detection apparatus according to claim 1 Symbol mounting, characterized in that it comprises a brightness gradient direction image creation means for creating a brightness gradient direction image seeking brightness gradient direction and the brightness gradient strength of 8 directions in each pixel of the input image. While the 8-direction HOG feature amount extraction unit extracts the 8-direction HOG feature amount based on the luminance gradient direction image,
Based on the luminance gradient direction image, the four-direction HOG feature amount extraction unit divides the luminance gradient strength in the intermediate direction with respect to the four directions out of the luminance gradient strengths in the eight directions into two adjacent directions to each component in the four directions. The person detection apparatus according to claim 2, wherein the four-way HOG feature value is extracted by synthesis. An image input means for inputting data of an image acquired by an imaging means for photographing the monitoring target area;
At least a processing setting unit for setting a plurality of object sizes corresponding to the size of a person appearing in the image and a person determination threshold value for determining the presence or absence of a person;
Object size setting means for sequentially selecting the plurality of object sizes and setting them as object sizes to be processed at present;
HOG feature quantity extraction means for extracting a HOG feature quantity, which is a feature quantity obtained by histogramating the gradient direction of luminance in a local area, based on the object size to be processed at the present time, for the image;
Identification means for judging the presence or absence of a person based on the HOG feature quantity extracted by the HOG feature quantity extraction means and the person judgment threshold;
A detection result aggregating unit that outputs the determination results of the presence / absence of a person based on all the object sizes together as a person detection result by collecting the determination results of the identification unit;
A plurality of the HOG feature amount extraction means are provided corresponding to a preset range of the object size to be processed, and a plurality of the identification means are provided corresponding to each of the HOG feature amount extraction means. A human detection device,
A luminance gradient direction image creating means for obtaining a luminance gradient direction image by obtaining eight luminance gradient directions and luminance gradient intensities in each pixel of the input image;
While the 8-direction HOG feature amount extraction unit extracts the 8-direction HOG feature amount based on the luminance gradient direction image,
Based on the luminance gradient direction image, the four-direction HOG feature amount extraction unit divides the luminance gradient strength in the intermediate direction with respect to the four directions out of the luminance gradient strengths in the eight directions into two adjacent directions to each component in the four directions. Combine and extract the 4-way HOG feature
A person detection apparatus characterized by the above.

Images