KR102653755B1 - System and method for collecting field image data sets for learning artificial intelligence image deep learning models - Google Patents

Abstract

The present invention builds a field image data set from field images captured by a collection image sensor installed on the field, so that an image deep learning fine-tuning model additionally learned from the field image data set can be created to suit the domestic road environment. , the accuracy and reliability of object detection can be further improved. In addition, by constructing a field image data set by selecting images from field images according to the shape of the object, size of the object, presence or absence of field background, shooting time, degree of overlap between objects, and degree of overlap between objects and background, field video data The reliability of the set can be improved.

Description

System and method for collecting field image data sets for learning artificial intelligence image deep learning models {System and method for collecting field image data sets for learning artificial intelligence image deep learning models}

The present invention relates to a system and method for collecting field image data sets for learning artificial intelligence image deep learning models. More specifically, field images taken of domestic road sites are filtered according to preset selection criteria to create a field image data set. This is about a system and method for collecting field video data sets for learning artificial intelligence video deep learning models built with .

In general, images from cameras installed on the road are used to determine traffic conditions or unexpected situations on the road. However, there is a problem in that it takes a long time because the manager must visually check the video from the camera and then guide the drivers.

Therefore, recently, there has been increasing interest in developing technology to acquire images using cameras installed on the road and detect and track objects using deep learning models. However, when using a pre-trained deep learning model, there is a problem in that accuracy decreases depending on the road environment.

Korean Patent No. 10-2453627

The purpose of the present invention is to provide a system and method for collecting field image data sets for learning an artificial intelligence image deep learning model that can collect field image data sets that reflect the characteristics of the field.

The system for collecting field image data sets for learning an artificial intelligence image deep learning model according to the present invention includes an image sensor installed at a plurality of preset sites to capture field images; The field images are filtered according to preset selection criteria, images that satisfy the selection criteria are collected into a selection image data set, and at least some of the images that are not selected as the selection image data set among the field images are noiseized. It includes a server that collects video data sets and mixes the selected video data set and the noise video data set at a preset ratio to construct a field video data set.

The server compares the shape of the object, the size of the object, the presence or absence of the scene background, the shooting time, the degree of overlap between objects, and the degree of overlap between objects and the background in the field images with selection criteria set differently for each, and selects the object. It includes a data set construction unit that selects image data sets.

The server learns a pre-stored general-purpose image data set to create an initial image deep learning model that derives the type of object in the image, and additionally learns the initial image deep learning model with the field image data set to determine the type of object in the image. It includes a model learning unit that generates an image deep learning fine-tuning model that reflects the characteristics and derives the type of object in the image.

The method of collecting field image data sets for learning an artificial intelligence image deep learning model according to the present invention includes a field image collection step of collecting field images taken by image sensors installed at a plurality of sites where a server is preset; An image selection step of filtering the field images according to preset selection criteria and collecting images satisfying the selection criteria into a selection image data set; A data set mixing step in which the server mixes the selected image data set and a noise image data set obtained by collecting at least some of the images that were not selected as the selected image data set in the image selection step at a preset ratio; It includes a data set storage step in which the server constructs the data set mixed in the data set mixing step into a field image data set.

The image selection step compares the shape of the object, the size of the object, the presence or absence of the scene background, the shooting time, the degree of overlap between objects, and the degree of overlap between objects and the background in the field images with selection criteria set differently for each, Selection is performed using the selected image data set.

The image selection step includes comparing the ratio of the shape of the object in the field image with a preset ratio, and selecting the object if the ratio of the shape of the object is greater than or equal to the set ratio.

In the image selection step, if the ratio in which the shape of the object appears is less than the set ratio and the size of the part of the object in which the shape appears is greater than or equal to a preset size, the image in which the object appears is excluded, and the image in which the shape of the object appears is excluded. If the ratio is less than the set ratio and the size of the part of the object is less than the preset size, a process of excluding only the corresponding object from the image in which the object appears is included.

The image selection step includes comparing the size of the object in the field image with a preset size, determining that the size of the object is greater than or equal to the set size, and selecting the object.

The image selection step includes comparing the background in the scene image with a background setting preset according to the scene, and selecting an image including the background setting.

The image selection step includes comparing the shooting time zone of the field video with a preset time zone, and selecting the corresponding image if the shooting time zone is within the set time zone.

In the image selection step, the degree of overlap between objects in the field image is compared with a preset overlap standard, and if the degree of overlap between objects satisfies the overlap standard, the image from which the object appears is selected, and the degree of overlap between objects is selected. If the overlapping criteria are not met, the process includes excluding the image in which the object appears.

In the image selection step, the degree of overlap between the object and the background in the field image is compared with a preset overlap ratio, and if the overlap between the object and the background is less than the set overlap ratio, the image in which the object appears is selected, If the degree of overlap between the object and the background is greater than the set overlap ratio, a process of excluding the image in which the object appears is included.

A method of collecting field image data sets for learning an artificial intelligence image deep learning model according to another aspect of the present invention includes a field image collection step of collecting field images taken by image sensors installed at a plurality of sites where servers are preset; ; An image selection step of filtering the field images according to preset selection criteria and collecting images satisfying the selection criteria into a selection image data set; A data set mixing step in which the server mixes the selected image data set and a noise image data set obtained by collecting at least some of the images that were not selected as the selected image data set in the image selection step at a preset ratio; The server includes a data set storage step of constructing the mixed data set in the data set mixing step into a field image data set, and the image selection step is to set the ratio of the shape of the object in the field image to a preset setting ratio. Compare with, and if the ratio of the shape of the object is more than the set ratio, the object is adopted, the ratio of the shape of the object is less than the set ratio, and the size of the part of the object is a preset size If it is above, the image in which the object appears is excluded, and if the proportion of the shape of the object is less than the set ratio and the size of the part of the object is less than the preset size, only the object is included in the image in which the object appears. A process of excluding, comparing the size of the object in the scene image with a preset size, and if the size of the object is larger than the set size, adopting the object, and adjusting the background in the scene image in advance according to the scene. A process of selecting an image including the setting background by comparing it with a set setting background, comparing the shooting time zone of the scene video with a preset setting time zone, and selecting the image if the shooting time zone is within the setting time zone. In the process, the degree of overlap between objects in the field image is compared with a preset overlap standard, and if the degree of overlap between objects satisfies the overlap standard, the image in which the object appears is adopted, and the degree of overlap between objects is compared with the overlap standard. If it is not satisfied, the process of excluding the image in which the object appears, comparing the degree of overlap between the object and the background in the field image with a preset overlap ratio, and if the overlap degree of the object and the background is less than the set overlap ratio, It includes a process of selecting an image in which the corresponding object appears and, if the degree of overlap between the object and the background is greater than the set overlap ratio, excluding the image in which the corresponding object appears.

The present invention builds an on-site image data set from on-site images captured by image collection sensors installed on-site, so that an image deep learning fine-tuning model additionally learned from the on-site image data set can be created to suit the domestic road environment. , the accuracy and reliability of object detection can be further improved.

In addition, by constructing a field image data set by selecting images from field images according to the shape of the object, size of the object, presence or absence of field background, shooting time, degree of overlap between objects, and degree of overlap between objects and background, field video data The reliability of the set can be improved.

Figure 1 is a diagram schematically showing the configuration of a system for collecting field image data sets for learning an artificial intelligence image deep learning model according to an embodiment of the present invention.
Figure 2 is a flowchart showing the steps of learning an artificial intelligence image deep learning model according to an embodiment of the present invention.
Figure 3 is a flowchart showing a method of collecting field image data sets for learning an image deep learning model according to an embodiment of the present invention.
Figure 4 shows an example of selection according to object shape when collecting a field image data set according to an embodiment of the present invention.
Figure 5 shows an example of selection according to object size when collecting a field image data set according to an embodiment of the present invention.
Figure 6 shows an example of selection according to the degree of overlap between objects when collecting a field image data set according to an embodiment of the present invention.
Figure 7 shows an example of selection according to the degree of overlap between an object and a background when collecting a field image data set according to an embodiment of the present invention.

Hereinafter, embodiments of the present invention will be described with reference to the attached drawings.

Figure 1 is a diagram schematically showing the configuration of a system for collecting field image data sets for learning an artificial intelligence image deep learning model according to an embodiment of the present invention.

Referring to FIG. 1, a system for collecting field image data sets for learning an artificial intelligence image deep learning model according to an embodiment of the present invention includes an image sensor 12 for collection and a server 30.

The collection image sensor 12 is an image sensor that is installed at a site preset by an administrator or the like and captures on-site images. A plurality of the image sensors 12 for collection are installed at a plurality of sites. The image collection sensor 12 is explained as an example of a CCTV installed on or around a road. In this embodiment, the collection image sensor 12 is explained as an example of being installed separately from a detection image sensor (not shown) that captures a detection image to be input to an image deep learning fine-tuning model described later. However, it is not limited to this, and some of the detection image sensors (not shown) may be used as the collection image sensor 12, and the detection image sensor (not shown) and the collection image sensor ( Of course, it is possible for 12) to be used equally.

The server 30 is a learning server that includes a data set construction unit 31 and a model learning unit 32.

The data set construction unit 31 selects the field images collected from the collection image sensors 12 and builds a field image data set. In this embodiment, the data set construction unit 31 is described as being included in the server 30, but the data set construction unit 31 is not limited to this, and may of course be configured separately from the server 30.

The field image data set is image data about the road environment, road conditions, and driving vehicles at the site where the collection image sensors 12 are installed.

The field image data set includes preset selection criteria for the size of the object, background, shooting time, degree of overlap between objects, and degree of overlap between objects and the background among the field images collected through the collection image sensors 12. In comparison, the selected images that satisfy the selection criteria and the noise images excluding the selected images among the field images collected through the collection image sensors 12 are mixed at a preset ratio. It is one data set. Here, the noise images are explained by way of example as samples of some images excluding the selected images among the field images collected through the image collection image sensors 12. For example, the mixing ratio of the selected images and the noise images is 1:1. The method of selecting the field images will be described in detail later.

The model learning unit 32 learns a general-purpose image data set pre-stored in the server 30 to create an initial image deep learning model that derives the type of object in the image.

Here, the general-purpose image data set is image data provided for general purposes by an overseas external organization, etc., and is pre-stored in the server 30 or a separate database (not shown).

In addition, the model learning unit 32 generates the image deep learning initial model and then additionally learns the image deep learning initial model with the field image data set to generate an image deep learning fine tuning model. do.

By additionally learning with the field image data set, the characteristics of the field where the collection image sensor 12 is installed can be reflected, and the type of object in the image can be derived more accurately.

The method of collecting a field image data set for learning an artificial intelligence image deep learning model according to an embodiment of the present invention configured as described above will be described as follows.

Figure 2 is a flowchart showing the steps of learning an artificial intelligence image deep learning model according to an embodiment of the present invention.

Referring to Figure 2, the artificial intelligence image deep learning model learning step according to an embodiment of the present invention includes an initial model learning step (S21), an on-site image data set generation step (S22), a fine-tuned model learning step (S23), and Includes a verification step (S24).

In the initial model learning step (S21), the model learning unit 32 of the server 30 learns the pre-stored general-purpose image data set to generate an initial image deep learning model for deriving the type of object in the image. This is the step.

The field image data set generation step (S22) is a step of collecting and generating the field image data set. In this embodiment, the field image data set generation step (S22) is described as an example after the initial model learning step (S21), but it is not limited to this and can also be performed before the initial model learning step (S21). Of course it is possible.

Figure 3 is a flowchart showing a method of collecting field image data sets for learning an image deep learning model according to an embodiment of the present invention.

Referring to FIG. 3, the field image data set generation step (S22) includes the field image collection step (S22-1), the image selection step (S22-2), the data set mixing step (S22-3), and the field image data. Includes a set storage step (S22-4).

The field image collection step (S22-1) is a process in which the server 30 collects field images captured by the image collection sensors 12 installed at a plurality of sites by a user or an administrator.

The image selection step (S22-2) is a process of comparing the collected field images with preset selection criteria and creating a selection image data set of images that satisfy the selection criteria.

In the image selection step (S22-2), the shape and size of the object in the scene image, the presence or absence of the scene background, the shooting time of the scene image, the degree of overlap between objects in the scene image, and the object in the scene image Select by checking the degree of overlap between the and background. The selection criteria are preset differently depending on the shape of the object, the size of the object, the presence or absence of the field background, the shooting time, the degree of overlap between the objects, and the degree of overlap between the object and the background.

The image selection process according to the shape of the object compares the ratio of the shape of the object in the field image with a preset ratio, and selects an object whose ratio is greater than or equal to the set ratio.

At this time, if the ratio of the shape of the object is less than the set ratio and only a part of the shape of the object is shown, the size of the object is compared with the first preset size.

That is, in the case where the shape of the object in the scene image is only partially shown, and the size of the object is larger than the first set size, the size of the object only partially shown is too large, so it is judged to be a noise image and the image in which the object appears is excluded. .

Meanwhile, if the shape of the object in the scene image is only partially shown, and the size of the object is less than the first set size, the size of the object only partially shown is small, and therefore only the corresponding object is excluded from the image.

Figure 4 shows an example of selection according to object shape when collecting a field image data set according to an embodiment of the present invention.

Referring to FIG. 4, since only a portion of the object, a car, was captured in the image and the size of the object in the image is very large, the image can be excluded and classified as a noise image.

Additionally, the image selection process according to the size of the object compares the size of the object with a second preset size, selects objects larger than the second set size and excludes objects smaller than the second set size. If the size of the object is less than the second set size, the object is excluded from the image because the object is too small to extract features.

Here, the second set size is preset to a size that can recognize the object and distinguish the type of object. Although the example is described as being set differently from the first setting size, it is not limited to this and the first setting size and the second setting size may be the same.

Figure 5 shows an example of selection according to object size when collecting a field image data set according to an embodiment of the present invention.

Referring to FIG. 5, a plurality of objects are included in the scene image, and among the plurality of objects, only objects larger than the second set size can be selected.

In addition, the image selection process according to the presence or absence of the scene background compares the background in the scene image with a background preset according to the scene, selects the image including the set background, and selects the image that does not include the set background. exclude The setting background is preset to a background that reflects the characteristics of each site.

For example, images collected from the collection image sensors 12 installed at a site with an overpass preferentially select images containing an overpass, and the images collected from the collection image sensors 12 installed at a beach site For the images collected from , images containing the sea are given priority. Therefore, the scene image data set can include images of various backgrounds from a plurality of scenes.

In addition, the image selection process according to the shooting time zone compares the shooting time zone of the on-site video with a preset setting time zone, selects the image if the shooting time zone is within the setting time zone, and the shooting time zone is within the setting time zone. If it deviates, the video is excluded. For example, the set time zone may be set as a time zone excluding a predetermined time zone from after sunset to before sunset. Accordingly, the field image data set includes images from multiple sites at various times, and more accurate image data can be used.

In addition, the image selection process according to the degree of overlap between objects is a process of determining and selecting whether the degree of overlap between the objects satisfies a preset overlap standard when a plurality of objects are included in the field image. The degree of overlap between the objects refers to the degree to which the objects overlap each other. The image selection process according to the degree of overlap between objects may be performed through a preset analysis algorithm. The overlapping criterion is explained as an example in which at least one of the four lines of a rectangular bounding box that detects and displays an object in the image is not completely obscured by an object visible in front. That is, if even one of the four lines of the bounding box of the first object is completely obscured by another object in front, it is determined that the overlap criterion is not satisfied and the image in which the object appears is excluded. When only the first object is excluded from the corresponding image, the first object can be recognized as the background, so the corresponding image is excluded for accurate learning.

Figure 6 shows an example of selection according to the degree of overlap between objects when collecting a field image data set according to an embodiment of the present invention.

Referring to Figure 6a, this is a case where two first and second objects (1) and (2) are detected in the image, and one of the four lines of the first bounding box indicating the first object (1) is It represents a state completely obscured by the second object (2). In this case, it is determined that the degree of overlap between the objects does not satisfy the overlap criteria, and the corresponding image is excluded.

Referring to FIG. 6B, a plurality of objects (1) and (2) are detected in the image, and since none of the four lines of the bounding box are completely obscured among the plurality of objects, the degree of overlap between the objects is It is determined that the above overlap criteria are satisfied, and the corresponding image is selected.

In addition, in the image selection process according to the degree of overlap between the object and the background, when an object in the field image is obscured by the background, the degree of overlap between the object and the background, which is the degree to which the object is obscured by the background, is set in advance. This is a selection process compared to the overlap ratio.

If the overlap between the object and the background is more than the set overlap ratio, the corresponding image is excluded, and if it is less than the set overlap ratio, the corresponding image is adopted. The image selection process according to the degree of overlap between the object and the background may be performed through a preset analysis algorithm. Here, the set overlap ratio is explained as an example of 60%.

Figure 7 shows an example of selection according to the degree of overlap between an object and a background when collecting a field image data set according to an embodiment of the present invention.

Referring to Figure 7, this shows a case where the object (O) detected in the image is obscured by the road sign (S), and if the degree to which the object (O) is obscured by the road sign (S) is more than 60%, the image is excluded. do.

In the image selection step (S22-2), the shape of the object, the size of the object, the presence or absence of the scene background, the shooting time of the scene image, the degree of overlap between objects in the scene image, and the difference between the object and the background in the scene image. Since selection is made according to selection items including overlap, the selection order can be changed and applied in various ways. In addition, in this embodiment, it is explained as an example that the video is selected only when all of the selection items are satisfied. However, the image is not limited to this, and it is of course possible to select images that satisfy at least some of the selection items.

The images selected and selected through the image selection step (S22-2) configured as above become a selected image data set. At this time, the selected images may be further selected to ensure a balanced ratio for each type of object.

Among the field images, images excluded through the image selection step (S22-2) are classified as a noise image data set.

After the image selection step (S22-2) is completed, the data set mixing step (S22-3) is performed.

The data set mixing step (S22-3) is a step of mixing the selected image data set and the noise image data set at a preset ratio. Here, the ratio is 1:1 as an example. However, the method is not limited to this, and if it is determined that the selected image data set alone is sufficient for further learning, the field image data set may be composed of only the selected image data set.

The field image data set saving step (S22-4) is a step of storing the data set mixed in the data set mixing step (S22-3) as the field image data set.

When the field image data set is created in the same manner as above, the fine-tuning model learning step (S23) is performed using the field image data set.

In the fine-tuning model learning step (S23), the model learning unit 32 additionally learns the initial image deep learning model with the field image data set to derive the type of object by reflecting the characteristics of the field. This is the step of creating a deep learning fine-tuning model. By additionally learning with the field image data set in the fine-tuning model learning step (S23), the type of object in the image can be more accurately derived by reflecting the characteristics of the field where the image sensor is installed.

In the verification step (S24), the verification data sets pre-stored in the image deep learning initial model and the image deep learning fine-tuning model are respectively input and the results derived from each are compared to determine the accuracy of the image deep learning fine-tuning model. This is the verification step. If the accuracy of the image deep learning fine-tuning model is less than the preset standard accuracy, selection criteria, etc. can be adjusted in the field image data set creation step.

As described above, in the present invention, by constructing a field image data set with field images of an actual field from which object information is to be derived, an image deep learning fine-tuning model generated by additional learning with the field image data set can view the characteristics of the field. Object information can be derived by accurately reflecting it.

The present invention has been described with reference to the embodiments shown in the drawings, but these are merely exemplary, and those skilled in the art will understand that various modifications and equivalent other embodiments are possible therefrom. Therefore, the true scope of technical protection of the present invention should be determined by the technical spirit of the attached patent claims.

12: Image sensor for collection 30: Server
31: Data set construction unit 32: Model learning unit

Claims (13)

An image sensor installed at a plurality of preset sites to capture field images;
The field images are filtered according to preset selection criteria, images that satisfy the selection criteria are collected into a selection image data set, and at least some of the images that are not selected as the selection image data set among the field images are noiseized. Comprising a server that collects video data sets and mixes the selected video data set and the noise video data set at a preset ratio to construct a field video data set,
The server is,
In the field images, the shape of the object, size of the object, presence or absence of field background, shooting time, degree of overlap between objects, and degree of overlap between objects and background are compared with selection criteria set differently for each, and the selected image data set is selected. Including a data set construction unit for selecting,
A collection system for field video data sets for learning artificial intelligence video deep learning models. delete An image sensor installed at a plurality of preset sites to capture field images;
The field images are filtered according to preset selection criteria, images that satisfy the selection criteria are collected into a selection image data set, and at least some of the images that are not selected as the selection image data set among the field images are noiseized. Comprising a server that collects video data sets and mixes the selected video data set and the noise video data set at a preset ratio to construct a field video data set,
The server is,
Create an initial video deep learning model that derives the types of objects in the video by learning a pre-stored general-purpose video data set, and further learn the initial video deep learning model with the field video data set to reflect the characteristics of the field. Comprising a model learning unit that generates an image deep learning fine-tuning model that derives the type of object in the image,
A collection system of field video data sets for learning artificial intelligence video deep learning models. A field image collection step in which a server collects field images captured by image sensors installed at a plurality of preset sites;
An image selection step in which the server filters the field images according to preset selection criteria and collects images satisfying the selection criteria into a selection image data set;
A data set mixing step in which the server mixes the selected image data set and a noise image data set obtained by collecting at least some of the images that were not selected as the selected image data set in the image selection step at a preset ratio;
A data set storage step in which the server constructs the mixed data set in the data set mixing step into a field image data set,
The image selection step is,
In the field images, at least some of the shape of the object, the size of the object, the presence or absence of the field background, the shooting time, the degree of overlap between objects, and the degree of overlap between the object and the background are compared with selection criteria set differently for each, and the selected image Selecting data sets,
Method of collecting field video data sets for learning artificial intelligence video deep learning models. delete In claim 4,
The image selection step is,
Compare the ratio of the shape of the object in the scene image with a preset ratio,
If the rate at which the shape of the object appears is greater than the set rate, including the process of selecting the image where the object appears,
Method of collecting field video data sets for learning artificial intelligence video deep learning models. In claim 6,
The image selection step is,
If the ratio of the shape of the object is less than the set ratio and the size of the part of the object is more than the preset size, exclude the image in which the object appears,
If the ratio of the shape of the object is less than the set ratio and the size of the part of the object is less than the preset size, including the process of excluding only the object from the image in which the object appears and adopting the image,
Method of collecting field video data sets for learning artificial intelligence video deep learning models. In claim 4,
The image selection step is,
By comparing the size of the object in the field image with a preset size,
The size of the object is greater than or equal to the set size, and comprising the process of selecting an image from which the object appears,
Method of collecting field video data sets for learning artificial intelligence video deep learning models. In claim 4,
The image selection step is,
Comprising the process of comparing the background in the scene video with a preset background according to the scene and selecting an image including the set background,
Method of collecting field video data sets for learning artificial intelligence video deep learning models. In claim 4,
The image selection step is,
By comparing the shooting time zone of the scene video with the preset setting time zone,
If the shooting time zone is within the set time zone, including the process of selecting the video,
Method of collecting field video data sets for learning artificial intelligence video deep learning models. In claim 4,
The image selection step is,
By comparing the degree of overlap between objects in the field image with a preset overlap standard,
If the degree of overlap between the objects satisfies the overlap criteria, the image in which the object appears is selected,
Including a process of excluding the image in which the object appears if the degree of overlap between the objects does not satisfy the overlap criterion,
Method of collecting field video data sets for learning artificial intelligence video deep learning models. In claim 4,
The image selection step is,
By comparing the degree of overlap between the object and the background in the field image with a preset overlap ratio,
If the degree of overlap between the object and the background is less than the set overlap ratio, the image from which the object appears is selected,
If the degree of overlap between the object and the background is greater than the set overlap ratio, including the process of excluding the image in which the object appears,
Method of collecting field video data sets for learning artificial intelligence video deep learning models. A field image collection step in which a server collects field images captured by image sensors installed at a plurality of preset sites;
An image selection step in which the server filters the field images according to preset selection criteria and collects images satisfying the selection criteria into a selection image data set;
A data set mixing step in which the server mixes the selected image data set and a noise image data set obtained by collecting at least some of the images that were not selected as the selected image data set in the image selection step at a preset ratio;
A data set storage step in which the server constructs the mixed data set in the data set mixing step into a field image data set,
The image selection step is,
The rate at which the shape of the object appears in the field image is compared with a preset setting ratio, and if the rate at which the shape of the object appears is greater than the set ratio, the image in which the object appears is selected, and the rate at which the shape of the object appears is If it is less than the set ratio and the size of the exposed part of the object is greater than or equal to a preset size, the image in which the object appears is excluded, and if the ratio of the shape of the object is less than the set ratio, the size of the exposed part of the object is excluded. If the size is less than the preset size, a process of excluding only the object in question from the image in which the object appears and adopting the image;
Comparing the size of the object in the field image with a preset size, and if the size of the object is greater than the set size, selecting the image in which the object appears;
A process of comparing the background in the scene video with a preset background according to the scene and selecting an image including the set background;
By comparing the shooting time zone of the scene video with the preset setting time zone,
If the shooting time zone is within the set time zone, a process of selecting the video,
The degree of overlap between objects in the field image is compared with a preset overlap standard, and if the degree of overlap between objects satisfies the standard for overlap, the image in which the object appears is adopted. If the degree of overlap between objects does not satisfy the standard for overlap, the image from which the object appears is selected. If not, the process of excluding the image in which the object appears,
The degree of overlap between the object and the background in the field image is compared with a preset overlap ratio, and if the overlap between the object and the background is less than the set overlap ratio, the image from which the object appears is selected, and the overlap between the object and the background is selected. If the degree is greater than the set overlap ratio, including the process of excluding the image in which the object appears,
Method of collecting field video data sets for learning artificial intelligence video deep learning models.