JP6989873B2 - System, image recognition method, and computer - Google Patents

Description

The present invention relates to a system, a method, and a computer for performing image recognition.

A technique for tracking an object such as a person or a vehicle from image data such as a moving image is known (see, for example, Patent Document 1).

Patent Document 1 describes "a normal example showing an acquisition unit for sequentially acquiring images, an extraction unit for extracting a partial region from an image and extracting a feature amount from the partial region, the extracted feature amount, and a target object". Whether or not the partial region is the target object is recognized based on the first recognition model including the feature amount of the negative case showing the feature amount of the target object and the feature amount of the positive case or the second recognition model including the feature amount of the positive case. It includes a recognition unit, an update unit that adds and updates the extracted feature amount to the first recognition model based on the recognition result, and an output unit that outputs the object area recognized as the target object. The recognition unit recognizes based on the first recognition model when the object area is output for the previous image acquired by the acquisition unit, and when the object area is not output for the previous image, the second recognition unit is used. Recognize based on the recognition model. "

Japanese Unexamined Patent Publication No. 2012-238119

In the conventional image recognition system, in order to track the tracking target, it is necessary to prepare a recognition model (discriminator) for recognizing the tracking target in advance. Generally, the recognition model needs to be generated by performing pre-learning using a large amount of training data, and it takes time to complete the pre-learning.

The present invention provides a system, a method, and a computer capable of promptly starting tracking of a tracked object without prior learning when instructed to start tracking the tracked object.

A typical example of the invention disclosed in the present application is as follows. That is, it is a system including a recognition server that executes recognition processing for tracking the tracking target included in the image data and a learning server that learns the classifier used for the recognition processing, and the recognition server is the first. It has 1 arithmetic unit, a 1st storage device connected to the 1st arithmetic apparatus, and a 1st interface connected to the 1st arithmetic apparatus, and the learning server is used in the 2nd arithmetic apparatus and the 2nd arithmetic apparatus. The recognition server has a second storage device to be connected and a second interface connected to the second arithmetic unit, and the recognition server receives initial image data to which the tracking target is designated, and a plurality of sizes smaller than the tracking target. The initial image data is divided into blocks, the feature amount of each block included in the initial image data is calculated, and a teacher signal indicating whether or not the block includes a part of the tracking target is used as the initial image. Initial learning data is generated by adding to the feature amount of each block included in the data, the initial learning data is transmitted to the learning server, and the learning server performs learning using the initial learning data. By doing so, a discriminator for classifying each block included in the received image data into either a block including a part of the tracking target or a block not including the tracking target is generated, and the discriminator is referred to as the discriminator. The data is transmitted to the recognition server, and the recognition server uses the classifier to execute the recognition process on the received image data and outputs the result of the recognition process.

According to one embodiment of the present invention, tracking of a tracking target can be started quickly without performing pre-learning using a large amount of learning data. Issues, configurations and effects other than those mentioned above will be clarified by the description of the following examples.

It is a figure which shows the configuration example of the computer system of Example 1. FIG. It is a figure which shows an example of the hardware configuration and software configuration of the recognition server of Example 1. FIG. It is a flowchart explaining the process of initial learning executed by the recognition module of Embodiment 1. FIG. It is a flowchart explaining the tracking process of the tracking target after the completion of the initial learning executed by the recognition module of the first embodiment. It is a figure which shows an example of the image data in which the tracking target area is set input to the recognition server of Example 1. FIG. It is a figure which shows an example of the learning data generated by the recognition server of Example 1. FIG. It is a flowchart explaining the process which the learning module of Example 1 executes.

Hereinafter, examples of the present invention will be described with reference to the accompanying drawings. The same reference numerals are given to the common configurations in each figure.

FIG. 1 is a diagram showing a configuration example of the computer system of the first embodiment.

The computer system includes a recognition server 100, a learning server 101, an image pickup device 105, and a database 106. The computer system may be configured to include either the image pickup apparatus 105 or the database 106.

The recognition server 100, the learning server 101, the image pickup device 105, and the database 106 are connected to each other directly or via a network. The present invention is not limited to the type of network. The types of networks include LAN (Local Area Network) and WAN (Wide Area Network). Further, the network connection method may be either wired or wireless.

The image pickup device 105 is a device such as a camera, and transmits the captured image as image data 150 to the recognition server 100 or the database 106. The database 106 stores the image data 150.

The recognition server 100 has a recognition module 110 that executes recognition processing. The recognition module 110 acquires the image data 150 from the image pickup apparatus 105 or the database 106, and executes the recognition process of the acquired image data 150. Further, the recognition module 110 generates learning data using the image data 150 and transmits it to the learning server 101.

The recognition process of this embodiment is a process for tracking the tracking target included in the image data.

The learning server 101 has a learning module 120 that performs learning by deep learning. The learning module 120 learns the classifier used for the recognition process using the learning data. Further, the learning module 120 transmits the learned classifier to the recognition server 100.

The classifier is represented as data such as a function or a matrix. In this embodiment, at the time when the image data 150 for which the tracking target is designated is input, the classifier for recognizing the tracking target has not been learned. As will be described later, the learning module 120 learns the classifier using the initial learning data generated based on the image data 150 to which the tracking target is designated.

In the following description, the image data 150 for which the tracking target is designated is referred to as the initial image data 150.

Although the recognition server 100 and the learning server 101 are represented as physically different computers, they may be realized by using a virtual computer running on one computer.

FIG. 2 is a diagram showing an example of a hardware configuration and a software configuration of the recognition server 100 of the first embodiment.

The recognition server 100 has an arithmetic unit 201, a storage device 202, a network interface 203, and an I / O interface 204 as hardware. The hardware is connected to each other via an internal bus or the like.

It is assumed that the hardware configuration of the learning server 101 is the same as that of the recognition server 100.

The arithmetic unit 201 executes a program stored in the storage device 202. The arithmetic unit 201 operates as a module that realizes a specific function by executing processing according to a program. In the following description, when the process is described with the module as the subject, it is shown that the arithmetic unit 201 is executing the program that realizes the module.

The storage device 202 stores the program executed by the arithmetic unit 201 and the information used by the program. The storage device 202 also includes a work area that the program temporarily uses. As the storage device 202, for example, a memory or the like can be considered.

The storage device 202 of the first embodiment stores a program that realizes the recognition module 110. Further, the image data 150 to be processed is temporarily stored in the storage device 202.

The network interface 203 is an interface for connecting to another device via a network.

The I / O interface 204 is an interface for connecting to an input device and an output device (not shown). The input device is, for example, a keyboard, a mouse, a touch panel, or the like. The output device is, for example, a touch panel, a display, or the like.

Regarding the recognition module 110 included in the recognition server 100, one module may be divided into a plurality of modules for each function.

3A and 3B are flowcharts illustrating the processing executed by the recognition module 110 of the first embodiment. FIG. 3A is a flowchart illustrating the initial learning process executed by the recognition module of the first embodiment. FIG. 3B is a flowchart illustrating the tracking process of the tracking target after the completion of the initial learning, which is executed by the recognition module of the first embodiment. FIG. 4 is a diagram showing an example of image data 150 in which a tracking target area is set, which is input to the recognition server 100 of the first embodiment. FIG. 5 is a diagram showing an example of a process of generating learning data according to the first embodiment.

First, the recognition module 110 accepts the input of the initial image data 150 (step S101). In this embodiment, it is assumed that the initial image data 150 is one.

Here, a method of designating a tracking target will be described with reference to FIG. The user refers to the image data 150 displayed on the output device, and specifies the tracking target by setting an area surrounding the tracking target. In the following description, the area where the tracking target exists is described as the tracking target area.

The recognition module 110 accepts the input of the initial image data 150 whose tracking target is designated by the tracking target area. At this time, the recognition module 110 holds the center position, size, and the like of the tracking target area as setting information of the tracking target area.

In the following description, the portion of the image data 150 that is not the tracking target is described as the background, and the area excluding the tracking target area is described as the background area.

The tracking target may be specified after the recognition module 110 receives the initial image data 150. For example, when the recognition module 110 receives the image data 150 for the first time, the recognition module 110 displays a screen as shown in FIG. 4 on the output device, and displays a display prompting the setting of the tracking target area.

In this embodiment, the number of initial image data 150 is one, but it may be plural. However, the number of initial image data 150 is sufficiently smaller than the number of learning data used in conventional machine learning.

Next, the recognition module 110 executes the initialization process (step S102).

Specifically, the recognition module 110 initializes a storage area for storing the blocks classified as the tracking target and a storage area for storing the blocks classified in the background. In the following description, the storage area for storing the blocks classified as the tracking target is described as the first storage area, and the storage area for storing the blocks classified as the background is described as the second storage area.

Next, the recognition module 110 generates initial learning data using the initial image data 150 and transmits it to the learning server 101 (step S103). Specifically, the recognition module 110 executes the following processing.

(Process 1) The recognition module 110 divides the initial image data 150 into regions of an arbitrary size as shown in FIG. In the following description, one area is referred to as a block. The size of the block is set to be larger than the pixel and smaller than the tracking target. For example, when the size of the image data is 640 × 480, the recognition module 110 generates 300 blocks by dividing the image data into 32 vertically and 32 horizontally. It is assumed that the number of divisions is set in advance. However, the number of divisions can be changed to any value. The rectangle in the block represents a pixel.

(Process 2) The recognition module 110 calculates the feature amount of each block. For example, as shown in the equation (1), the combination of the RGB values of each pixel included in the block is calculated as the feature amount _fi of the block. In this case, the feature amount F of the image data 150 is expressed by the equation (2).

Here, i is a block identification number and takes a value from 1 to n. Note that n is an integer larger than 1. x and y are coordinates indicating the relative positions of pixels in the block.

(Process 3) The recognition module 110 gives a teacher signal indicating that the tracking target is a tracking target to the block included in the tracking target area. Further, the recognition module 110 gives a teacher signal indicating that the background is a background to the block included in the background area. In this embodiment, when a part of the block is included in the tracking target area, the block is treated as a block included in the tracking target area. In the following description, the teacher signal indicating that it is a tracking target is described as a first teacher signal, and the teacher signal indicating that it is a background is described as a second teacher signal.

Here, when the teacher signal ti given to the block whose identification number is _i is defined as in the equation (3), the initial learning data S is given as in the equation (4).

(Process 4) The recognition module 110 transmits the initial learning data S represented by the equation (4) to the learning server 101. The above is the description of step S103.

Next, the recognition module 110 determines whether or not the initial learning is completed (step S104).

Specifically, the recognition module 110 determines whether or not the classifier generated by the learning server 101 has been received. The classifier of this example classifies blocks as either tracked or background.

If it is determined that the initial learning has not been completed, the recognition module 110 shifts to the waiting state until the initial learning is completed.

When it is determined that the initial learning is completed, the recognition module 110 acquires new image data 150 (step S105) and starts the block loop processing (step S106).

Specifically, the recognition module 110 divides the acquired image data 150 into blocks and calculates the feature amount of each block. Further, the recognition module 110 selects a target block from the blocks. Further, the recognition module 110 sets the search area based on the setting information of the tracking target area. In this embodiment, the center position of the search area is set to be the same as the center position of the tracking target area, and the size of the search area is set to be larger than the size of the tracking target area. For example, the size of the search area is set to 1.5 times the size of the tracked area. The recognition module 110 temporarily holds the setting information of the search area.

The recognition module 110 determines whether or not the target block is included in the search area (step S107).

Specifically, the recognition module 110 determines whether or not the target block is included in the search area based on the position of the target block in the image data 150 and the setting information of the search area.

When it is determined that the target block is not included in the search area, the recognition module 110 classifies the target block as a background and stores the target block in the second storage area (step S111). After that, the recognition module 110 proceeds to step S112.

When it is determined that the target block is included in the search area, the recognition module 110 executes the recognition process (step S108).

Specifically, the recognition module 110 inputs the feature amount of the target block to the discriminator, and determines whether or not the target block includes a part of the tracking target based on the value obtained from the discriminator. If the target block contains part of the tracked object, the target block is classified as tracked. If the target block does not contain part of the tracked object, the target block is classified as background. The result of the classification enables tracking of the tracking target in the image data.

When the target block is classified as the background, the recognition module 110 stores the target block in the second storage area (step S111). After that, the recognition module 110 proceeds to step S112.

When the target block is classified as a tracking target, the recognition module 110 stores the target block in the first storage area (step S110). After that, the recognition module 110 proceeds to step S112.

In step S112, the recognition module 110 determines whether or not the processing is completed for all the blocks (step S112).

If it is determined that the processing is not completed for all the blocks, the recognition module 110 returns to step S106, selects a new target block, and executes the same processing.

When it is determined that the processing is completed for all the blocks, the recognition module 110 updates the tracking target area (step S113).

Specifically, the recognition module 110 estimates the tracking target area based on the distribution of the blocks classified into the tracking target in the image data 150. The recognition module 110 updates the setting information of the tracking target area based on the estimation result.

For example, the area surrounding all the blocks classified as the tracking target is estimated as the tracking target area. Further, the recognition module 110 calculates the center of gravity of the block classified as the tracking target, and estimates a rectangle of an arbitrary size based on the center of gravity as the tracking target area. The method for estimating the tracking target area is an example and is not limited to this.

Next, the recognition module 110 outputs the recognition result of the tracking target (step S114).

For example, the recognition module 110 generates display information for displaying the recognition result of the tracking target and outputs the display information to the output device. The display information includes, for example, data for displaying a block determined to be a tracking target, a tracking target area, and the like.

Next, the recognition module 110 generates additional learning data using the recognized image data 150 and transmits it to the learning server 101 (step S115).

Specifically, the recognition module 110 generates additional learning data by adding a first teacher signal to the block determined to be the tracking target and adding a second teacher signal to the block determined to be the background. The structure of the additional learning data is the same as that of the initial learning data S.

Next, the recognition module 110 determines whether or not the tracking of the tracking target has been completed (step S116). For example, when the tracking end instruction is received, the recognition module 110 determines that the tracking is completed.

When it is determined that the tracking of the tracking target is completed, the recognition module 110 ends the process.

If it is determined that the tracking of the tracking target has not been completed, the recognition module 110 returns to step S105 and executes the same process. The recognition module 110 executes the same process by using the classifier that is sequentially updated by the learning server 101. When returning to step S105, the recognition module 110 may shift to the waiting state until the updated classifier is received.

FIG. 6 is a flowchart illustrating a process executed by the learning module 120 of the first embodiment.

When the learning server 101 receives the initial learning data or the additional learning data from the recognition server 100, the learning server 101 starts the process described below. In the following description, when the initial learning data and the additional learning data are not distinguished, they are described as learning data.

The learning module 120 performs learning by deep learning using the learning data (step S201). By this learning, it is possible to learn the tracking target and the background in block units.

The present invention is not limited to the deep learning method used. For example, it is conceivable to use RBM (Restricted Boltzmann Machine).

Next, the learning module 120 generates or updates the classifier based on the learning result, and transmits the classifier to the recognition server 100 (step S202). After that, the learning module 120 ends the process.

Specifically, when the initial learning data is received, the learning module 120 generates a discriminator and transmits the generated discriminator to the recognition server 100. Further, when the additional learning data is received, the learning module 120 updates the classifier and transmits the updated classifier to the recognition server 100.

In conventional learning, a classifier (recognition model) for recognizing the entire tracked object has been generated. In order to generate such a classifier, it is necessary to perform learning using a large amount of learning data. Therefore, there is a problem that it is costly to prepare a large amount of learning data and the learning time is very long.

On the other hand, the learning server 101 of this embodiment performs learning by deep learning using the feature amount of the block obtained by dividing one initial learning data. That is, the learning server 101 learns a plurality of blocks included in the tracking target area as tracking targets. Therefore, it is not necessary to prepare a large amount of learning data, and the learning time is short. Learning using additional learning data can also be performed at high speed.

In the conventional recognition process, the entire tracking target is recognized. Since the entire tracking target is treated as one feature quantity, the discriminator generated by learning using only the image data input for the first time has very low recognition accuracy. Therefore, even if the classifier is used, rapid tracking of the tracking target cannot be started.

On the other hand, the recognition server 100 of this embodiment recognizes the tracking target for the block included in the search area. It is considered that the feature amount of the block has a small change amount with respect to the movement of the tracking target and the change in the shape of the tracking target. Therefore, it is possible to maintain a certain recognition accuracy even when the tracking target is recognized immediately after the tracking target is designated. Therefore, rapid tracking of the tracked object can be started using the classifier generated using the initial training data. Further, since the block to be recognized is only the block included in the search area, real-time tracking of the tracking target is possible.

Since the discriminator is updated by the learning server 101 even during the continuation of tracking, it is clear that the recognition accuracy increases with the passage of time.

As described above, according to the present embodiment, it is possible to realize a system, an image recognition method, and a computer that quickly start tracking of a tracking target without performing learning using a large amount of learning data.

The present invention is not limited to the above-described embodiment, and includes various modifications. Further, for example, the above-described embodiment describes the configuration in detail in order to explain the present invention in an easy-to-understand manner, and is not necessarily limited to the one including all the described configurations. Further, it is possible to add, delete, or replace a part of the configuration of each embodiment with other configurations.

Further, each of the above configurations, functions, processing units, processing means and the like may be realized by hardware by designing a part or all of them by, for example, an integrated circuit. The present invention can also be realized by a software program code that realizes the functions of the examples. In this case, a storage medium in which the program code is recorded is provided to the computer, and the processor included in the computer reads out the program code stored in the storage medium. In this case, the program code itself read from the storage medium realizes the function of the above-described embodiment, and the program code itself and the storage medium storing it constitute the present invention. Examples of the storage medium for supplying such a program code include a flexible disk, a CD-ROM, a DVD-ROM, a hard disk, an SSD (Solid State Drive), an optical disk, a magneto-optical disk, a CD-R, and a magnetic tape. Non-volatile memory cards, ROMs, etc. are used.

In addition, the program code that realizes the functions described in this embodiment can be implemented in a wide range of programs or script languages such as assembler, C / C ++, perl, Shell, PHP, and Java (registered trademark).

Further, by distributing the program code of the software that realizes the functions of the embodiment via the network, the program code is stored in a storage means such as a hard disk or a memory of a computer or a storage medium such as a CD-RW or a CD-R. The processor included in the computer may read and execute the program code stored in the storage means or the storage medium.

In the above-described embodiment, the control lines and information lines show what is considered necessary for explanation, and do not necessarily indicate all the control lines and information lines in the product. All configurations may be interconnected.

100 Recognition server 101 Learning server 105 Imaging device 106 Database 110 Recognition module 120 Learning module 150 Image data 201 Arithmetic device 202 Storage device 203 Network interface 204 I / O interface

Claims (8)

A system including a recognition server that executes a recognition process for tracking a tracking target included in image data and a learning server that learns a classifier used for the recognition process.
The recognition server has a first arithmetic unit, a first storage device connected to the first arithmetic unit, and a first interface connected to the first arithmetic unit.
The learning server has a second arithmetic unit, a second storage device connected to the second arithmetic unit, and a second interface connected to the second arithmetic unit .
The recognition server is
The initial image data specified by the tracking target is received, and the initial image data is divided into a plurality of blocks having a size smaller than the tracking target.
The feature amount of each block included in the initial image data is calculated, and the feature amount is calculated.
Initial learning data is generated by applying a teacher signal indicating whether or not the block includes a part of the tracking target to the feature amount of each block included in the initial image data.
The initial learning data is transmitted to the learning server, and the initial learning data is transmitted to the learning server.
The learning server is
By performing learning using the initial learning data, identification that classifies each block included in the received image data into either a block containing a part of the tracking target or a block not including the tracking target. Generate a vessel,
Send the classifier to the recognition server and
The recognition server is
Using the classifier, the recognition process for the received image data is executed.
A system characterized by outputting the result of the recognition process. The system according to claim 1.
The first storage device stores setting information of a tracking target area, which is an area including the tracking target of the image data.
The recognition server is
When new image data is received after receiving the classifier from the learning server, the new image data is divided into the plurality of blocks.
The feature amount of each block included in the new image data is calculated, and the feature amount is calculated.
Based on the setting information of the tracking target area, the search area, which is the target area of the recognition process, is set.
Select the target block from the plurality of blocks included in the new image data, and select the target block.
It is determined whether or not the target block is a block included in the search area, and it is determined.
If it is determined that the target block is not a block included in the search area, the target block is classified into a block that does not include a part of the tracking target.
When it is determined that the target block is a block included in the search area, the feature amount of the target block is input to the classifier.
Based on the output of the classifier, the target block is classified into either a block containing a part of the tracking target or a block not including a part of the tracking target.
A new tracked area is estimated based on the distribution of blocks containing a part of the tracked object.
A system characterized in that the setting information of the tracking target area is updated based on the estimation result of the new tracking target area. The system according to claim 2.
The learning server is
When the additional learning data generated by the recognition server is received, the discriminator is updated by performing learning using the additional learning data.
The updated classifier is sent to the recognition server and
The recognition server is
The additional learning data is generated by adding the teacher signal corresponding to the result of the classification to the feature quantities of the plurality of blocks included in the new image data.
The additional learning data is transmitted to the learning server, and the data is transmitted to the learning server.
Using the updated classifier, the recognition process for the received image data is executed.
A system characterized by outputting the result of the recognition process. An image recognition method performed by a system with multiple computers.
The system includes a recognition server that executes a recognition process for tracking a tracking target included in image data and a learning server that learns a classifier used for the recognition process.
The recognition server has a first arithmetic unit, a first storage device connected to the first arithmetic unit, and a first interface connected to the first arithmetic unit.
The learning server has a second arithmetic unit, a second storage device connected to the second arithmetic unit, and a second interface connected to the second arithmetic unit.
The image recognition method is
A step in which the recognition server receives the initial image data to which the tracking target is specified and divides the initial image data into a plurality of blocks having a size smaller than the tracking target.
A step in which the recognition server calculates the feature amount of each block included in the initial image data, and
A step of generating initial learning data by applying a teacher signal indicating whether or not the recognition server is a block including a part of the tracking target to the feature amount of each block included in the initial image data. When,
A step in which the recognition server transmits the initial learning data to the learning server,
When the learning server performs learning using the initial learning data , each block included in the received image data is either a block containing a part of the tracking target or a block not including the tracking target. Steps to generate a classifier to classify into
A step in which the learning server transmits the classifier to the recognition server,
A step in which the recognition server executes the recognition process for the received image data using the classifier.
An image recognition method comprising the step of outputting the result of the recognition process by the recognition server. The image recognition method according to claim 4.
The first storage device stores setting information of a tracking target area indicating an area including the tracking target of the image data.
The image recognition method is
When the recognition server receives the new image data after receiving the classifier, the step of dividing the new image data into the plurality of blocks and the step of dividing the new image data into the plurality of blocks.
A step in which the recognition server calculates the feature amount of each block included in the new image data, and
A step in which the recognition server sets a search area, which is an area to be recognized, based on the setting information of the tracking target area.
A step in which the recognition server selects a target block from a plurality of blocks included in the new image data.
A step in which the recognition server determines whether or not the target block is a block included in the search area.
When the recognition server determines that the target block is not a block included in the search area, the step of classifying the target block into a block not including the tracking target, and
When the recognition server determines that the target block is a block included in the search area, the feature amount of the target block is input to the classifier, and the target block is based on the output of the classifier. Is classified into either a block containing a part of the tracking target or a block not including the tracking target.
A step in which the recognition server estimates a new tracking target area based on the distribution of blocks containing a part of the tracking target.
An image recognition method comprising the step of updating the setting information of the tracking target area based on the estimation result of the new tracking target area. The image recognition method according to claim 5.
A step in which the recognition server generates additional learning data by adding the teacher signal corresponding to the result of the classification to the feature quantities of the plurality of blocks included in the new image data.
A step in which the recognition server transmits the additional learning data to the learning server,
When the learning server receives the additional learning data, the step of updating the classifier by performing learning using the additional learning data, and
A step in which the learning server sends the updated classifier to the recognition server,
A step in which the recognition server executes the recognition process for the received image data by using the updated classifier.
An image recognition method comprising the step of outputting the result of the recognition process by the recognition server. It is a computer that executes a recognition process that recognizes the tracking target included in the image data.
It has an arithmetic unit, a storage device connected to the arithmetic unit, and an interface connected to the arithmetic unit.
The arithmetic unit is
As the information for designating the tracking target, the initial image data in which the tracking target area including the tracking target is set is received, and the setting information of the tracking target area is stored in the storage device.
The initial image data is divided into a plurality of blocks having a size smaller than the tracking target, and the initial image data is divided into a plurality of blocks.
The feature amount of each block included in the initial image data is calculated, and the feature amount is calculated.
Initial learning data is generated by applying a teacher signal indicating whether or not the block includes a part of the tracking target to the feature amount of each block included in the initial image data.
The initial learning data is output to a learning unit that learns the classifier used for the recognition process.
New image data after receiving a classifier that classifies each block included in the image data generated by the learning unit into either a block containing a part of the tracking target or a block not including the tracking target. Is received, the new image data is divided into the plurality of blocks, and the new image data is divided into the plurality of blocks.
The feature amount of each block included in the new image data is calculated, and the feature amount is calculated.
Based on the setting information of the tracking target area, the search area, which is the target area of the recognition process, is set.
Select the target block from the plurality of blocks included in the new image data, and select the target block.
It is determined whether or not the target block is a block included in the search area, and it is determined.
When it is determined that the target block is not a block included in the search area, the target block is classified into a block not including the tracking target.
When it is determined that the target block is a block included in the search area, the feature amount of the target block is input to the classifier.
Based on the output of the classifier, the target block is classified into either a block containing a part of the tracking target or a block not including the tracking target.
A new track target area is estimated based on the distribution of blocks containing the track target, and a new track target area is estimated.
Based on the result of the estimation of the new tracking target area, the setting information of the tracking target area is updated.
A computer characterized by outputting the result of the recognition process including the result of the classification. The computer according to claim 7.
The arithmetic unit is
Additional learning data is generated by adding the teacher signal corresponding to the result of the classification to the feature quantities of the plurality of blocks included in the new image data.
The additional learning data is output to the learning unit, and the data is output to the learning unit.
A computer characterized in that the recognition process for the received image data is executed by using the classifier updated by the learning unit.

Images