JP7177609B2 - Image recognition device, image recognition method, machine learning model providing device, machine learning model providing method, machine learning model generating method, and machine learning model device - Google Patents

Description

The disclosed embodiments relate to an image recognition device, an image recognition method, a machine learning model providing device, a machine learning model providing method, a machine learning model generating method, and a machine learning model device.

Conventionally, image data used for image recognition processing is collected from multiple image recognition devices, and machine learning for image recognition processing is performed using the accumulated image data as learning data, and the parameters used for image recognition processing are updated. There is a system that provides the image recognition device with the image recognition device.

JP 2015-135552 A

However, the conventional image recognition device may not be able to perform image recognition of the recognition target accurately depending on the relative speed between the imaging device that captures the recognition target and the recognition target.

One aspect of the embodiments has been made in view of the above, and enables accurate image recognition of a recognition target regardless of the relative speed between an imaging device that captures an image of the recognition target and the recognition target. An object of the present invention is to provide an image recognition device, an image recognition method, a machine learning model providing device, a machine learning model providing method, a machine learning model generating method, and a machine learning model device.

An image recognition device according to an aspect of an embodiment includes an acquisition unit, a determination unit, a storage unit, and a determination unit. The acquisition unit acquires an image in which a recognition target is captured from the imaging device. A determination unit determines a relative speed between the imaging device and the recognition object. The storage unit stores, for each speed range, a machine learning model for deriving a recognition result that a subject in the image is the recognition target object from an image of the recognition target object. The determination unit determines the subject in the image by employing the machine learning model in the speed range corresponding to the relative speed determined by the determination unit from among the plurality of machine learning models.

An image recognition device, an image recognition method, a machine learning model providing device, a machine learning model providing method, a machine learning model generating method, and a machine learning model device according to one aspect of an embodiment are an imaging device that captures an image of a recognition target and a recognition target. It is possible to accurately perform image recognition of a recognition object regardless of the relative speed with respect to the object.

FIG. 1 is an explanatory diagram showing an outline of an image recognition method according to an embodiment. FIG. 2 is a block diagram showing an example of the configuration of the image recognition device according to the embodiment. FIG. 3A is an explanatory diagram illustrating a procedure for generating a machine learning model according to the embodiment; FIG. 3B is an explanatory diagram illustrating a procedure for generating a machine learning model according to the embodiment; FIG. 3C is an explanatory diagram illustrating a procedure for generating a machine learning model according to the embodiment; FIG. 4 is an explanatory diagram illustrating an example of image recognition processing according to the embodiment. 5 is a flowchart illustrating an example of processing executed by a control unit of the image recognition device according to the embodiment; FIG.

Hereinafter, embodiments of an image recognition device, an image recognition method, a machine learning model providing device, a machine learning model providing method, a machine learning model generating method, and a machine learning model device will be described in detail with reference to the accompanying drawings. In addition, this invention is not limited by embodiment shown below. In the following, an image recognition apparatus and an image recognition method for recognizing an object appearing in an image of the surroundings of a vehicle obtained from an imaging device mounted on the vehicle will be described as an example.

FIG. 1 is an explanatory diagram showing an outline of an image recognition method according to an embodiment. The image recognition device 1 according to the embodiment acquires an image of the surroundings of the vehicle from an imaging device mounted on the vehicle. Here, for example, as shown in FIG. 1A, the image recognition device 1 is a vehicle (hereinafter referred to as "own vehicle") in which the imaging device and the image recognition device 1 are mounted. A case where an image 101 showing a vehicle 100 is acquired will be described as an example.

When a moving recognition object such as another vehicle 100 is captured, the image 101 may include another vehicle 100 having a shape slightly different from the shape of the other vehicle 100 seen by humans. For example, when another vehicle 100 moving at high speed is imaged by an imaging device with a relatively slow shutter speed, blurring occurs in the image of the other vehicle 100 in the image 101 as shown in FIG. There is

In addition, when an image capturing apparatus with a relatively slow shutter speed captures an image of a subject moving in the lateral direction at high speed, the subject may be captured in an image with a shape that extends in the lateral direction more than the actual shape of the subject.

Specifically, the imaging device sequentially drives (exposes) the imaging elements arranged in the upper row of an imaging region in which a plurality of imaging elements are arranged in a matrix to capture an image of one frame. Therefore, the subject imaged in the upper part of the imaging area has already moved in the horizontal direction when it is imaged in the lower part of the imaging area.

As a result, a so-called rolling shutter phenomenon (hereinafter referred to as a rolling phenomenon) occurs in an image, in which an object that is longer in the horizontal direction than its actual shape is captured. The degree of such rolling phenomenon and the magnitude of blurring vary depending on the difference in relative speed between the imaging device and the subject.

For this reason, for example, in order to generate a machine learning model for deriving an image recognition result that the subject is another vehicle 100 from an image 101 in which another vehicle 100 is captured, the degree of rolling phenomenon and the magnitude of blurring must be determined. Machine learning using a huge number of different images as teaching materials is required.

Further, even if a machine learning model is generated by machine learning using a huge number of images as teaching materials, an accurate image recognition result that the subject is another vehicle 100 from the image 101 showing another vehicle 100 can be obtained. In some cases, it cannot be derived.

For example, when recognizing a vehicle in an image with a rolling phenomenon or blurring using a machine learning model, even if an image similar to that image is included in the machine learning teaching material, the rolling phenomenon or blurring phenomenon may not Many images with different degrees of blurring are included.

Therefore, in the determination by the machine learning model, the likelihood of the determination result that the subject in the image is a vehicle is low, the subject in the image is not determined to be a vehicle, and an accurate image recognition result cannot be derived. Sometimes.

Therefore, when the image recognition device 1 according to the embodiment acquires the image 101, first, the relative speed between the imaging device and the recognition target (here, another vehicle 100) is determined. Then, the image recognition device 1 determines, for example, in which range the relative speed is included among the first speed range, the second speed range, and the third speed range.

Here, the first speed range is the slowest speed range, and the second speed range is faster than the first speed range and slower than the third speed range. The image recognition device 1 determines, for example, the moving speed of the subject in the image as the relative speed. Note that the moving speed determination method is not limited to this.

At this time, for example, when the image recognition device 1 determines that the relative speed is included in the second speed range as shown in FIG. 1(b), the image recognition processing shown in FIG. 1(c) is performed. Specifically, the image recognition device 1 has, for example, a first speed range machine learning model 1-1, a second speed range machine learning model 1-2, and a third speed range machine learning model 1-3. It has three machine learning models.

The first speed range machine learning model 1-1 is, for example, a machine learning model specialized for image recognition of a vehicle whose relative speed to the subject is within the first speed range. The second speed range machine learning model 1-2 is, for example, a machine learning model specialized for image recognition of a vehicle whose relative speed to the subject is within the second speed range.

The third speed range machine learning model 1-3 is, for example, a machine learning model specialized for image recognition of a vehicle whose relative speed to the subject is within the third speed range. Note that the number of machine learning models provided in the image recognition device 1 is not limited to three as long as it is plural.

Moreover, below, when referring to an unspecified machine learning model among the three machine learning models, it is simply described as a machine learning model. An example of the procedure for generating these three machine learning models will be described later with reference to FIGS. 3A, 3B, and 3C.

As described above, when the image recognition apparatus 1 determines that the relative velocity with respect to the subject is included in the second velocity range, the image recognition apparatus 1 transfers the image to the machine learning model 1-2 for the second velocity range among the three machine learning models. Enter 101. As a result, the image recognition device 1 can derive the correct image recognition result 102 that the subject is another vehicle 100 .

Next, an example of the configuration of the image recognition device 1 according to the embodiment will be described with reference to FIG. FIG. 2 is a block diagram showing an example of the configuration of the image recognition device 1 according to the embodiment. As shown in FIG. 2 , the image recognition device 1 is connected to an imaging device 41 , radar 42 , steering angle sensor 43 and vehicle control device 44 .

Note that the image collection classified by relative speed 110 is an image collection including a plurality of images in which the object to be recognized is shown. Each image included in the image collection classified by relative speed 110 is provided with information indicating the type of subject in the image and the relative speed between the subject and the imaging device at the time of imaging.

The image collection classified by relative speed 110 is used by the image recognition apparatus 1 as teaching materials for supervised machine learning of image recognition processing. An example of the image collection classified by relative speed 110 will be described later with reference to FIGS. 3A, 3B, and 3C together with the description of the procedure for generating the machine learning model.

The imaging device 41 is, for example, an in-vehicle camera that is installed at the front of the vehicle on which the image recognition device 1 is mounted and captures an image of the front of the vehicle. Note that the imaging device 41 may be an in-vehicle camera that is provided at the side or rear of the vehicle and captures an image of the surroundings of the vehicle. The imaging device 41 outputs the captured image to the image recognition device 1 .

The radar 42, for example, radiates millimeter-wave transmission waves to the surroundings of the vehicle, receives the reflected waves of the transmission waves reflected by the target, and calculates the distance from the vehicle to the target based on the transmission waves and the reception waves. It is a millimeter-wave radar that detects the angle at which a target exists and the relative speed of the target to the vehicle. The radar 42 outputs the detected relative velocity to the image recognition device 1 . The steering angle sensor 43 is a sensor that detects the steering angle of the vehicle. The steering angle sensor 43 outputs the detected steering angle to the image recognition device 1 .

The vehicle control device 44 is, for example, an ECU (Electronic Control Unit) that controls the entire vehicle. For example, when the image recognition device 1 recognizes an object that may hinder the running of the own vehicle, the vehicle control device 44 controls the speed, steering, and braking of the vehicle to prevent the vehicle from being disturbed. avoid things.

The image recognition device 1 also includes a control unit 2 and a storage unit 3 . The storage unit 3 is, for example, an information storage device such as data flash. The storage unit 3 stores a first speed range machine learning model 1-1, a second speed range machine learning model 1-2, ... an n-th speed range machine learning model 1-n (n is a natural number of 3 or more ) are stored as n machine learning models.

The control unit 2 includes a microcomputer having a CPU (Central Processing Unit), a ROM (Read Only Memory), a RAM (Random Access Memory), and various circuits.

The control unit 2 includes a learning unit 21, an acquisition unit 22, a determination unit 23, and a determination unit 24 that function when the CPU executes a program stored in the ROM using the RAM as a work area. Note that the control unit 2 may be configured by hardware such as an ASIC (Application Specific Integrated Circuit) or an FPGA (Field Programmable Gate Array).

The learning unit 21, the acquisition unit 22, the determination unit 23, and the determination unit 24 included in the control unit 2 implement or execute the information processing operation described below. Note that the internal configuration of the control unit 2 is not limited to the configuration shown in FIG. 2, and may be another configuration as long as it performs the information processing described later.

The learning unit 21 acquires a plurality of images (for example, thousands of images) for each relative velocity range and object type from the collection of images classified by relative velocity 110 . Then, the learning unit 21 uses the acquired images as teaching materials for supervised machine learning to create a machine learning model (for example, a first speed range for image recognition of a vehicle) for each relative speed range and object type. Machine learning model 1-1, etc.) is generated and stored in the storage unit 3.

Here, an example of the machine learning model generation procedure according to the embodiment will be described with reference to FIGS. 3A, 3B, and 3C. 3A, 3B, and 3C are explanatory diagrams showing the procedure for generating a machine learning model according to the embodiment.

Here, a case of generating a machine learning model for image recognition of a vehicle included in an image will be described. The learning unit 21 can also generate machine learning models for image recognition of motorcycles, bicycles, pedestrians, etc., in addition to vehicles.

As shown in FIG. 3A, when the learning unit 21 generates the machine learning model 1-1 for the first speed range in which the relative speed between the subject and the imaging device 41 is the slowest, the learning unit 21 selects the first speed range from the image collection 110 classified by relative speed. Images 11-1, 11-2, etc. of the range image collection 10-1 are obtained.

The images included in the first speed range image collection 10-1 have the slowest relative speed between the subject and the imaging device 41, so rolling phenomena and blurring occur as in images 11-1 and 11-2. do not have.

The learning unit 21 machine-learns the features of these images, so that when an image of a vehicle whose relative speed to the imaging device is within the first speed range is input, an image indicating that the subject is the vehicle is input. A first speed range machine learning model 1-1 for outputting recognition results is generated.

Further, as shown in FIG. 3B, when the learning unit 21 generates the machine learning model 1-2 for the second speed range in which the relative speed between the subject and the imaging device 41 is the second fastest after the first speed range, the relative The images 12-1, 12-2, etc. of the second speed range image collection 10-2 are acquired from the speed-specific image collection 110. FIG.

The images included in the second speed range image collection 10-2 are slightly different, such as images 12-1 and 12-2, because the relative speed between the subject and the imaging device 41 is faster than that in the first speed range. is rolling or blurring.

The learning unit 21 machine-learns the features of these images, so that when an image of a vehicle whose relative speed with respect to the imaging device is within the second speed range is input, an image indicating that the subject is a vehicle is input. A second speed range machine learning model 1-2 for outputting recognition results is generated.

Further, as shown in FIG. 3C, when the learning unit 21 generates the n-th speed range machine learning model 1-n in which the relative speed between the object and the imaging device 41 is the fastest, Images 13-1, 13-2, etc. of n speed range image collection 10-n are obtained.

Since the images included in the n-th speed range image collection 10-n have the fastest relative speed between the subject and the imaging device 41, large shakes and remarkable rolling phenomena occur as in images 13-1 and 13-2. It has occurred.

The learning unit 21 machine-learns the features of these images, so that when an image of a vehicle whose relative speed to the imaging device is within the n-th speed range is input, an image indicating that the subject is a vehicle is input. A machine learning model 1-n for the n-th speed range that outputs recognition results is generated.

In this way, the learning unit 21 obtains information for each speed range based on a plurality of images in which the type of authentication object moving at a relative speed within each speed range input for each speed range of the relative speed is captured. Generate machine learning models. As a result, the machine learning model for each speed range does not reflect the influence of the captured image of the authentication target moving at a relative speed that is not included in each speed range.

In addition, since the learning unit 21 does not need to identify and learn the speed of the subject from images used as teaching materials by performing machine learning on images for each speed range, the number of images used as teaching materials can be greatly reduced. can.

The above-described machine learning model generation procedure has been described assuming that the frame rate (shutter speed) is fixed. Normalize the range according to the shutter speed.

For example, if a certain frame rate is set as the reference frame rate and the frame rate of the imaging device 41 is the reference frame rate, the reference speed range corresponding to the reference frame rate is multiplied by 1 as a coefficient to set the speed range. .

If the frame rate is other than the reference frame rate, the speed range is obtained by multiplying the reference speed range by a coefficient (reference frame rate/imaging frame rate). Thereby, the learning unit 21 can generate a machine learning model of an image for each appropriate speed range according to the frame rate set in the imaging device 41 .

Also, in images used as teaching materials for machine learning, the movement speed of a recognition target moving within the image varies depending on the zoom ratio. For this reason, it is desirable to normalize the zoom magnification of images used as teaching materials so that the distance from the own vehicle to the recognition target in the image is the reference distance.

Similarly, for the image that the image recognition device 1 (determining unit 24, which will be described later) actually recognizes, the zoom magnification is normalized so that the distance from the own vehicle to the recognition target in the image becomes the reference distance. It is desirable to keep Note that the distance to the recognition target used when normalizing the image zoom can be obtained from the radar 42 .

Returning to FIG. 2, the description of the processing units other than the learning unit 21 included in the control unit 2 will proceed. The acquisition unit 22 sequentially acquires images of the vehicle periphery from the imaging device 41 and outputs the images to the determination unit 23 . The determination unit 23 determines the relative speed between the subject of the image input from the acquisition unit 22 and the imaging device 41 .

The determination unit 23 determines the moving speed of the subject in the image as the relative speed between the subject and the imaging device 41 . Thereby, the determination unit 23 can determine relative velocity that directly affects blurring and rolling phenomenon of the subject rather than relative velocity based on the physical positional relationship between the subject and the imaging device 41 .

The determining unit 23 determines, for example, the moving speed of the position of the subject between the images of a plurality of frames continuously captured by the imaging device 41 as the relative speed between the subject and the imaging device 41 . Thereby, the determination unit 23 can accurately determine the relative speed between the subject and the imaging device 41 in the image.

At this time, in the image input from the imaging device 41, the moving speed of the recognition target moving within the image varies depending on the zoom magnification. Therefore, it is desirable that the determination unit 23 normalizes the zoom magnification of the image input from the imaging device 41 so that the distance from the own vehicle to the recognition target in the image becomes the reference distance. The determination unit 23 can acquire from the radar 42 the distance to the recognition target used when normalizing the zoom of the image.

In addition, the position where the recognition target is captured between consecutive frames also changes depending on the frame rate (shutter speed) of the imaging device 41 . Therefore, when the frame rate of the imaging device 41 is variable, it is desirable that the determination unit 23 normalizes the moving speed of the authentication object according to the frame rate set in the imaging device 41 .

The determination unit 23 can also determine the relative speed between the vehicle and the subject input from the radar 42 as the relative speed between the subject and the imaging device 41, for example. Accordingly, the determination unit 23 can determine the relative speed between the subject and the imaging device 41 without performing processing for determining the relative speed.

However, if the determination unit 23 determines the relative speed input from the radar 42 as the relative speed between the subject and the imaging device 41, it may not be possible to accurately determine the moving speed of the subject in the image. For example, when a vehicle makes a right or left turn at an intersection, the imaging direction of the imaging device 41 changes suddenly, so the subject may move in the image at a faster moving speed than the relative speed between the subject and the imaging device 41 .

Therefore, when determining the relative velocity input from the radar 42 as the relative velocity between the subject and the imaging device 41 , the determination unit 23 determines the relative velocity input from the radar 42 based on the change speed of the imaging direction of the imaging device 41 . Compensate for relative velocity.

For example, based on the steering angle of the vehicle input from the steering angle sensor 43, the determination unit 23 estimates the change speed of the traveling direction of the vehicle, that is, the change speed of the imaging direction. Based on this, the relative velocity input from the radar 42 is corrected.

Thereby, the determination unit 23 can determine a moving speed close to the accurate moving speed of the subject in the image as the relative speed between the subject and the imaging device 41 . The determination unit 23 then outputs the determined relative velocity between the subject and the imaging device 41 and the image input from the imaging device 41 to the determination unit 24 .

When recognizing an object in an image input from the determination unit 23, the determination unit 24 selects a speed range including the relative speed input from the determination unit 23 from among the plurality of machine learning models stored in the storage unit 3. A machine learning model for is adopted to determine the subject.

Here, an example of image recognition processing by the determination unit 24 will be described with reference to FIG. 4 as well. FIG. 4 is an explanatory diagram illustrating an example of image recognition processing according to the embodiment. As shown in FIG. 4A, for example, when an image 104 including an image 103 of a vehicle in which blur occurs is input from the determination unit 23, the determination unit 24 divides the image 104 into a plurality of divided regions. .

Then, when the relative speed input from the determination unit 23 is included in the second speed range, the determination unit 24 selects the second speed range machine learning model 1-2 as shown in FIG. adopt. The second speed range machine learning model 1-2 conceptually comprises processing layers having a three-layer structure, an input layer 2a, an intermediate layer 2b, and an output layer 2c, as shown in FIG. 4(b). In addition, although FIG. 4A shows the case where the intermediate layer 2b is one layer, the intermediate layer 2b may be provided in plural layers.

The input layer 2a, the intermediate layer 2b, and the output layer 2c are provided with a plurality of nodes 2z that concurrently determine whether or not the image of each divided area is part of the vehicle. Each node 2z of the input layer 2a is connected to all nodes 2z of the intermediate layer 2b. Each node 2z of the intermediate layer 2b is connected to all nodes 2z of the output layer 2c. Thus, the second speed range machine learning model 1-2 has a neural network structure in which the nodes 2z are connected to each other.

The determination unit 24 inputs the image of each divided area of the image 104 to each node 2z of the input layer 2a. Each node 2z of the input layer 2a determines whether or not the image of each divided region sequentially input is a part of the vehicle. Output to node 2z of layer 2b.

Similarly, each node 2z of the intermediate layer 2b determines whether or not the image of each divided area is a part of the vehicle. to node 2z of .

Each node 2z of the output layer 2c determines whether the image of each divided area is a part of the vehicle or not, and outputs a determination result 2d for each divided area. In addition, FIG. 4B shows the determination result 2d of the divided area in the fourth row from the bottom in the image 104 .

In the example shown in FIG. 4(b), the second speed range machine learning model 1-2, for the fourth row from the bottom of the image 104, divides the area other than the images of the left and right ends of the vehicle (one part of the vehicle). part).

The second speed range machine learning model 1-2 determines whether or not the images of all divided areas in the image 104 are part of the vehicle. Based on the determination result 2d for each divided area, the determining unit 24 derives an image recognition result 102 indicating that the subject is a vehicle, as shown in FIG. 4(c).

At this time, the machine learning model 1-2 for the second speed range is generated by machine learning using only the image of the subject whose relative speed with respect to the imaging device 41 is within the second speed range as teaching materials. It is possible to derive an accurate image recognition result that is a vehicle.

After that, the determination unit 24 determines whether or not the image-recognized vehicle hinders the running of the own vehicle. The determination unit 24 determines whether or not there is an obstacle to running of the own vehicle based on the image-recognized position of the vehicle relative to the own vehicle.

In addition to the vehicle, the determination unit 24 also determines whether or not the subject, such as a pedestrian or a falling object on the road surface, interferes with the running of the vehicle. When determining that the image-recognized object interferes with the running of the vehicle, the determination unit 24 outputs information indicating the position of the image-recognized object to the vehicle control device 44 .

When the information indicating the position of the image-recognized object is input from the determination unit 24, the vehicle control device 44 performs speed control, steering control, and braking control of the vehicle to avoid danger to the own vehicle.

Next, an example of processing executed by the control unit 2 of the image recognition device 1 according to the embodiment will be described with reference to FIG. FIG. 5 is a flowchart showing an example of processing executed by the control unit 2 of the image recognition device 1 according to the embodiment.

The control unit 2 executes the process shown in FIG. 5 each time the imaging device 41 captures an image of the surroundings of the vehicle. Specifically, when an image around the vehicle is captured by the imaging device 41, the control unit 2 first acquires the image from the imaging device 41 (step S101).

Subsequently, the control unit 2 determines the relative speed between the imaging device 41 and the recognition object (step S102). After that, the control unit 2 selects a machine learning model of the speed range including the determined relative speed (step S103).

Then, the control unit 2 inputs the image to the selected machine learning model (step S104), and determines the subject based on the output of the machine learning model (step S105). Subsequently, the control unit 2 determines whether or not notification to the vehicle control device 44 is necessary (step S106).

At this time, based on the position of the subject with respect to the vehicle, the type of the subject, and the like, the control unit 2 determines that the notification is necessary when determining that the subject interferes with the running of the own vehicle, and determines that the subject does not interfere. If so, it is determined that notification is unnecessary.

Then, when the control unit 2 determines that the notification is not necessary (step S106, No), the process ends. If the control unit 2 determines that notification is necessary (step S106, Yes), the control unit 2 notifies the vehicle control device 44 of the position of the subject relative to the vehicle (step S107), and terminates the process.

In the above-described embodiment, an image recognition apparatus and an image recognition method for recognizing an object appearing in an image obtained by imaging the surroundings of a vehicle obtained from an imaging device mounted on the vehicle have been described as an example. is an example.

For example, the functions of the control unit 2 and the storage unit 3 included in the image recognition apparatus 1 may be provided in an information providing apparatus provided in a center capable of wireless communication with a plurality of vehicles. In such a case, the information providing device acquires images for each relative speed from a plurality of vehicles, performs machine learning, and creates and stores machine learning models 1-1 to 1-n for the first to n-th speed ranges.

In addition, identification data for each speed range machine learning model is added to the first to n-th speed range machine learning models 1-1 to 1-n by, for example, adding them to the data header, and image recognition is performed. It will be used for selection processing of machine learning models for each speed range in image recognition processing by a processing device or the like.

When the information providing device receives images for determination captured from each vehicle, the information providing device uses the same method as the image recognition device 1 described above to generate the first to n-th speed range machine learning models 1-1 to 1- A machine learning model selected from n is adopted to determine a recognition target. Then, the information providing device returns the determination result of the recognition target to the vehicle that is the transmission source of the image.

Further, the information providing device may be configured to have the functions of the learning unit 21 and the storage unit 3 included in the image recognition device 1 described above. In such a configuration, the information providing device acquires images for each relative speed from a plurality of vehicles, performs machine learning, and creates and stores machine learning models 1-1 to 1-n for the first to n-th speed ranges. do.

Then, the information providing device creates a machine learning model 1 for the first to n-th speed ranges according to the speed information specified by the image recognition device provided in the vehicle for determining the subject in the image. -1 to 1-n are selected and provided to the image recognition device of the vehicle. According to such an information providing device, the processing load on the vehicle side can be greatly reduced.

In addition, the image recognition device and image recognition method according to the embodiments can be used in an image captured at any place by any image capturing device, such as a camera or a security camera that captures an image of a product conveyed by a belt conveyor. It can be applied to object image recognition.

Further, in the above-described embodiment, the case where the image recognition device 1 includes the learning unit 21 has been described, but the image recognition device according to the embodiment does not necessarily include the learning unit 21 . The image recognition device includes a first speed range machine learning model 1-1, a second speed range machine learning model 1-2, . . . an n-th speed range machine learning model 1- By storing n, the image recognition processing described above can be performed.

Further effects and modifications can be easily derived by those skilled in the art. Therefore, the broader aspects of the invention are not limited to the specific details and representative embodiments so shown and described. Accordingly, various changes may be made without departing from the spirit or scope of the general inventive concept defined by the appended claims and equivalents thereof.

1 image recognition device 2 control unit 21 learning unit 22 acquisition unit 23 determination unit 24 determination unit 3 storage unit 1-1 machine learning model for first speed range 1-2 machine learning model for second speed range 1-3 third speed Machine learning model for range 41 Imaging device 42 Radar 43 Rudder angle sensor 44 Vehicle control device 110 Image collection by relative speed

Claims (11)

an acquisition unit that acquires an image in which a recognition target is captured from an imaging device;
a determination unit that determines the relative speed between the imaging device and the recognition object;
a storage unit that stores, for each relative velocity range, a machine learning model for deriving a recognition result that the subject in the image is the recognition target object from an image in which the recognition target object is captured;
a determining unit that determines the subject in the image by adopting the machine learning model of the relative velocity range according to the relative velocity determined by the determining unit from among the plurality of machine learning models. image recognition device. Each machine learning model is
2. The image recognition device according to claim 1, wherein the machine learning model is specialized for image recognition of a recognition object in a state of relative velocity in each of the relative velocity ranges. The determination unit is
3. The image recognition device according to claim 1, wherein the relative speed is determined based on the moving speed of the subject in the image. The determination unit is
4. The image recognition device according to claim 3 , wherein the relative speed is determined based on a moving speed of the position of the subject between images of a plurality of frames successively captured by the imaging device. The determination unit is
3. The image recognition device according to claim 1, wherein the relative speed is corrected based on a change speed of an imaging direction of the imaging device. The machine learning model for each relative speed range is generated based on a plurality of images in which the recognition target object whose type is known to move at the relative speed within each relative speed range input for each relative speed range is captured. The image recognition device according to any one of claims 1 to 5 , further comprising a learning unit for generating. an acquisition step of acquiring an image in which a recognition target is captured from an imaging device;
a determination step of determining the relative speed between the imaging device and the recognition object;
a storage step of storing, for each relative velocity range, a machine learning model for deriving a recognition result that the subject in the image is the recognition target from an image in which the recognition target is captured;
and determining a subject in the image by employing the machine learning model of the relative velocity range according to the relative velocity determined by the determining step from among the plurality of machine learning models. image recognition method. a storage unit that stores, for each relative velocity range, a machine learning model for deriving a recognition result that the subject in the image is the recognition target from an image of the recognition target captured by an imaging device;
a providing unit that selects from a plurality of the machine learning models the machine learning model of the relative speed range according to speed information specified by an image recognition device that determines a subject in an image, and provides the machine learning model to the image recognition device; A machine learning model providing device comprising: a storage step of storing, for each relative velocity range, a machine learning model for deriving a recognition result that the subject in the image is the recognition target from an image of the recognition target captured by an imaging device;
a providing step of selecting from a plurality of the machine learning models the machine learning model of the relative speed range according to speed information specified by an image recognition device for determining a subject in an image and providing the machine learning model to the image recognition device; A method for providing a machine learning model, comprising: A method for generating a machine learning model used for image recognition processing by an image recognition device,
a model generation step of generating a machine learning model for each relative speed range by learning images of a moving object moving at a speed in each relative speed range in a plurality of relative speed ranges;
an identification data addition step of adding identification data for selectively using the machine learning model for each relative speed range;
A machine learning model generation method, comprising: The image recognition device is a machine learning model device used for image recognition processing,
a plurality of machine learning model units for each relative speed range generated by learning images of a moving object moving at a speed in each relative speed range in a plurality of relative speed ranges;
an identification data section added to each relative velocity range machine learning model section for selectively using the respective relative velocity range machine learning model section;
A machine learning model device comprising:

Images