JP7136849B2 - LEARNING DATA GENERATION METHOD, HUMAN DETECTION MODEL LEARNING METHOD, AND LEARNING DATA GENERATOR - Google Patents

Description

The present disclosure relates to a learning data generation method, a human detection model learning method, and a learning data generation device.

A human detection technology has been proposed for detecting a person positioned around an industrial vehicle for safety. For example, Patent Literature 1 discloses a perimeter monitoring system that captures an image of the periphery of an industrial vehicle, detects a helmet as an image representing a characteristic part of a person, and extracts a person candidate image.

JP 2017-101420 A

By the way, human detection systems for industrial vehicles are difficult to ensure quality at low cost because the number of sales is small and the conditions of use differ depending on the customer's environment. For example, in Patent Document 1, a helmet is used for human detection as an image representing a characteristic part of a person. Such human detection technology may cause over-detection or omission of detection when applied to work places where people are not wearing helmets, work places where there are many obstacles, places where there are many people, etc. there is a risk of In other words, unless the human detection system is optimized for the customer's environment, it may not be possible to ensure sufficient detection accuracy.

As a method of optimizing the human detection system for the customer's environment, it is conceivable to use a human detection model that performs deep learning. However, in order to create a robust human detection model using deep learning, it is necessary to train the human detection model using various human images taken in the real environment as learning data. Furthermore, it takes time and effort to create correct information (correct person's position information) for human images. For example, it takes a lot of effort to actually place people in various places in the real environment, label them as correct information, and verify whether they can be detected.

Therefore, it is required to reduce the labor for generating various learning data including correct information and to shorten the generation time. It is conceivable to increase the learning data in a short time by generating a large number of virtual human images using only CG (computer graphics) technology. It is difficult to ensure detection accuracy.

In view of the circumstances described above, an object of the present disclosure is to shorten the generation time for generating various types of learning data including correct information, and to generate learning data capable of improving human detection accuracy.

A method for generating learning data according to the present disclosure includes:
A method for generating learning data for a human detection model that performs machine learning, comprising:
a step of generating a three-dimensional point cloud environment model based on actual images captured while traveling in a work place;
generating a three-dimensional synthetic model in which one or more human models are arranged in a three-dimensional space represented by the environment model;
generating the learning data based on the three-dimensional synthetic model;
including.

The human detection model learning method according to the present disclosure includes:
a step of generating learning data by the method for generating learning data;
a step of learning a human detection model using the generated learning data;
including.

The learning data generation device according to the present disclosure includes:
A device for generating learning data for a human detection model that performs machine learning,
an environment model generation unit that generates a three-dimensional point cloud environment model based on actual images captured while traveling in a work place;
a data synthesizing unit that generates a three-dimensional synthesized model in which one or more human models are arranged in a three-dimensional space indicated by the environment model;
a data generation unit that generates the learning data based on the three-dimensional synthetic model;
Prepare.

Advantageous Effects of Invention According to the present disclosure, it is possible to shorten the generation time for generating various types of learning data including correct information, and generate learning data capable of improving human detection accuracy.

1 is a block diagram schematically showing a functional configuration of a human detection system including a learning data generation device and a human detection device according to an embodiment; FIG. 1 is a block diagram schematically showing the hardware configuration of a learning data generation device according to an embodiment; FIG. 4 is a flow chart showing the procedure of a method for generating learning data according to an embodiment; FIG. 2 is a schematic diagram showing learning data generated by a learning data generation device according to an embodiment; FIG. 4 is a schematic diagram showing an example of an actual image acquired by a learning data generation device according to an embodiment; FIG. 4 is a schematic diagram showing an example of image data to which attributes are added by the learning data generation device according to the embodiment;

Several embodiments will now be described with reference to the accompanying drawings. However, the dimensions, materials, shapes, relative arrangements, etc. of the components described as embodiments or shown in the drawings are not intended to limit the scope of the invention, but are merely illustrative examples. .
For example, expressions denoting relative or absolute arrangements such as "in a direction", "along a direction", "parallel", "perpendicular", "center", "concentric" or "coaxial" are strictly not only represents such an arrangement, but also represents a state of relative displacement with a tolerance or an angle or distance to the extent that the same function can be obtained.
For example, expressions such as "identical", "equal", and "homogeneous", which express that things are in the same state, not only express the state of being strictly equal, but also have tolerances or differences to the extent that the same function can be obtained. It shall also represent the existing state.
For example, expressions that express shapes such as squares and cylinders do not only represent shapes such as squares and cylinders in a geometrically strict sense, but also include irregularities and chamfers to the extent that the same effect can be obtained. The shape including the part etc. shall also be represented.
On the other hand, the expressions "comprising", "comprising", "having", "including", or "having" one component are not exclusive expressions excluding the presence of other components.

(Overall configuration of human detection system)
The configuration of the industrial vehicle human detection system 400 according to one embodiment will be described below. FIG. 1 schematically shows a functional configuration of a human detection system 400 including a learning data generation device 100 according to an embodiment, a human detection device 200, and an industrial vehicle 300 equipped with the human detection device 200. It is a block diagram showing.

The human detection device 200 includes a human detection model 210 for detecting a person included in a captured image, and an output unit 220 for outputting the human detection result. The human detection device 200 is configured to allow a human detection model 210 that performs deep learning to learn learning data (learning image) including correct information, and to detect a person included in an actual captured image using the learning data. be done. The human detection device 200 is mounted on an industrial vehicle 300 such as a forklift. The input is configured to detect people around the industrial vehicle 300 . When the human detection model 210 detects a person around the industrial vehicle 300 , the output unit 220 outputs the detection result to the notification unit 320 of the industrial vehicle 300 . The notification unit 320 is configured to notify the passenger of the risk of contact between the industrial vehicle 300 and a person by means of screen display, warning light, warning sound, or the like. The learning data generation device 100 is a device for generating learning data for the human detection model 210 of the human detection device 200 .

FIG. 2 is a block diagram schematically showing the hardware configuration of the learning data generation device 100 according to one embodiment. The learning data generation device 100 includes a processor 72 such as a CPU (Central Processing Unit) or a GPU (Graphics Processing Unit), a RAM (Random Access Memory) 74, a ROM (Read Only Memory) 76, and a HDD (Hard Disk Drive). 78 , an input I/F 80 and an output I/F 82 , which are connected to each other via a bus 84 using a computer. The human detection device 200 may also have a configuration similar to this.

The learning data generation device 100 and the human detection device 200 implement various functions described later by the processor 72 executing programs stored in memories such as the ROM 76 and the RAM 74 . The learning data generation device 100 and the human detection device 200 may be integrated or separated. If it is separate, the learning data generation device 100 may be an information processing device used in a learning environment outside the industrial vehicle 300 . In this case, as shown in FIG. 1 , only the trained human detection device 200 is mounted on the industrial vehicle 300 .

As shown in FIG. 1, the learning data generation device 100 is configured to acquire actual images captured while an industrial vehicle 300 equipped with a human detection device 200 is traveling in a work place where the human detection device 200 is actually operated. an actual image acquisition unit 110, an environment model generation unit 120 that generates an environment model consisting of a three-dimensional point cloud based on information such as the acquired actual image and the acquisition position of the actual image; A data synthesizing unit 130 that generates a three-dimensional synthetic model in which one or more virtual human models are arranged in the original space, and an image of the three-dimensional synthetic model observed from a predetermined viewpoint position as learning data (learning data). and a data generation unit 140 that generates a data image).

The real image acquisition unit 110 is configured to acquire an actual image of the work place from, for example, the imaging unit 310 provided in the industrial vehicle 300 or a test vehicle for data measurement. Acquisition of the actual image may be performed, for example, by running a test vehicle before introduction of the industrial vehicle 300, or may be acquired while performing actual cargo handling work after the introduction of the industrial vehicle 300. The environment model generation unit 120 extracts feature points of the work place by, for example, V-SLAM (Visual Simultaneous Localization and Mapping) technology or SfM (Structure from Motion) technology, and obtains the actual image acquired by the actual image acquisition unit 110 and its acquisition position. Based on such information, it is configured to generate an environment model consisting of a three-dimensional point cloud. The three-dimensional point group that constitutes the environment model may include, for example, in addition to the three-dimensional coordinates of the extracted feature points, additional information such as hue, saturation, and brightness related to the three-dimensional coordinates. .

The data synthesizing unit 130 is configured to arrange (synthesize) pre-stored virtual human models in the environment model to generate a three-dimensional synthetic model. The human model is generated by CG technology, and may be defined as a three-dimensional point group of characteristic points of a person in the same way as the environment model, or defined as a surface of an object representing the existence range of the person. It may be a thing, specifically a three-dimensional object created by 3D modeling software. It is assumed that attribute information as a person, which will be described later, is given in advance to the human model arranged by the data synthesizing unit 130 . As human models, it is preferable to prepare patterns of people with various clothes, postures, heights, and genders, assuming a work place where the industrial vehicle 300 is actually operated. The human model is appropriately enlarged, reduced, rotated, and color-adjusted according to the place of placement, and then combined with the environment model. Based on the travel history of the industrial vehicle 300 when the actual image is acquired, the data synthesizing unit 130 prioritizes routes in which the industrial vehicle 300 travels frequently over routes in which the industrial vehicle 300 travels less frequently as a region for arranging the human model. may be configured to select to The data synthesizing unit 130 may be configured to preferentially select an area in which a large number of point groups or surfaces indicating a person's attribute are extracted in an attribute assignment process, which will be described later, and arrange the human model.

Based on the 3D synthetic model generated by the data synthesizing unit 130, the data generation unit 140 generates an image obtained by observing the 3D synthetic model from a predetermined viewpoint position as learning data (learning image). The image assumed here is an image in which the human model placed by the data synthesizing unit 130 is reflected around the industrial vehicle 300, and the learning data (learning It is used for deep learning of the human detection model 210 as a human image). When generating a learning image from a three-dimensional synthetic model, the position in the work place represented by the environment model, the installation height, installation angle, angle of view, focal length, etc. of the imaging unit 310 provided in the industrial vehicle 300 viewpoint position information is considered.

The learning data generating device 100 includes an attribute assigning unit 150 configured to assign attributes to at least a part of the three-dimensional point cloud that constitutes the environment model, and an object surface or object in attribute assignment. A processing unit 160 configured to perform processing for rearranging, deleting, or duplicating a point cloud forming a surface on an object-by-object basis; A correct answer information setting unit 170 configured to perform labeling as correct answer information when learning the model 210 is further provided.

The attribute assigner 150 may be configured to assign attributes by a semantic segmentation technique. The attribute assignment unit 150 obtains attributes of each 3D point group in the environment model. Attributes are, for example, people, forklifts, trucks, shelves, roads, buildings, and the like. The attribute assigning unit 150 may connect points with the same attribute to form a mesh surface. For example, the attribute assigning unit 150 may be configured to group and identify a surface of an object when groups of points assigned with the same attribute are continuously arranged. That is, the attribute assigning section 150 may identify the surface of the object based on the assigned attribute. An object to which the attribute assigning unit 150 assigns an attribute may be a person who is a detection target of the human detection device 200, or may be an article, a structure, an animal, or the like which is not a detection target.

When acquiring the actual image for generating the environment model, there are cases where people are reflected in the image. When the training data is generated, a situation arises in which a person reflected in the background is not labeled as correct information, while the arranged human model is labeled as correct information. It may affect detection accuracy. Therefore, a person who appears in the real image is assigned an attribute and treated as a person. As a result, the point cloud representing the person or the surface of the object captured in the real image is processed so as not to be treated as the background of the environment model (three-dimensional synthetic model). Since it is already known that the human model arranged by the data synthesizing unit 130 is a person, the point cloud or the surface of the object that constitutes the human model is automatically Attributes as a person are given.

The processing unit 160 may be configured to receive an editing operation for an object to which attributes have been assigned. This makes it possible to identify object surfaces or point groups thereof included in the environment model and the three-dimensional composite model for each object, and to edit, such as rearrangement change, deletion, or duplication, for each object. Therefore, it is possible to edit the background and correct answer information in generating the learning data. For example, an object (including a person) placed in an area where no object usually exists can be deleted, or an object (including a person) can be moved or added to a location where an object often exists. For example, the processing unit 160 can remove attributes of a person who appears in the real image on the environment model, or move the person as a new person model to an arbitrary location in the environment model. is also possible. As a result, there is no need to exclude people during travel for constructing a three-dimensional environment model.

The correct information setting unit 170 is configured to label, as correct information, a point cloud or a surface of an object to which an attribute indicating a person is assigned in the three-dimensional synthetic model. This labeling information is also reflected in the learning images generated by the data generator 140 based on the three-dimensional synthetic model. Note that the correct information setting unit 170 may be configured to perform labeling of correct information after generating a learning image based on the three-dimensional synthetic model.

(Method of generating learning data)
A method of generating learning data according to an embodiment will be described below. FIG. 3 is a flow chart showing procedures of a method for generating learning data according to an embodiment.

As shown in FIG. 3, the learning data generation device 100 (actual image acquisition unit 110) acquires an actual image captured while traveling in a work place (step S1). In order to generate a dense 3D point cloud environment model, the real images are preferably acquired in a number of laps around the work site. For example, by superimposing three-dimensional point clouds obtained by multiple runs, it is possible to generate a dense three-dimensional point cloud environment model. Also, in order to generate an environment model of a dense three-dimensional point group, a real image for environment model generation may be acquired by an imaging device having a higher resolution than the image used for human detection. The actual image may be captured in a normal business environment at the work site.

The learning data generation device 100 (environment model generation unit 120) generates a three-dimensional point cloud environment model based on the acquired real image (step S2). The learning data generation device 100 (attribute assigning unit 150) assigns attributes to at least a part of the three-dimensional point group of the environment model (step S3). The learning data generation device 100 (attribute assigning unit 150) identifies the surface of the object based on this attribute.

Here, the learning data generation device 100 (attribute assigning unit 150) determines whether or not there is an object with an attribute indicating a person (step S4). When it is determined that there is an object with an attribute indicating a person (step S4; Yes), the learning data generation device 100 (processing unit 160) performs processing on the object or the point group of the object (arrangement change, deletion, or duplication) is determined (step S5). Specifically, the learning data generation device 100 (processing unit 160) determines whether an object indicating a person's attribute or a point group of the object should be excluded from the environment model or moved to an arbitrary location. . If it is determined that there is no object with the attribute indicating a person (step S4; No), the learning data generation device 100 skips step S5.

The learning data generation device 100 (data synthesizing unit 130) arranges one or more human models in the three-dimensional space represented by the environment model to generate a three-dimensional synthetic model (step S6). Here, an attribute indicating a person is given in advance to the point group or the surface of the object that constitutes the human model to be arranged. The learning data generation device 100 (correct answer information setting unit 170) labels an object indicating a person's attribute or a point group of the object as correct answer information (step S7). The learning data generation device 100 (data generation unit 140) generates learning data labeled with correct information based on the three-dimensional synthetic model in which the human model is arranged (step S8). Based on the travel history of the industrial vehicle 300 at the time of acquiring the actual image, the human model may be placed near a route with high travel frequency (for example, a position within 5 m from the travel route). Human models may be preferentially placed in places where there is a lot of foot traffic (for example, entrances and exits of buildings and rooms). Such a place with a lot of foot traffic may be identified based on the number or frequency of detection of objects indicating human attributes in step S4.

(Concrete example of environmental model)
A specific example of the environment model will be described below.

FIG. 4 is a schematic diagram showing an example of an environment model generated by the environment model generation unit 120 of the learning data generation device 100 according to one embodiment. FIG. 5 is a schematic diagram showing an example of an environment model to which attributes are assigned by the attribute assigning unit 150 of the learning data generating apparatus 100 according to one embodiment.

The example shown in FIG. 4 is an environment model generated based on an actual image taken in a warehouse. The environment model includes people 40, cargo 50, walls 60 and floor 70. If a new human model is placed in the environment model in this state, the placed human model is assigned a human attribute and labeled as correct information. Therefore, it is necessary to assign an attribute to the person 40 to exclude it from the environment model, or move it to an arbitrary location. Note that if no special operation is required for the placement of the person 40 to whom attributes have been assigned, the point cloud or the surface of the object that constitutes the person 40 may be labeled as it is as the correct information.

FIG. 5 schematically shows a state after the attribute assigning unit 150 assigns attributes to the environment model shown in FIG. A person 40, a load 50, a wall 60, and a floor 70, which are objects included in the environment model, are given different attributes. In this figure, they are distinguished by hatching in order to visualize the assigned attributes. However, such image processing is not essential. The attribute assigning unit 150 may assign an attribute as meta information, or may add a code or identifier to each region of the three-dimensional point group.

As shown in FIG. 5, attributes are assigned to the surfaces of the object so that they can be identified. By assigning attributes, the processing unit 160 can edit, such as changing the layout, deleting, and adding, on an object-by-object basis. As for the area of the person 40, if the arrangement does not cause a sense of incongruity when considering the state of the industrial vehicle 300 when it is actually running, it may be labeled as correct information as it is, or the arrangement as it is may be different from that during actual operation. If there is a discrepancy, it may be removed from the environment model, or it may be moved to a different location after processing such as enlargement, reduction, or rotation.

(Specific example of learning data)
A specific example of the learning data will be described below.

FIG. 6 is a schematic diagram showing learning data (learning image) generated by the data generating unit 140 of the learning data generating apparatus 100 according to one embodiment. The learning image shown in FIG. 6 is generated by the data generation unit 140 based on a three-dimensional synthetic model obtained by synthesizing the environment model generated by the environment model generation unit 120 with the human model by the data synthesis unit 130. It is. In FIG. 6, the position of the imaging unit 310 installed above the industrial vehicle 300 is assumed as the viewpoint position when generating the learning image from the three-dimensional synthetic model. A vehicle 10 parked on the road and a person 20 existing around the industrial vehicle are included. This image of the person 20 is obtained as a result of observing the person model placed in the environment model by the data synthesizing unit 130 from a predetermined viewpoint position. Since the learning data generation device 100 assigns attributes to the human model in the three-dimensional synthetic model and labels it as correct information, it is possible to easily identify a region including the person 20 in the generated learning image as correct information. , and input to the human detection model 210 as learning data.

The present disclosure is not limited to the above-described embodiments, and includes modifications of the above-described embodiments and modes in which a plurality of embodiments are appropriately combined. For example, the method of generating learning data is not limited to the example shown in FIG. The order of each step may be changed, and some steps may be omitted. The learning data generated by the learning data generation method described above is provided to the human detection model 210, and the human detection model 210 executes machine learning (deep learning) based on the learning data. For deep learning means, for example, a method such as SSD (Single Shot Multibox Detector) can be adopted, but the method is not limited to this, and any object detection algorithm may be used. The trained human detection model uses, for example, a captured image acquired at a work place as input information, the position of an object included in the captured image (specified by a rectangle), the class (person, vehicle, etc.), and the probability of belonging to that class. and the like may be used as output information.

(summary)
The contents described in each of the above embodiments are understood as follows, for example.

(1) The method of generating learning data according to the present disclosure includes:
A method for generating learning data for a human detection model (210) that performs machine learning,
a step of generating a three-dimensional point cloud environment model based on actual images captured while traveling in a work place;
generating a three-dimensional composite model in which one or more human models (20) are arranged in a three-dimensional space indicated by the environment model;
generating the learning data based on the three-dimensional synthetic model;
including.

According to the above method, since the learning data is generated with the environment model based on the actual image as the background, it is possible to reduce the divergence from the background in the work place, that is, the customer's environment. As a result, it is possible to generate learning data capable of improving human detection accuracy. In addition, when arranging the human model (20) in the image space of the environment model, it is possible to generate various learning data including correct information by devising the position and number of the arrangement. As a result, the time required to generate training data can be significantly reduced compared to the case where correct information is generated by collecting actual images of each person's placement pattern, and deep learning can be performed using a variety of training data including correct information. becomes.

(2) In some embodiments, in the method described in (1) above,
Attributing attributes to at least a portion of the three-dimensional point cloud of the environment model and identifying surfaces of objects based on the attributes.

Since the 3D point cloud is composed of discrete points, if the human model (20) is incorporated therein to generate a composite image (3D composite model), there is a risk that part of the background will be missing or transparent. . Learning data generated based on such a three-dimensional synthetic model may lead to deterioration of the human detection function. In this respect, according to the above method, since the surface of the object is identified, the possibility that part of the background is missing or transparent can be reduced.

(3) In some embodiments, in the method of (2) above,
determining whether there is a surface (40) of said point cloud or said object with said attribute indicating a person in said environment model.

When acquiring the actual image for generating the environment model, there are cases where people are reflected in the image. When the training data is generated, a situation arises in which a person reflected in the background is not labeled as correct information, while the arranged human model is labeled as correct information. It may affect detection accuracy. In this regard, according to the above method, the point cloud showing the person reflected in the real image or the surface of the object is not treated as the background of the environment model (three-dimensional synthetic model), and the real image is reproduced in the environment where people are present. Since it may be acquired, an environment model can be generated while performing normal work.

(4) In some embodiments, in the method described in (3) above,
It includes a step of performing a process of rearranging, deleting, or duplicating the surface of the object or the point group forming the surface of the object for each object.

According to the above method, it is possible to identify the surface of an object included in the environment model or its point group in units of objects, and edit such as changing the layout, deleting, duplicating, etc. in units of objects. Therefore, it is possible to delete an object (including a person) placed in an area where no object normally exists, or to move or add an object (including a person) to a location where an object often exists.

(5) In some embodiments, in the method of (3) or (4) above,
In the step of generating the three-dimensional synthetic model, the human model is arranged by preferentially selecting the point group to which the attribute representing the person or the region in which many surfaces of the object are detected.

According to the above method, the human detection model is learned by preferentially arranging the human model in high-risk areas (e.g., roads where people often pass, near building entrances and exits, places where people work, etc.). can efficiently generate learning data. Therefore, generation of useless learning data can be suppressed.

(6) In some embodiments, in the method according to any one of (1) to (5) above,
In the step of generating the three-dimensional synthetic model, based on the travel history when the actual image is acquired, the human model (20) is placed on a route with a high travel frequency rather than a route with a low travel frequency. choose preferentially.

According to the above method, real images are acquired while the vehicle is traveling in normal business, and the human model is arranged on a route that the vehicle frequently travels during the traveling. Therefore, generation of useless learning data can be suppressed.

(7) A method for learning a human detection model (210) according to the present disclosure includes:
a step of generating learning data by the method of generating learning data according to any one of (1) to (6) above;
a step of learning a human detection model (210) using the generated training data;
including.

According to the above method, it is possible to shorten the generation time for generating various learning data including correct answer information. In addition, since the human detection model (210) learns various learning data including correct information, the human detection accuracy can be improved.

(8) The learning data generation device (100) according to the present disclosure is
A learning data generation device (100) for a human detection model (210) that performs machine learning,
an environment model generation unit (120) that generates a three-dimensional point cloud environment model based on actual images captured while traveling in a work place;
a data synthesizing unit (130) for generating a three-dimensional synthesized model in which one or more human models (20) are arranged in a three-dimensional space indicated by the environment model;
a data generation unit (140) that generates the learning data based on the three-dimensional model;
Prepare.

According to the above configuration, since the learning data is generated with the environment model based on the actual image as the background, it is possible to reduce the divergence from the background in the work place, that is, the customer's environment. As a result, it is possible to generate learning data capable of improving human detection accuracy. In addition, when arranging the human model (20) in the image space of the environment model, it is possible to generate various learning data including correct information by devising the position and number of the arrangement. As a result, the time required to generate training data can be significantly reduced compared to the case where correct information is generated by collecting actual images of each person's placement pattern, and deep learning can be performed using a variety of training data including correct information. becomes.

10 vehicles 20 people (human model)
40 people 50 cargo 60 walls 70 floor 72 processor 74 RAM
76 ROMs
78 HDDs
80 Input I/F
82 Output I/F
84 bus 100 learning data generation device 110 real image acquisition unit 120 environment model generation unit 130 data synthesis unit 140 data generation unit 150 attribute assignment unit 160 processing unit 170 correct information setting unit 200 human detection device 210 human detection model 220 output unit 300 industrial vehicle 310 imaging unit 320 notification unit 400 human detection system

Claims (8)

A method for generating learning data for a human detection model that performs machine learning, comprising:
a step of generating a three-dimensional point cloud environment model based on an actual image captured by an imaging unit installed in the industrial vehicle while the industrial vehicle is running at a work site;
generating a three-dimensional synthetic model in which one or more human models are arranged in a three-dimensional space represented by the environment model;
A learning image obtained when the three-dimensional synthetic model in which the human model is arranged is observed from a viewpoint position assuming the installation position of the imaging unit in the industrial vehicle, using the viewpoint position information of the imaging unit. a step of generating as the learning data;
How to generate training data including 2. The method of generating learning data according to claim 1, further comprising the step of assigning attributes to at least a portion of the three-dimensional point cloud of the environment model, and identifying surfaces of objects based on the attributes. 3. The method of generating learning data according to claim 2, further comprising the step of determining whether or not there is a surface of said point group or said object to which said attribute indicating a person is assigned in said environment model. 4. The method of generating learning data according to claim 3, further comprising a step of changing, deleting, or duplicating the surface of the object or the group of points forming the surface of the object for each object. 4. In the step of generating the three-dimensional synthetic model, the human model is arranged by preferentially selecting an area in which many faces of the point group or the object to which the attribute indicating the person is assigned are detected. Or the method of generating learning data according to 4. In the step of generating the three-dimensional synthetic model, a route with a high travel frequency is prioritized over a route with a low travel frequency as a region for arranging the human model based on the travel history when the actual image is acquired. 6. The method of generating learning data according to any one of claims 1 to 5. generating learning data by the method for generating learning data according to any one of claims 1 to 6;
a step of learning a human detection model using the generated learning data;
Training methods for human detection models, including A device for generating learning data for a human detection model that performs machine learning,
an environment model generation unit that generates a three-dimensional point cloud environment model based on an actual image captured by an imaging unit installed in the industrial vehicle while the industrial vehicle is running at a work site;
a data synthesizing unit that generates a three-dimensional synthesized model in which one or more human models are arranged in a three-dimensional space indicated by the environment model;
A learning image obtained when the three-dimensional synthetic model in which the human model is arranged is observed from a viewpoint position assuming the installation position of the imaging unit in the industrial vehicle, using the viewpoint position information of the imaging unit. a data generation unit that generates the learning data;
A learning data generation device comprising:

Images