Abnormal posture detection apparatus

Abstract

To provide an abnormal posture detection apparatus configured to prevent adverse effect on detection processing without requiring manual labeling work or image data photography or limiting detectable abnormal postures.SOLUTION: An abnormal posture detection apparatus includes: a skeletal position acquisition unit which acquires three-dimensional skeletal position of a person; a vector acquisition unit which acquires a perpendicular vector, a right-leg vector and a left-leg vector based on the three-dimensional skeletal position; a calculation unit which calculates a right-leg angle formed by the perpendicular vector and the right-leg vector, and a left-leg angle formed by the perpendicular vector and the left-leg vector; a pre-processing unit which performs pre-processing to adjust the three-dimensional skeletal position to be aligned to a predefined direction and size; and a determination unit which determines whether the person is in an abnormal posture, using a learning model generated by machine learning using the pre-processed three-dimensional skeletal position, the right-leg angle and the left-leg angle.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present disclosure relates to an abnormal posture detection device. [Background technology]

[0002] Conventionally, there is known a technique for detecting the posture of a person based on skeletal data of the person.

[0003] For example, Patent Document 1 discloses a technology in which a label indicating an abnormal posture is assigned to skeletal data showing the three-dimensional skeletal position obtained from image data of a person, and a learning model is trained using the labeled skeletal data, and the learning model is used to detect abnormal postures of a person.

[0004] Furthermore, for example, Non-Patent Document 1 discloses a technology for detecting abnormal postures of a person by using a three-dimensional skeleton position estimated by a learning model to train another learning model by machine learning. [Prior art documents] [Patent documents]

[0005] [Patent Document 1] Japanese Patent Publication No. 2022-55077 [Non-patent literature]

[0006] [Non-Patent Document 1] Takamune Harashina and Yuji Yamauchi, "Abnormal Motion Detection Based on 3D Posture Estimation Information," IEICE General Conference, 2021 Summary of the Invention [Problem to be solved by the invention]

[0007] However, the technology of Patent Document 1 has the following three problems. First, the task of manually labeling skeletal data is time-consuming. Second, when image data from which the 3D skeletal position is obtained is prepared by photography, a person must actually assume an abnormal pose in front of the camera, which is dangerous. Third, because only labeled abnormal poses are detected, the detectable abnormal poses are limited.

[0008] On the other hand, in the technology of Non-Patent Document 1, for example, when preparing image data for learning by photographing a person, even if the person has the same posture, the three-dimensional skeleton position will differ depending on the positional relationship between the camera and the person (in other words, the orientation of the person relative to the camera). As a result, there is a concern that this may have an adverse effect on the detection process (for example, different detection results for the same posture, etc.).

[0009] An object of one embodiment of the present disclosure is to provide an abnormal posture detection device that can solve the problems in the above-mentioned Patent Document 1 and Non-Patent Document 1, and more specifically, to provide an abnormal posture detection device that can prevent adverse effects on the detection process without requiring manual labeling or photographing of abnormal postures, without limiting the abnormal postures that can be detected. [Means for solving the problem]

[0010] an abnormal posture detection device according to one embodiment of the present disclosure, comprising: a skeletal position acquisition unit that acquires a three-dimensional skeletal position of a person; a vector acquisition unit that acquires, based on the three-dimensional skeletal position, a vertical vector that is perpendicular to a floor surface on which the person is located, a right foot vector that indicates the orientation of the person's right foot, and a left foot vector that indicates the orientation of the person's left foot; a calculation unit that calculates the right foot angle formed by the vertical vector and the right foot vector and the left foot angle formed by the vertical vector and the left foot vector; a preprocessing unit that performs a first preprocessing to align the three-dimensional skeletal position to a specified orientation and a second preprocessing to align the three-dimensional skeletal position to a specified size; and a determination unit that uses a learning model to determine whether the person has an abnormal posture and controls a predetermined notification device to issue a notification to that effect if the abnormal posture is detected, wherein the learning model is created by machine learning learning data that associates the right foot angle and the left foot angle calculated based on the three-dimensional skeletal position of a person with a normal posture with the three-dimensional skeletal position of the person with the normal posture after the first preprocessing. [Effects of the Invention]

[0011] According to the present disclosure, there is no need to manually label or photograph abnormal postures, and the detectable abnormal postures are not limited, making it possible to prevent adverse effects on the detection process. [Brief explanation of the drawings]

[0012] [Figure 1] FIG. 1 is a block diagram illustrating a configuration of an abnormal posture detection system according to an embodiment of the present disclosure. [Figure 2] FIG. 1 is a diagram illustrating an example of a three-dimensional skeletal position according to an embodiment of the present disclosure. [Figure 3] 1 is a diagram illustrating an example of a vertical vector, a right foot vector, and a left foot vector according to an embodiment of the present disclosure. [Figure 4] FIG. 1 illustrates an example of a three-dimensional skeleton position after preprocessing according to an embodiment of the present disclosure. [Figure 5] 1 is a flowchart illustrating a learning process of an abnormal posture detection device according to an embodiment of the present disclosure. [Figure 6]1 is a flowchart illustrating a detection process of an abnormal posture detection device according to an embodiment of the present disclosure. DETAILED DESCRIPTION OF THE INVENTION

[0013] Hereinafter, embodiments of the present disclosure will be described with reference to the drawings.

[0014] First, the configuration of an abnormal posture detection system 1 according to this embodiment will be described with reference to Fig. 1. Fig. 1 is a block diagram showing the configuration of the abnormal posture detection system 1.

[0015] 1, the abnormal posture detection system 1 includes an abnormal posture detection device 10, a camera 20, and a notification device 30. The abnormal posture detection device 10 is connected to each of the camera 20 and the notification device 30 so as to be able to communicate with each other.

[0016] The camera 20 captures an image of the entire body of a person. The camera 20 outputs image data obtained by capturing the image to the abnormal posture detection device 10.

[0017] Examples of locations where camera 20 may be installed include workplaces such as factories, and people photographed by camera 20 (in other words, people to be detected) include, but are not limited to, people in the workplace (e.g., workers).

[0018] The abnormal posture detection device 10 creates a learning model by machine learning normal postures of a person based on image data input from the camera 20. Details of this will be described later.

[0019] Furthermore, the abnormal posture detection device 10 detects an abnormal posture of a person based on the image data input from the camera 20 and the learning model. If an abnormal posture is detected, the abnormal posture detection device 10 controls the notification device 30 to notify the fact. Details of this will be described later.

[0020] In this embodiment, the abnormal posture refers to, but is not limited to, a posture during an action that may lead to an accident or injury (for example, falling, walking while looking at a smartphone, walking with hands in pockets, inappropriate work actions, etc.). On the other hand, the normal posture refers to, but is not limited to, a posture during an action other than the above actions (for example, standing still upright, walking without looking away, appropriate work actions, etc.).

[0021] The notification device 30, under the control of the abnormal posture detection device 10, issues a notification that a person has an abnormal posture. The notification device 30 may be installed at the location where the camera 20 is installed, or at a location separate from the camera 20. Examples of the notification device 30 include a display, a lamp, and a speaker. Examples of the notification mode by the notification device 30 include displaying an image on a display, turning on a lamp, and outputting sound from a speaker. These may be used in combination.

[0022] The configuration of the abnormal posture detection system 1 has been described above.

[0023] Next, the configuration of the abnormal posture detection device 10 according to this embodiment will be described with reference to FIG.

[0024] Although not shown in the drawings, the abnormal posture detection device 10 has, as hardware, for example, a CPU (Central Processing Unit), a ROM (Read Only Memory) that stores computer programs, a RAM (Random Access Memory) that is a working memory, etc. Each unit described below is realized by the CPU reading out a computer program from the ROM and executing it in the RAM.

[0025] As shown in FIG. 1, the abnormal posture detection device 10 includes a skeleton position acquisition unit 110, a vector acquisition unit 120, a calculation unit 130, a preprocessing unit 140, a creation unit 150, and a determination unit 160.

[0026] The skeleton position acquisition unit 110 acquires the three-dimensional skeleton position of a person based on image data from the camera 20.

[0027] This acquisition method can be performed using known techniques. For example, a method of capturing an RGB image of a person and inputting it into a machine learning model capable of estimating a three-dimensional skeletal position, a method of preparing a pair of a depth image and an RGB image of a person, inputting the RGB image into a machine learning model capable of estimating a two-dimensional skeletal position, acquiring the two-dimensional skeletal position, and then associating the two-dimensional skeletal position with the depth image, or a method of preparing a depth image of a person and inputting it into a machine learning model capable of estimating a two-dimensional skeletal position or a three-dimensional skeletal position on the depth image may be used.

[0028] Here, the three-dimensional skeleton position acquired by the skeleton position acquisition unit 110 will be described with reference to Fig. 2. Fig. 2 is a diagram showing an example of the three-dimensional skeleton position.

[0029] In FIG. 2, the area above the pelvic center, the right groin, and the left groin corresponds to the person's upper body, and the area below the pelvic center, the right groin, and the left groin corresponds to the person's lower body. In the following description, the upper three-dimensional skeletal position of the lower body is referred to as the "upper skeletal position," and the lower three-dimensional skeletal position of the lower body is referred to as the "lower skeletal position." Examples of upper skeletal positions include the right groin and the left groin, or the pelvic center. Examples of lower skeletal positions include the right foot and the left foot. The right foot and the left foot may be any part below the ankle, and may be, for example, the ankle, the toe, or the heel.

[0030] An example of a three-dimensional skeletal position has been described above. Now, we will return to the description of FIG.

[0031] Based on the three-dimensional skeleton position acquired by the skeleton position acquisition unit 110, the vector acquisition unit 120 acquires a vertical vector perpendicular to the floor surface on which the person is standing, a right foot vector indicating the direction of the person's right foot, and a left foot vector indicating the direction of the person's left foot.

[0032] Here, the vertical vector, right foot vector, and left foot vector acquired by the vector acquisition unit 120 will be described with reference to Fig. 3. Fig. 3 is a diagram showing an example of the vertical vector, right foot vector, and left foot vector in a virtual three-dimensional space coordinate system. The three-dimensional skeletal position shown in Fig. 3 indicates the three-dimensional skeletal position when a person is in a normal posture (for example, when standing upright and stationary).

[0033] The solid arrow a is an example of a vertical vector. As shown in Fig. 3, the vertical vector is a vector that points in the vertical direction from the floor surface (xy plane) on which the person is standing.

[0034] The dotted arrow b is an example of a right foot vector. As shown in Figure 3, the right foot vector is a vector that points from the right hip, which is the upper skeletal position, to the right foot, which is the lower skeletal position.

[0035] The dotted arrow c is an example of a left foot vector. As shown in Figure 3, the left foot vector is a vector that points from the base of the left foot, which is the upper skeletal position, to the left foot, which is the lower skeletal position.

[0036] When the upper skeleton position is the center of the pelvis, the right foot vector is a vector directed from the center of the pelvis to the right foot, and the left foot vector is a vector directed from the center of the pelvis to the left foot.

[0037] Above, examples of various vectors have been explained. Now, we will return to the explanation of FIG.

[0038] The calculation unit 130 calculates the angle between the vertical vector and the right foot vector (hereinafter referred to as the right foot angle), and the angle between the vertical vector and the left foot vector (hereinafter referred to as the left foot angle).

[0039] The preprocessing unit 140 performs preprocessing (parallel / rotational movement processing, corresponding to first preprocessing) to align the three-dimensional skeleton positions acquired by the skeleton position acquisition unit 110 to a specified orientation by translation or rotation. Furthermore, the preprocessing unit 140 performs scaling processing (enlargement / reduction processing, corresponding to second preprocessing) to align the three-dimensional skeleton positions acquired by the skeleton position acquisition unit 110 to a specified size.

[0040] An example of preprocessing of aligning the three-dimensional skeleton position in a specified direction by the preprocessing unit 140 will now be described with reference to Fig. 4. Fig. 4 is a diagram showing an example of the three-dimensional skeleton position after preprocessing in a virtual three-dimensional space coordinate system.

[0041] First, if the pelvic center is not at the origin (x=0, y=0, z=0), the preprocessing unit 140 translates the pelvic center to position it at the origin as shown in FIG.

[0042] Next, if the base of the neck is not on the positive z-axis, the pre-processing unit 140 rotates the base of the neck to position it on the positive z-axis as shown in FIG.

[0043] Next, if the left shoulder is not on the xz plane, the preprocessing unit 140 rotates the left shoulder to position it on the xz plane as shown in FIG.

[0044] In this way, the pre-processing unit 140 performs a process of aligning the three-dimensional skeletal positions (specifically, skeletal positions of the torso, for example, the base of the neck, left shoulder, right shoulder, chest, spine, pelvic center, base of the right foot, and base of the left foot) with a specified position (for example, an axis, plane, or point) in the virtual three-dimensional space coordinate system by translating or rotating at least three of them. As a result, the three-dimensional skeletal positions acquired by the skeletal position acquisition unit 110 are always aligned with a specified orientation. Furthermore, the pre-processing unit 140 performs a scaling process (enlargement / reduction process, equivalent to the second pre-processing) on ​​the three-dimensional skeletal positions to align them with a specified size in the virtual three-dimensional space coordinate system.

[0045] The reason for performing the above-mentioned preprocessing will now be explained. If a person with the same normal posture is photographed from different angles and multiple 3D skeletal positions are obtained based on the image data obtained thereby, the 3D skeletal positions will be different from one another. For example, a 3D skeletal position obtained based on image data of a person with a normal posture facing forward relative to the camera will be different from a 3D skeletal position obtained based on image data of the same person with a normal posture facing backward relative to the camera. If a learning model is trained on the 3D skeletal positions of the normal posture of a person facing forward relative to the camera and then a detection process is performed, the person facing forward relative to the camera will be determined to have a normal posture, while the person facing backward relative to the camera will be determined to have an abnormal posture, even though they are in the same normal posture.

[0046] Furthermore, if a person with the same normal posture is photographed at different camera distances and multiple 3D skeletal positions are obtained based on the image data obtained, the 3D skeletal positions will be different from each other. For example, a 3D skeletal position obtained based on image data of a person photographed 1 m from the camera will be different from a 3D skeletal position obtained based on image data of the same person with a normal posture photographed 3 m away from the camera. If a learning model is trained on the 3D skeletal position of a person with a normal posture at a distance of 1 m from the camera and then a detection process is performed, the person facing forward toward the camera will be determined to have a normal posture, while the person 3 m away from the camera will be determined to have an abnormal posture, even though they are both in the same normal posture. As such, without the above preprocessing, the detection results will be adversely affected.

[0047] In contrast, in this embodiment, the orientation of the 3D skeleton positions is aligned by the preprocessing, so that all the 3D skeleton positions can be learned as the same, thereby preventing the above-mentioned adverse effects on the detection results.

[0048] An example of pre-processing has been described above. Now, we will return to the explanation of FIG.

[0049] The creation unit 150 creates a learning model that has learned the three-dimensional skeleton position of a normal posture, and that is used in the abnormal posture detection process.

[0050] This learning model is created by machine learning learning data that associates the right foot angle and left foot angle calculated based on the three-dimensional skeletal position of a person with a normal posture with the pre-processed three-dimensional skeletal position of the person with a normal posture. In other words, the creation unit 150 causes the learning model to learn the right foot angle and left foot angle calculated based on the three-dimensional skeletal position of a person with a normal posture and the pre-processed three-dimensional skeletal position of the person with a normal posture as learning data. Therefore, the creation unit 150 may also be called the "learning unit 150."

[0051] Here, the reason for learning the right foot angle and the left foot angle will be explained. When the above-mentioned preprocessing is performed on the three-dimensional skeleton position, different postures may be treated as the same posture, making it difficult to distinguish between them. For example, after the above-mentioned preprocessing, a posture in which the user is lying on the floor (an example of an abnormal posture) may be treated as the same as a posture in which the user is standing upright on the floor (an example of a normal posture).

[0052] In contrast, in this embodiment, the right foot angle and left foot angle are learned, so information indicating the positional relationship between the person and the floor (in other words, the posture of the person relative to the floor) is not lost, thereby preventing adverse effects on the detection process.

[0053] The determination unit 160 determines whether or not a person has an abnormal posture using a learning model. Specifically, the determination unit 160 detects a 3D skeleton position that has a large difference in features from the 3D skeleton position of a learned normal posture as an abnormal posture. Note that a known technique can be applied to this detection method, so a detailed description will be omitted.

[0054] Then, the determination unit 160 controls the notification device 30 so that, if a person has an abnormal posture, it issues a notification indicating that the person has an abnormal posture (hereinafter referred to as a notification that an abnormal posture exists). As a result, the notification device 30 issues a notification that an abnormal posture exists by, for example, displaying an image, outputting a sound, turning on a lamp, etc.

[0055] The configuration of the abnormal posture detection device 10 has been described above.

[0056] The operation (learning process and detection process) of the abnormal posture detection device 10 will be described below.

[0057] First, the learning process performed by the abnormal posture detection device 10 will be described with reference to Fig. 5. Fig. 5 is a flowchart showing the flow of the learning process.

[0058] 5 starts, for example, when image data from camera 20 is input to abnormal posture detection device 10. The image data here is data obtained by photographing a person in a normal posture. Also, the following steps S1 to S5 are assumed to be performed repeatedly.

[0059] First, the skeleton position acquisition unit 110 acquires the three-dimensional skeleton position of a person in a normal posture based on input image data (step S1).

[0060] Next, the vector acquisition unit 120 acquires a vertical vector, a right foot vector, and a left foot vector based on the three-dimensional skeleton position acquired in step S1 (step S2).

[0061] Next, the calculation unit 130 calculates the right foot angle and the left foot angle based on the vertical vector, the right foot vector, and the left foot vector acquired in step S2 (step S3).

[0062] Next, the preprocessing unit 140 performs preprocessing on the three-dimensional skeleton position acquired in step S1 (step S4).

[0063] Next, the creation unit 150 trains a learning model by machine learning using learning data that associates the right foot angle and left foot angle calculated in step S3 with the three-dimensional skeleton position preprocessed in step S4 (step S5).

[0064] The learning process performed by the abnormal posture detection device 10 has been described above.

[0065] Next, the detection process performed by the abnormal posture detection device 10 will be described with reference to Fig. 6. Fig. 6 is a flowchart showing the flow of the detection process.

[0066] 6 starts, for example, when image data from camera 20 is input to abnormal posture detection device 10. The image data here is data obtained by photographing a person in either a normal posture or an abnormal posture. Furthermore, the following steps S11 to S15 are assumed to be performed repeatedly.

[0067] First, the skeleton position acquisition unit 110 acquires the three-dimensional skeleton position of a person based on the input image data (step S11).

[0068] Next, the vector acquisition unit 120 acquires a vertical vector, a right foot vector, and a left foot vector based on the three-dimensional skeleton position acquired in step S11 (step S12).

[0069] Next, the calculation unit 130 calculates the right foot angle and the left foot angle based on the vertical vector, the right foot vector, and the left foot vector acquired in step S12 (step S13).

[0070] Next, the preprocessing unit 140 performs preprocessing on the three-dimensional skeleton position acquired in step S11 (step S14).

[0071] Next, the determination unit 160 determines whether or not there is an abnormal posture by inputting the right foot angle and left foot angle calculated in step S13 and the three-dimensional skeleton position preprocessed in step S14 into the learning model described above (step S15).

[0072] If there is no abnormal posture (step S15: NO), the flow returns to step S11. Then, the processing from step S11 onwards is carried out based on the next input new image data.

[0073] On the other hand, if an abnormal posture is present (step S15: YES), the abnormal posture detection device 10 controls the alarm device 30 to issue a notification that an abnormal posture is present (step S16). As a result, the alarm device 30 issues a notification that an abnormal posture is present.

[0074] The detection process performed by the abnormal posture detection device 10 has been described above.

[0075] As described above, abnormal posture detection device 10 of this embodiment comprises skeletal position acquisition unit 110 that acquires the three-dimensional skeletal position of a person; vector acquisition unit 120 that acquires, based on the three-dimensional skeletal position, a vertical vector that is perpendicular to the floor surface on which the person is located, a right foot vector that indicates the direction of the person's right foot, and a left foot vector that indicates the direction of the person's left foot; calculation unit 130 that calculates the right foot angle formed by the vertical vector and the right foot vector, and the left foot angle formed by the vertical vector and the left foot vector; pre-processing unit 140 that performs pre-processing to align the three-dimensional skeletal position to a specified orientation and size; and judgment unit 160 that uses a learning model to determine whether or not the person has an abnormal posture, and controls a specified notification device to notify the person if an abnormal posture is detected.The learning model is characterized in that it is created by machine learning learning data that associates the right foot angle and left foot angle calculated based on the three-dimensional skeletal position of a person with a normal posture with the three-dimensional skeletal position of a person with a normal posture that has been pre-processed.

[0076] In other words, the abnormal posture detection device 10 of this embodiment uses label-free learning to train a machine learning model using only normal postures, and therefore has the advantages of eliminating the need for manual labeling or photographing of abnormal postures, and of not limiting the abnormal postures that can be detected (in other words, it can detect any abnormal posture whose characteristics deviate from those of the normal posture).

[0077] Furthermore, the abnormal posture detection device 10 of this embodiment performs preprocessing to align the orientation and size of the acquired three-dimensional skeleton position, so that the same posture can be treated as the same three-dimensional skeleton position, thereby preventing adverse effects on the detection process.

[0078] Normally, when the above-mentioned preprocessing is performed, there is a problem that the original orientation of the person (the positional relationship between the floor and the person) is lost. However, the abnormal posture detection device 10 of this embodiment associates the right foot angle and left foot angle with the three-dimensional skeleton position after preprocessing and provides them to the learning model, so the original orientation of the person is not lost.

[0079] The present disclosure is not limited to the above-described embodiment, and various modifications are possible without departing from the spirit of the present disclosure. Modifications will be described below.

[0080] [Variation 1] In the embodiment, the skeleton position acquiring unit 110 acquires a three-dimensional skeleton position based on captured image data, but the present invention is not limited to this.

[0081] For example, a sensor capable of measuring the three-dimensional skeletal position may be attached to the person, and the skeletal position acquisition unit 110 may receive data indicating the measurement results from the sensor and acquire the three-dimensional skeletal position based on the data.

[0082] Furthermore, for example, in the case of learning, the skeleton position acquisition unit 110 may receive image data of a person in a normal posture created using CG (Computer Graphics) from the outside, and acquire the three-dimensional skeleton position based on the image data.

[0083] Furthermore, for example, in the case of learning, the skeleton position acquisition unit 110 itself may directly acquire the three-dimensional skeleton position by creating the three-dimensional skeleton position using CG.

[0084] [Variation 2] Although the abnormal posture detection device 10 has the creation unit 150 and performs the detection process using a learning model created thereby, the present invention is not limited to this. For example, the abnormal posture detection device 10 may use the learning model created externally. In this case, the abnormal posture detection device 10 does not need to include the creation unit 150. [Industrial Applicability]

[0085] The abnormal posture detection device of the present disclosure is useful in general technology for detecting human posture. [Explanation of symbols]

[0086] 1. Abnormal posture detection system 10 Abnormal posture detection device 20 Camera 30 Alarm device 110 Skeleton position acquisition unit 120 Vector Acquisition Unit 130 Calculation Unit 140 Pretreatment section 150 Creation Department 160 Judgment section

Claims

1. a skeleton position acquisition unit for acquiring a three-dimensional skeleton position of a person; a vector acquisition unit that acquires, based on the three-dimensional skeleton position, a vertical vector that is perpendicular to a floor surface on which the person is located, a right foot vector that indicates the direction of the person's right foot, and a left foot vector that indicates the direction of the person's left foot; a calculation unit that calculates a right foot angle formed by the vertical vector and the right foot vector, and a left foot angle formed by the vertical vector and the left foot vector; a pre-processing unit that performs a first pre-processing for aligning the three-dimensional skeleton position to a specified orientation and a second pre-processing for aligning the three-dimensional skeleton position to a specified size; a determination unit that determines whether or not the person has an abnormal posture using a learning model, and controls a predetermined notification device to notify the person of the abnormal posture if the abnormal posture is detected; The learning model is The learning data is created by machine learning, in which the right foot angle and the left foot angle calculated based on the three-dimensional skeletal position of a person in a normal posture are associated with the three-dimensional skeletal position of the person in a normal posture that has been subjected to the first preprocessing. Abnormal posture detection device.

2. the three-dimensional skeleton position includes an upper three-dimensional skeleton position and a lower three-dimensional skeleton position of a lower body of the person, the upper three-dimensional skeletal position is the base of the right foot and the base of the left foot, or the center of the pelvis; the lower three-dimensional skeletal positions are a part below the right ankle and a part below the left ankle, the right foot vector is a vector directed from either the base of the right foot or the center of the pelvis to a portion below the right ankle, The left foot vector is a vector directed from either the base of the left foot or the center of the pelvis to a portion below the left ankle. The abnormal posture detection device according to claim 1 .

3. The first pretreatment is a process of translating or rotating at least three of the three-dimensional skeletal positions included in the torso of the person to align them with specified positions in a virtual three-dimensional space coordinate system; The abnormal posture detection device according to claim 1 or 2.

Images