Full-space visual behavior discrimination method based on machine vision

Abstract

The invention discloses a full-space visual behavior discrimination method based on machine vision, and the method comprises the following steps: S1, obtaining a video stream, carrying out the recognition of a human body in the video stream, and building a coordinate system for each frame of image in the video stream; s2, segmenting a human body in the video streams to obtain video streams of all parts of the human body; s3, carrying out key point calibration and line calibration on each part of the human body of each frame of image in the segmented video stream, wherein key points comprise joint points; s4, identifying the calibrated video stream of each part to obtain the change condition of the key point of each part; and S5, assembling the video streams of all the parts according to the common key points to obtain the assembled video stream of the whole action of the human body. According to the invention, the movement condition of each part is identified independently, so that the accuracy of overall identification of the human body is improved.

Description

Technical Field

[0001] The invention relates to a behavior identification method, in particular to a behavior identification method based on full-space vision of machine vision. Background Art

[0002] Behavior recognition technology is a key research issue in the field of computer vision. It has gone through a development process from manually designed feature representation to deep learning feature expression. Early behavior recognition methods mainly relied on manually extracted features, such as feature description methods based on silhouettes, spatiotemporal interest points, and human joints. Although these methods perform well in certain specific scenarios, they are often difficult to cope with complex and changeable practical application scenarios.

[0003] With the rise of deep learning technology, the field of behavior recognition has ushered in new development opportunities. Deep learning automatically learns feature representations through neural networks, significantly improving the accuracy and robustness of behavior recognition. However, existing technologies only recognize static human behavior, such as whether a person is wearing a helmet, and cannot perform full-body recognition of the human body, thus missing detailed movements of the human body and reducing the accuracy of human behavior judgment. Summary of the invention

[0004] The purpose of the present invention is to provide a method for distinguishing behavior based on full-space visual machine vision to solve the problem of low accuracy of robot behavior distinction in the prior art.

[0005] A method for distinguishing visual behaviors in full space based on machine vision comprises the following steps:

[0006] S1. Obtain a video stream, identify a human body in the video stream, and establish a coordinate system for each frame of the video stream;

[0007] S2. Segment the human body in the video stream to obtain video streams of various parts of the human body;

[0008] S3. After segmentation, each frame of the video stream is calibrated with key points and lines for each part of the human body, wherein the key points include joint points;

[0009] S4. Identify the calibrated video streams of each part to obtain the changes of the key points of each part;

[0010] S5. Assemble the video streams of various parts according to the common key points to obtain the assembled video stream of the overall human body movement.

[0011] Furthermore, step S2 segments the human body in the video stream to obtain video streams of the torso, left arm, right arm, left hand, right hand, left leg and right leg respectively.

[0012] Furthermore, the specific method of calibrating the joint points of various parts of the human body in step S3 is:

[0013] The left shoulder joint, right shoulder joint, left hip joint and right hip joint are calibrated on the torso image; the left shoulder joint, left elbow joint and left wrist joint are calibrated on the left arm image; the right shoulder joint, right elbow joint and right wrist joint are calibrated on the right arm image; the left wrist joint, the fingertips of each finger, the first joint point, the second joint point and the palm joint are calibrated on the left hand image; the right wrist joint, the fingertips of each finger, the first joint point, the second joint point and the palm joint are calibrated on the right hand image; the left hip joint, left knee joint, left ankle joint, left heel and left toe are calibrated on the left leg image; the right hip joint, right knee joint, right ankle joint, right heel and right toe are calibrated on the right leg image.

[0014] Furthermore, the specific method of performing line calibration on various parts of the human body in step S3 is:

[0015] Connect the left shoulder joint of the torso to the right shoulder joint and the left hip joint, and connect the right hip joint to the left hip joint and the right shoulder joint, to form a quadrilateral representing the torso; connect the left shoulder joint of the arm to the left elbow joint, and the left elbow joint to the left wrist joint, to form a line representing the left arm; connect the right shoulder joint of the arm to the right elbow joint, and the right elbow joint to the right wrist joint, to form a line representing the right arm; connect the wrist joints of each hand to the palm joints of each finger, and each hand The palm joint of the finger is connected to the second joint, the second joint is connected to the first joint, and the first joint is connected to the fingertip; connect the left hip joint to the left knee joint, connect the left knee joint to the left ankle joint, connect the left ankle joint to the left heel and the left toe respectively, and connect the left heel to the left toe to form a line representing the left leg; connect the right hip joint to the right knee joint, connect the right knee joint to the right ankle joint, connect the right ankle joint to the right heel and the right toe respectively, and connect the right heel to the right toe to form a line representing the right leg.

[0016] Furthermore, step S4 recognizes the calibrated video stream, and obtains the change of the key points of the human body in the following specific manner:

[0017] According to the coordinate system established for each frame image in the video stream before the human body in the video stream is segmented, the coordinates of the key points on each frame image after segmentation are determined; according to the coordinates of the same key point on consecutive frame images, the relative distance change of the same key point in the entire video stream is determined; according to the relative distance of the same key point and the interval time between two adjacent frame images, the movement speed change of the same key point in the entire video stream and the change of the angle between adjacent lines in the entire video stream are determined.

[0018] Furthermore, the method further includes step S6 of analyzing and determining the behavior of the human body through the changes of key points in the assembled video stream:

[0019] The relative distance change, moving speed change, and angle change between adjacent lines of key points of at least one part are input into a behavior recognition model, and the behavior of the human body is semantically annotated and output. The behavior recognition model is pre-trained according to defined actions.

[0020] The present invention divides various parts of the human body, calibrates the joints of the arms, legs, hands and trunk of the segmented parts, and divides the various parts of the human body more finely. Through the calibration of joints and the connection of adjacent joints, the movement of the human body is more intuitive and the identification of various parts of the human body is more accurate. The relative distance and speed of the joints are determined by the coordinates of the joints in the video stream, and the changes in the joints of each part are determined according to the changes in the relative distance and the movement speed of each joint of each part, and finally integrated into the overall movement of the human body. Through the methods of separate identification and overall synthesis, the final recognition result is more accurate. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 It is a schematic diagram of the present invention for performing joint point calibration and line calibration on various parts of the human body. DETAILED DESCRIPTION

[0022] The present invention is further described below.

[0023] The present invention provides a method for distinguishing visual behaviors in a full-space based on machine vision, which specifically comprises the following steps:

[0024] S1. Obtain a video stream and identify a human body in the video stream.

[0025] The video stream can be captured by a surveillance camera or motion capture device. Before recognizing a human body in the video stream, it needs to be preprocessed, including adjusting the resolution, normalizing the frame rate, and correcting the color, to ensure the quality and consistency of the video stream, and to facilitate the subsequent recognition of human movements. The human body recognition model is used to recognize the human body in the video stream. The human body recognition model is pre-trained. Specifically, a large number of video streams including human bodies are collected, and the human bodies in the video streams are labeled to obtain samples, and the samples are input into the human body recognition model for training.

[0026] After the human body is recognized by the human body recognition model, a coordinate system is established for each frame of the video stream. For example, the recognized human body can be framed by a rectangular frame, and a point on the rectangular frame is used as the coordinate origin to establish the coordinate system. During subsequent human body movements, the rectangular frame does not move. The upper left corner of each frame can also be used as the coordinate origin.

[0027] S2. Segment the human body in the video stream to obtain video streams of various parts of the human body.

[0028] After the human body in the video stream is identified by the human body recognition model, the video stream is sent to the segmentation module. The segmentation module performs matrix segmentation on the video stream according to various parts of the human body. When segmenting, each frame of the image needs to be segmented to obtain the segmented video stream. The video stream is segmented in a matrix according to various parts of the human body, specifically: the segmentation module is based on various parts of the human body, such as the trunk, left arm, right arm, left hand, right hand, left leg and right leg, wherein the left leg includes the left thigh, left calf and left foot, and the right leg includes the right thigh, right calf and right foot; each frame of the image is divided into seven parts, namely, the trunk, left arm, right arm, left hand, right hand, left leg and right leg, and the final video stream also obtains the video stream corresponding to the seven parts of the trunk, left arm, right arm, left hand, right hand, left leg and right leg. The present invention can also segment the head to obtain a video stream of the head.

[0029] According to the coordinate system established before segmentation, a coordinate system is established for each frame image in the video stream of each part after segmentation, so that the coordinate systems of the images before and after segmentation remain consistent.

[0030] S3. Perform key point calibration and line calibration on each part of the human body in the segmented video stream.

[0031] The coordinate systems of the images of each part after segmentation are the same.

[0032] like Figure 1 As shown, after segmenting the various parts of the human body in the video stream, the calibration models of the various parts are used to perform key point calibration and line calibration on the various parts of the human body in each frame of the video stream. The calibration models include a torso calibration model, a left arm calibration model, a right arm calibration model, a left hand calibration model, a right hand calibration model, a left leg calibration model, and a right leg calibration model. The calibration models of the various parts are obtained by pre-training, specifically: a large number of video streams of various parts are collected, and each frame of the video stream of each part is manually labeled with key points and lines, and the labeled results are input as samples into the calibration model of the corresponding part for training. Key points include joint points, but are not limited to joint points. This embodiment takes joint points as an example to introduce the key point calibration of the present invention in detail.

[0033] Combination Figure 1 , the left shoulder joint point 24, the right shoulder joint point 23, the left hip joint point 47 and the right hip joint point 56 are calibrated on the torso image through the torso calibration model; the left shoulder joint point of the torso is connected to the right shoulder joint point and the left hip joint point respectively, and the right hip joint point is connected to the left hip joint point and the right shoulder joint point respectively, to form a quadrilateral representing the torso.

[0034] The left shoulder joint point 24, the left elbow joint point 25 and the left wrist joint point 26 are calibrated on the left arm image through the left arm calibration model; the left shoulder joint point and the left elbow joint point of the left arm are connected, and the left elbow joint point and the left wrist joint point are connected to form a line representing the left arm. It can be seen that the left arm image and the torso image have a common joint point, namely, the left shoulder joint point 24.

[0035] The right arm calibration model is used to calibrate the right shoulder joint point 23, the right elbow joint point 22 and the right wrist joint point 5 on the right arm image; the right shoulder joint point and the right elbow joint point of the right arm are connected, and the right elbow joint point and the right wrist joint point are connected to form a line representing the right arm. The right arm image and the torso image have a common joint point, namely, the right shoulder joint point 23.

[0036] The left wrist joint 26, the fingertip 30, the first joint 29, the second joint 28 and the palm joint 27 of the thumb, the fingertip 34, the first joint 33, the second joint 32 and the palm joint 31 of the index finger, the fingertip 35, the first joint 36, the second joint 37 and the palm joint 38 of the middle finger, the fingertip 42, the first joint 41, the second joint 40 and the palm joint 39 of the ring finger, and the fingertip 43, the first joint 44, the second joint 45 and the palm joint 46 of the little finger are calibrated on the left hand image through the left hand calibration model; the wrist joints of the left hand are connected to the palm joints of the thumb, index finger and little finger respectively, and the palm joints of the index finger, middle finger, ring finger and little finger are connected in sequence, the palm joints of each finger are connected to the second joint, the second joint is connected to the first joint, and the first joint is connected to the fingertip. There is a common joint point between the left hand image and the left arm image, namely, the left wrist joint point 26 .

[0037] The right wrist joint 5, the fingertip 1, the first joint 2, the second joint 3 and the palm joint 4 of the thumb, the fingertip 18, the first joint 19, the second joint 20 and the palm joint 21 of the index finger, the fingertip 17, the first joint 16, the second joint 15 and the palm joint 14 of the middle finger, the fingertip 10, the first joint 11, the second joint 12 and the palm joint 13 of the ring finger, and the fingertip 9, the first joint 8, the second joint 7 and the palm joint 6 of the little finger are calibrated on the right hand image through the right hand calibration model; the wrist joints of the left hand are connected to the palm joints of the thumb, index finger and little finger respectively, and the palm joints of the index finger, middle finger, ring finger and little finger are connected in sequence, and the palm joints of each finger are connected to the second joint, the second joint is connected to the first joint, and the first joint is connected to the fingertip. There is a common joint between the right hand image and the right arm image, namely: the right wrist joint 5.

[0038] The left hip joint 47, the left knee joint 48, the left ankle joint 49, the left heel 50 and the left toe 51 are calibrated on the left leg image by the left leg calibration model; the left hip joint is connected to the left knee joint, the left knee joint is connected to the left ankle joint, the left ankle joint is connected to the left heel and the left toe respectively, and the left heel is connected to the left toe to form a line representing the left leg. There is a common joint point between the left leg image and the torso image, namely: the left hip joint 47.

[0039] The right hip joint 56, right knee joint 55, right ankle joint 54, right heel 52 and right toe 53 are calibrated on the right leg image through the right leg calibration model; the right hip joint is connected to the right knee joint, the right knee joint is connected to the right ankle joint, the right ankle joint is connected to the right heel and the right toe respectively, and the right heel is connected to the right toe. There is a common joint point between the right leg image and the torso image, namely: the right hip joint 56.

[0040] In addition to calibrating joint points, the present invention can also calibrate key points of other positions of the human body. For example, the midpoint between two adjacent joint points can be calibrated.

[0041] The present invention can also draw the outline of the head, and perform grid calibration on the head after the outline is drawn to obtain a grid-shaped head, which is convenient for subsequent identification of the head.

[0042] S4. Identify the calibrated video streams of each part to obtain changes in the joint points of each part.

[0043] The video stream of each part corresponds to a certain part of the human body. According to the coordinate system established in each frame image in the video stream of each part, the change of the coordinates of the joint points in each frame image in the video stream is determined, and the change of the joint points of each part in the video stream is determined according to the change of the coordinates of the joint points in each frame image.

[0044] The changes of the joint points include the changes in the relative distance and movement speed of the same joint point in different frame images, as well as the changes in the angles between the lines and lines between adjacent joint points.

[0045] According to the coordinate changes of the joint points in the continuous frame images, the relative distance changes and speed changes of the corresponding joint points are calculated. The specific method for determining the relative distance changes of the joint points is: according to the coordinates of the joint points in the previous frame image and the coordinates in the next frame image, the relative distance of the same joint point in the two adjacent frames of images is determined, and so on, the change of the relative distance of the same joint point in the entire video stream can be determined. The specific method for determining the change of the moving speed of the joint point is: according to the relative distance and time interval of the same joint point in the two adjacent frames of images, the moving speed of the same joint point can be determined, and so on, the change of the moving speed of the same joint point in the entire video stream can be determined.

[0046] The specific method of determining the angle change between lines is as follows: according to the coordinates of the joint points at both ends of the line on each frame image, the vector of the line is determined, the angle between two adjacent lines is calculated using the dot product formula, and according to the size of the angle between the two lines in the two adjacent frames of images, the angle change of the adjacent lines in the two adjacent frames of images is determined, and so on, so as to determine the angle change of the adjacent lines in the entire video stream. For example, the angle between the line of the first joint point and the second joint point of the finger and the line of the second joint point and the palm joint point changes from 180° to 60°.

[0047] When analyzing the changes in the trunk joints, it is also necessary to determine the changes in the posture inclination angle. The specific method is: based on the changes in the left shoulder joint and the right shoulder joint, determine the changes in the angle between the lines of the left shoulder joint and the right shoulder joint and the x-axis or y-axis, and determine the changes in the posture inclination angle of the human body.

[0048] Each part corresponds to an operator of the part, and the operator is used to calculate and analyze the displacement changes of the joint points of the corresponding part to obtain the changes of the joint points.

[0049] The present invention can also identify the human head and determine the head movement trajectory and speed. Specifically, the head movement trajectory is determined by the change of the coordinates of the intersection of the head grids in two adjacent frames of images, and the head movement speed is determined according to the interval time between the two adjacent frames of images.

[0050] S5. Assemble the video streams of various parts according to the common joint points to obtain the assembled video stream of the overall human body movement.

[0051] When synthesizing the movements of various parts, the synthesis is performed based on the same time of the video stream, and the adjacent body parts are synthesized based on the common joints of the images of different parts.

[0052] The identified parts are assembled according to the time points and common joints to obtain an assembled video stream. Specifically, the images of the seven parts with the same frame number are first assembled according to the common joints to obtain an assembled overall image, and so on, to generate the overall video stream according to the time sequence.

[0053] The specific method of assembling the seven parts according to the common joint points is: using the right shoulder joint point to connect the right arm image and the torso image, using the left shoulder joint point to connect the left arm image and the torso image, using the left wrist joint point to connect the left hand image and the left arm image, using the right wrist joint point to connect the right hand image and the right arm image, using the left hip joint point to connect the left leg image and the torso image, and using the right hip joint point to connect the right leg image and the torso image.

[0054] S6. Analyze and determine the behavior of the human body through the changes of the joint points in the assembled video stream.

[0055] After assembling the video streams of various parts and linking the changes in the joints of each part, the entire movement and posture changes of the human body are formed.

[0056] The present invention can also use the behavior recognition model to recognize the behavior in a specific environment.

[0057] The action recognition model is pre-trained.

[0058] For example, based on the model's application scenarios and safety requirements, a series of rules for dangerous actions are defined. More specifically, in industrial scenarios, actions such as limbs approaching dangerous equipment quickly, losing balance and falling into dangerous areas are defined as dangerous behaviors. In sports training scenarios, excessive twisting of joints and actions beyond the normal range of motion may be considered dangerous behaviors. These dangerous actions are input into the behavior recognition model for training, and the behavior recognition model can be used to quickly identify dangerous actions later.

[0059] For example, a series of standard movements can also be defined according to needs. More specifically, for a set of gymnastics, first perform the standard movement method of each movement. Or in medical rehabilitation, standard movements are given to rehabilitation trainees during training. These standard movements are input into the recognition model for training, and later the behavior recognition model can be used to quickly identify whether the movement is standard and qualified. When determining human behavior, human behavior can be identified only by one part, or by combining the movements of at least two parts to determine human behavior. For example, when the present invention is applied to hand rehabilitation training, it is only necessary to judge the degree of hand grip; when identifying dangerous behaviors, it is necessary to identify the overall movement of the human body to determine whether the human behavior is dangerous, such as the head and body moving quickly toward a dangerous object at the same time.

[0060] When only one part is to be identified, the human behavior is determined by the recognition model of the part according to the changes of the joints of the part, and semantic annotation is performed. The relative distance change and speed change of each joint point of the part to be identified in the video stream, as well as the line and angle change between adjacent joint points, are input into the recognition model of the part for identification.

[0061] The recognition model for each part of the human body is pre-trained. A large number of video streams of different parts of the human body performing various actions are collected. The changes in the joints of different parts and the corresponding action categories are divided into training sets and test sets as well as the corresponding behavioral semantics, and then input into the corresponding part recognition model for training. The recognition model of each part learns the complex mapping relationship between the changes in the joints and the behavioral semantics.

[0062] The recognition models of various parts include torso recognition model, left arm recognition model, right arm recognition model, left hand recognition model, right hand recognition model, left leg recognition model and right leg recognition model. The recognition models of various parts are pre-trained, and the characteristics of different actions are learned according to the distance changes between different joints and the angle changes of adjacent lines.

[0063] The torso recognition model is used to extract features based on the motion trajectory and speed changes between the joints on the torso, as well as the angles of the lines connecting the joints and the adjacent lines, to identify the changes in the joints of the human torso.

[0064] The left arm behavior recognition model is used to learn the characteristics of the left arm and obtain the behavior of the left arm according to the changes in the left arm joints.

[0065] The right arm behavior recognition model is used to learn the characteristics of the right arm and obtain the behavior of the right arm according to the changes in the right arm joints.

[0066] The left-hand behavior recognition model is used to learn the characteristics of the left hand and obtain the behavior of the left hand according to the changes in the left-hand joints.

[0067] The right-hand behavior recognition model is used to learn the characteristics of the right hand and obtain the behavior of the right hand according to the changes in the right-hand joints.

[0068] The left leg behavior recognition model is used to determine the left leg trunk movement based on the angle change between the line between the left hip joint and the left knee joint and the line between the left knee joint and the left ankle joint, that is, the angle change between the thigh and the calf, and the relative distance change between the feature points on the left leg. The left foot movement is determined based on the relative distance change between the left ankle joint, the left heel joint and the left toe joint of the foot. The left leg behavior is determined based on the left leg trunk movement and the left foot movement.

[0069] The right leg behavior recognition model is used to determine the right leg trunk movement based on the angle changes between the line between the right hip joint and the right knee joint and the line between the right knee joint and the right ankle joint, as well as the motion trajectory and movement speed changes of the joints on the right leg. The right foot movement is determined based on the motion trajectory and movement speed of the right ankle joint, right heel joint and right toe joint of the foot. The right leg behavior is determined based on the right leg trunk movement and the right foot movement.

[0070] The head recognition model can also be used to determine the behavior of the head based on the movement trajectory and speed of the head.

[0071] When the changes in the joints of at least two parts are needed to determine the behavior of the human body, after determining the individual changes in each part, it is necessary to analyze how these changes work together to form a complete action.

[0072] The parts that need to be considered are considered as relevant parts. For example, when shooting a three-pointer, the torso, left arm, right arm, left hand, right hand, left leg, and right leg need to work together. Combine the individual behaviors of each part to jointly determine the behavior of the human body.

[0073] When human behavior is identified through one part, it is only necessary to use the corresponding recognition model to determine whether the corresponding part has any dangerous behavior, and the recognition result for each part is more accurate; when human behavior is identified through at least two parts, the final human behavior is more accurate by combining the behaviors of each part.

Claims

1. A method for distinguishing visual behaviors based on full-space machine vision, characterized in that: The steps include: S1. Obtain a video stream, identify a human body in the video stream, and establish a coordinate system for each frame of the video stream; S2. Segment the human body in the video stream to obtain video streams of various parts of the human body; S3. After segmentation, each frame of the video stream is calibrated with key points and lines for each part of the human body, wherein the key points include joint points; S4. Identify the calibrated video streams of each part to obtain the changes of the key points of each part; S5. Assemble the video streams of various parts according to the common key points to obtain the assembled video stream of the overall human body movement.

2. The method for distinguishing visual behaviors based on full-space machine vision according to claim 1, characterized in that: Step S2 segments the human body in the video stream to obtain video streams of the torso, left arm, right arm, left hand, right hand, left leg and right leg respectively.

3. The method for distinguishing visual behaviors based on full-space machine vision according to claim 2, characterized in that: The specific method of calibrating the joint points of various parts of the human body in step S3 is: The left shoulder joint, right shoulder joint, left hip joint and right hip joint are calibrated on the torso image; the left shoulder joint, left elbow joint and left wrist joint are calibrated on the left arm image; the right shoulder joint, right elbow joint and right wrist joint are calibrated on the right arm image; the left wrist joint, the fingertips of each finger, the first joint point, the second joint point and the palm joint are calibrated on the left hand image; the right wrist joint, the fingertips of each finger, the first joint point, the second joint point and the palm joint are calibrated on the right hand image; the left hip joint, left knee joint, left ankle joint, left heel and left toe are calibrated on the left leg image; the right hip joint, right knee joint, right ankle joint, right heel and right toe are calibrated on the right leg image.

4. The method for distinguishing visual behaviors based on full-space machine vision according to claim 3, characterized in that: The specific method of performing line calibration on various parts of the human body in step S3 is: Connect the left shoulder joint of the torso to the right shoulder joint and the left hip joint, and connect the right hip joint to the left hip joint and the right shoulder joint, forming a quadrilateral representing the torso; Connect the left shoulder joint of the arm to the left elbow joint, and connect the left elbow joint to the left wrist joint to form a line representing the left arm; Connect the right shoulder joint of the arm to the right elbow joint, and connect the right elbow joint to the right wrist joint to form a line representing the right arm; Connect the wrist joint of each hand to the palm joint of each finger, connect the palm joint of each finger to the second joint, connect the second joint to the first joint, and connect the first joint to the fingertip; connect the left hip joint to the left knee joint, connect the left knee joint to the left ankle joint, connect the left ankle joint to the left heel and left toe, and connect the left heel to the left toe to form a line representing the left leg; Connect the right hip joint point to the right knee joint point, connect the right knee joint point to the right ankle joint point, connect the right ankle joint point to the right heel and right toe respectively, and connect the right heel and right toe to form a line representing the right leg.

5. The method for distinguishing visual behaviors based on full-space machine vision according to claim 1, characterized in that: Step S4 recognizes the calibrated video stream, and obtains the specific method of the change of the key points of the human body as follows: According to the coordinate system established for each frame image in the video stream before the human body in the video stream is segmented, the coordinates of the key points on each frame image after segmentation are determined; according to the coordinates of the same key point on consecutive frame images, the relative distance change of the same key point in the entire video stream is determined; according to the relative distance of the same key point and the interval time between two adjacent frame images, the movement speed change of the same key point in the entire video stream and the change of the angle between adjacent lines in the entire video stream are determined.

6. The method for distinguishing visual behaviors based on full-space machine vision according to claim 5, characterized in that: The step S6 also includes analyzing and determining the behavior of the human body through the changes of key points in the assembled video stream: The relative distance change, moving speed change, and angle change between adjacent lines of key points of at least one part are input into a behavior recognition model, and the behavior of the human body is semantically annotated and output. The behavior recognition model is pre-trained according to defined actions.

Images