Personnel fall behavior recognition method based on deep learning

Abstract

The present invention relates to the field of image recognition technology, and in particular to a method for identifying a person's falling behavior based on deep learning, comprising: obtaining skeleton points of a person in a road monitoring image; obtaining a marking direction angle according to a marking object; obtaining a main body direction angle according to the skeleton points; obtaining a main balance according to the marking direction angle and the main body direction angle; obtaining a stable feature degree according to the skeleton points; obtaining an overall state index according to the main balance and stable feature degree; obtaining a falling probability according to the overall state index; obtaining a person's falling behavior recognition result according to the falling probability combined with the overall state index. The present invention identifies a person's falling behavior by using the falling probability, thereby improving the accuracy of the person's falling behavior recognition result.

Description

Technical Field

[0001] The present invention relates to the technical field of image recognition, and in particular to a method for identifying a person's falling behavior based on deep learning. Background Art

[0002] On public transportation roads, falls are a major threat to vulnerable groups such as the elderly, often causing serious physical injuries and even affecting life safety. Therefore, it is necessary to use cameras to identify and monitor road falls, which will help to immediately issue an alarm when a fall occurs, thereby speeding up response time and providing timely assistance.

[0003] At present, the overall movement posture of the human body is often used to identify the fall behavior of a person. However, since the movement of the human body is relatively complex, the fall behavior is more manifested in the movement characteristics of each limb of the falling person. Identifying the fall behavior of a person only based on the overall movement posture of the human body will lead to a decrease in the accuracy of the fall behavior recognition results. Summary of the invention

[0004] The present invention provides a method for identifying human falling behavior based on deep learning to solve the existing problems.

[0005] The method for identifying falling behavior of a person based on deep learning of the present invention adopts the following technical scheme:

[0006] An embodiment of the present invention provides a method for identifying a person's falling behavior based on deep learning, the method comprising the following steps:

[0007] Acquire a continuous multi-frame road monitoring image, wherein the road monitoring image contains a number of people; acquire the shoulder, elbow, wrist, hip, knee, and ankle skeleton points of each person in the road monitoring image;

[0008] The road monitoring image includes a number of markers, and the marker direction angle is obtained according to the marker; the main direction angle of each person's body is obtained according to the relative position relationship between the shoulder and hip skeleton points of each person; the main balance of each person is obtained according to the difference between the marker direction angle and the main direction angle of each person's body;

[0009] According to the positional relationship of each person's elbow, wrist, knee, and ankle skeleton points, the stability characteristic degree of each person is obtained; according to the main balance and stability characteristic degree of each person, the overall state index of each person is obtained;

[0010] According to the difference in the overall status index of each person and other persons in each frame of the road monitoring image, the relative abnormality index of each person in each frame of the road monitoring image is obtained; according to the difference in the change and fluctuation of the overall status index between the persons in multiple consecutive frames of the road monitoring image, the fall probability of each person in each frame of the road monitoring image is obtained in combination with the relative abnormality index;

[0011] According to the fall probability and in combination with the overall state index, a person's fall behavior recognition result is obtained.

[0012] Furthermore, the step of obtaining the marking direction angle according to the marking object includes the following specific steps:

[0013] The road monitoring image is input into the sign recognition neural network to obtain the main direction angle of each sign in the road monitoring image; the average of the main direction angles of all the signs is recorded as the sign direction angle.

[0014] Furthermore, the method for obtaining the main direction angle of the body is as follows:

[0015] For any person in the road monitoring image, obtain the midpoint of the line connecting the two shoulder skeleton points and the midpoint of the line connecting the two hip skeleton points of the person, and record the angle corresponding to the line connecting the two midpoints as the main direction angle of the person's body.

[0016] Furthermore, the method for obtaining the main balance is as follows:

[0017] For any person, the cosine value of the difference between the main direction angle of the person's body and the marked direction angle is recorded as the main balance of the person.

[0018] Furthermore, the method for obtaining the degree of stability characteristics is as follows:

[0019] The pixel point in the lower left corner of the road monitoring image is taken as the origin, and the horizontal direction is Axis, longitudinal direction Axis, establish a rectangular coordinate system;

[0020] For any person in the road monitoring image, the sine value of the angle between the line connecting the elbow skeleton point and the wrist skeleton point and the line connecting the ankle skeleton point and the knee skeleton point is recorded as the direction stability value of the person; The normalized result of the absolute value of the coordinate difference is recorded as the distance stability value of the person; the product of the person's direction stability value and the distance stability value is recorded as the degree of stability characteristics of the person.

[0021] Furthermore, the method for obtaining the overall status index is as follows:

[0022] For any person, the product of the degree of the person's main balance and stability characteristics is recorded as the person's overall status index.

[0023] Furthermore, the relative abnormality index is obtained as follows:

[0024] For any person in any frame of road monitoring image, the normalized result of the mean of the difference between all persons in the frame of road monitoring image and the overall state index of the person is recorded as the relative abnormality index of the person in the frame of road monitoring image.

[0025] Furthermore, the method for obtaining the fall probability is as follows:

[0026] Record any frame of road monitoring image as a target monitoring image, record any person in the target monitoring image as a target person, record the previous frame of road monitoring image as a reference monitoring image of the target monitoring image, record the absolute value of the difference between the overall state index of the target person in the target monitoring image and its reference monitoring image as the state change of the target person in the target monitoring image; calculate the mean of the state change of all persons in the target monitoring image, and record the absolute value of the difference between the state change of the target person in the target monitoring image and the mean as the relative state change degree of the target person in the target monitoring image;

[0027] According to the fluctuation difference of the overall state index between the personnel in the continuous multiple frames of road monitoring images, the relative state fluctuation degree of each person in each frame of road monitoring images is obtained;

[0028] The linear normalization result of the product of the relative state change degree, the relative state fluctuation degree and the relative abnormality index of the target person in the target monitoring image is recorded as the fall probability of the target person in the target monitoring image.

[0029] Furthermore, the method for obtaining the relative state fluctuation degree is as follows:

[0030] For any person, the variance of the overall state index of the person in all frames of road monitoring images containing the person is recorded as the state variance index of the person; the mean of the state variance index of all persons in the target monitoring image is calculated, and the absolute value of the difference between the state variance index of the target person and the mean is recorded as the relative state fluctuation degree of the target person in the target monitoring image.

[0031] Furthermore, the step of obtaining the person's falling behavior recognition result based on the falling probability and the overall state index includes the following specific steps:

[0032] For any frame of road monitoring image, if there are two or more persons in the road monitoring image, when the probability of falling in the road monitoring image is greater than or equal to the preset fall threshold, it is judged that a person has fallen in the road monitoring image; when the probability of falling for all persons in the road monitoring image is less than the preset fall threshold, it is judged that no one has fallen in the road monitoring image; if there is only one person in the road monitoring image, when the overall state index of the person is less than or equal to the preset state threshold, it is judged that the person has fallen; when the overall state index of the person is greater than the preset state threshold, it is judged that the person is normal.

[0033] The beneficial effects of the technical solution of the present invention are as follows: the present invention obtains the marking direction angle according to the marking object in the road monitoring image, analyzes the difference between the marking direction angle and the main direction angle of the person's body, obtains the main balance of the person, and preliminarily judges the balance of the person; because the movement of the human body is relatively complex, the probability of falling obtained only by the main balance is inaccurate. The present invention analyzes the movement characteristics of the person's forearm and calf through the relative position relationship of the skeleton points, further judges the possibility of the person falling based on the limb movement characteristics, and obtains the overall state index of the person; by comparing the overall state index of each person with other persons in the road monitoring image, and combining the changes in the overall state index of each person in multiple consecutive frames of road monitoring images, a more reliable probability of falling is obtained. So far, the reliable probability of falling of the present invention is used to identify the falling behavior of the person, which improves the accuracy of the result of the recognition of the falling behavior of the person. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0035] Figure 1 This is a flowchart of the steps of the method for identifying falling behavior of a person based on deep learning of the present invention. DETAILED DESCRIPTION

[0036] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined purpose of the invention, the specific implementation method, structure, characteristics and effects of the person falling behavior recognition method based on deep learning proposed by the present invention are described in detail as follows in combination with the accompanying drawings and preferred embodiments. In the following description, different "one embodiment" or "another embodiment" does not necessarily refer to the same embodiment. In addition, specific features, structures or characteristics in one or more embodiments may be combined in any suitable form.

[0037] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.

[0038] The specific scheme of the method for identifying falling behavior of a person based on deep learning provided by the present invention is described in detail below with reference to the accompanying drawings.

[0039] See also Figure 1 , which shows a flowchart of a method for identifying a person's falling behavior based on deep learning provided by an embodiment of the present invention, the method comprising the following steps:

[0040] Step S001: Acquire multiple frames of continuous road monitoring images; acquire the shoulder, elbow, wrist, hip, knee, and ankle skeleton points of each person in the road monitoring images.

[0041] It should be noted that the purpose of this embodiment is to determine the probability of each person falling and to identify and monitor road falling behavior by analyzing the changes and differences in the relative positions of the skeleton points of the people in multiple consecutive frames of road monitoring images, so multiple consecutive frames of road monitoring images are first collected.

[0042] Specifically, for a road section where road fall behavior recognition monitoring is required, a road monitoring camera is used to continuously collect road monitoring images of the road section to obtain a continuous plurality of frames of road monitoring images, wherein the road monitoring images contain a number of persons;

[0043] Input multiple continuous frames of road monitoring images into the posture recognition neural network model to obtain the shoulder, elbow, wrist, hip, knee, and ankle skeleton points of each person appearing in the monitoring in each frame of the road monitoring image; the neural network model in this embodiment is poseNet, and other models can be used in other embodiments, which are not limited in this embodiment. The method for obtaining the data set is: collect a number of continuous multiple frames of road images through a road monitoring camera, manually mark the shoulder, elbow, wrist, hip, knee, and ankle skeleton points of each person in all road images, and take all the continuous multiple frames of road images and all the skeleton points of all people in each frame of the road image as a data set. The neural network is trained using the above-obtained data set, and the loss function used is the mean square error loss function. The specific training process is well known and will not be described in detail in this embodiment.

[0044] Step S002: Obtain the marking direction angle according to the marker; obtain the main body direction angle of each person according to the relative position relationship of the shoulder and hip skeleton points of each person; obtain the main balance of each person according to the difference between the marking direction angle and the main body direction angle of each person.

[0045] It should be noted that in the road monitoring image, when a person falls, the body direction of the person will tilt, so the main balance of the person is judged according to the main direction of the person's body. When judging the main balance of a person, it is necessary to combine the surrounding vertical markers to the ground, such as street lights, railings and other vertical facilities and buildings, and use the direction of the markers perpendicular to the ground as a reference to judge whether the person's body is balanced. First, the marking direction angle is obtained according to the marker, and then the main direction angle of each person's body is obtained according to the relative position relationship of the shoulder and hip skeleton points of each person, and the main balance of each person is obtained according to the difference between the marking direction angle and the main direction angle of each person's body.

[0046] It should be further explained that the main direction of a person's body will change when the person falls, so it is necessary to determine the main direction angle of the person's body when walking normally. When a person walks normally, the main direction of the body is perpendicular to the ground, and there are many signs perpendicular to the ground in the road monitoring images, such as street lights, railings and other vertical facilities and buildings. The main direction of these signs perpendicular to the ground is the same as the main direction of the person's body when walking normally, so the sign direction angle is first obtained based on the sign.

[0047] Preferably, in one embodiment of the present invention, obtaining the marking direction angle according to the marking object specifically includes: taking the pixel point at the lower left corner of the road monitoring image as the origin, and the horizontal direction is Axis, longitudinal direction Axis, establish a rectangular coordinate system; input the road monitoring image into the sign recognition neural network, and obtain the main direction angles of several signs in the road monitoring image, where the main direction angle is the clockwise direction from the straight line where the main body of the sign is located to The angle in the positive direction of the axis. The neural network model for identification in this embodiment is the VGG model. Other models can be used in other embodiments, which are not limited in this embodiment. The method for obtaining the data set is: collecting a number of road images through a road monitoring camera, manually marking the markers in the road images and the main direction angle of each marker, taking all the road images and the markers in the road images and the main direction angles of the markers as the data set, and using the above-obtained data set to train the neural network. The loss function used is the cross entropy loss function. The specific training process is well known and will not be described in detail in this embodiment; the average of the main direction angles of all the markers is recorded as the marker direction angle.

[0048] It should be noted that before comparing the marked direction angle with the main direction angle of the person's body, the main direction angle of the person's body needs to be obtained. The main direction of the person's body is determined by the upper part of the body, that is, the direction from the neck to the hip of the human body, so the acquired shoulder and hip skeleton points are analyzed to obtain the main direction angle of the body represented by the shoulder and hip skeleton points.

[0049] Preferably, in one embodiment of the present invention, the main direction angle of the body of each person is obtained according to the relative position relationship between the shoulder and hip skeleton points of each person, specifically comprising: for any person in the road monitoring image, obtaining the midpoint of the line connecting the two shoulder skeleton points and the midpoint of the line connecting the two hip skeleton points of the person, and extending the straight line connecting the two midpoints clockwise to The angle with respect to the positive direction of the axis is recorded as the main direction angle of the person's body.

[0050] It should be further explained that the smaller the difference between the main direction angle of the person's body and the marked direction angle is, the better the person's main balance is.

[0051] Preferably, in one embodiment of the present invention, the main balance of each person is obtained according to the difference between the marked direction angle and the main direction angle of each person's body, specifically including: for any person, the cosine value of the difference between the main direction angle of the person's body and the marked direction angle is recorded as the main balance of the person.

[0052] Step S003: Obtain the stability characteristic degree of each person according to the positional relationship of the elbow, wrist, knee and ankle skeleton points of each person; obtain the overall state index of each person according to the main balance and stability characteristic degree of each person.

[0053] It should be noted that the possibility of a person falling can only be preliminarily judged based on the main balance of the person. Since the movement of the human body is relatively complex, it is necessary to further judge whether the person has fallen in combination with the elbow, wrist, knee, and ankle skeleton points to increase the credibility of the judgment result. When a person falls more seriously, in order to bear the impact, the forearm will touch the ground when the person falls to the ground, and the calf will also touch the ground. At the same time, the person's hand will be closer to the ground. Therefore, the degree of stability characteristics of each person can be obtained based on the positional relationship of the elbow, wrist, knee, and ankle skeleton points of each person, and then the overall state index of each person can be judged in combination with the main balance.

[0054] Preferably, in one embodiment of the present invention, considering that when a person falls, the forearm will touch the ground in order to bear the impact, and the calf will also touch the ground, so the direction of the line connecting the elbow skeleton point and the wrist skeleton point is closer to the direction of the line connecting the ankle skeleton point and the knee skeleton point, and the wrist skeleton point and the ankle skeleton point are closer. The closer the coordinates are, the greater the possibility of a person falling. Therefore, the method for obtaining the degree of stability features includes: according to the positional relationship of the elbow, wrist, knee, and ankle skeleton points of each person, the degree of stability features of each person is obtained, specifically including: for any person in the road monitoring image, the sine value of the angle between the line connecting the elbow skeleton point and the wrist skeleton point and the line connecting the ankle skeleton point and the knee skeleton point is recorded as the direction stability value of the person; the sine value of the angle between the wrist skeleton point and the ankle skeleton point is recorded as the direction stability value of the person; The linear normalization result of the absolute value of the coordinate difference is recorded as the distance stability value of the person; the product of the person's direction stability value and the distance stability value is recorded as the degree of stability characteristics of the person.

[0055] It should be noted that the greater the degree of stability of a person's characteristics and the higher the main balance, the higher the person's overall status index.

[0056] Preferably, in one embodiment of the present invention, the overall status index of each person is obtained according to the main balance and stability characteristics of each person, specifically including: for any person, the product of the main balance and stability characteristics of the person is recorded as the overall status index of the person.

[0057] Step S004: According to the difference in the overall status index of each person and other persons in each frame of the road monitoring image, the relative abnormality index of each person in each frame of the road monitoring image is obtained; according to the change difference and fluctuation difference of the overall status index between the persons in multiple consecutive frames of road monitoring images, combined with the relative abnormality index, the fall probability of each person in each frame of the road monitoring image is obtained.

[0058] It should be noted that the overall state index of each person in the road monitoring image is different. The smaller the overall state index of a person is compared with the overall state index of other persons in the road monitoring image, the greater the possibility of the person falling; and in the continuous multi-frame road monitoring image, the overall state index of the person is constantly changing. The greater the degree of change of the overall state index of the person, the greater the possibility of the person falling. Therefore, the accuracy of the judgment result is further improved according to the difference and change of the overall state index. First, according to the difference in the overall state index of each person and other persons in each frame of the road monitoring image, the relative abnormality index of each person in each frame of the road monitoring image is obtained, and then according to the change difference and fluctuation difference of the overall state index between the persons in the continuous multi-frame road monitoring image, combined with the relative abnormality index, the probability of falling of each person in each frame of the road monitoring image is obtained.

[0059] Preferably, in one embodiment of the present invention, the relative abnormality index of each person in each frame of road monitoring image is obtained according to the difference in the overall status index between each person and other persons in each frame of road monitoring image, specifically including: for any person in any frame of road monitoring image, the linear normalization result of the mean of the difference between the overall status index of all persons in the frame of road monitoring image and the person is recorded as the relative abnormality index of the person in the frame of road monitoring image.

[0060] As an example, the relative anomaly index is formulated as:

[0061] ;

[0062] In the formula, For the Frame road monitoring image Relative abnormality index of individual personnel; For the The number of people in the frame road monitoring image; For the Frame road monitoring image The overall status index of the individual; No. Frame road monitoring image The overall status index of a person, is a linear normalization function, and the normalization object is Frame road surveillance images of all people .

[0063] It should be noted that, when the overall status index of a person relative to other persons in the road monitoring image is smaller, the relative abnormality index of the person is larger.

[0064] It should be noted that when a person falls, the overall state index corresponding to the person will decrease rapidly, causing the fluctuation degree of the overall state index corresponding to the person in multiple consecutive frames of road monitoring images to change. Therefore, the probability of falling for each person in each frame of road monitoring images is calculated.

[0065] Preferably, in one embodiment of the present invention, according to the change difference and fluctuation difference of the overall status index between persons in multiple consecutive frames of road monitoring images, combined with the relative abnormality index, the fall probability of each person in each frame of road monitoring images is obtained, specifically including:

[0066] Any frame of road monitoring image is recorded as the target monitoring image, any person in the target monitoring image is recorded as the target person, the previous frame of road monitoring image of the target monitoring image is recorded as the reference monitoring image of the target monitoring image, and the absolute value of the difference between the overall state index of the target person in the target monitoring image and its reference monitoring image is recorded as the state change of the target person in the target monitoring image; it should be noted that if the target person does not exist in the reference monitoring image of the target monitoring image, the state change of the target person in the target monitoring image is not calculated; the mean of the state change of all persons in the target monitoring image is calculated, and the absolute value of the difference between the state change of the target person in the target monitoring image and the mean is recorded as the relative state change degree of the target person in the target monitoring image;

[0067] According to the fluctuation difference of the overall state index between the personnel in the continuous multiple frames of road monitoring images, the relative state fluctuation degree of each person in each frame of road monitoring images is obtained;

[0068] The linear normalization result of the product of the relative state change degree, the relative state fluctuation degree and the relative abnormality index of the target person in the target monitoring image is recorded as the fall probability of the target person in the target monitoring image.

[0069] Specifically, the method for obtaining the relative state fluctuation degree includes:

[0070] For any person, the variance of the overall state index of the person in all frames of road monitoring images containing the person is recorded as the state variance index of the person; the mean of the state variance index of all persons in the target monitoring image is calculated, and the absolute value of the difference between the state variance index of the target person and the mean is recorded as the relative state fluctuation degree of the target person in the target monitoring image.

[0071] As an example, the probability of falling is formulated as:

[0072] ;

[0073] In the formula, The falling probability of the target person in the target monitoring image; The state change amount of the target person in the target monitoring image; is the mean value of the state change of all persons in the target monitoring image; is the state variance index of the target person; is the mean of the state variance index of all persons in the target monitoring image; The relative abnormality index of the target person in the target monitoring image; is the absolute value function.

[0074] It should be noted that It indicates the relative status change degree of the target person in the target monitoring image. The larger the value, the better. Indicates the relative state fluctuation degree of the target person in the target monitoring image.

[0075] Step S005: According to the fall probability and in combination with the overall state index, a person's fall behavior recognition result is obtained.

[0076] It should be noted that after obtaining the fall probability of each person in each frame of the road monitoring image, it is necessary to judge the fall behavior of the person according to the fall probability.

[0077] Specifically, for any frame of road monitoring image, if there are two or more persons in the road monitoring image, when the probability of falling in the road monitoring image is greater than or equal to the preset falling threshold When the probability of falling of all persons in the road monitoring image is less than the preset falling threshold, , it is determined that no one has fallen in the road monitoring image; if there is only one person in the road monitoring image, when the overall state index of the person is less than or equal to the preset state threshold When the overall status index of the person is greater than the preset status threshold When , it is judged that the person is normal; among them, , This embodiment is described using this as an example.

[0078] So far, the present invention is completed.

[0079] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for identifying falling behavior of a person based on deep learning, characterized in that: The method comprises the following steps: Acquire a continuous multi-frame road monitoring image, wherein the road monitoring image contains a number of people; acquire the shoulder, elbow, wrist, hip, knee, and ankle skeleton points of each person in the road monitoring image; The road monitoring image includes a number of markers, and the marker direction angle is obtained according to the marker; the main direction angle of each person's body is obtained according to the relative position relationship between the shoulder and hip skeleton points of each person; the main balance of each person is obtained according to the difference between the marker direction angle and the main direction angle of each person's body; According to the positional relationship of each person's elbow, wrist, knee, and ankle skeleton points, the stability characteristic degree of each person is obtained; according to the main balance and stability characteristic degree of each person, the overall state index of each person is obtained; According to the difference in the overall status index of each person and other persons in each frame of the road monitoring image, the relative abnormality index of each person in each frame of the road monitoring image is obtained; according to the difference in the change and fluctuation of the overall status index between the persons in multiple consecutive frames of the road monitoring image, the fall probability of each person in each frame of the road monitoring image is obtained in combination with the relative abnormality index; According to the fall probability and in combination with the overall state index, a person's fall behavior recognition result is obtained; The specific steps of obtaining the marking direction angle according to the marking object are as follows: Input the road monitoring image into the sign recognition neural network to obtain the main direction angle of each sign in the road monitoring image; the average of the main direction angles of all the signs is recorded as the sign direction angle; The method for obtaining the main direction angle of the body is as follows: For any person in the road monitoring image, obtain the midpoint of the line connecting the two shoulder skeleton points and the midpoint of the line connecting the two hip skeleton points of the person, and record the angle corresponding to the line connecting the two midpoints as the main direction angle of the person's body; The main balance acquisition method is as follows: For any person, the cosine value of the difference between the main direction angle of the person's body and the marked direction angle is recorded as the main balance of the person; The method for obtaining the degree of stability characteristics is as follows: The pixel point in the lower left corner of the road monitoring image is taken as the origin, and the horizontal direction is Axis, longitudinal direction Axis, establish a rectangular coordinate system; For any person in the road monitoring image, the sine value of the angle between the line connecting the elbow skeleton point and the wrist skeleton point and the line connecting the ankle skeleton point and the knee skeleton point is recorded as the direction stability value of the person; The normalized result of the absolute value of the coordinate difference is recorded as the distance stability value of the person; the product of the direction stability value and the distance stability value of the person is recorded as the degree of stability characteristics of the person; The method for obtaining the overall status index is as follows: For any person, the product of the degree of the person's main balance and stability characteristics is recorded as the person's overall status index.

2. The method for identifying falling behavior of a person based on deep learning according to claim 1, characterized in that: The method for obtaining the relative abnormality index is as follows: For any person in any frame of road monitoring image, the normalized result of the mean of the difference between all persons in the frame of road monitoring image and the overall state index of the person is recorded as the relative abnormality index of the person in the frame of road monitoring image.

3. The method for identifying falling behavior of a person based on deep learning according to claim 1 is characterized in that: The method for obtaining the fall probability is as follows: Record any frame of road monitoring image as a target monitoring image, record any person in the target monitoring image as a target person, record the previous frame of road monitoring image as a reference monitoring image of the target monitoring image, and record the absolute value of the difference between the overall state index of the target person in the target monitoring image and its reference monitoring image as the state change of the target person in the target monitoring image; Calculate the mean of the state changes of all persons in the target monitoring image, and record the absolute value of the difference between the state change of the target person in the target monitoring image and the mean as the relative state change degree of the target person in the target monitoring image; According to the fluctuation difference of the overall state index between the personnel in the continuous multiple frames of road monitoring images, the relative state fluctuation degree of each person in each frame of road monitoring images is obtained; The linear normalization result of the product of the relative state change degree, the relative state fluctuation degree and the relative abnormality index of the target person in the target monitoring image is recorded as the fall probability of the target person in the target monitoring image.

4. The method for identifying falling behavior of a person based on deep learning according to claim 3 is characterized in that: The method for obtaining the relative state fluctuation degree is as follows: For any person, the variance of the overall state index of the person in all frames of road monitoring images containing the person is recorded as the state variance index of the person; the mean of the state variance index of all persons in the target monitoring image is calculated, and the absolute value of the difference between the state variance index of the target person and the mean is recorded as the relative state fluctuation degree of the target person in the target monitoring image.

5. The method for identifying falling behavior of a person based on deep learning according to claim 1, characterized in that: The specific steps of obtaining the person's falling behavior recognition result according to the falling probability and in combination with the overall state index are as follows: For any frame of road monitoring image, if there are two or more persons in the road monitoring image, when the probability of falling in the road monitoring image is greater than or equal to the preset fall threshold, it is judged that a person has fallen in the road monitoring image; when the probability of falling for all persons in the road monitoring image is less than the preset fall threshold, it is judged that no one has fallen in the road monitoring image; if there is only one person in the road monitoring image, when the overall state index of the person is less than or equal to the preset state threshold, it is judged that the person has fallen; when the overall state index of the person is greater than the preset state threshold, it is judged that the person is normal.

Images