An elderly fall risk assessment method and system
The machine learning model identifies the posture of the elderly from the camera video data, evaluates the risk of falls, solves the problem of inconvenient sensor wear, and achieves efficient and accurate fall detection.
Patent Information
- Application Number
- CN201911412802.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2019-12-31
- Publication Date
- 2025-08-05
- Estimated Expiration
- 2039-12-31
AI Technical Summary
In the prior art, the elderly’s fall detection system relies on sensors to cause inconvenience in wearing and interference in movement, which is difficult to popularize.
Using machine learning models, surveillance video data is obtained through cameras, the posture of the elderly is identified and evaluated, and the dependence on sensors is reduced. The network model of human posture estimation is used to obtain joint coordinate data and evaluate the risk of falling.
It effectively reduces the use of sensors, reduces the trouble of wearing and interference in actions, and improves the accuracy and real-time detection.
Smart Images
Figure CN111243229B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of machine learning, and specifically, to a method and system for evaluating the fall risk of the elderly based on a machine learning model. Background Art
[0002] With the gradual increase in the proportion of the elderly population, more and more elderly communities have emerged. According to statistics, the injuries caused by falls are one of the main reasons threatening the healthy life of the elderly. At present, the prevention of falls in the elderly mainly relies on manual monitoring, which has many inconveniences and is difficult to apply to all elderly people.
[0003] In order to solve the above technical problems, the patent document with the publication number CN103027687A discloses an elderly fall detection and alarm system based on a 3D accelerometer and a gyroscope. It detects the posture of the elderly through multiple sensors. Although it can well judge whether the elderly have a fall risk and give an alarm, it is difficult to popularize among many elderly people; extensive use of detection sensors may interfere with the activities of the elderly, and there is also the problem of troublesome wearing.
[0004] The patent document with the publication number CN106539587A discloses a fall risk assessment and monitoring system and method based on multiple motion sensors. It uses a trained model to judge the fall risk of the elderly, but it still detects the walking posture of the elderly based on sensors, which is likely to interfere with the actions of the elderly and is also troublesome to wear. Summary of the Invention
[0005] The main object of the present invention is to provide a method and system for evaluating the fall risk of the elderly, which evaluates the fall risk by obtaining the human body posture based on a machine learning model, effectively reducing the usage amount of detection sensors, and reducing the trouble caused by wearing and the influence on the actions of the elderly.
[0006] To achieve the above object, the method for evaluating the fall risk of the elderly provided by the present invention includes the following steps:
[0007] Receiving step: receiving the monitoring video data obtained by a camera installed in the monitored place;
[0008] Preprocessing step: cropping a human body image containing at least human joints from each frame of the monitoring video data, and processing the human body image into a human body image to be recognized with a predetermined standard size;
[0009] Recognition step: obtaining the joint coordinate data in the human body image to be recognized based on a pre-trained human body pose estimation network model;
[0010] Evaluation step: Based on the joint coordinate data of consecutive multiple frames of images, evaluate the risk of the elderly falling.
[0011] As can be seen from the above solution, based on video data, use the existing human pose estimation network model to obtain the human joint coordinate data in each frame of image, and evaluate the risk of the elderly falling based on this coordinate data. It can effectively reduce the detection sensors arranged on the human joints, thereby effectively reducing the movement interference and trouble caused by wearing.
[0012] The specific solution is to install multiple cameras for monitoring from different perspectives in the monitored area. For the monitoring video data obtained by each camera, perform the recognition step and the evaluation step separately; if the data from more than one perspective evaluates that there is a risk of falling, it is determined that the elderly person has a risk of falling. Effectively reduce the probability of missed judgment.
[0013] The preferred solution is that the consecutive multiple frames of images are such that the deviation between the joint point coordinates in any two adjacent frames within them is less than the first preset threshold, and / or the coordinate deviation of the smallest rectangular frame that can enclose all joint points is less than the second preset threshold. It can effectively reduce the recognition work due to misjudgment of the human body image.
[0014] The preferred solution is that the step of processing the human body image into a human body image to be recognized with a predetermined standard size includes scaling and filling the cropped human body image into a human body image to be recognized with a predetermined standard size.
[0015] A further solution is to scale the human body image proportionally so that one pair of sides coincides with the boundary of the predetermined standard size, and the other pair of sides coincides with the boundary of the predetermined standard size or there is a gap space, and then fill the existing gap space to obtain a standard image to be recognized.
[0016] The preferred solution is that the preprocessing step includes obtaining the human body image from each frame of image based on a human body detector. Obtaining the human body image based on an existing human body detector, the technology is mature.
[0017] The preferred solution is that the preprocessing step includes the following steps:
[0018] Boundary determination step: In the image of the monitoring video data, based on the joint point coordinate data in the previous frame of image, generate a cropping boundary box for the human body image in the current frame of image; the cropping boundary box is the boundary of the intersection area between the expanded rectangular border obtained by expanding the four sides of the smallest rectangular frame that can enclose all joint points in the previous frame of image outward and the current frame of image.
[0019] Normalization step: Crop the current human body image from the current frame image based on the cropping border, and scale the cropped current human body image proportionally so that one pair of sides coincides with the predetermined standard size boundary, and the other pair of sides coincides with the predetermined standard size boundary or there is a gap space, and then fill the existing gap space to obtain a standard image to be recognized.
[0020] As can be seen from the above solution, the iterative strategy required for detecting a human body image that at least includes human body joint points is optimized, the update speed of the bounding box is accelerated, which provides guarantee for real-time output of joint coordinate data, so as to better reduce the delay of the elderly fall risk assessment and ensure the accuracy and real-time of control.
[0021] A further solution includes a sampling inspection step:
[0022] Expand the bounding box of the currently sampled current frame image outward by a predetermined ratio to form a temporary bounding box, and use the boundary of the intersection area between the area enclosed by the temporary bounding box and the currently sampled frame image as the sampling inspection bounding box;
[0023] Crop the sampled human body image from the currently sampled current frame image based on the sampling inspection bounding box, and scale and fill the sampled human body image into a reference comparison image;
[0024] Based on the human pose estimation network model, obtain the initial coordinates of the human joints in the reference comparison image, and transform the initial coordinates into reference coordinates in the coordinate system of the current frame image;
[0025] If the deviation between the standard coordinates and the reference coordinates of the currently sampled current frame image exceeds the preset threshold, then replace the standard coordinates with the reference coordinates as the human joint coordinate data of the currently sampled current frame image.
[0026] As can be seen from the above solution, it can effectively avoid error accumulation due to too much difference in the coordinate data of the front and back two frame images, and improve the accuracy of this method.
[0027] To achieve the above another purpose, the elderly fall risk assessment system provided by the present invention includes a processor and a memory, and the memory stores a computer program. It is characterized in that when the computer program is executed by the processor, it can implement the steps of the elderly fall risk assessment method described in any of the above technical solutions. Description of the Drawings
[0028] Figure 1 It is the workflow diagram of the risk assessment method in the embodiment of the present invention;
[0029] Figure 2 It is the schematic diagram of the position layout and viewing angle orientation of multiple cameras in the same monitored place in the embodiment of the present invention;
[0030] Figure 3 This is a schematic diagram of the process of cropping a human body image that at least includes human joints in an embodiment of the present invention. Among them, Figure 3 (a) is the cropped human body image, Figure 3 (b) is the human body image after being scaled proportionally; Figure 3 (c) is to obtain the gap space between the scaled human body image and the predetermined standard size; Figure 3 (d) is the human body image to be recognized obtained by filling the gap space.
[0031] The present invention will be further described below in conjunction with the embodiments and their accompanying drawings. Specific embodiments
[0032] Embodiment
[0033] As Figure 1 shown, the method for assessing the risk of the elderly falling in the present invention includes a receiving step S1, a preprocessing step S2, an identification step S3, and an assessment step S4. The specific process is as follows:
[0034] Receiving step S1: Receive the monitoring video data obtained by the camera installed in the monitored place.
[0035] As Figure 2 shown, a plurality of cameras 1 with different viewing angles are arranged in the monitored place 01. The more viewing angles of the cameras arranged, the smaller the probability of missed detection. Among them, the monitored place can be a yard, a room, an exercise place, etc., and usually requires as few obstacles as possible.
[0036] Preprocessing step S2: Crop a human body image that at least includes human joints from each frame of the monitoring video data, and process the human body image into a human body image to be recognized with a predetermined standard size.
[0037] To obtain a human body image that at least includes human joints from the image, it can be based on existing and relatively mature human body detectors. For example, the deep convolutional neural network YOLOv2 adopted by the patent document with the publication number CN109345504A can be used, or the deep convolutional neural network Deep CNN disclosed in the patent document with the publication number CN106778614A can also be used. Additionally, a method combining traditional HOG feature extraction and SVM classification can be used.
[0038] Although these existing human detector technologies are relatively mature, their calculations are relatively complex and require more computing resources. In addition, since it is only necessary to identify human joints when judging whether there is a risk of falling through changes in body posture, the following technical solution can be adopted. However, since it requires the use of existing human detectors for initialization and calibration, it is not suitable for situations where the elderly are continuously active for a long time within the camera's field of view. In this case, the preprocessing step S2 specifically includes a bounding box initialization step S21, a boundary determination step S22, and a normalization step S23. The specific process is as follows:
[0039] In the step S21 of initializing the bounding box, the bounding box of the human body image in the image at the beginning of the video data is calibrated.
[0040] When the human body bounding box determination program starts running, there is no historical joint point coordinate data for the bounding box to quickly update. This is when the bounding box needs to be initialized.
[0041] The calibration is to automatically calibrate the existing detector and end the initialization process by random sampling. For the existing detector, the deep convolutional neural network YOLOv2 adopted in the patent document with publication number CN109345504A can be used, or the deep convolutional neural network Deep CNN disclosed in the patent document with publication number CN106778614A can be used. You can also use some traditional HOG feature extraction and SVM classification methods.
[0042] The trigger for starting calibration using existing human detectors can be triggered by additional human detection equipment. For example, there are fewer obstructions in the monitored place, so that the elderly can be photographed by all cameras when entering the place. At this time, a card swiping device or a human infrared sensor can be set at the entrance, that is, when a human body is detected entering the monitored place, all cameras are started for monitoring, and the initial image in the monitoring video data is automatically calibrated using the existing detector. WIFI probes, Bluetooth probes, etc. deployed in the monitored place can also be used to obtain a predetermined mobile terminal carried by the elderly. The WIFI or Bluetooth of the terminal is in the turned-on state, so that the corresponding MAC address pre-stored in the identification library can be obtained to determine whether the elderly have entered the monitored place; human detection sensors can also be deployed on each camera or in the adjacent area to start monitoring and perform initial calibration when a human body is detected around it.
[0043] In each loop of the program, a new image is captured. After basic preprocessing such as denoising it, it is then handed over to a human body detector, which is the HOG+SVM human body detector in this embodiment. This human body detector detects the picture and obtains a set of multiple bounding boxes that may contain a human body. Then, duplicate removal operations need to be performed on the multiple bounding boxes to obtain the best bounding box. Therefore, by means of non-maximum suppression, the bounding box with the highest confidence, that is, the one most likely to contain a human body, is selected from this set of bounding boxes as the finally selected bounding box for the current image. This "selected bounding box" is not the final result of initialization, but only the optimal result for the current image.
[0044] After the bounding box initialization program ends, as Figure 3 shown, the coordinates (x0, y0) of the upper left vertex of the passed-out square box, as well as the width w and height h of the bounding box, these parameters represent this bounding box.
[0045] Boundary determination step S22, in the image of the monitored video data, based on the joint point coordinate data in the previous frame image, generate a cropping bounding box for the human body image in the current frame image; this cropping bounding box is the boundary of the intersection area between the enlarged rectangular border obtained by expanding the four sides of the smallest rectangular box that can enclose all the joint points in the previous frame image outwards and the current frame image.
[0046] In the initial stage, the current human body image is cropped from the current initial frame image based on the bounding box obtained by initialization. After initialization, the cropping bounding box is determined based on the joint coordinate data in the previous frame image.
[0047] Normalization step S23, as Figure 3 shown, scale and fill the current human body image into a standard image to be recognized.
[0048] Whether it is the bounding box obtained by initialization or update, after cropping the corresponding image slice containing the human body, operations of further scaling and filling are required to make it meet the size requirements of the input image of the convolutional neural network. The specific processing process includes the following steps:
[0049] First, calculate the ratio of the height / width of the bounding box h / w, and compare it with the required height / width ratio h net / w net at the input end of the neural network. If the former is greater than the latter, scale the bounding box by height to align with the input end of the network, that is, scale it to make h = h net , otherwise scale the bounding box by width to align with the input end of the network, that is, scale it to make w = w net; Then, move the scaled image slices to the middle area of the network input image; Finally, for the other areas of the network input image except the previous image slices, it can be filled with pure black [0, 0, 0] or pure gray [128, 128, 128] in the RGB color, or other colors can also be used, preferably colors that are easy to distinguish from the content part, and using solid colors for filling can avoid noise interference and is helpful for the calculation of the neural network. As Figure 3 shown is an example of an image operation after the former is greater than.
[0050] That is, scale the current human body image proportionally so that a pair of sides coincide with the standard size boundary, and the other pair of sides coincide with the standard size boundary or there are gap spaces, and then fill the existing gap spaces to obtain the standard image to be recognized.
[0051] Assume that the updated joint coordinates can be represented by J i (i = 1 to N), where N is the number of joints. Each J i is a three-dimensional vector corresponding to the XYZ coordinates of the joint. First, traverse all joints J i , and find the maximum and minimum values X max , X min , Y max , Y min in the X and Y directions of the image plane. Through these parameters, the smallest bounding box containing all the joints to be measured of the human body can be obtained.
[0052] X max = MIN(MAX(J i [0]; i = 1 to N), W)
[0053] X min = MAX(MIN(J i [0]; i = 1 to N), 0)
[0054] Y max = MIN(MAX(J i [1]; i = 1 to N), H)
[0055] Y min = MIN(MAX(J i [1]; i = 1 to N), 0)
[0056] w0 = X max - X min
[0057] h0 = Y max - Y min
[0058] W and H are the width and height of the input image respectively; w0 and h0 are the width and height of the minimum bounding box respectively.
[0059] Then, based on this minimum bounding box, a certain length is extended in both the horizontal and vertical directions to serve as the updated bounding box. The extension of the length is symmetric. For example, in the horizontal direction, the left and right sides of the box are extended by the same length. In the implementation of this system, the width of the box is extended by 0.4 times in the horizontal direction, and the height of the box is extended by 0.2 times in the vertical direction, that is, each side in the horizontal direction is extended by 0.2 times the width, and each side in the vertical direction is extended by 0.1 times the height. Finally, a bounding box with width w and height h is obtained. That is, in this embodiment, the minimum rectangular box is enlarged proportionally in its four directions.
[0060] If the current human body is located at the edge of the image, resulting in the boundary of the extended box exceeding the boundary of the image itself, the value here is changed to the corresponding value of the image boundary. That is, in this embodiment, regardless of whether there is a boundary exceeding the image, the boundary of the intersection area between the region enclosed by the proportionally enlarged rectangular box and the current frame image constitutes the cropped bounding box.
[0061] At this time, the following verification method can be used to determine whether the initialization process has achieved the preset purpose and decide whether to terminate the initialization process and proceed with the subsequent steps:
[0062] Step S41, expand the bounding box of the current frame image used for initialization into a verification bounding box according to a predetermined ratio, and the boundary of the intersection area between the region enclosed by the verification bounding box and the verification frame image constitutes the aforementioned verification bounding box.
[0063] Step S42, crop the human body verification image from the current frame image to be verified based on the verification bounding box, and scale and fill the human body verification image into a reference comparison image.
[0064] Step S43, based on the human body pose estimation network model, obtain the initial coordinates of the human body joints in the reference comparison image, and transform these initial coordinates into the reference coordinates in the coordinate system of the current frame image; and based on the initial coordinates obtained by using the human body pose estimation network model from the standard image to be recognized obtained in the aforementioned initialization step, and transform these initial coordinates into the standard coordinates in the coordinate system of the current frame image.
[0065] In this embodiment, the human body pose estimation network model uses VNect based on ResNet-50.
[0066] Step S44, if the deviation between the standard coordinates and the reference coordinates of the current frame image to be sampled is less than or equal to the preset threshold, the initialization is completed.
[0067] In the post-initialization stage, the bounding box of the current frame is obtained as the recognition result based on the joint coordinates of the previous frame image, the bounding box is determined, and the human body image containing at least the human body joints is cropped.
[0068] During this process, the obtained standard image to be identified can be spot-checked to prevent data deviation caused by the enlarged bounding box failing to completely surround the relevant nodes. The specific process includes the following steps:
[0069] In step S51 , the boundary frame of the current frame image being sampled is expanded to form a temporary boundary frame according to a predetermined ratio, and the boundary of the intersection area of the temporary boundary frame and the sampled frame image constitutes the sampled boundary frame.
[0070] Step S52 : cropping the sampled body image from the sampled current frame image based on the sampled boundary box, and scaling and filling the sampled body image as a reference comparison image.
[0071] Step S53: Based on the human posture estimation network model, the initial coordinates of the human body joints in the reference comparison image are obtained, and the initial coordinates are transformed into reference coordinates in the coordinate system of the current frame image.
[0072] Step S54: if the deviation between the standard coordinates of the current frame image being sampled and its reference coordinates exceeds a preset threshold, the standard coordinates are replaced by the reference coordinates as the human body joint coordinate data of the current frame image.
[0073] Identification step S3, based on the pre-trained human posture estimation network model, obtains the initial coordinates of the human joints in the standard image to be identified, and transforms the initial coordinates into standard coordinates in the coordinate system of the current frame image as the human joint coordinate data for subsequent processing.
[0074] In this embodiment, a convolutional neural network trained in advance for human posture estimation is used. In this embodiment, VNect based on ResNet-50 is used. Of course, other trained networks can also be used to obtain the initial coordinates of human joints in the standard image to be identified.
[0075] Since the acquired coordinates are coordinate data in the basic coordinate system of the standard image to be identified, the coordinates of the joint points are restored to the coordinate data in the coordinate system of the cropped previous human body image according to the reverse of the aforementioned scaling method. Then, the relationship between the coordinates of the previous human body image and the image before cropping is used to restore the corresponding coordinate data to the coordinate data in the coordinate system of the current frame image.
[0076] Evaluation step S4: evaluating the risk of the elderly falling based on the joint coordinate data of multiple consecutive frames of images.
[0077] Its evaluation is based on an existing evaluation model, which is not the inventive point of this application and will not be elaborated here.
[0078] After completing the initialization step, the above receiving step S1, preprocessing step S2, recognition step S3, and evaluation step S4 are repeated. And in the preprocessing step S2, the coordinate data of the key points required for determining the bounding box is based on the joint coordinate data obtained from the processing of the previous frame, without the need to use an existing human body detector, thereby improving the detection real-time performance. It can be the last frame image for stopping the initialization process or the previous frame image after normal processing.
[0079] In the above embodiment, since a bounding box for characterizing the human body image needs to be obtained, and the smallest rectangle box that can enclose all the key points cannot completely cover the entire human body image, the bounding box needs to be enlarged according to an empirical value and with an additional redundancy. Usually, the left and right sides of the bounding box are enlarged proportionally, and the upper and lower sides can be enlarged correspondingly according to the body type of the person object. The enlargement ratio is processed according to the actual situation and is not limited to the specific data in the above embodiment.
[0080] In the above embodiment, the preprocessing step, recognition step, and evaluation step are performed separately on the monitoring video data obtained by each camera; if the data from more than one perspective evaluates that there is a risk of falling, it is determined that the elderly person has a risk of falling, thereby effectively avoiding the problem of missed detection.
[0081] In order to avoid the person in the image walking out of the camera and then re-entering the monitoring range of the camera, after completing the recognition target, the image is verified. The specific solution is that for consecutive multiple frames of images, the deviation between the coordinate of the key points in any two adjacent frames is less than the first preset threshold, and / or the coordinate deviation of the smallest rectangle box that can enclose all the key points is less than the second preset threshold. The two preset thresholds are set according to the actual situation. For example, in the current shooting speed, the normal change range of the joint coordinate data of the monitored object during walking is used as the threshold.
[0082] In the present invention, an evaluation system is also disclosed. The system includes a processor and a memory. When a computer program stored in the memory is executed by the processor, the steps of the corresponding method can be implemented, which will not be elaborated here.
[0083] In the present invention, mainly an evaluation is performed on a specific object and a single object, and its recognition model is trained for this object; in addition, an existing technology recognition model for general objects can also be used. At this time, face recognition technology can be combined to monitor the corresponding object, so that multiple objects can be monitored and evaluated synchronously.
Claims
1. A method for assessing the risk of falls in the elderly, characterized in that: The following steps are involved: A receiving step of receiving surveillance video data acquired by a camera installed in a monitored location; a WIFI probe or a Bluetooth probe is arranged in the monitored location for acquiring a MAC address of a predetermined mobile terminal carried by an elderly person; the WIFI or Bluetooth of the predetermined mobile terminal is turned on, and based on this, a corresponding MAC address pre-stored in an identification library is acquired to determine that the elderly person has entered the monitored location; upon detecting that the elderly person has entered the monitored location, the camera is activated to acquire the surveillance video data; A preprocessing step of cropping a human body image containing at least human joints from each frame of the monitoring video data, and processing the human body image into a human body image to be identified of a predetermined standard size; The recognition step is to obtain the joint coordinate data in the human body image to be recognized based on the pre-trained human posture estimation network model; the step of obtaining the joint coordinate data in this step includes obtaining the initial coordinates of the human joints in the human body image to be recognized, and transforming the initial coordinates into standard coordinates in the coordinate system of the current frame image, and the standard coordinates are the joint coordinate data in this step; Evaluation step: assess the risk of falling of the elderly based on joint coordinate data of continuous multi-frame images; The preprocessing step includes: (1) a boundary determination step, in which, in the image of the monitoring video data, a cropping boundary box for the human body image in the current frame image is generated based on the joint point coordinate data in the previous frame image; the cropping boundary box is obtained by expanding the four sides of the minimum rectangular box that can enclose all the joint points of the previous frame image outward, and the boundary of the intersection area of the expanded rectangular frame and the current frame image is obtained; (2) a standardization step, based on the cropping boundary box, the current human body image is cropped from the current frame image, and the cropped current human body image is scaled proportionally to a pair of side edges coincide with the boundary of the predetermined standard size, and the other pair of side edges coincide with the boundary of the predetermined standard size or there is a gap space, and then the existing gap space is filled to obtain the human body image to be identified of standard size; In the preprocessing step, a human body image bounding box in an initial portion of the surveillance video data is calibrated using an existing detector; and then the human body image bounding box in a subsequent portion of the surveillance video data is determined using the boundary determination step and the standardization step; The step of calibrating the bounding box of the human image in the initial part of the surveillance video data using the existing detector includes the following steps: detecting the image using the human detector to obtain a set of multiple bounding boxes that may contain human bodies; and performing a deduplication operation on the multiple bounding boxes to select the bounding box with the highest confidence from the set of bounding boxes by non-maximum suppression as the final selected bounding box of the current image; The method for assessing the risk of falls of the elderly includes a sampling step: (1) expanding the boundary frame of the current frame image to be sampled into a temporary boundary frame according to a predetermined ratio, the boundary frame of the current frame image is a cropping boundary frame determined based on the boundary determination step of the preprocessing step, and the boundary of the intersection area of the area enclosed by the temporary boundary frame and the current sampled frame image constitutes a sampling boundary frame; (2) based on the sampling boundary frame, a sampling human body image is cropped from the current frame image to be sampled, and the sampling human body image is scaled and filled into a reference comparison image, the scaling and filling step includes scaling the sampling human body image to a pair of sides that are proportional to the preprocessed image. The boundaries of the predetermined standard size coincide with each other, and the other pair of sides coincide with the boundaries of the predetermined standard size or there is a gap space, and then the existing gap space is filled to obtain the standard size of the reference comparison image; (3) based on the human posture estimation network model, the initial coordinates of the human joints in the reference comparison image are obtained, and the initial coordinates are transformed into the reference coordinates in the coordinate system of the current frame image; (4) if the deviation between the standard coordinates of the current frame image being sampled and its reference coordinates exceeds a preset threshold, the standard coordinates are replaced by the reference coordinates as the human joint coordinate data of the current frame image being sampled.
2. The method for assessing the risk of falls in the elderly according to claim 1, wherein: A plurality of cameras for monitoring from different viewing angles are installed in the monitored location, and the identification step and the evaluation step are performed separately for the monitoring video data acquired by each camera; If data from more than one perspective assesses that the elderly person is at risk of falling, then the elderly person is deemed to be at risk of falling.
3. The method for assessing the risk of falls in the elderly according to claim 2, wherein: In order to avoid the situation where the person in the image walks out of the monitoring range of the camera and then walks back into the monitoring range of the camera, the deviation between the coordinates of the joint points in any two adjacent frames of the continuous multi-frame image is less than a first preset threshold, and / or the coordinate deviation of the minimum rectangular box that can enclose the relevant nodes is less than a second preset threshold.
4. The method for assessing the risk of falls in the elderly according to claim 1, wherein: In order to avoid the situation where the person in the image walks out of the monitoring range of the camera and then walks back into the monitoring range of the camera, the deviation between the coordinates of the joint points in any two adjacent frames of the continuous multi-frame image is less than a first preset threshold, and / or the coordinate deviation of the minimum rectangular box that can enclose the relevant nodes is less than a second preset threshold.
5. A method for assessing the risk of falls of the elderly, comprising a processor and a memory, wherein the memory stores a computer program, characterized in that: When the computer program is executed by the processor, it can implement the steps of the method for assessing the risk of falls of the elderly as described in any one of claims 1 to 4.