Anti-falling monitoring method and system for dementia patient

Abstract

The invention relates to the technical field of medical assistance, in particular to an anti-falling monitoring method and system for a dementia patient. An anti-falling monitoring system for a dementia patient comprises a dementia patient monitoring module, a motion parameter data construction module, a motion control analysis module and an anti-falling intervention module. Video data of a space where a dementia patient is located is continuously collected, motion parameters of key nodes of the trunk and the lower limbs are extracted in combination with skeleton topology and time sequence constraint, a motion control constraint space reflecting the human body collaborative motion relation is constructed, and then the effective freedom degree of motion control is quantified; the falling state of the human body stability before falling is recognized in advance; when it is detected that the effective freedom degree of motion control is lower than an instability threshold value, intervention measures are triggered in time.

Description

Technical Field

[0001] This invention relates to the field of medical assistive technology, specifically to a method and system for monitoring and preventing falls in dementia patients. Background Technology

[0002] Existing fall prevention technologies for the elderly, especially those with dementia, mainly rely on wearable sensors, pressure sensing devices, or post-fall video playback analysis. Wearable devices require long-term use by patients, which can easily lead to monitoring interruptions due to poor patient compliance, misoperation, or loss, and can also interfere with patients' daily activities. Solutions based on environmental sensing or simple video analysis are mostly focused on the recognition and alarm after a fall occurs, usually judging by superficial features such as sudden changes in body posture, sudden increases in speed, or whether the body has touched the ground. They are difficult to capture in time the gradual decline in the body's motor control ability before a fall, resulting in problems such as delayed warnings and high false alarm rates. Summary of the Invention

[0003] This invention achieves stable acquisition of the motor state of dementia patients over a short period of time by continuously collecting video data in scenarios such as wards, homes, or nursing homes, and sampling and grayscale processing of video frames based on a preset time window. By introducing skeleton topology rules and temporal constraints, it identifies, connects, and tracks key motor nodes such as the trunk, hip joint, knee joint, and ankle joint in the video frame sequence, constructing high-dimensional motion parameter data containing displacement, velocity, and acceleration, thereby accurately representing the real motor behavior of dementia patients over a continuous time period. Furthermore, it constructs a covariance matrix based on the motion parameter data and extracts a change consistency index, and obtains a motion control constraint space by combining least squares iterative optimization, and quantifies the effective degrees of freedom that the human body can independently adjust the stability of movement in the current state through singular value decomposition. By comparing the effective degrees of freedom of motion control with the instability threshold, it can identify the downward trend of the patient's motor stability ability before the patient falls, and promptly trigger voice prompts or manual intervention.

[0004] This invention provides a method for monitoring and preventing falls in dementia patients, comprising: Video data is sampled based on a preset time length to obtain a video frame sequence; Based on skeleton topology and temporal constraints, the video frame sequence is processed to obtain motion parameter data of key motion nodes corresponding to dementia patients; The motion control constraint space is determined based on the motion parameter data, and the effective degrees of freedom for motion control are determined in the motion control constraint space based on the motion parameter data. The effective degrees of freedom of motor control are compared with the instability threshold. If the effective degrees of freedom of motor control are less than the instability threshold, intervention measures are taken for the corresponding dementia patients. If the effective degrees of freedom of motor control are not less than the instability threshold, no operation is performed.

[0005] As a preferred aspect, the video frame sequence is processed based on skeleton topology and temporal constraints to obtain motion parameter data of key motion nodes corresponding to dementia patients, specifically including the following steps: Human region detection is performed on all video frames in the video frame sequence, and all the obtained human region images are combined into a human region image sequence, with the human region images labeled by the timestamps corresponding to the video frames. For each human region image in the human region image sequence, the following operations are performed: the human region image is fed into the key part annotation network for processing to obtain the annotation information corresponding to several human structural regions in the human region image. The annotation information includes the coordinates of the region center and the corresponding annotation label. Under the constraint of the skeleton topology rules, the motion key nodes are connected based on the annotation labels corresponding to the human structural regions to construct the skeleton topology graph. Based on the annotation information and skeleton topology map corresponding to human body region images, key motion nodes are tracked, and motion parameter data of key motion nodes corresponding to dementia patients are constructed.

[0006] As a preferred aspect, human region detection is performed on all video frames within the video frame sequence, and all the obtained human region images are combined into a human region image sequence, specifically including the following steps: For each pixel in a video frame, perform the following steps: obtain the pixel value of the pixel in all video frames, and fit a Gaussian distribution model corresponding to the pixel for all pixel values. The Gaussian distribution model includes the observation mean and observation variance. For each video frame in the video frame sequence, the following steps are performed: Each pixel in the video frame is traversed, and the pixel value is matched against its corresponding Gaussian distribution model. The matching method is to compare the absolute value of the difference between the pixel value and the observed mean with the product of the observed variance and the empirical coefficient. If the absolute value of the difference between the pixel value and the observed mean is less than the product of the observed variance and the empirical coefficient, the match is considered successful. If the pixel value matches the corresponding Gaussian distribution model, the pixel is marked as a background pixel. If the pixel value does not match the corresponding Gaussian distribution model, the pixel is marked as a non-background pixel. Then, connected component analysis is performed on all non-background pixels in the video frame, and the rectangular image corresponding to the largest connected component is marked as the human body region image. The pixel values ​​of pixels in the human body region image that do not belong to the largest connected component are set to 0. All the obtained human body region images are then combined into a human body region image sequence.

[0007] As a preferred aspect, motion key nodes are tracked based on the annotation information and skeleton topology map corresponding to human body region images, and motion parameter data of motion key nodes corresponding to dementia patients are constructed, specifically including the following steps: Perform a union operation on the labels corresponding to all skeleton topology diagrams to obtain the label itemset; Iterate through the annotation labels in the annotation label set. For each annotation label, perform the following: Iterate through the human region images. Based on the annotation label, search the annotation information corresponding to the current human region image. If a matching annotation label exists in the annotation information corresponding to the current human region image, select the region center coordinates in the annotation information of the human region image as the positioning coordinates, and annotate the current human region image using the timestamp corresponding to the human region image sequence. If no matching annotation label exists in the annotation information corresponding to the current human region image, search the annotation information corresponding to the adjacent human region images based on the annotation label. Record the human region images with matching annotation labels as the positioning human region images. Based on the motion key nodes corresponding to the annotation labels, search the skeleton topology graph corresponding to the positioning human region image. If a motion key node corresponding to the annotation label is found in the skeleton topology graph corresponding to the positioning human region image... The labels corresponding to adjacent motion key nodes exist in the annotation information of the current human region image. These motion key nodes are designated as localization motion key nodes. The difference between the region center coordinates of the localization motion key node in the annotation information of the current human region image and the region center coordinates in the annotation information of the localization human region image is recorded as the fitting difference. The region center coordinates in the annotation information of the localization human region image are then used as calibration coordinates. The sum of the calibration coordinates and the fitting difference is recorded as the candidate fitting coordinates. All candidate fitting coordinates are summed and averaged to obtain the fitted coordinates, which are then labeled using the timestamp corresponding to the current human region image in the human region image sequence. The localization coordinates or fitted coordinates are arranged and concatenated according to their sequence number in the human region image sequence to form a label data sequence set. The label data sequence set includes the label coordinates arranged according to their sequence number in the human region image sequence. For each set of labeled data sequences, the distance between adjacent labeled coordinates is recorded as the displacement, and the displacement is labeled using the timestamp corresponding to the latter of the adjacent labeled coordinates. All displacements are arranged and concatenated according to the timestamps to form a displacement vector. The interval between the labeled coordinate and the timestamp corresponding to the first labeled coordinate in the labeled data sequence is recorded as the interval time corresponding to the labeled coordinate. With the interval time as the independent variable and the displacement as the dependent variable, a curve function is fitted. After performing the first derivative, the data corresponding to the interval time is used as the velocity. After performing the second derivative, the data corresponding to the interval time is used as the acceleration. All velocities are arranged and concatenated according to the interval time to form a velocity vector, and all accelerations are arranged and concatenated according to the interval time to form an acceleration vector. The motion displacement vector, motion velocity vector, and motion acceleration vector corresponding to the labeled data sequence set are concatenated from top to bottom to form local motion parameter data. Then, the local motion parameter data corresponding to all the labeled data sequence sets are combined to form overall motion parameter data. Overall motion parameter data is actually a matrix data. Finally, the transpose of overall motion parameter data is recorded as motion parameter data.

[0008] As a preferred aspect, the motion control constraint space is determined based on motion parameter data, and the effective degrees of freedom for motion control are determined in the motion control constraint space based on the motion parameter data, specifically including the following: Construct the covariance matrix corresponding to the motion parameter data, and perform eigenvalue decomposition on the covariance matrix corresponding to the motion parameter data. The ratio of the largest eigenvalue to the sum of all eigenvalues ​​is denoted as the consistency of change. Each row of the motion parameter data is denoted as a motion parameter vector. The product of the motion parameter vector and the motion control constraint matrix is ​​denoted as the fit variation consistency. The mean square error between all fit variation consistency and variation consistency is calculated and denoted as the mapping error. The direction of the mapping error is minimized. The motion control constraint matrix is ​​iteratively optimized by the least squares method until the termination condition is met. The motion control constraint matrix is ​​output. The transpose of the motion control constraint matrix is ​​the motion control constraint space. Singular value decomposition is performed on the motion control constraint space, and the number of non-zero singular values ​​is the effective degree of freedom of motion control.

[0009] As a preferred approach, the network is trained for labeling key parts, specifically in the following ways: Obtain several key part annotation samples, and annotate the key part annotation samples with annotation information. Combine all the annotated key part annotation samples into a key part annotation set, and train the key part annotation network using the key part annotation set.

[0010] This invention also provides a fall prevention monitoring system for dementia patients, comprising: The dementia patient monitoring module is used to sample video data based on a preset time length to obtain a video frame sequence; The motion parameter data construction module is used to process video frame sequences based on skeleton topology and temporal constraints to obtain motion parameter data of key motion nodes corresponding to dementia patients. The motion control analysis module is used to determine the motion control constraint space based on motion parameter data, and to determine the effective degrees of freedom of motion control in the motion control constraint space based on the motion parameter data. The fall prevention intervention module compares the effective degrees of freedom of motor control with the instability threshold. If the effective degrees of freedom of motor control is less than the instability threshold, intervention measures are taken for the corresponding dementia patient. If the effective degrees of freedom of motor control is not less than the instability threshold, no operation is performed.

[0011] The present invention has the following advantages: This invention achieves stable acquisition of the motor state of dementia patients over a short period of time by continuously collecting video data in scenarios such as wards, homes, or nursing homes, and sampling and grayscale processing of video frames based on a preset time window. By introducing skeleton topology rules and temporal constraints, it identifies, connects, and tracks key motor nodes such as the trunk, hip joint, knee joint, and ankle joint in the video frame sequence, constructing high-dimensional motion parameter data containing displacement, velocity, and acceleration, thereby accurately representing the real motor behavior of dementia patients over a continuous time period. Furthermore, it constructs a covariance matrix based on the motion parameter data and extracts a change consistency index, and obtains a motion control constraint space by combining least squares iterative optimization, and quantifies the effective degrees of freedom that the human body can independently adjust the stability of movement in the current state through singular value decomposition. By comparing the effective degrees of freedom of motion control with the instability threshold, it can identify the downward trend of the patient's motor stability ability before the patient falls, and promptly trigger voice prompts or manual intervention. Attached Figure Description

[0012] Figure 1 This is a schematic diagram of the fall prevention monitoring system for dementia patients used in an embodiment of the present invention. Detailed Implementation

[0013] To enable those skilled in the art to better understand the technical solutions of this invention, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of this invention.

[0014] Example 1: A method for monitoring and preventing falls in dementia patients, comprising: The system continuously collects video data from the space where the dementia patient is located using a camera. It should be noted that, in order to monitor dementia patients, video is typically collected from the space where the patient is located. This space includes wards, homes, and nursing homes. The video data is sampled based on a preset time length to obtain a video frame sequence. The preset time length is usually set by the operator in the background configuration file, typically 5 seconds. Sampling refers to selecting all video frames in the video data according to fixed rules, such as a uniform sampling method of 25 frames per second. The video frames are in image form and have undergone grayscale processing. The resulting video frame sequence includes several video frames, which can be used to analyze the state of the dementia patient for the preset time length. By processing video frame sequences based on skeleton topology and temporal constraints, motion parameter data of key motion nodes corresponding to dementia patients can be obtained. It should be noted that in human motion control, the movements of the trunk, hip joint, knee joint, and ankle joint are not completely synchronous, but have a certain degree of independent adjustment capability, which is called redundant degrees of freedom in motion control. By processing video frame sequences through skeleton topology and temporal constraints, the key motion nodes of dementia patients in video frames can be identified. These key motion nodes correspond to the trunk, hip joint, knee joint, and ankle joint. Then, the position, velocity, acceleration, and other motion parameters of these key motion nodes are analyzed to form motion parameter data, which is used to analyze the motion control mode of dementia patients and thus implement fall prevention measures for dementia patients. The motion control constraint space is determined based on motion parameter data, and the effective degrees of freedom of motion control are determined in the motion control constraint space based on the motion parameter data. Here, the motion control constraint space refers to the parameter action space composed of the influence relationship of each motion parameter on the motion stability capability under the constraint of the cooperative motion relationship corresponding to the motion parameter data. It represents the ways in which the human body can use to regulate motion stability in the time period corresponding to the video frame sequence. The effective degrees of freedom of motion control refers to the number of adjustment dimensions that can independently affect motion stability under the constraint of the motion control constraint space. The effective degrees of freedom of motor control are compared with the instability threshold, which is set by the operator. If the effective degrees of freedom of motor control are less than the instability threshold, it indicates that the adjustment methods that the human body can use under the current motor control constraint space can no longer maintain motor stability, and it is in a state of degree-of-freedom collapse. In this state, multiple human body parts or segments will be highly synchronized, causing the original independent instability adjustment methods to disappear, and the human body will appear as a rigid structure. At the data level, the motion correlation of human body parts is significantly increased. Due to the decline in perception, dementia patients are more difficult to recover from the state of degree-of-freedom collapse and are more likely to fall in the future. Intervention measures are taken for the corresponding dementia patients. These intervention measures include voice prompts for caregivers to intervene for the corresponding dementia patients, such as providing assistance. If the effective degrees of freedom of motor control are not less than the instability threshold, no operation is performed. Based on skeleton topology and temporal constraints, the video frame sequence is processed to obtain motion parameter data of key motion nodes corresponding to dementia patients. The specific steps include the following: Human region detection is performed on all video frames in the video frame sequence, and all the obtained human region images are combined into a human region image sequence, with the human region images labeled by the timestamps corresponding to the video frames. For each human region image in the human region image sequence, the following operations are performed: the human region image is fed into a key part annotation network for processing, obtaining annotation information corresponding to several human structural regions in the human region image. Human structural regions include regions corresponding to parts such as the torso, hip joint, knee joint, and ankle joint. The annotation information includes the region center coordinates and corresponding annotation labels. The annotation labels are the human structural regions such as the torso, hip joint, knee joint, and ankle joint in the human region image. The region center coordinates are the pixel coordinates of the center of the region where the annotation label is located in the human region image, generally obtained by averaging the pixel coordinates corresponding to all pixels in the region where the annotation label is located. Using human body structural regions as key nodes for motion, and under the constraints of skeleton topology rules, the key nodes are connected based on the corresponding labels of the human body structural regions to construct a skeleton topology graph. It should be noted that the skeleton topology rules here refer to the allowed connections between human body structural regions, such as the torso region and the hip joint region, the knee joint region and the ankle joint region. More specifically, left and right side rules can be introduced when setting labels, such as left hip joint, right hip joint. Then the skeleton topology rules also include allowing only same-side connections. These skeleton topology rules are set by the developers based on expert experience; the key part annotation network adopts the YOLOv8 model. Based on the annotation information and skeleton topology map corresponding to human body region images, key motion nodes are tracked, and motion parameter data of key motion nodes corresponding to dementia patients are constructed. Human region detection is performed on all video frames within a video frame sequence, and all the obtained human region images are combined into a human region image sequence. The specific steps include the following: For each pixel in a video frame, the following steps are performed: obtain the pixel value across all video frames, and fit a Gaussian distribution model for each pixel value. The Gaussian distribution model includes the observed mean and observed variance. The observed mean of a pixel is used to characterize the typical value of that pixel in the background state, and the observed variance of a pixel is used to characterize the fluctuation range of that pixel in the background state. It should be noted that all video frames here are not limited to a preset time length, but are placed in a longer time dimension. Alternatively, the Gaussian distribution model for each pixel can be continuously maintained in the background. For each video frame in the video frame sequence, the following steps are performed: Each pixel in the video frame is traversed, and the pixel value is matched against its corresponding Gaussian distribution model. The matching method compares the absolute value of the difference between the pixel value and the observed mean with the product of the observed variance and an empirical coefficient. The empirical coefficient is typically set to 2-3, corresponding to a confidence level of 95%-99%. If the absolute value of the difference between the pixel value and the observed mean is less than the product of the observed variance and the empirical coefficient, the match is considered successful. If the pixel value matches the corresponding Gaussian distribution model, the pixel is marked as a background pixel. If the pixel value does not match the corresponding Gaussian distribution model, the pixel is marked as a non-background pixel. Then, connected component analysis is performed on all non-background pixels in the video frame, and the rectangular image corresponding to the largest connected component is marked as the human body region image. Pixels in the human body region image that do not belong to the largest connected component are set to 0. All the obtained human body region images are then combined into a human body region image sequence. Based on the annotation information and skeleton topology map corresponding to human body region images, key motion nodes are tracked, and motion parameter data of key motion nodes corresponding to dementia patients are constructed. The specific steps include the following: Perform a union operation on the labels corresponding to all skeleton topology diagrams to obtain the label itemset; Iterate through the annotation labels in the annotation label set. For each annotation label, perform the following: Iterate through the human region image, and based on the annotation label, search the annotation information corresponding to the current human region image. If a matching annotation label exists in the annotation information corresponding to the current human region image, select the region center coordinates in the annotation information corresponding to the human region image as the positioning coordinates, and annotate the current human region image using the timestamp corresponding to the human region image sequence. If no matching annotation label exists in the annotation information corresponding to the current human region image, search the annotation information corresponding to the adjacent human region images based on the annotation label. Here, "adjacent" means adjacent in the human region image sequence. Record the human region images with matching annotation labels as the positioning human region images. Based on the motion key nodes corresponding to the annotation labels, search the skeleton topology graph corresponding to the positioning human region images. If the annotation label corresponding to the motion key node corresponding to the annotation label exists in the annotation information corresponding to the current human region image in the skeleton topology graph corresponding to the positioning human region image, record the motion key node as the positioning motion key node. The method involves using a node to calculate the difference between the coordinates of the center of the current human region image and the coordinates of the center of the corresponding labeled information. The center coordinates of the labeled information are then used as calibration coordinates. The sum of the calibration coordinates and the fitting difference is recorded as the candidate fitting coordinates. Since there will be several key nodes during the query process, there will also be several candidate fitting coordinates. All candidate fitting coordinates are summed and averaged to obtain the fitting coordinates, which are then labeled using the timestamp corresponding to the current human region image in the human region image sequence. The positioning coordinates or fitting coordinates are arranged and concatenated according to their sequence number in the human region image sequence to form a labeled data sequence set. This set includes the labeled coordinates arranged according to their sequence number in the human region image sequence. It should be noted that during image recognition, the displacement of some human structural regions may not be recognized or the recognition results may be inaccurate at certain angles. In practical applications, multiple cameras can be set up in the area where dementia patients reside to prevent falls from multiple perspectives. For each set of labeled data sequences, the distance between adjacent labeled coordinates is recorded as the motion displacement, and labeled using the timestamp corresponding to the latter of the adjacent labeled coordinates. In particular, the timestamp corresponding to the first labeled coordinate in the labeled data sequence does not correspond to a motion displacement, so all motion displacements are arranged and concatenated according to the timestamps to form a motion displacement vector. The interval between the labeled coordinate and the timestamp corresponding to the first labeled coordinate in the labeled data sequence is recorded as the interval time corresponding to the labeled coordinate. In particular, with the timestamp corresponding to the first labeled coordinate in the labeled data sequence as the origin, the interval time as the independent variable, and the motion displacement as the dependent variable, a curve function is fitted by polynomial fitting. After performing the first derivative, the data corresponding to the interval time is used as the motion velocity, and after performing the second derivative, the data corresponding to the interval time is used as the motion acceleration. All motion velocities are arranged and concatenated according to the interval time to form a motion velocity vector, and all motion accelerations are arranged and concatenated according to the interval time to form a motion acceleration vector. The motion displacement vector, motion velocity vector, and motion acceleration vector corresponding to the labeled data sequence set are concatenated from top to bottom to form local motion parameter data. Then, the local motion parameter data corresponding to all the labeled data sequence sets are combined to form overall motion parameter data. The overall motion parameter data is actually a matrix data. The transpose of the overall motion parameter data is then recorded as motion parameter data. The motion control constraint space is determined based on motion parameter data, and the effective degrees of freedom for motion control are determined within the motion control constraint space based on the motion parameter data. Specifically, this includes the following: Construct the covariance matrix corresponding to the motion parameter data, and perform eigenvalue decomposition on the covariance matrix corresponding to the motion parameter data. The ratio of the largest eigenvalue to the sum of all eigenvalues ​​is denoted as the consistency of change. Each row of the motion parameter data is denoted as a motion parameter vector. The product of the motion parameter vector and the motion control constraint matrix is ​​denoted as the fit variation consistency. The mean square error between all fit variation consistency and variation consistency is calculated and denoted as the mapping error. The direction of minimizing the mapping error is used to iteratively optimize the motion control constraint matrix using the least squares method until the termination condition is met. The termination condition can be set to the mapping error being less than a pre-set threshold or to satisfying a certain number of iterations. The motion control constraint matrix is ​​output. The transpose of the motion control constraint matrix is ​​the motion control constraint space. The motion control constraint space is essentially a numerical matrix, with columns corresponding to motion parameters. The motion control constraint space is obtained through the influence relationship between motion parameters and variation consistency, reflecting the ways in which the human body can regulate motion stability. Singular value decomposition is performed on the motion control constraint space. The number of non-zero singular values ​​is the effective degree of freedom of motion control. When the effective degree of freedom of motion control is lower than the instability threshold, it indicates that the influence of multiple motion parameters on the stability control task tends to be ineffective or jointly constrained, resulting in an insufficient number of ways to independently adjust posture stability, which makes it easier to fall in the future. Training the network for key component annotation includes the following methods: Several key part annotation samples are obtained, including human body region images constructed from video frames actually captured by the operator. The key part annotation samples are annotated with annotation information, which is the contour division annotation made by the operator based on different structures of the human body region images. All annotated key part annotation samples are combined into a key part annotation set. The key part annotation network is trained using the key part annotation set. During training, the loss value is constructed based on the difference between the annotation information corresponding to each region output by the key part annotation network and the actual annotation information. The construction method can be MSE. The key part annotation network is iteratively optimized by gradient descent to minimize the loss value. If the accuracy of the key part annotation network meets the expectations, the training of the key part annotation network is completed.

[0015] This application achieves stable acquisition of the motor state of dementia patients over a short period of time by continuously collecting video data in scenarios such as wards, homes, or nursing homes, and sampling and grayscale processing of video frames based on a preset time window. By introducing skeleton topology rules and temporal constraints, it identifies, connects, and tracks key motor nodes such as the trunk, hip joint, knee joint, and ankle joint in the video frame sequence, constructing high-dimensional motion parameter data containing displacement, velocity, and acceleration, thereby accurately representing the real motor behavior of dementia patients over a continuous time period. Furthermore, it constructs a covariance matrix based on the motion parameter data and extracts a change consistency index, and obtains a motor control constraint space by combining least squares iterative optimization, and quantifies the effective degrees of freedom that the human body can independently adjust the stability of movement in the current state through singular value decomposition. By comparing the effective degrees of freedom of motor control with the instability threshold, it can identify the downward trend of the patient's motor stability ability before the patient falls, and promptly trigger voice prompts or manual intervention.

[0016] Example 2: A fall prevention monitoring system for dementia patients, such as... Figure 1 As shown, it includes: The dementia patient monitoring module is used to continuously collect video data of the space where the dementia patient is located via a camera. It should be noted that in order to monitor dementia patients, video is usually collected from the space where the dementia patient is located. This space includes wards, homes, and nursing homes. The video data is sampled based on a preset time length to obtain a video frame sequence. The preset time length is generally set by the operator in the background configuration file, usually set to 5 seconds. Sampling refers to selecting all video frames in the video data according to fixed rules, such as a uniform sampling method of 25 frames per second. The video frames are in image form and have undergone grayscale processing. The resulting video frame sequence includes several video frames, which can be used to analyze the state of the dementia patient for the preset time length. The motion parameter data construction module is used to process video frame sequences based on skeleton topology and temporal constraints to obtain motion parameter data of key motion nodes corresponding to dementia patients. It should be noted that in human motion control, the movements of the trunk, hip joint, knee joint, and ankle joint are not completely synchronous, but have a certain degree of independent adjustment capability, which is called redundant degrees of freedom in motion control. By processing the video frame sequence through skeleton topology and temporal constraints, the key motion nodes of dementia patients in the video frames can be identified. These key motion nodes correspond to the trunk, hip joint, knee joint, and ankle joint. Then, the position, velocity, acceleration, and other motion parameters of these key motion nodes are analyzed to form motion parameter data, so as to analyze the motion control mode of dementia patients and thus implement fall prevention measures for dementia patients. The motion control analysis module is used to determine the motion control constraint space based on motion parameter data, and to determine the effective degrees of freedom of motion control in the motion control constraint space based on the motion parameter data. Here, the motion control constraint space refers to the parameter action space composed of the influence relationship of each motion parameter on the motion stability ability under the constraint of the cooperative motion relationship corresponding to the motion parameter data. It represents the ways in which the human body can use to regulate motion stability in the time period corresponding to the video frame sequence. The effective degrees of freedom of motion control refers to the number of adjustment dimensions that can independently affect motion stability under the constraint of the motion control constraint space. The fall prevention intervention module compares the effective degrees of freedom of motor control with the instability threshold, which is set by the operator. If the effective degrees of freedom of motor control are less than the instability threshold, it indicates that the body's adjustment methods under the current motor control constraints are no longer sufficient to maintain stability, resulting in a state of degree-of-freedom collapse. In this state, multiple body parts or segments become highly synchronized, causing the previously independent instability adjustment methods to disappear, and the body becomes a rigid structure. At the data level, the motion correlation of body parts is significantly increased. Dementia patients, due to their decreased sensory abilities, have more difficulty recovering from the state of degree-of-freedom collapse and are more prone to subsequent falls. Intervention measures are taken for corresponding dementia patients, including voice prompts to caregivers to intervene, such as providing assistance. If the effective degrees of freedom of motor control are not less than the instability threshold, no action is taken.

[0017] It should be understood that those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims. Parts not described in detail in this specification are prior art known to those skilled in the art.

Claims

1. A method for monitoring and preventing falls in dementia patients, characterized in that, include: Video data is sampled based on a preset time length to obtain a video frame sequence; Based on skeleton topology and temporal constraints, the video frame sequence is processed to obtain motion parameter data of key motion nodes corresponding to dementia patients; The motion control constraint space is determined based on the motion parameter data, and the effective degrees of freedom for motion control are determined in the motion control constraint space based on the motion parameter data. The effective degrees of freedom of motor control are compared with the instability threshold. If the effective degrees of freedom of motor control are less than the instability threshold, intervention measures are taken for the corresponding dementia patients. If the effective degrees of freedom of motor control are not less than the instability threshold, no operation is performed.

2. The method for monitoring and preventing falls in dementia patients according to claim 1, characterized in that, Based on skeleton topology and temporal constraints, the video frame sequence is processed to obtain motion parameter data of key motion nodes corresponding to dementia patients. The specific steps include the following: Human region detection is performed on all video frames in the video frame sequence, and all the obtained human region images are combined into a human region image sequence, with the human region images labeled by the timestamps corresponding to the video frames. For each human region image in the human region image sequence, the following operations are performed: the human region image is fed into the key part annotation network for processing to obtain the annotation information corresponding to several human structural regions in the human region image. The annotation information includes the coordinates of the region center and the corresponding annotation label. Under the constraint of the skeleton topology rules, the motion key nodes are connected based on the annotation labels corresponding to the human structural regions to construct the skeleton topology graph. Based on the annotation information and skeleton topology map corresponding to human body region images, key motion nodes are tracked, and motion parameter data of key motion nodes corresponding to dementia patients are constructed.

3. The method for monitoring and preventing falls in dementia patients according to claim 2, characterized in that, Human region detection is performed on all video frames within a video frame sequence, and all the obtained human region images are combined into a human region image sequence. The specific steps include the following: For each pixel in a video frame, perform the following steps: obtain the pixel value of the pixel in all video frames, and fit a Gaussian distribution model corresponding to the pixel for all pixel values. The Gaussian distribution model includes the observation mean and observation variance. For each video frame in the video frame sequence, the following steps are performed: Each pixel in the video frame is traversed, and the pixel value is matched against its corresponding Gaussian distribution model. The matching method is to compare the absolute value of the difference between the pixel value and the observed mean with the product of the observed variance and the empirical coefficient. If the absolute value of the difference between the pixel value and the observed mean is less than the product of the observed variance and the empirical coefficient, the match is considered successful. If the pixel value matches the corresponding Gaussian distribution model, the pixel is marked as a background pixel. If the pixel value does not match the corresponding Gaussian distribution model, the pixel is marked as a non-background pixel. Then, connected component analysis is performed on all non-background pixels in the video frame, and the rectangular image corresponding to the largest connected component is marked as the human body region image. The pixel values ​​of pixels in the human body region image that do not belong to the largest connected component are set to 0. All the obtained human body region images are then combined into a human body region image sequence.

4. The method for monitoring and preventing falls in dementia patients according to claim 3, characterized in that, Based on the annotation information and skeleton topology map corresponding to human body region images, key motion nodes are tracked, and motion parameter data of key motion nodes corresponding to dementia patients are constructed. The specific steps include the following: Perform a union operation on the labels corresponding to all skeleton topology diagrams to obtain the label itemset; Iterate through the annotation labels in the annotation label set. For each annotation label, perform the following: Iterate through the human region images. Based on the annotation label, search the annotation information corresponding to the current human region image. If a matching annotation label exists in the annotation information corresponding to the current human region image, select the region center coordinates in the annotation information of the human region image as the positioning coordinates, and annotate the current human region image using the timestamp corresponding to the human region image sequence. If no matching annotation label exists in the annotation information corresponding to the current human region image, search the annotation information corresponding to the adjacent human region images based on the annotation label. Record the human region images with matching annotation labels as the positioning human region images. Based on the motion key nodes corresponding to the annotation labels, search the skeleton topology graph corresponding to the positioning human region image. If a motion key node corresponding to the annotation label is found in the skeleton topology graph corresponding to the positioning human region image... The labels corresponding to adjacent motion key nodes exist in the annotation information of the current human region image. These motion key nodes are designated as localization motion key nodes. The difference between the region center coordinates of the localization motion key node in the annotation information of the current human region image and the region center coordinates in the annotation information of the localization human region image is recorded as the fitting difference. The region center coordinates in the annotation information of the localization human region image are then used as calibration coordinates. The sum of the calibration coordinates and the fitting difference is recorded as the candidate fitting coordinates. All candidate fitting coordinates are summed and averaged to obtain the fitted coordinates, which are then labeled using the timestamp corresponding to the current human region image in the human region image sequence. The localization coordinates or fitted coordinates are arranged and concatenated according to their sequence number in the human region image sequence to form a label data sequence set. The label data sequence set includes the label coordinates arranged according to their sequence number in the human region image sequence. For each set of labeled data sequences, the distance between adjacent labeled coordinates is recorded as the displacement, and the displacement is labeled using the timestamp corresponding to the latter of the adjacent labeled coordinates. All displacements are arranged and concatenated according to the timestamps to form a displacement vector. The interval between the labeled coordinate and the timestamp corresponding to the first labeled coordinate in the labeled data sequence is recorded as the interval time corresponding to the labeled coordinate. With the interval time as the independent variable and the displacement as the dependent variable, a curve function is fitted. After performing the first derivative, the data corresponding to the interval time is used as the velocity. After performing the second derivative, the data corresponding to the interval time is used as the acceleration. All velocities are arranged and concatenated according to the interval time to form a velocity vector, and all accelerations are arranged and concatenated according to the interval time to form an acceleration vector. The motion displacement vector, motion velocity vector, and motion acceleration vector corresponding to the labeled data sequence set are concatenated from top to bottom to form local motion parameter data. Then, the local motion parameter data corresponding to all the labeled data sequence sets are combined to form overall motion parameter data. Overall motion parameter data is actually a matrix data. Finally, the transpose of overall motion parameter data is recorded as motion parameter data.

5. The method for monitoring and preventing falls in dementia patients according to claim 4, characterized in that, The motion control constraint space is determined based on motion parameter data, and the effective degrees of freedom for motion control are determined within the motion control constraint space based on the motion parameter data. Specifically, this includes the following: Construct the covariance matrix corresponding to the motion parameter data, and perform eigenvalue decomposition on the covariance matrix corresponding to the motion parameter data. The ratio of the largest eigenvalue to the sum of all eigenvalues ​​is denoted as the consistency of change. Each row of the motion parameter data is denoted as a motion parameter vector. The product of the motion parameter vector and the motion control constraint matrix is ​​denoted as the fit variation consistency. The mean square error between all fit variation consistency and variation consistency is calculated and denoted as the mapping error. The direction of the mapping error is minimized. The motion control constraint matrix is ​​iteratively optimized by the least squares method until the termination condition is met. The motion control constraint matrix is ​​output. The transpose of the motion control constraint matrix is ​​the motion control constraint space. Singular value decomposition is performed on the motion control constraint space, and the number of non-zero singular values ​​is the effective degree of freedom of motion control.

6. The method for monitoring and preventing falls in dementia patients according to claim 5, characterized in that, Training the network for key component annotation includes the following methods: Obtain several key part annotation samples, and annotate the key part annotation samples with annotation information. Combine all the annotated key part annotation samples into a key part annotation set, and train the key part annotation network using the key part annotation set.

7. A fall prevention monitoring system for dementia patients, characterized in that, The system employs a fall prevention monitoring method for dementia patients as described in any one of claims 1-6, comprising: The dementia patient monitoring module is used to sample video data based on a preset time length to obtain a video frame sequence; The motion parameter data construction module is used to process video frame sequences based on skeleton topology and temporal constraints to obtain motion parameter data of key motion nodes corresponding to dementia patients. The motion control analysis module is used to determine the motion control constraint space based on motion parameter data, and to determine the effective degrees of freedom of motion control in the motion control constraint space based on the motion parameter data. The fall prevention intervention module compares the effective degrees of freedom of motor control with the instability threshold. If the effective degrees of freedom of motor control is less than the instability threshold, intervention measures are taken for the corresponding dementia patient. If the effective degrees of freedom of motor control is not less than the instability threshold, no operation is performed.

Images