An elderly posture early warning method and system based on visual detection

Abstract

This invention relates to the field of visual detection technology and discloses a method and system for early warning of elderly posture based on visual detection. The method includes: acquiring a continuous image sequence through a camera and performing posture estimation to obtain the coordinates of key parts and movement trajectories; calculating a dynamic behavior feature vector based on this; extracting periodic features and change patterns of the behavior; comparing the obtained patterns with a behavior database; generating potential abnormal signals when the change frequency exceeds a threshold; fusing historical behavior data to obtain an abnormality confidence level; and classifying early warning levels and triggering alarms based on the confidence level. This method can accurately identify abnormal postures, improve monitoring accuracy and real-time performance, and quickly provide early warnings after an elderly person experiences an abnormal situation.

Description

Technical Field

[0001] This invention relates to the field of visual detection technology, and in particular to a method and system for early warning of elderly posture based on visual detection. Background Technology

[0002] It is particularly important to learn how to use image recognition to monitor and provide early warnings of abnormal behaviors in the elderly in real time.

[0003] One existing technology relies on low-cost devices such as distributed infrared motion sensors, door magnetic sensors, pressure pad sensors, and sound sensors deployed within living spaces. It monitors environmental changes such as human movement trajectories, dwell time in specific areas, door and window opening / closing status, bed and chair usage, and abnormal sounds. These changes are combined with preset temporal rules or basic pattern matching, such as prolonged absence from the bathroom, extended periods away from bed at night, or unusual stillness in the kitchen, to infer potential anomalies. This method has significant drawbacks: first, the recognition granularity is coarse, unable to distinguish between specific postures and intentions, leading to a high false alarm rate (e.g., both falling and sitting trigger a stillness alarm); second, it is sensitive to environmental interference, easily affected by pet activity, moving objects, or everyday noise; and third, the behavior is highly ambiguous, with the same sensor data potentially corresponding to multiple normal or abnormal scenarios, lacking contextual understanding capabilities.

[0004] In summary, existing technologies have the problem of difficulty in identifying interference caused by obstacles or the movement of other objects, leading to false alarms and missed alarms in elderly posture warnings. Summary of the Invention

[0005] This invention provides a visual detection-based method and system for early warning of elderly postures, enabling the monitoring of elderly activities and effectively providing early warning of abnormal behaviors.

[0006] In a first aspect, to address the aforementioned technical problems, the present invention provides a visual detection-based method for early warning of elderly postures, comprising: By capturing continuous images of elderly people's activities through a camera, a continuous image sequence is obtained. The posture of the continuous image sequence is estimated to obtain the coordinate data of key parts and the movement trajectory information. Based on the coordinate data of the key parts and the motion trajectory information, the posture vector is calculated to obtain the dynamic behavior feature vector; Based on the dynamic behavior feature vector, periodic features and change patterns are extracted to obtain behavior pattern features; By comparing the behavioral pattern features with a preset behavioral database, when the frequency of behavioral changes in the behavioral pattern features exceeds a preset threshold for behavioral change frequency, a potential abnormal signal is obtained. The potential abnormal signals are fused with historical behavior data in context to obtain the confidence level of abnormal behavior. When the confidence level of abnormal behavior is higher than the preset confidence level threshold, anomaly detection analysis is performed to obtain a warning level classification. Based on the aforementioned warning level classification, an alarm mechanism is triggered to send specific notification content to the device, completing the real-time response process.

[0007] Secondly, the present invention provides a posture warning system for the elderly based on visual detection, comprising: The data acquisition module is used to acquire continuous images of elderly people's activities through a camera, obtain a continuous image sequence, perform posture estimation on the continuous image sequence, and obtain key part coordinate data and movement trajectory information; The trajectory calculation module is used to calculate the posture vector based on the coordinate data of the key parts and the motion trajectory information to obtain the dynamic behavior feature vector. The behavior confirmation module is used to extract periodic features and change patterns based on the dynamic behavior feature vector to obtain behavior pattern features; An anomaly detection module is used to compare the behavior pattern features with a preset behavior database. When the frequency of behavior change of the behavior pattern features exceeds a preset behavior change frequency threshold, a potential anomaly signal is obtained. The historical integration module is used to perform contextual fusion of the potential abnormal signals with historical behavior data to obtain the confidence level of abnormal behavior. The early warning classification module is used to perform anomaly detection and analysis and obtain an early warning level classification when the confidence level of abnormal behavior is higher than a preset confidence threshold. The real-time early warning module is used to classify the early warning levels, trigger the alarm mechanism, send specific notification content to the device, and complete the real-time response process.

[0008] Thirdly, the present invention also provides an electronic device including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor, when executing the computer program, implements the visual detection-based posture warning method for the elderly as described in any one of the above.

[0009] Fourthly, the present invention also provides a computer-readable storage medium comprising a stored computer program, wherein, when the computer program is executed, it controls the device on which the computer-readable storage medium is located to execute the visual detection-based posture warning method for the elderly as described above.

[0010] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention acquires continuous image sequences of the elderly person's activity area using a camera. After denoising and illumination correction preprocessing, a pose estimation algorithm is used to extract coordinate data of key human body parts. Movement trajectory information is generated through trajectory tracking and smoothing. When the frequency of detected behavioral changes exceeds a threshold, a potential abnormal signal is generated. Contextual fusion analysis is performed using historical behavioral data to calculate the confidence level of abnormal behavior. Based on the confidence level and duration, a warning level is determined, ultimately triggering a corresponding alarm mechanism to send a notification to the target device. This fusion and utilization of historical information can effectively distinguish between individual daily behavioral habits and true abnormal states, thereby significantly reducing the false alarm rate and improving the accuracy of anomaly identification.

[0011] (2) This invention extracts key point coordinates from image sequences based on a pose estimation algorithm and obtains continuous action information through trajectory tracking. When the real-time behavior pattern deviates significantly from the normal pattern in the database, it indicates a possible abnormal state. By introducing historical behavior data for contextual fusion, it effectively distinguishes between individual behavioral habitual differences and genuine anomalies. Finally, based on multi-dimensional evaluation of signal strength and duration, a graded early warning mechanism is established. This introduction of historical data for contextual fusion can adapt to different behavior patterns and improve the specificity of recognition.

[0012] (3) This invention, through non-contact monitoring of elderly behavior and multi-level algorithm fusion, can continuously optimize individual behavior models through historical data analysis, thereby reducing false alarm rates. This combination of posture estimation, temporal analysis, and context fusion accurately identifies abnormal behaviors and enables real-time responses, effectively improving the safety and reliability of elderly activity monitoring.

[0013] (4) This invention constructs a complete multi-level fusion detection framework, which connects real-time image processing, high-precision pose estimation, temporal behavior modeling, historical data comparison, and dynamic early warning response into an adaptive closed-loop system, ensuring seamless flow from visual detection to nursing decision-making, and deep collaboration among modules rather than isolated operation. At the same time, the system can continuously learn and optimize the behavior model of each individual, dynamically adjust the recognition threshold, and make the monitoring capability increasingly match the user's real state as the usage time increases, thus realizing truly personalized intelligent service. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of the visual detection-based posture early warning method for the elderly provided in the first embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of the visual detection-based posture warning system for the elderly provided in the second embodiment of the present invention. Detailed Implementation

[0015] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0016] Reference Figure 1 The first embodiment of the present invention provides a method for early warning of elderly posture based on visual detection, including the following steps: S11, continuously capture images of the elderly person's activities through a camera to obtain a continuous image sequence, perform posture estimation on the continuous image sequence to obtain key part coordinate data and movement trajectory information; S12, calculate the posture vector based on the coordinate data of the key parts and the motion trajectory information to obtain the dynamic behavior feature vector; S13, extract periodic features and change patterns based on the dynamic behavior feature vector to obtain behavior pattern features; S14, compare the behavior pattern features with a preset behavior database, and when the behavior change frequency of the behavior pattern features exceeds a preset behavior change frequency threshold, a potential abnormal signal is obtained; S15, perform contextual fusion of the potential abnormal signals with historical behavior data to obtain the confidence level of abnormal behavior; S16. When the confidence level of abnormal behavior is higher than the preset confidence level threshold, anomaly detection analysis is performed to obtain the warning level classification. S17, based on the aforementioned warning level classification, trigger the alarm mechanism, send specific notification content to the device, and complete the real-time response process.

[0017] In step S11, the continuous image sequence of the elderly person's activities is obtained by acquiring continuous image sequences through a camera. The pose of the continuous image sequence is then estimated to obtain key body coordinate data and motion trajectory information, including: Image data of elderly people's activities are collected by camera equipment to obtain an initial image sequence. The initial image sequence is then denoised and illumination corrected to obtain a continuous image sequence. Based on the continuous image sequence, key component coordinate data and key component location information are extracted to obtain a preliminary coordinate set; The coordinate data of the key parts in the preliminary coordinate set are tracked to obtain position change information. When the change amplitude of the position change information exceeds a preset change amplitude threshold, the position change information is smoothed to obtain smooth trajectory information. A dynamic trajectory diagram is generated based on the smooth trajectory information, and the coordinate data of key parts and the motion trajectory information are obtained.

[0018] First, image data of the elderly's activity areas is continuously collected using camera equipment. Specifically, high-definition cameras are installed in the activity room of a nursing home to cover the areas where the elderly spend their daily time. The cameras capture images at a frequency of 30 frames per second, forming a continuous image sequence. These images may be affected by changes in ambient light or equipment noise, therefore preprocessing is required. During image preprocessing, mean filtering is used to remove noise from the images, reducing noise caused by equipment vibration or dust. Simultaneously, histogram equalization is used for illumination correction to ensure image clarity under different lighting conditions, resulting in a continuous image sequence.

[0019] Among them, histogram equalization technology for illumination correction can adjust the gray level of each gray level in the gray-scale histogram. Number of pixels Divide by the total number of pixels in the image, M×N, to obtain the normalized probability density for each gray level. Then, starting from the minimum gray level, the probability density of each gray level is sequentially accumulated, that is, for the current gray level... Its cumulative distribution function value Equal to all less than or equal to The sum of the gray level probability densities is expressed as: ; The current gray level, with a value range of [value range missing]. ; A normalized grayscale histogram, i.e., grayscale levels The probability of occurrence, calculation formula , grayscale level is The number of pixels, This represents the total number of pixels.

[0020] Then, a normalization mapping is performed to convert the cumulative distribution function value... Multiply by the maximum grayscale value allowed in the image To obtain the initial rounded mapping value Finally, pixel value replacement is performed, iterating through each pixel of the original image and directly replacing its original grayscale value with the corresponding calculated value. This completes a deterministic mapping from the original grayscale distribution to the new grayscale distribution. The processed image sequence can provide a more reliable data foundation for subsequent analysis and significantly improve the accuracy of pose estimation.

[0021] Then, for pose estimation of continuous image sequences, a pre-trained model can be used to extract the coordinate data of key parts of the elderly body. For example, it has been fully trained using human pose training datasets such as COCO and MPII, and can detect human key points from images with various poses, clothing and backgrounds, focusing on the head, shoulders, elbows, knees and other positions, and outputs the two-dimensional coordinates of these parts in each frame.

[0022] For each frame in a continuous image sequence, the image is read into memory and converted into the input format required by the model. For example, the image size is adjusted to 384x384 pixels using bilinear interpolation, and the pixel values ​​are normalized to the floating-point range of [0,1] by dividing by 255.0. The floating-point coordinates of each target pixel in the original image are calculated, and the four neighboring pixels around the coordinate are taken. The decimal part of the coordinate is used as the weight, which is (u,v). Then, for horizontal interpolation, the weights of the upper pixels are (1-u) and u, and the weights of the lower pixels are (1-u) and u. For vertical interpolation, the weights (1-v) and v are used to perform a weighted average of the horizontal interpolation results to obtain the final pixel value. The values ​​are then standardized by subtracting the mean [0.485, 0.456, 0.406] from the ImageNet dataset statistics and dividing by the standard deviation [0.229, 0.224, 0.225], and then converted into floating-point tensors in CHW format.

[0023] The preprocessed tensor is input into a pre-trained model, which extracts feature maps through convolutional neural network layers. The initial convolutional layer uses a stride of 2 to reduce spatial resolution, while subsequent convolutional layers use a stride of 1 for finer feature extraction. The keypoint detection head outputs a heatmap and offset for each keypoint. The argmax operation is used to parse the 2D coordinates of the keypoints from the heatmap, and non-maximum suppression is applied to remove duplicate detections. For each frame, the coordinates of keypoints such as the head, shoulders, elbows, and knees are extracted (e.g., head coordinates (200, 300), and their confidence scores are recorded. Through frame-by-frame analysis, a preliminary coordinate set is formed. The advantage of this method is that it can capture the body posture of elderly people in real time, laying the foundation for subsequent trajectory tracking, while reducing misjudgments caused by image blur.

[0024] Next, a time-series analysis of the trajectory tracking of key parts is performed based on the preliminary coordinate set, focusing primarily on the displacement changes of key parts between consecutive frames. When the knee coordinates are (150, 400) in the first frame and (160, 410) in the second frame, the displacement change is calculated based on the Euclidean distance formula, i.e., for the coordinates of the same key part in two consecutive frames... and displacement change The calculated displacement change was approximately 14 pixels. Since the preset displacement change threshold, obtained by fitting historical abnormal behavior displacement change data, is within 20 pixels, this indicates the change is within the normal range. However, when a frame suddenly jumps to (200, 450), the calculated displacement change is 70.7 pixels, exceeding the preset displacement change threshold. This is considered abnormal data and requires smoothing through interpolation using subsequent coordinate values ​​to generate more continuous motion trajectory information, resulting in smoothed trajectory information. This smoothing process effectively avoids trajectory breaks caused by brief occlusions or recognition errors, improving data reliability.

[0025] Finally, when generating dynamic trajectory maps, timestamp data can be incorporated to map the smoothed trajectory information onto a timeline. For example, if an elderly person walks from one corner of an activity room to another within 10 minutes, the trajectory map can visually display the movement path of their head and knees, reflecting walking speed and stability. This information can be integrated into feature data, such as an average walking speed of 0.5 meters per second. Specifically, the movement paths of each part are drawn in chronological order using lines of different colors, and the direction of time progression is reflected through gradient colors or changes in point density. An animation engine is used to render the data frame by frame, connecting historical trajectory points in each frame to form a continuous path. Key points at the current moment are highlighted, and real-time displacement data can be overlaid to enhance the visualization effect, ultimately yielding the coordinate data of key parts and the movement trajectory information. Generating such dynamic trajectory maps helps nursing home staff monitor the activity status of the elderly in real time, promptly detect abnormal behaviors such as the risk of falls, and thus improve safety.

[0026] In step S12, the step of calculating the posture vector based on the coordinate data of the key parts and the motion trajectory information to obtain the dynamic behavior feature vector includes: Spatial location information is extracted from the coordinate data of the key parts and the changes in the joint parts are analyzed to obtain a preliminary joint dataset. By analyzing the displacement velocity using the preliminary joint dataset and the temporal information of the motion trajectory information, the continuously changing displacement values ​​are obtained. When the value of the continuous displacement change exceeds the preset threshold for continuous displacement change, the abnormal displacement value is smoothed and corrected to obtain the adjusted displacement data. Based on the adjusted displacement data and the preliminary joint dataset, a comprehensive body posture vector is calculated to obtain a dynamic behavior feature vector.

[0027] First, based on the key component coordinate data and motion trajectory information, when analyzing the spatial position information of joints, a data comparison tool can be used to compare the key point coordinates in each frame of the image frame by frame to obtain a preliminary dataset of joint angles. For example, the pixel coordinates of the shoulder, elbow, and wrist can be obtained from the frame-by-frame coordinate data. For the first frame, the shoulder coordinates are S(A1,A2), the elbow coordinates are E(B1,B2), and the wrist coordinates are W(C1,C2), then the vector... The vector pointing from the shoulder to the elbow is represented as (B1-A1, B2-A2). The distance from the elbow to the wrist is represented as (C1-B1, C2-B2). Elbow joint angle. Calculated using the vector dot product formula: ; The results are usually converted to degrees, and for the second frame, the angle is calculated using the corresponding coordinates. Angle change for and The difference in angle is used to represent the change in elbow position. Positive values ​​indicate increased bending, while negative values ​​indicate increased extension. In a monitoring scenario in a nursing home activity room, a camera captures an elderly person's elbow coordinates as (100, 200) in one frame and (110, 210) in the next frame. The shoulder coordinates change from (90, 180) to (100, 190), and the wrist coordinates change from (115, 215) to (120, 225). By calculating the angle changes, the initial elbow joint angle change is determined to be 10 degrees. This frame-by-frame analysis helps construct a dataset containing multiple joint angle changes, providing a foundation for subsequent behavioral feature analysis.

[0028] Subsequently, trajectory analysis is performed on the preliminary joint dataset to calculate displacement velocity. Based on the temporal changes in the movement trajectory, the positional movement of the knee joint in consecutive frames can be observed. For example, if the knee joint's coordinates are (150, 400) in the first frame and (155, 405) in the second frame, the displacement is approximately 7 pixels. Considering a time interval of 1 / 30 of a second, and assuming a 1-meter distance in an activity room corresponds to 1000 pixels in the image, the preliminary estimated displacement velocity is approximately 0.2 meters per second. In this way, continuously changing displacement values ​​can be obtained, reflecting the movement status of the elderly.

[0029] Next, in one frame, the speed suddenly reaches 2 meters per second. This continuous change in displacement exceeds the preset threshold for continuous displacement change derived from historical walking speeds of elderly individuals. If the normal walking speed range is 0.1 to 0.8 meters per second, this data can be identified as abnormal. In this case, a smoothing correction is performed on the abnormal value. This can be achieved by averaging the values ​​of the preceding and following frames, adjusting the speed to 0.5 meters per second. The corrected data needs further evaluation to determine if it conforms to the dynamic behavioral characteristics range, such as whether it closely approximates the daily walking speed characteristics of elderly individuals. This processing method helps ensure the reasonableness of the data.

[0030] Finally, when integrating the adjusted displacement velocity data with the preliminary joint angle dataset, a comprehensive vector describing body posture is generated through data fusion. For example, information such as an elbow joint angle change of 10 degrees and a knee joint displacement velocity of 0.5 meters per second is integrated into a multi-dimensional vector containing data in both angle and velocity dimensions, i.e., a dynamic behavior feature vector. This comprehensive vector can fully reflect the dynamic behavior characteristics of the elderly, facilitating subsequent analysis of their activity status.

[0031] In step S13, the step of extracting periodic features and change patterns based on the dynamic behavior feature vector to obtain behavior pattern features includes: The behavior intervals and change frequencies within a complete behavior cycle are obtained from the dynamic behavior feature vector. Based on the behavior interval and the change frequency, a dynamic trend analysis is performed on the dynamic behavior feature vector to obtain the dynamic trend change; When the dynamic trend change exceeds the preset dynamic trend change threshold, the dynamic behavior feature vector is smoothed and corrected to obtain the adjusted behavior sequence. Behavioral pattern matching is performed based on the adjusted behavioral sequence to obtain behavioral pattern features.

[0032] First, in the monitoring scenario of the nursing home activity room, the dynamic behavioral feature vector is processed by temporal segmentation to obtain the behavioral intervals and frequency of change within a complete behavioral cycle. The principle of temporal segmentation is to divide a continuous behavioral data segment into multiple smaller segments according to time or action characteristics, in order to analyze each behavioral cycle in more detail. For example, in a 5-minute monitoring data segment, two main behaviors of elderly people are identified: walking and sitting. Temporal segmentation divides the data into walking and sitting segments. In the walking segment, the behavioral interval might be one step every 2 seconds, while the frequency of change reflects the switching between fast and slow steps. This segmentation process helps to capture behavioral features more accurately.

[0033] Subsequently, fluctuation detection is performed on the behavioral intervals and frequency of change to obtain the main dynamic trend changes. The principle of fluctuation detection is to determine the stability of behavior by analyzing the fluctuations of behavioral data on the time axis. For example, in a walking segment, if the elderly person's step interval suddenly changes from 2 seconds to 1 second, fluctuation detection will record the increase in the frequency of this change and analyze whether the fluctuation amplitude is abnormal in combination with the time dimension to obtain the dynamic trend change.

[0034] Next, when the dynamic trend changes exceed a preset threshold based on the deviation of the median and absolute median of historical behavioral data marked as normal activity over a 30-day period—for example, if historical data shows normal step interval fluctuations within 0.5 seconds, but actual fluctuations reach 1 second—it will be marked as a potential anomaly. This analytical approach helps to promptly identify irregular changes in behavior. For instance, when the fluctuation amplitude and temporal pattern exceed the historical maximum fluctuation amplitude or historical temporal pattern threshold, the behavioral sequence can be corrected through smoothing. The purpose of smoothing is to reduce abrupt changes in the data and ensure the continuity of the sequence. In the walking segment mentioned above, if there is an abnormal jump in the step interval data, smoothing will adjust the data based on the average of the preceding and following time points, making the behavioral sequence more consistent with actual walking patterns, resulting in an adjusted behavioral sequence. The continuity and trajectory consistency of the adjusted sequence better reflect the true behavioral state of the elderly, providing a reliable basis for subsequent analysis.

[0035] Finally, behavior pattern matching is performed based on the adjusted behavior sequence, comparing the behavior sequence with the predetermined pattern to determine whether the behavior switching point meets expectations. The principle of pattern matching is to compare the current behavior data with known standard behavior templates. Behavior pattern templates mainly come from publicly available human activity datasets, such as Kinetics, UCF101, which includes actions such as walking, sitting, standing, and falling, or MobiAct, which is specifically designed for the behavior of the elderly. These datasets provide a large number of standardized behavior sequences through motion capture or crowdsourced annotation, as well as local historical data collected in nursing home scenarios and annotated by experts, forming a behavior pattern library that conforms to specific environments.

[0036] For example, if an elderly person suddenly sits down after walking, the system will check whether this transition point conforms to the common "walk-sit" pattern. Specifically, this is based on time-domain statistics included in the dynamic behavior feature vector, such as the vertical displacement rate of the centroid and the rate of change of the knee flexion angle. When matching patterns, a multidimensional Euclidean distance similarity is calculated; the formula for multidimensional weighted Euclidean distance similarity is... , and are the mean and standard deviation of the i-th feature in the historical data, respectively, used for standardization; Let be the weight coefficient for this dimension, satisfying The weights are derived by averaging the weights of relevant dimensions based on historical pattern matching data. For example, historical pattern matching shows a weight of 0.35 for the vertical displacement rate of the centroid, 0.2 for the rate of change of the knee flexion angle, 0.25 for the horizontal movement speed of the hip joint, and 0.2 for the standard deviation of the trunk tilt angle. When the similarity score between two dimensions exceeds the minimum threshold based on the similarity score of historical pattern matching, the behavior switching point is considered valid. If the match is successful, the corresponding behavior pattern features will be generated. This comparative analysis helps to accurately identify behavior types and provide nursing home staff with clear behavior reports.

[0037] In step S14, the behavior pattern features are compared with a preset behavior database. When the frequency of behavior changes in the behavior pattern features exceeds a preset behavior change frequency threshold, a potential abnormal signal is obtained, including: The feature distribution is obtained from the behavioral pattern features to determine the range of behavioral change frequency. The behavioral deviation is obtained by comparing the frequency range of behavioral changes with a preset behavioral database and calculating the deviation. When the behavioral deviation exceeds a preset behavioral deviation threshold, the behavioral pattern features are classified and labeled according to the specific deviation to obtain an abnormal signal classification label. A secondary screening analysis is performed based on the abnormal signal classification labels to identify potential abnormal signals.

[0038] First, time-series segment processing is performed based on behavioral pattern characteristics to obtain the distribution of dynamic behavioral features. The processing groupes behavioral data according to time or feature dimensions and analyzes the behavioral change patterns within each segment. For example, in a 10-minute monitoring data segment, the elderly person's behavior is divided into three segments: walking, standing, and sitting. The walking segment lasts 4 minutes, the standing segment lasts 2 minutes, and the sitting segment lasts 4 minutes. By analyzing the characteristic distribution of each segment, the fluctuation range of behavioral change frequency is determined; for example, the step interval in the walking segment fluctuates between 1 and 3 seconds.

[0039] Subsequently, the fluctuation range of behavioral changes was compared with a pre-defined database of normal behaviors. This database contains data on various behavioral categories of elderly individuals during daily activities, such as walking, standing, and sitting, covering characteristic parameters and statistical indicators (including minimum, maximum, mean, and standard deviation) such as stride intervals, duration, and frequency fluctuation range. This database is built upon a large amount of historical normal behavior data, forming a typical fluctuation range benchmark to support accurate comparison. The deviation between the current data and the database is calculated; for example, if the fluctuation range of the current walking segment is 1 to 3 seconds, the deviation from the typical range (0.5 to 2.5 seconds) in the database is 0.5 seconds.

[0040] When a behavioral deviation exceeds a preset threshold of 0.3 seconds by 0.5 seconds, the behavioral pattern is flagged. The preset threshold of 0.3 seconds is determined based on the statistical distribution of the fluctuation range of step intervals in historical normal walking data; for example, it can be set based on the standard deviation of this fluctuation range. Anomalies are initially classified according to the degree of behavioral deviation to determine the behavioral risk level. For example, a deviation exceeding the threshold by 0.3 seconds but not exceeding 0.6 seconds is classified as low risk; exceeding 0.6 seconds but not exceeding 1.0 second is classified as medium risk; and exceeding 1.0 second is classified as high risk. The preliminary classification results are summarized to provide a basis for subsequent analysis.

[0041] Finally, a secondary screening analysis is performed on the preliminary classification results, and the final confirmation result of potential abnormal signals is determined by combining contextual data. The verification process incorporates more contextual data, such as the elderly's walking data over the past week, to analyze whether the current abnormality is an isolated event. The specific process includes: calculating the statistical benchmarks (such as moving average, standard deviation, and normal fluctuation range) of behavioral characteristics such as step intervals within the period, and comparing the feature values ​​of the current abnormal segment with historical benchmarks. If the feature values ​​exceed the historical normal range, the continuity of adjacent behavioral segments before and after the segment is further analyzed to rule out equipment interference or temporary external influences, and the frequency of similar abnormalities within the same period is checked. If the abnormality is an isolated event and there are no consecutive abnormalities in adjacent behavioral segments, its behavioral risk level is adjusted to low risk; if the number of times the abnormal feature occurs within the same period exceeds a preset frequency threshold, or if time series analysis reveals a clear trend change (for example, using linear regression analysis to determine that the abnormal indicator shows an upward trend over time), it is confirmed as a potential abnormal signal, and the final confirmation result is generated and the occurrence pattern, timestamp, and associated contextual data are recorded. This secondary screening effectively reduces misjudgments and improves the reliability of abnormality identification.

[0042] In step S15, the step of contextually fusing the potential abnormal signals with historical behavior data to obtain the confidence level of abnormal behavior includes: Obtain the time series distribution data of the potential abnormal signals, perform preliminary matching of abnormal behaviors based on the time series distribution data, and obtain the pattern comparison results; The correlation information between the pattern comparison results and historical pattern comparison results is integrated to obtain consistent data; When the consistency data deviates from the preset consistency data threshold range, the pattern comparison result is deeply verified to obtain the behavior deviation situation. Based on the behavior deviation situation, a classification label is obtained to obtain a corrected classification label. The confidence level of the behavioral feature distribution is quantified based on the modified classification label to obtain the confidence level of abnormal behavior.

[0043] First, time-series distribution data corresponding to potential abnormal signals are extracted from historical behavioral records. For example, if an elderly person's walking data from the past month is recorded, and an abnormal signal is identified in a certain segment of the walking time series, such as a sudden increase in step intervals, a preliminary comparison and matching is performed. The distribution characteristics of this data are compared with typical patterns in historical records to obtain initial results of behavioral pattern comparison. For example, the walking data time series segment corresponding to the suddenly increased step intervals of the currently identified potential abnormal signal is used for feature calculation to obtain distribution feature values ​​such as mean and standard deviation. For each dimension of the feature vector x to be calculated, its standardized value is... The calculation formula is ,in and These are the mean and standard deviation of the numerical set formed by the i-th statistical feature itself. Then, the current standardized feature vector and mean vector, derived from the mean and standard deviation of the stride interval, are calculated using Mahalanobis distance. If the result exceeds the 95% confidence level based on the chi-square distribution, it indicates that the current data has deviated significantly in a statistical sense.

[0044] For example, the time-series database records show that its historical step intervals are stable between 0.9 and 1.1 seconds, with a mean of 1.0 second and a standard deviation of 0.1 seconds. In a walking sequence this week, the step interval suddenly increased to 1.8 seconds, lasting for 10 minutes. After feature extraction, the current segment's distribution features are a mean of 1.5 seconds and a standard deviation of 0.3 seconds. The standardized feature vector is Z=[5,12], which is significantly different from the historical baseline. The deviation of the feature vector from the historical distribution parameters was calculated using Mahalanobis distance. The covariance matrix, based on data from the past 90 days, yielded a Mahalanobis distance of 25.1. This value exceeds the 95% confidence threshold for a chi-square distribution with 2 degrees of freedom, indicating that the current extended step interval is statistically significantly deviating from the typical pattern.

[0045] The typical behavioral pattern data mainly comes from sensor data collected over a long period of time by wearable devices, daily monitoring records of nursing homes, and clinical research datasets. It is stored in a time series database and records behavior types, timestamps, feature values, and statistical indicators in a structured table format, supporting rapid querying and comparative analysis.

[0046] Subsequently, the pattern comparison results are integrated with contextual data and related information from historical behavior records. For example, the behavior of the elderly at other times of the day, or environmental factors, are used to determine consistency data. Specifically, the initial results show that the step interval has increased to 5 seconds, while the normal range in historical records is 2 to 3 seconds. Further analysis is needed. If there are contextual factors such as fatigue or medication on that day, and if historical data shows factors affecting the step interval such as medication use, labels are generated for walking after medication or meals on that day. All walking data within 30 minutes of medication use are statistically analyzed from historical records, and the distribution of their step intervals is calculated. If the mean increases by 0.4 seconds and the standard deviation increases by 0.2 seconds, then a threshold of approximately 0.4 seconds needs to be added to the normal step range to obtain consistent data.

[0047] Next, when the consistency data deviates from the preset threshold range obtained based on historical behavior—for example, if the step interval still exceeds the corrected deviation threshold of 2 seconds after threshold correction—a secondary filtering mechanism is triggered for in-depth verification. This secondary filtering mechanism delves into long-term trends in historical data, such as whether similar deviations have occurred in the past week, and combines this with the persistence characteristics of the current data to ultimately generate corrected classification labels for the abnormal signals.

[0048] The process involves querying the historical database using the current anomaly type as the key to extract all anomaly events with the same context label within the past period. The frequency, average deviation, and duration of these events are then obtained, and their statistical distribution is constructed. For example, if the "prolonged step interval after medication" event occurred three times in the past seven days, with an average extension of 2.8 seconds and lasting approximately 10 minutes, the current event with a deviation of 5 seconds and a duration of 15 minutes will be compared to this distribution and subjected to persistence analysis. Complete time-series data within a continuous window before and after the current anomaly point are extracted, and simple linear fitting is used to determine whether the anomaly is a transient pulse or a trend. The linear fitting formula is as follows: , ,in Let i be the time independent variable corresponding to the i-th observation data point. The observed value of the dependent variable corresponding to the i-th observation data point The total number of observation data points participating in the fitting. for The arithmetic mean, Observation sequence The arithmetic mean, The intercept parameter is the line segment of the fitted line.

[0049] A continuous time series subset is extracted before and after the detected outlier. Least squares linear regression is applied independently to both subsets to fit two straight lines representing local trends. The slope parameters β1 and β2 are extracted for each. The nature of the outlier is inferred by analyzing the sign and relative magnitude of these two slopes. If β1 and β2 have opposite signs and their absolute values ​​both exceed an empirical threshold set based on historical fluctuations, it indicates a sharp reversal in direction before and after the outlier, classifying this as a random anomaly. If β1 and β2 have the same sign or their absolute values ​​are both low, it indicates that the data maintained a consistent direction or a stable state before and after the outlier, thus escalating to a "risk anomaly."

[0050] The empirical threshold calculation originates from a large number of time series segments labeled "normal" or "no significant abnormality." For each segment, the same piecewise linear fitting method is applied between any two consecutive points to obtain a large sample library of "slope pairs (β1, β2) under normal conditions." Subsequently, the distribution of the difference in absolute values ​​of the slopes |β1-β2| and the frequency of slope sign changes are calculated for these samples. The fitting method for β is consistent throughout, so it will not be elaborated further.

[0051] Finally, based on the revised classification labels, the distribution of behavioral characteristics is quantified using confidence scores to determine the confidence scores of abnormal behaviors. The confidence scoring principle is based on a comprehensive evaluation of multiple dimensions of behavioral characteristics, such as duration and deviation magnitude. Five core feature values ​​are extracted from the revised events and normalized, such as deviation magnitude feature value F1, duration feature value F2, and historical frequency feature value F3.

[0052] Specifically, F1 is calculated by dividing the absolute difference between the current observation and the personalized dynamic baseline by the standard deviation of historical normal fluctuations; F2 is calculated by dividing the current anomalous duration by the historical typical duration of similar events; and F3 is calculated by statistically analyzing the frequency of similar anomalous events over the past week and dividing by a threshold fitted based on the historical frequency of similar anomalous events. A defined weight is assigned to each feature. The total weight is 1. The weights assigned to the deviation magnitude and duration are determined by querying the historically consistent deviation degree and duration weights, which are used for historical frequency feature weights. Then, the historical weight of 0.2 is used, for example, the initial weight. The result is (0.5, 0.3, 0.2).

[0053] The three eigenvalues ​​for this event are F1 = 0.8, F2 = 0.3, and F3 = 0.6. The final confidence score is calculated using the formula... The calculation yields a value between 0 and 1, typically multiplied by 100% and presented as a percentage. For example, a calculated value of 0.61 corresponds to a 61% confidence level. For instance, correcting the label to "to be observed" will, combined with characteristics such as a deviation of 2 seconds and a duration of only 1 minute, prompt staff to pay attention but not to intervene immediately. This quantitative processing helps to more accurately assess the likelihood of abnormal behavior.

[0054] In step S16, when the confidence level of the abnormal behavior is higher than a preset confidence threshold, anomaly detection analysis is performed to obtain a warning level classification, including: When the confidence level of the abnormal behavior is higher than the preset confidence level threshold, the abnormal signal strength and abnormal duration of the abnormal behavior are obtained from the preset historical record database. The abnormal signal strength and abnormal duration are compared and analyzed to obtain the signal fluctuation characteristic value. When the signal fluctuation characteristic value exceeds the preset signal fluctuation threshold, the behavior frequency data and time distribution data of the abnormal behavior are integrated and analyzed to obtain the abnormal behavior deviation, and the behavior risk level is determined based on the abnormal behavior deviation. Based on the aforementioned behavioral risk level, abnormal behavior is compared with historical behavior records and the degree of correlation is calculated to obtain a comprehensive index of abnormal behavior. Abnormal behaviors are classified and processed according to comprehensive indicators to obtain warning level classifications.

[0055] First, when the confidence level of the abnormal behavior exceeds a preset confidence threshold derived from the historical minimum confidence deviation, the signal strength and duration data of the abnormal behavior are retrieved from the historical data database and subjected to time series comparison processing. This primarily involves comparing the current behavior data with the time distribution of historical data to extract the signal fluctuation range characteristic value. For example, the signal fluctuation characteristic value is obtained by calculating the relative deviation between the current parameter value and the historical baseline value. For instance, the heart rate deviation is the difference between the current heart rate and the historical average heart rate, divided by the historical average heart rate. If an elderly person's sleep data from the past 30 days is recorded, and a sudden increase in heart rate signal strength is observed one night for a duration of 2 hours, this is recorded, and the signal fluctuation characteristic value is obtained.

[0056] Secondly, through comparative analysis, the signal fluctuation characteristic value shows that the heart rate deviates by 15% from the historical baseline value, exceeding the preset threshold of 10% obtained based on historical heart rate data statistics. Specifically, the signal fluctuation characteristic value is calculated by subtracting the historical baseline value from the current monitored parameter value, dividing by the historical baseline value, and then multiplying by 100%. The historical baseline value is obtained by taking the arithmetic mean of parameter data under normal conditions within a historical period.

[0057] By integrating context, the signal fluctuation characteristics are fused with environmental factors (such as temperature index, calculated based on the daily average temperature; activity intensity index, based on the ratio of daytime activity to historical baseline activity intensity) and temporal distribution characteristics (such as abnormal duration) through multi-source data fusion. The historical baseline activity intensity is obtained by statistically analyzing the activity intensity data of historical normal days to obtain the 50th quantile as the benchmark value.

[0058] When calculating the comprehensive score, each feature must first be standardized before a weighted summation. Specifically, the signal fluctuation feature value itself is defined as a relative deviation. For example, the heart rate deviation is the difference between the current heart rate and the historical average heart rate divided by the historical average heart rate, so it can be directly used as a dimensionless parameter. The temperature index is calculated based on historical temperature data using a z-score standardized value, which is obtained by subtracting the historical average temperature from the current temperature value and then dividing by the historical temperature standard deviation. The activity intensity index is itself the ratio of daytime activity to historical baseline activity intensity, which is already dimensionless, but to maintain consistency, it is also calculated based on historical data using a z-score standardized value. The abnormal duration is obtained by dividing by the historical typical duration to obtain a dimensionless ratio.

[0059] The comprehensive score is calculated by weighted summation, which is obtained by multiplying the standardized signal fluctuation characteristic value, temperature index, activity intensity index and abnormal duration by their respective weights and then summing them. The weights are determined based on historical data regression analysis. The weights w1 to w4 are determined by multiple linear regression analysis and satisfy the normalization condition w1+w2+w3+w4=1.

[0060] When the comprehensive score is below the first preset threshold, the behavioral risk level is determined to be low-risk; when the comprehensive score is between the first and second preset thresholds, it is determined to be medium-risk; and when the comprehensive score is above the second preset threshold, it is determined to be high-risk. The behavioral risk level is further mapped to the warning level classification: low-risk corresponds to blue warning, medium-risk corresponds to yellow warning, and high-risk corresponds to orange or red warning.

[0061] Specifically, when the following combination of characteristics occurs, for example, the change in the frequency of abnormal behavior per unit time calculated based on posture estimation data compared to the historical moving average exceeds a preset standard deviation threshold (e.g., 3 standard deviations set based on the distribution of historical data), the activity intensity index (calculated based on joint movement speed and amplitude) is lower than a preset low activity threshold (e.g., 0.5), and the ratio of the abnormal duration to a preset duration threshold based on historical data statistics (e.g., the 95th percentile of historical duration) is greater than 2, then the comprehensive score obtained by weighted summation will exceed the second preset threshold, and the behavior risk level will be judged as high risk.

[0062] If abnormal behavioral signals are detected during the nighttime period (defined as 10 PM to 6 AM), and the activity intensity index shows that the current activity frequency is at a historical high, such as an activity intensity index greater than 0.8, the activity intensity index value will have a large negative contribution to the comprehensive score calculation, thereby lowering the comprehensive score and causing the result to fall between the first preset threshold and the second preset threshold. In this case, the behavioral risk level is judged as medium risk.

[0063] Next, an analysis comparing individual operational habits, historical records, and current status is conducted to determine the degree of matching of related events, ultimately forming a comprehensive index of abnormal behavior. The data fusion process integrates multi-source information, including individual work-rest patterns based on historical data analysis, and correlation analysis between current abnormal behavior characteristics and events in historical behavior records. Work-rest patterns are quantified by calculating the standard deviation of historical work-rest time datasets; a larger standard deviation indicates more irregular work-rest patterns. Event correlation is measured by calculating the Euclidean distance between the current abnormal behavior feature vector and the historical event feature vectors; a smaller distance indicates a higher correlation.

[0064] For example, historical records show that the elderly person occasionally experienced elevated heart rate due to high activity levels, but this usually lasted no more than one hour. However, this abnormality lasted for two hours. Based on the core characteristic of the current abnormal behavior—a heart rate deviation of 15%—cases with similar deviations were matched from the historical database. The proportion of cases confirmed as genuine abnormalities was calculated, resulting in an initial probability assessment of 65%. A context-adjusted factor was introduced to correct the initial probability, with the high-intensity activity factor determined to be 0.9 based on historical data. The probability correction employed a historical statistical adjustment method, multiplying the initial probability of 65% by the activity factor of 0.9 to obtain an adjusted probability of 58.5%. Given that the historical deviation factor of 1.2 reflects the significant impact of the persistent abnormality, it was multiplied by the adjusted probability, resulting in a preliminary comprehensive index value of 70.2%. To ensure the index value falls within a reasonable probabilistic interpretation range, a final index value of 100% was set when the calculated result exceeded 100%. Therefore, the comprehensive index for abnormal behavior was 70%, indicating a high probability of abnormality requiring further monitoring.

[0065] The allocation of weight coefficients is determined based on the information gain method, which quantifies the importance of each feature in historical data by calculating its contribution to the classification of abnormal behavior, thereby allocating weights and ensuring that the sum of all weight coefficients is one.

[0066] Finally, based on the comprehensive index of abnormal behavior and combined with feedback information, a classification decision-making process is performed to determine the warning level classification result of the abnormal behavior. The classification decision is based on a comprehensive judgment of the comprehensive index of abnormal behavior and real-time feedback, such as whether the elderly showed any discomfort that evening as recorded by staff. By analyzing the real-time collected heart rate and activity data, the relative percentage deviation from the individual's historical baseline is calculated, and the duration of abnormal fluctuations is added as a time-weighted factor. These parameters are then input into a preset normalization function.

[0067] The normalization function is specifically designed to calculate the relative deviation H between the current value of a monitoring parameter such as heart rate and activity level and the individual's historical dynamic baseline value. This deviation is calculated by subtracting the absolute value of the baseline value from the current value and dividing by the baseline value. The original duration of this deviation is then obtained and converted into a relative duration ratio T. This ratio is calculated by dividing the current duration by the typical duration baseline value of the parameter under historical conditions, resulting in a dimensionless, standardized relative duration ratio T.

[0068] The outlier of an individual item is calculated using a piecewise linear function. Here, v and u are preset weighting coefficients, representing the contribution intensity of instantaneous deviation and duration, respectively. The relative deviation H, weights v and u are obtained based on historical data regression analysis. Then, the maximum value of the abnormality of multiple parameters is taken to generate an initial value between 0 and 100, namely the abnormality index Ei. Manual verification feedback is introduced as an adjustment coefficient k. This coefficient is discretized according to the actual confirmed condition of the elderly: k=0.5 when no discomfort is confirmed, k=1.0 when there is slight discomfort, and k=1.5 when there is severe discomfort or when contact cannot be made. This design allows subjective judgment to linearly adjust the weight of objective data.

[0069] It should be noted that the typical duration baseline value needs to be extracted from historical data in a preset historical record database to determine the duration distribution of normal states, and the typical value is determined through statistical calculation. For example, through the anomaly detection analysis in step S16, only the duration corresponding to events marked as "normal" or "low risk" is selected. Data sources include long-term monitored sensor data and duration records in historical behavioral pattern characteristics, such as the duration of routine behaviors like walking and sitting. Obvious anomalies need to be excluded during screening to ensure that the typical duration baseline value reflects the typical state. For example, trajectory smoothing in step S11 can be used to remove extreme values ​​caused by transient interference.

[0070] Furthermore, statistical analysis is performed on the filtered historical duration dataset. For example, the median can be calculated because it is not sensitive to outliers and better represents typical values, indicating that 50% of historical events have durations below this value. Alternatively, the mean can be calculated. If the data distribution is approximately normal, the arithmetic mean can be used as the baseline value for the typical duration, but it should be noted that the mean is easily affected by extreme values. Another option is to use a specific percentile, which, depending on the application scenario, can be set to the median or the 60th percentile of the historical duration to balance typicality and conservatism.

[0071] The statistical process should be based on individualized data to ensure that the baseline typical duration reflects the specific behavioral habits of the elderly individual. For example, for a particular elderly person, the median duration of heart rate deviations can be calculated from their historical data over the past 30 days as a personalized baseline typical duration. After setting the initial baseline typical duration, its representativeness needs to be verified through backtesting with historical data. For example, check whether the baseline typical duration makes the ratio T of the relative duration of most normal events close to 1 to indicate a typical value. Because behavioral patterns may change over time, the baseline typical duration should be updated dynamically. For example, it can be updated monthly to reflect the latest behavioral trends. When updating, a sliding window method should be used, retaining only the data from the most recent 3 months to ensure the timeliness of the baseline typical duration.

[0072] Subsequently, the algorithm multiplies the anomaly index by the adjustment coefficient to obtain the early warning decision score. This product operation reflects the coupling effect of subjective and objective factors, ensuring that the final score includes both basic information about data anomalies and real-time corrections from on-site verification. Finally, the decision score is divided into intervals through threshold comparison: a blue warning is issued when S < 30, a yellow warning when 30 ≤ S < 60, an orange warning when 60 ≤ S < 90, and a red warning when S ≥ 90. For example, if feedback indicates that the elderly person had no obvious discomfort that evening, and the calculated final score is 50, then the warning level is classified as a yellow warning, indicating attention rather than emergency intervention. This tiered approach helps to rationally allocate monitoring resources.

[0073] In step S17, the process of triggering an alarm mechanism based on the warning level classification, sending specific notification content to the device, and completing the real-time response includes: The triggering conditions corresponding to the warning level category are obtained from the preset warning level database. When the warning level category exceeds the preset warning level threshold, the alarm activation status is confirmed according to the triggering conditions. Based on the alarm activation status, the notification information and message content are integrated to obtain the specific notification content; The notification is sent to the target device according to the specific content, completing the real-time response process.

[0074] First, the trigger conditions corresponding to the pre-set warning level classification are retrieved from the pre-set warning level database. When an elderly person's abnormal heart rate is judged as a medium warning, the pre-set trigger condition in the database is that medium or higher levels require notification of the caregiver. During the message distribution phase, the warning level is compared with the conditions. If the current heart rate exceeds the normal heart rate range by 20%, the activation requirements are met, and the process proceeds to the next step. When determining the activation status of the alarm mechanism, unnecessary alarms are filtered according to the specific context to avoid wasting resources. For example, if the aforementioned medium warning occurs during the day and the elderly person is detected to be engaged in daily activities, the condition filtering tool will combine the time and activity status to confirm that the alarm mechanism needs to be activated, but the priority is non-urgent, requiring only a reminder rather than an emergency call.

[0075] Secondly, regarding the integration of notification information, the warning level, specific circumstances, and recommended measures are combined into complete content during the notification generation stage. The content format is adjusted according to the characteristics of the notification channel during this stage. For example, if the notification needs to be sent to caregivers via SMS, concise content will be generated, such as "An elderly person experienced an abnormal heart rate at 14:00 today, lasting for 30 minutes; attention is advised," ensuring the information is clear, easy to read, and provides specific notification content.

[0076] Finally, during the notification sending phase, matching data for the target device is obtained to ensure that the information is delivered to the correct recipient. For example, if the monitoring personnel are using a specific model of mobile phone, the device type and status will be confirmed to ensure that the push distribution tool can accurately deliver the notification and record the sending status, such as "sent" or "failed to send". Tracking the sending status can be done in real time using time monitoring combined with timestamps and feedback status. Time monitoring will record the time the notification is sent and track the recipient's response. For example, if the notification is sent at 14:05, the system will check at 14:15 whether it has received confirmation feedback from the monitoring personnel. If not, it will be marked as "pending response" and will continue to track at subsequent time points until a response confirmation is received.

[0077] In summary, this invention discloses a visual detection-based method for early warning of elderly postures. By accurately identifying abnormal postures, it improves the accuracy and real-time performance of monitoring, and solves the problem of false alarms and missed alarms in existing technologies due to the difficulty in identifying interference caused by obstacles or moving objects. Reference Figure 2 The second embodiment of the present invention provides a visual detection-based posture warning system for the elderly, comprising: The data acquisition module is used to acquire continuous images of elderly people's activities through a camera, obtain a continuous image sequence, perform posture estimation on the continuous image sequence, and obtain key part coordinate data and movement trajectory information; The trajectory calculation module is used to calculate the posture vector based on the coordinate data of the key parts and the motion trajectory information to obtain the dynamic behavior feature vector. The behavior confirmation module is used to extract periodic features and change patterns based on the dynamic behavior feature vector to obtain behavior pattern features; An anomaly detection module is used to compare the behavior pattern features with a preset behavior database. When the frequency of behavior change of the behavior pattern features exceeds a preset behavior change frequency threshold, a potential anomaly signal is obtained. The historical integration module is used to perform contextual fusion of the potential abnormal signals with historical behavior data to obtain the confidence level of abnormal behavior. The early warning classification module is used to perform anomaly detection and analysis and obtain an early warning level classification when the confidence level of abnormal behavior is higher than a preset confidence threshold. The real-time early warning module is used to classify the early warning levels, trigger the alarm mechanism, send specific notification content to the device, and complete the real-time response process.

[0078] It should be noted that the visual detection-based elderly posture warning system provided in this embodiment of the invention is used to execute all the process steps of the visual detection-based elderly posture warning method in the above embodiment. The working principles and beneficial effects of the two are one-to-one, so they will not be described again.

[0079] This invention also provides an electronic device. The electronic device includes a processor, a memory, and a computer program stored in the memory and executable on the processor, such as a notification generation program. When the processor executes the computer program, it implements the steps in the various embodiments of the vision detection-based elderly posture warning method described above, for example... Figure 1 The step S11 shown. Alternatively, when the processor executes the computer program, it implements the functions of each module / unit in the above-described device embodiments, such as the data acquisition module.

[0080] For example, the computer program may be divided into one or more modules / units, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules / units may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the computer program in the electronic device.

[0081] The electronic device may be a desktop computer, laptop, handheld computer, or smart tablet, etc. The electronic device may include, but is not limited to, a processor and memory. Those skilled in the art will understand that the above components are merely examples of electronic devices and do not constitute a limitation on the electronic device. It may include more or fewer components than described above, or combine certain components, or different components. For example, the electronic device may also include input / output devices, network access devices, buses, etc.

[0082] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the electronic device, connecting all parts of the electronic device via various interfaces and lines.

[0083] The memory can be used to store the computer programs and / or modules. The processor implements various functions of the electronic device by running or executing the computer programs and / or modules stored in the memory and by calling data stored in the memory. The memory may mainly include a program storage area and a data storage area. The program storage area may store the operating system, at least one application program required for a function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created according to the use of the mobile phone (such as audio data, phonebook, etc.). In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0084] Wherein, if the modules / units integrated in the electronic device are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium can be appropriately added or removed according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electrical carrier signals and telecommunication signals.

[0085] It should be noted that the device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the device embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0086] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A method for early warning of elderly posture based on visual detection, characterized in that, include: By capturing continuous images of elderly people's activities through a camera, a continuous image sequence is obtained. The posture of the continuous image sequence is estimated to obtain the coordinate data of key parts and the movement trajectory information. Based on the coordinate data of the key parts and the motion trajectory information, the posture vector is calculated to obtain the dynamic behavior feature vector; Based on the dynamic behavior feature vector, periodic features and change patterns are extracted to obtain behavior pattern features; By comparing the behavioral pattern features with a preset behavioral database, when the frequency of behavioral changes in the behavioral pattern features exceeds a preset threshold for behavioral change frequency, a potential abnormal signal is obtained. The potential abnormal signals are fused with historical behavior data in context to obtain the confidence level of abnormal behavior. When the confidence level of the abnormal behavior is higher than the preset confidence level threshold, anomaly detection analysis is performed to obtain a warning level classification. Based on the aforementioned warning level classification, an alarm mechanism is triggered, specific notification content is sent to the device, and a real-time response process is completed.

2. The method for early warning of elderly posture based on visual detection according to claim 1, characterized in that, The process involves capturing continuous images of elderly individuals' activities using a camera to obtain a continuous image sequence, performing pose estimation on the continuous image sequence to obtain key body coordinate data and movement trajectory information, including: Image data of elderly people's activities are collected by camera equipment to obtain an initial image sequence. The initial image sequence is then denoised and illumination corrected to obtain a continuous image sequence. Based on the continuous image sequence, key component coordinate data and key component location information are extracted to obtain a preliminary coordinate set; The coordinate data of the key parts in the preliminary coordinate set are tracked to obtain position change information. When the change amplitude of the position change information exceeds a preset change amplitude threshold, the position change information is smoothed to obtain smooth trajectory information. A dynamic trajectory diagram is generated based on the smooth trajectory information, and the coordinate data of key parts and the motion trajectory information are obtained.

3. The visual detection-based posture early warning method for the elderly according to claim 1, characterized in that, The step of calculating the posture vector based on the coordinate data of the key parts and the motion trajectory information to obtain the dynamic behavior feature vector includes: Spatial location information is extracted from the coordinate data of the key parts and the changes in the joint parts are analyzed to obtain a preliminary joint dataset. By analyzing the displacement velocity using the preliminary joint dataset and the temporal information of the motion trajectory information, the continuously changing displacement values ​​are obtained. When the value of the continuous displacement change exceeds the preset threshold for continuous displacement change, the abnormal displacement value is smoothed and corrected to obtain the adjusted displacement data. Based on the adjusted displacement data and the preliminary joint dataset, a comprehensive body posture vector is calculated to obtain a dynamic behavior feature vector.

4. The method for early warning of elderly posture based on visual detection according to claim 1, characterized in that, The step of extracting periodic features and change patterns from the dynamic behavior feature vector to obtain behavior pattern features includes: The behavior intervals and change frequencies within a complete behavior cycle are obtained from the dynamic behavior feature vector. Based on the behavior interval and the change frequency, a dynamic trend analysis is performed on the dynamic behavior feature vector to obtain the dynamic trend change; When the dynamic trend change exceeds the preset dynamic trend change threshold, the dynamic behavior feature vector is smoothed and corrected to obtain the adjusted behavior sequence. Behavioral pattern matching is performed based on the adjusted behavioral sequence to obtain behavioral pattern features.

5. The method for early warning of elderly posture based on visual detection according to claim 1, characterized in that, The process involves comparing the behavioral pattern features with a preset behavioral database. When the frequency of behavioral changes in the behavioral pattern features exceeds a preset behavioral change frequency threshold, a potential abnormal signal is obtained, including: The feature distribution is obtained from the behavioral pattern features to determine the range of behavioral change frequency. The behavioral deviation is obtained by comparing the frequency range of behavioral changes with a preset behavioral database and calculating the deviation. When the behavioral deviation exceeds a preset behavioral deviation threshold, the behavioral pattern features are classified and labeled according to the specific deviation to obtain an abnormal signal classification label. A secondary screening analysis is performed based on the abnormal signal classification labels to identify potential abnormal signals.

6. The method for early warning of elderly posture based on visual detection according to claim 1, characterized in that, The step of contextually fusing the potential abnormal signals with historical behavior data to obtain the abnormal behavior confidence level includes: Obtain the time series distribution data of the potential abnormal signals, perform preliminary matching of abnormal behaviors based on the time series distribution data, and obtain the pattern comparison results; The correlation information between the pattern comparison results and historical pattern comparison results is integrated to obtain consistent data; When the consistency data deviates from the preset consistency data threshold range, the pattern comparison result is deeply verified to obtain the behavior deviation situation. Based on the behavior deviation situation, a graded label is obtained to obtain a corrected classification label. The confidence level of the behavioral feature distribution is quantified based on the modified classification label to obtain the confidence level of abnormal behavior.

7. The method for early warning of elderly posture based on visual detection according to claim 1, characterized in that, When the confidence level of the abnormal behavior is higher than a preset confidence threshold, anomaly detection analysis is performed to obtain a warning level classification, including: When the confidence level of the abnormal behavior is higher than the preset confidence level threshold, the abnormal signal strength and abnormal duration of the abnormal behavior are obtained from the preset historical record database. The abnormal signal strength and abnormal duration are compared and analyzed to obtain the signal fluctuation characteristic value. When the signal fluctuation characteristic value exceeds the preset signal fluctuation threshold, the behavior frequency data and time distribution data of the abnormal behavior are integrated and analyzed to obtain the abnormal behavior deviation, and the behavior risk level is determined based on the abnormal behavior deviation. Based on the aforementioned behavioral risk level, abnormal behavior is compared with historical behavior records and the degree of correlation is calculated to obtain a comprehensive index of abnormal behavior. Abnormal behaviors are classified and processed according to comprehensive indicators to obtain warning level classifications.

8. The method for early warning of elderly posture based on visual detection according to claim 1, characterized in that, The process of classifying warning levels, triggering an alarm mechanism, sending specific notification content to the device, and completing a real-time response includes: The triggering conditions corresponding to the warning level category are obtained from the preset warning level database. When the warning level category exceeds the preset warning level threshold, the alarm activation status is confirmed according to the triggering conditions. Based on the alarm activation status, the notification information and message content are integrated to obtain the specific notification content; The notification is sent to the target device according to the specific content, completing the real-time response process.

9. A posture warning system for the elderly based on visual detection, characterized in that, include: The data acquisition module is used to acquire continuous images of elderly people's activities through a camera, obtain a continuous image sequence, perform posture estimation on the continuous image sequence, and obtain key part coordinate data and movement trajectory information; The trajectory calculation module is used to calculate the posture vector based on the coordinate data of the key parts and the motion trajectory information to obtain the dynamic behavior feature vector. The behavior confirmation module is used to extract periodic features and change patterns based on the dynamic behavior feature vector to obtain behavior pattern features; An anomaly detection module is used to compare the behavior pattern features with a preset behavior database. When the frequency of behavior change of the behavior pattern features exceeds a preset behavior change frequency threshold, a potential anomaly signal is obtained. The historical integration module is used to perform contextual fusion of the potential abnormal signals with historical behavior data to obtain the confidence level of abnormal behavior. The early warning classification module is used to perform anomaly detection and analysis and obtain an early warning level classification when the confidence level of abnormal behavior is higher than a preset confidence threshold. The real-time early warning module is used to classify the early warning levels, trigger the alarm mechanism, send specific notification content to the device, and complete the real-time response process.

Images