Old people falling early warning system based on motion trail analysis

Abstract

The invention relates to the technical field of motion detection, in particular to an old people falling early warning system based on motion trail analysis, which comprises a gravity center identification module, a trend identification module, an inertia pushing module, a phase judgment module and a grade response module. According to the method, the vertical coordinates of the key points of the waist and the two feet in the image frame are acquired, and the height difference change is continuously tracked, so that the potential falling initial state can be identified from the static posture, the vector included angles among the three groups of bone points of the hip knee, the knee ankle and the shoulder pelvis are subjected to sequence analysis, and whether synchronous disintegration occurs in the coordination action is identified; the identified trend sequence is mapped according to the risk level, a differential response mechanism is applied, a high-density corresponding relation is constructed between the behavior trend and the alarm level, graded response and time sequence triggering of the falling risk are achieved, the accurate identification capability of the falling precursor action is enhanced, the sensitivity and discrimination of the alarm response are also improved, and the safety of the falling risk is improved. And misjudgment and missed judgment are effectively avoided.

Description

Technical Field

[0001] The present invention relates to the technical field of motion detection, and in particular to a fall warning system for the elderly based on motion trajectory analysis. Background Art

[0002] The field of motion detection technology involves identifying and tracking the movement of objects, particularly humans, within images or video sequences. The core of this technology lies in acquiring continuous image sequences and analyzing information such as the target object's position, direction, and speed to identify its motion patterns and state changes. Motion detection technology utilizes methods such as image frame difference analysis, target outline extraction, and optical flow calculation to dynamically track the target object and determine whether its behavior is abnormal. The overall technical system encompasses image acquisition, feature extraction, and temporal behavior analysis, and is widely applicable in scenarios such as video surveillance, security, intelligent transportation, and human-computer interaction.

[0003] Among them, the elderly fall warning system refers to a system that uses image recognition and motion trajectory analysis technology to conduct real-time monitoring and abnormal movement recognition for falls in the elderly during daily activities. The technical matters covered by this patent subject include locating key points of the human skeleton through image sequences, calculating the human motion trajectory by the changes in the position of key points in consecutive frames, and then analyzing the fall trend by combining the vertical displacement changes, speed mutations and angle offset characteristics of the trajectory, and generating warning signals accordingly. The system uses a depth camera to collect image data, combines skeleton tracking methods to identify key points, and then uses mathematical modeling to calculate the trajectory curve to complete the recognition and warning of the precursory movements of falls.

[0004] Existing fall detection technologies rely on changes in the trajectory and velocity of skeletal keypoints as the primary basis for judgment. However, these methods lack sufficient sensitivity to detect early abnormal signals in continuous behavior. Due to a lack of detailed analysis of the coordination between multiple keypoints, it is difficult to distinguish between sporadic movements and sustained imbalances, resulting in low recognition accuracy for falls involving slow falls or gradual downward shifts of the center of gravity. Existing solutions rely on fixed threshold models to identify abnormal movements, which cannot adapt to the non-standard postures of elderly people in the context of diverse daily movements. Using single-point velocity changes at the shoulder or knee as an indicator can easily be distracted by behaviors such as bending over to pick up an object or briefly sitting or lying down, leading to false positives. Existing technologies lack a coordinated mapping between fall risk and response mechanisms, lacking effective differentiation between different levels of fall risk and providing targeted real-time intervention. For example, the same behavioral characteristics may represent different risk levels in different contexts, but existing systems can only output a single warning signal and lack a hierarchical processing strategy, limiting the system's reliability in practical applications. Summary of the Invention

[0005] The purpose of the present invention is to solve the shortcomings of the existing technology and propose a fall warning system for the elderly based on motion trajectory analysis.

[0006] To achieve the above objectives, the present invention adopts the following technical solutions: a fall warning system for the elderly based on motion trajectory analysis, comprising:

[0007] The center of gravity recognition module obtains continuous image frames of elderly people in indoor activity spaces, extracts the vertical coordinates of the waist and feet key points in the image frames, performs consistency judgment on the direction of difference changes between adjacent frames, and generates continuous segments of center of gravity offset;

[0008] The trend recognition module extracts the spatial position path of the elderly person's waist key points within three adjacent frames based on the continuous segment of the center of gravity offset, detects whether there is an initial reverse switch, and determines the fluctuation continuity of the center of gravity change trend before and after the reversal point, thereby generating a center of gravity offset slope reversal mark frame segment;

[0009] The inertial shift module calls the center of gravity offset slope reversal mark frame segment, extracts the vertical positioning sequence of the pelvis, shoulder, and knee key points, determines whether there is a segment with consistent direction among the three points, and calculates the segment length and switching frequency to generate the center of gravity inertial co-directional shift state;

[0010] The phase judgment module extracts the motion vector sequences of the three groups of bone point pairs in the vertical and anterior-posterior axis directions of the hip and knee, knee and ankle, and shoulder and pelvis according to the time frame segment marked by the co-directional movement state of the center of gravity inertia, analyzes the changes in the angles of the bone point direction vectors between adjacent frames, and generates a center of gravity collaborative splitting fluctuation trend segment.

[0011] As a further solution of the present invention, the center of gravity offset continuous segment segment includes a height change amplitude sequence, a change direction consistency label, and a continuous frame segment number; the center of gravity offset slope reversal mark frame segment includes a reverse switching frame index, a trend fluctuation continuity label, and a waist path segment number; the center of gravity inertia co-directional movement state includes the co-directional segment length, the number of direction maintenance times, and the switching event frequency; the center of gravity collaborative split fluctuation trend segment includes an angle change trend sequence, a bone point collaborative state identifier, and an action deviation area number.

[0012] As a further solution of the present invention, the center of gravity identification module includes:

[0013] The image coordinate extraction submodule obtains continuous image frames of the elderly in the indoor activity space, detects the key points of the waist and feet in the image frames, extracts the vertical coordinates of the key points of the waist and feet in the image frames, and generates a vertical coordinate sequence of the waist and feet;

[0014] The key point difference calculation submodule sequentially extracts the coordinate height differences of the same key points between the differentiated frames based on the waist and foot vertical coordinate sequence, calculates the average vertical distance between the waist point and the foot points in the image frame, and obtains the waist and foot height difference sequence;

[0015] The difference change judgment submodule determines whether the difference change direction between adjacent image frames is consistent based on the waist-foot height difference sequence. If the difference change direction in consecutive image frames is the same, the segments are connected, the change consistency score is calculated, and combined with the sequence continuity, a continuous segment of the center of gravity offset is generated.

[0016] As a further solution of the present invention, the trend identification module includes:

[0017] The path extraction submodule extracts the spatial positions of the elderly person's waist key points in three adjacent frames based on the center of gravity offset continuous segment, calculates the two-dimensional spatial path formed by the three points, obtains the speed direction value between every two frames, and obtains the speed change direction sequence;

[0018] The reverse identification submodule calls the direction vector changes between adjacent frame segments in the speed change direction sequence, compares the angle value between the real-time direction vector and the previous direction vector to see whether a reverse switch occurs, determines whether there is an initial direction reversal, selects the frame segment where the angle changes from an acute angle to an obtuse angle as a reverse point segment, records the position of the direction switching critical point, and obtains the first reverse switching frame segment;

[0019] The trend determination submodule adopts the first reverse switching frame segment to obtain the center of gravity position coordinates of two consecutive frames before and after the reversal point, identifies the center of gravity longitudinal axis offset difference sequence, makes a difference judgment on the sequence fluctuation direction, calculates the fluctuation amplitude evaluation value, and compares the fluctuation amplitude evaluation value with the center of gravity fluctuation judgment benchmark. If the judgment is valid, it is determined to be a trend reversal frame segment, and a center of gravity offset slope reversal mark frame segment is obtained.

[0020] As a further solution of the present invention, the inertial movement module includes:

[0021] The center of gravity offset marking submodule calls the center of gravity offset slope reversal marking frame segment, detects the vertical change trend between adjacent frames for the center of gravity position point of each frame in the marked frame segment, calculates the offset slope value based on the vertical difference change rate of adjacent frames, and identifies the frame segment position where the slope changes from positive to negative or from negative to positive, thereby generating a center of gravity reversal frame sequence;

[0022] The key point extraction submodule calls the center of gravity inversion frame sequence to extract the vertical positioning values ​​of three key points of the pelvis, shoulder, and knee in each frame, generates a vertical positioning sequence in consecutive frames for the key points, determines the upward or downward trend based on the direction of change of the vertical position coordinates of the key points in the consecutive frames, and generates a vertical direction sequence set;

[0023] The direction sequence screening submodule selects the direction identifiers of the three key points of the pelvis, shoulder, and knee within the same time period based on the vertical direction sequence set, determines whether they are in the same direction, records the frame segments of continuous same-direction state, and counts the frame length of the segment and the frequency of direction switching between the previous and subsequent same-direction segments. The total number of continuous frames and the number of switching times of each same-direction movement are recorded to generate the center of gravity inertia co-directional movement state.

[0024] As a further solution of the present invention, the phase judgment module includes:

[0025] The bone point vector extraction submodule extracts the continuous spatial displacement of the three bone point pairs of hip and knee, knee and ankle, and shoulder and pelvis in the vertical and anterior-posterior directions based on the time frame segment marked by the co-directional displacement state of the center of gravity inertia, obtains the motion vector sequence at each frame time point, and calculates the unit time displacement of adjacent bone points in different directions to obtain the bone point direction vector sequence;

[0026] The angle change trend identification submodule calls the bone point direction vector sequence, calculates the angle value change between the bone point direction vectors of the same group at adjacent time points, identifies the angle change direction, analyzes the angle change trend of each group of bone points in the full frame segment, and obtains the angle change trend parameter group;

[0027] The collaborative fluctuation identification submodule analyzes the degree of movement synchronization and the state of coordination based on the angle change trend parameter group and whether the directions of the angle change curves at the same frame time points between the bone points are consistent and whether the change rates are in a similar range. It calculates the coordination amplitude of the synchronization angle, calls the segment continuity on the time frame axis, extracts the segment boundaries of the change direction, and obtains the center of gravity collaborative splitting fluctuation trend segment.

[0028] As a further solution of the present invention, the system further includes a level response module:

[0029] The level response module calls the center of gravity collaborative splitting fluctuation trend segment and sets it as mild, moderate, and severe fall risk levels respectively. The differentiated levels are triggered in the corresponding frame segments and time-mapped. The correspondence between the differentiated levels and the buzzer ringing frequency, flashing light frequency, and notification command delay is evaluated to generate a center of gravity fall risk response time period set.

[0030] The center of gravity fall risk response period set includes a risk level label, a buzzer frequency value, a flashing light cycle code, and a notification delay parameter.

[0031] As a further solution of the present invention, the level response module includes:

[0032] The time segmentation identification submodule collaboratively splits the fluctuation trend segments based on the center of gravity, extracts the time frame segments of the differentiated fluctuation segments, distinguishes the time frame segments by frame sequence numbers, identifies the interruption points of the fluctuation continuity and divides the frames into segments, calculates the time span value by the change in the fluctuation characteristics between consecutive frames, and obtains the time frame segment interval;

[0033] The risk level judgment submodule calls the time frame interval and, based on a two-dimensional parameter set consisting of a center of gravity fluctuation amplitude value and a fluctuation rate value, judges the belonging to the fluctuation interval corresponding to the mild, moderate, and severe fall risk levels, and obtains a risk level classification trend;

[0034] The response parameter setting submodule combines and arranges the correspondence between the frame segments and the buzzer, flashing light, and notification instructions according to the risk level classification trend, maps the time axis according to the start and end time of the time frame segment, summarizes the time line and arranges the sequence to generate a center of gravity fall risk response time period set.

[0035] Compared with the prior art, the advantages and positive effects of the present invention are:

[0036] In the present invention, by obtaining the vertical coordinates of the key points of the waist and feet in the image frame and continuously tracking the changes in their height difference, continuous monitoring of the center of gravity shift trend is achieved, which helps to identify the potential starting state of a fall from a static posture. The path of the key points of the waist in three frames is constructed as a sequence of speed change directions, and the reversal points and their forward and backward fluctuation trends are identified. This can effectively analyze the precursors of critical unbalanced movements and provide dynamic support for the judgment of fall trends. In the continuous directional consistency analysis of the pelvis, shoulders, and knees, by calculating the co-directionality and continuation length of their displacement directions, the continuous evolution stage of the inertial downward tilt process is captured, and the structured recognition of the actual fall motion pattern is strengthened. In the recognition process, the vector angles between the three groups of bone points of the hip and knee, knee and ankle, and shoulder and pelvis are further analyzed in sequence to identify whether there is synchronization disintegration in their coordinated actions, and to explore the fluctuation trend of the action coordination between multiple key points, providing multi-dimensional cross-validation conditions for critical action recognition. Identified trend sequences are mapped by risk level and differentiated response mechanisms are applied. This establishes a high-density correspondence between behavioral trends and alert levels, enabling a graded response and time-series triggering for fall risks. The overall process, informed by time series data, integrates multiple factors, including key point spatial positioning, dynamic directional changes, movement coordination trends, behavioral inertia characteristics, and response mapping mechanisms. This not only enhances the ability to accurately identify pre-fall motions, but also improves the sensitivity and discrimination of alert responses, effectively avoiding misjudgments and missed detections. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 is a system flow chart of the present invention;

[0038] Figure 2 This is a flow chart of the center of gravity identification module in the present invention;

[0039] Figure 3 This is a flow chart of the trend identification module in the present invention;

[0040] Figure 4 This is a flow chart of the inertial shift module in the present invention;

[0041] Figure 5 This is a flow chart of the phase judgment module in the present invention;

[0042] Figure 6 This is a flow chart of the level response module in the present invention. DETAILED DESCRIPTION

[0043] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0044] In the description of the present invention, it should be understood that the terms "length," "width," "up," "down," "front," "back," "left," "right," "vertical," "horizontal," "top," "bottom," "inside," "outside," and the like, indicating positions or relationships, are based on the positions or relationships shown in the accompanying drawings and are intended only to facilitate the description of the present invention and simplify the description. They do not indicate or imply that the devices or elements referred to must have a specific orientation, be constructed, or operate in a specific orientation. Therefore, they should not be construed as limiting the present invention. Furthermore, in the description of the present invention, "plurality" means two or more, unless otherwise expressly and specifically defined.

[0045] See also Figure 1 , a fall warning system for the elderly based on motion trajectory analysis includes:

[0046] The center of gravity recognition module obtains continuous image frames of elderly people in indoor activity spaces, extracts the vertical coordinates of the waist and feet key points in the image frames, calculates the height difference between the two types of points in the image frame sequence, performs consistency judgment on the direction of the difference change between adjacent frames, and generates continuous sections of center of gravity offset;

[0047] The trend recognition module extracts the spatial position path of the elderly person's waist key points within three adjacent frames based on the continuous segment of the center of gravity offset. It then establishes a speed change direction sequence, detects whether there is an initial reverse switch, and determines the fluctuation continuity of the center of gravity change trend before and after the reversal point, generating a center of gravity offset slope reversal marker frame segment.

[0048] The inertial movement module uses the center of gravity offset slope reversal marker frame segment to extract the vertical positioning sequence of the pelvis, shoulder, and knee key points. It extracts the direction of each point in consecutive frames, determines whether there is a segment with consistent direction among the three points, and calculates the segment length and switching frequency to generate the center of gravity inertial co-directional movement state.

[0049] The phase judgment module extracts the motion vector sequences of the three bone point pairs (hip and knee, knee and ankle, and shoulder and pelvis) in the vertical and anteroposterior directions based on the time frame segments marked by the co-directional inertial motion of the center of gravity. It analyzes the changes in the angles of the bone point direction vectors between adjacent frames, identifies the angle change trends of each bone point group at each frame time point, analyzes the synergy between the bone point movements, and generates the center of gravity synergy splitting fluctuation trend segment.

[0050] The level response module calls the center of gravity collaborative splitting fluctuation trend segment, which is set as mild, moderate and severe fall risk levels respectively. The differentiated levels are triggered to respond in the corresponding frame segments and time-mapped. The correspondence between the differentiated levels and the buzzer ringing frequency, flashing light frequency and notification command delay is evaluated to generate a center of gravity fall risk response time period set.

[0051] The continuous segment of the center of gravity offset includes the height change amplitude sequence, the change direction consistency label, and the continuous frame segment number. The center of gravity offset slope reversal mark frame segment includes the reverse switching frame index, the trend fluctuation continuation label, and the waist path segment number. The center of gravity inertia co-directional movement state includes the co-directional segment length, the number of direction maintenance times, and the switching event frequency. The center of gravity coordinated split fluctuation trend segment includes the angle change trend sequence, the bone point coordinated state identifier, and the action deviation area number. The center of gravity fall risk response time period set includes the risk level label, the buzzer frequency value, the flashing light cycle code, and the notification delay parameter.

[0052] See also Figure 2 , the center of gravity recognition module includes:

[0053] The image coordinate extraction submodule obtains continuous image frames of the elderly in the indoor activity space, detects the key points of the waist and feet in the image frames, extracts the vertical coordinates of the key points of the waist and feet in the image frames, and generates a vertical coordinate sequence of the waist and feet;

[0054] Table 1 Image frame key points and time parameters

[0055]

[0056] Obtain continuous image frames of the elderly in the indoor activity space. The acquisition of image frames can be based on a fixed-installed surveillance camera device, which continuously collects data at a frame rate of 25 frames per second. Set the acquisition of 5 frames of images, and the corresponding frame numbers are 1 to 5. Combined with the characteristics of the elderly's indoor activities, the position of the elderly's waist and feet in each frame of the image is located through the image recognition algorithm. The positioning method uses the human body key point recognition network model to identify the pixels of the torso and lower limb ends in the image. For example, in image frame 1, the longitudinal pixel of the waist key point is 540, and the longitudinal pixel of the foot key point is 960, and the coordinate extraction operation is performed. , each frame of the image is converted into a grayscale image for clarity enhancement, and the human shape area is framed by regional segmentation. The vertical coordinate values ​​are then calculated from the human shape area to obtain the longitudinal coordinate values ​​of the waist and feet in each frame, forming a coordinate pair data sequence, which are recorded as the waist coordinate sequence and the foot coordinate sequence respectively. The corresponding relationship between the image frame sequence and the key point coordinates is constructed. For example, the waist coordinates in image frame 2 are 545 pixels, and the feet are 963 pixels. Similarly, each image frame generates a set of vertical coordinate values ​​of the waist and feet, which are used for subsequent difference calculation processing to obtain the waist and foot vertical coordinate sequence.

[0057] The key point difference calculation submodule extracts the coordinate height difference of the same key point between the differentiated frames based on the waist and foot vertical coordinate sequence, calculates the average vertical distance between the waist point and the foot points in the image frame, and obtains the waist and foot height difference sequence;

[0058] Extract the pixel coordinate difference of the waist and feet in each frame from image frame 1 to image frame 5. The calculation process is based on the vertical height difference formed by subtracting the coordinates of the feet from the waist coordinates in the image frame. For example, in image frame 1, the waist pixel is 540 and the foot pixel is 960. The vertical difference is 540-960=-420. The obtained value represents the relative relationship between the center of gravity height of the human body and the support surface in this frame; in image frame 2, it is 545 and 963 respectively, and the difference is -418. Repeat the calculation for image frame 5. The differences between the key points in image frames 3 and 5 are -417, -415 and -414 respectively. The sorted height difference sequence is {-420, -418, -417, -415, -414}. At the same time, a corresponding array structure of the image frame number and the difference sequence is established for subsequent change trend judgment. The difference structure reflects the vertical change trend of the waist center of gravity relative to the foot support surface when the elderly are standing or walking in consecutive image frames, forming a waist-foot height difference sequence.

[0059] The difference change judgment submodule judges whether the difference change direction between adjacent image frames is consistent based on the waist-foot height difference sequence. If the difference change direction in consecutive image frames is the same, the segments are connected using the formula:

[0060]

[0061] Calculate the change consistency score and combine it with the sequence continuity to generate the center of gravity shift continuous segment;

[0062] Among them, S k represents the change consistency score, d i represents the waist-to-foot height difference of real-time image frame i, d i-1 Represents the height difference between the waist and feet in the previous frame i-1, v i Represents the change in the vertical coordinate of the waist point of the real-time image frame, a i Represents the acceleration change value of the vertical coordinates of the two foot points in the real-time image frame i, Δt i represents the frame time difference between real-time image frame i and the previous frame, and n is the number of real-time image frames;

[0063] The benefit of the formula is that by introducing the waist variation v i , the acceleration change of both feet a i and frame time difference Δt i The composite operation improves the stability and continuity of the center of gravity movement direction judgment, has high recognition accuracy in multi-frame continuously changing scenes, and effectively copes with local jitter caused by shooting angle changes or image noise;

[0064] The formula is used to calculate the consistency of waist height difference changes between adjacent frames in the image frame sequence, that is, the change consistency score S k Its rationality is reflected in the multiple fusion and weighted processing of spatial and temporal change information between image frames. In the formula, the difference term (d i -d i-1 ) captures the changing trend of the height difference between frames. If the changing direction is consistent, the result is positive. If the deviation is, it is offset, providing a basis for direction judgment; multiply by The speed and acceleration information of the waist movement in the image are introduced, so that the change intensity is not only affected by the difference change, but also by the weighted degree of dynamic behavior change. At the same time, Δt i+1 The time factor is normalized. The structural design of the entire formula reflects the high emphasis on "continuity" and "directional consistency" in time series image analysis. It can effectively eliminate occasional local jumps caused by camera angle, jitter or noise, and achieve a comprehensive judgment of motion stability and trend continuity. It is suitable for tracking and identifying stable targets in complex video sequences.

[0065] Determine whether the direction of difference change between image frames is consistent. When performing consistency connection, calculate the direction of the difference increment between each two frames. If the change signs are the same, the direction is considered consistent. The following formula is used to calculate the change consistency score:

[0066]

[0067] Among them, d i The height difference between the waist and feet in the current frame is obtained from the previous difference sequence, such as d3 in the third frame.

[0068] =-417, d2 =-418, the difference between the two is 1;

[0069] v i v3 is the change in the waist longitudinal coordinate in the current frame, i.e., the waist coordinate change value. For example, the waist coordinate in the third frame is 550, and in the second frame it is 545, so v3 = 550 - 545 = 5;

[0070] a i is the change in the longitudinal acceleration of the two feet coordinates, calculated by the difference method. For example, the two feet pixels in the third frame and the second frame are 967 and 963 respectively, the change is 4, and the difference between the previous frame and the first frame is 3, so the acceleration is 1;

[0071] Δt i is the frame time difference, the acquisition is set to 0.04 seconds, so Δt3 = 0.04, substitute it into the calculation formula;

[0072]

[0073] The change consistency score between frame segments is calculated accordingly to form a score sequence {5.00, 5.77, 5.83, 5.77}. The change consistency score indicates the degree of directional continuity of the waist height difference change in the image sequence. The larger the value, the more consistent the movement trend of adjacent frames, indicating that the center of gravity change is more stable and coherent. If the sequence value is greater than a set direction consistency threshold, it is judged to be direction consistent. The direction consistency threshold is set according to the experimental results in the actual scene. The setting standard is a score value of not less than 5.5, which is derived from the data statistics of 30 elderly people in their daily activities. When three consecutive score values ​​in the sequence are greater than 5.5, it is regarded as the center of gravity movement stage in the same direction. The frame segments are connected to obtain the continuous segment of the center of gravity offset.

[0074] Table 2 Image frame key points and time parameters:

[0075] Image frame number Waist coordinates (pixels) Foot coordinates (pixels) Frame time difference (seconds) 1 540 960 0.04 2 545 963 0.04 3 550 967 0.04 4 555 970 0.04 5 561 975 0.04

[0076] As shown in Table 2, the longitudinal pixel values ​​of the key points of the waist and feet in five consecutive frames and the corresponding frame time differences are displayed. By combining the data in this table to perform change trend recognition, we can accurately determine the stable movement sections of the elderly in the sequence and form continuous sections with center of gravity shift.

[0077] See also Figure 3 , the trend identification module includes:

[0078] The path extraction submodule extracts the spatial position of the elderly person's waist key points in three adjacent frames based on the continuous segment of the center of gravity offset, calculates the two-dimensional spatial path formed by the three points, obtains the speed direction value between every two frames, and obtains the speed change direction sequence;

[0079] The spatial path extraction of the key points of the elderly's waist in three frames of images is achieved by collecting the key point coordinates in continuous video images and constructing a sequence of spatial coordinate changes. The operation starts with obtaining the data of three adjacent frames of image in the monitoring, and extracting the coordinate values ​​of the corresponding waist position in each frame. For example, in image frames 101, 102, and 103, the pixel coordinates of the waist points P1 (200, 90.2), P2 (202, 89.6), and P3 (204, 90.1) are extracted respectively. The coordinates of the key points in each pair of adjacent frames are subtracted to obtain the horizontal and vertical displacements of the key points in the image space, that is, the first group of displacements is P2-P1=(2,-0.6), and the second group is P3-P2=(2,0.5). The direction of the spatial change path is recorded based on the difference, and then the image is recorded through By setting the inter-frame time interval to 1 frame (such as a frame rate of 25fps and a time of 0.04 seconds), a direction sequence is formed. Since the waist key point is affected by gait and body swing during the monitoring process, its path presents periodic oscillations. The record of spatial direction will directly reflect its movement trend. In the process of obtaining the direction sequence, care should be taken to avoid key point drift due to image jitter. Frame selection should be based on the stable interval of the key point. When the waist point cannot be identified in a certain frame due to occlusion or detection failure, the system should set a tolerance mechanism to automatically skip the frame or use interpolation to process the missing value, and construct a complete direction sequence such as [(2, -0.6), (2, 0.5)]. This will serve as the direct basis for subsequent direction judgment and trend identification. This sequence is the speed change direction sequence.

[0080] The reverse identification submodule calls the direction vector changes between adjacent frames in the speed change direction sequence, compares the angle value between the real-time direction vector and the previous direction vector to see if there is a reverse switch, determines whether there is an initial direction reversal, selects the frame segment where the angle changes from acute angle to obtuse angle as the reverse point segment, records the position of the direction switching critical point, and obtains the first reverse switching frame segment;

[0081] The direction change is judged to detect whether a reverse switch occurs. The process extracts the change of the current vector and the previous vector from the direction sequence. The system judges the angle change trend based on the difference in the orientation of the two direction vectors in the plane. The previous vector is set to (2, -0.6) and the latter vector is set to (2, 0.5). Although the horizontal direction is the same, the longitudinal direction changes from descending to ascending, reflecting that the direction of movement has changed. The system performs a logical comparison based on the change of the coordinate components between the two directions to determine whether the sign of the longitudinal quantity changes to determine whether it constitutes a reverse switching point. If the detection result shows that the direction switch occurs for the first time, that is, the direction of the previous sequence is consistent, and this point is the first direction mutation point, then this frame segment is marked as a reverse key frame segment.

[0082] At the same time, in order to avoid misjudgment, the comparison of the displacement size between the two frames is introduced as an additional judgment condition in the judgment process. If the absolute value of the vertical difference between the previous and next frames exceeds the set minimum movement threshold, it is confirmed that the switch is of practical significance. In implementation, it is set to detect that the vertical direction changes from negative to positive at frame 103, and the amplitude is 1.1 pixel units, which exceeds the minimum perception threshold. Then, this point is recorded as the frame segment where the direction reversal occurs for the first time, and the first reverse switching frame segment is obtained.

[0083] The trend determination submodule uses the first reverse switching frame segment to obtain the center of gravity position coordinates of two consecutive frames before and after the reversal point, identify the center of gravity longitudinal axis offset difference sequence, and make a difference judgment on the sequence fluctuation direction using the formula:

[0084]

[0085] Calculate the fluctuation amplitude evaluation value, compare the fluctuation amplitude evaluation value with the center of gravity fluctuation judgment benchmark, and if the judgment is correct, determine it as a trend reversal frame segment, and obtain the center of gravity offset slope reversal mark frame segment;

[0086] Where SA represents the fluctuation amplitude assessment value, Δy1 represents the difference in the longitudinal coordinates of the center of gravity between the frame before the reversal point and the two frames before it, and Δy2 represents the difference in the longitudinal coordinates of the center of gravity between the frame after the reversal point and the two frames after it;

[0087] The formula is used to calculate the estimated fluctuation amplitude along the longitudinal axis of the center of gravity. Its rationality lies in accurately quantifying the magnitude of the center of gravity change trend in image frames before and after reverse switching, assisting in determining whether an abnormal reversal in the direction of the action has occurred. By introducing the longitudinal coordinate differences of the center of gravity between the front and rear segments, Δy1 and Δy2, the formula constructs a symmetrical structure centered on longitudinal displacement. The geometric magnitude of the fluctuation amplitude is calculated using the sum of squares to increase sensitivity to drastic changes. The denominator, Δy1+Δy2+1, achieves normalization, avoiding fluctuation amplification caused by frame differences or a baseline value close to zero. This ensures that the overall value is both directionally aware and maintains amplitude stability. This design embodies the "locally significant change + global normalization judgment" strategy used in dynamic behavior analysis and can effectively identify key action features such as trend reversals, high-speed movements, or sudden posture changes in consecutive frames. This plays a key role in determining whether the action has completed reverse switching and whether it should be marked as a trend reversal frame. The formula considers the comparative relationship between change amplitude and directional continuity, enhancing the model's accuracy in identifying action switching points in complex sequences.

[0088] Parameter meaning and formula calculation derivation process:

[0089] Δy1 represents the difference in longitudinal coordinates of the center of gravity between two consecutive frames in the previous frame of the reversal point, and Δy2 represents the difference in longitudinal coordinates of the center of gravity between two consecutive frames in the next frame of the reversal point. In this process data, the Y coordinate values ​​are extracted from consecutive frames numbered 101 to 106 for calculation. The specific extraction method is:

[0090] The coordinates of the center of gravity of the waist in frame 101 are Y1=90.2, frame 102 is Y2=89.6, frame 103 is Y3=90.1, frame 104 is Y4=91.3, frame 105 is Y5=90.7, and frame 106 is Y6=89.9. The units are all centimeters. The coordinates of the key points are recorded and converted into measurement values ​​through image processing. The reversal point is selected as frame 103, the corresponding pre-reversal frame segment is frames 101 to 103, and the post-reversal frame segment is frames 104 to 106. The corresponding values ​​are as follows:

[0091] Δy1=Y3-Y2=90.1-89.6=0.5;

[0092] Δy2=Y5-Y4=90.7-91.3=-0.6;

[0093] Calculate the molecular part:

[0094] |Δy1-Δy2|=|0.5-(-0.6)|=|1.1|=1.1;

[0095]

[0096] Numerator total:

[0097] 1.1+0.781=1.881;

[0098] Calculate the denominator:

[0099] Δy1+Δy2+1=0.5+(-0.6)+1=0.9;

[0100] Substitute into the formula for calculation:

[0101]

[0102] The results show that the longitudinal displacement fluctuations of the frame segments before and after the reversal point have high asymmetry and amplitude intensity. The fluctuation amplitude evaluation value represents the amplitude of the change in the vertical coordinate of the center of gravity in the two images before and after the reverse switching. The larger the value, the more violent the fluctuation of the center of gravity, reflecting a clear turning point in the movement trend. When compared with the preset fluctuation reference standard, the fluctuation amplitude evaluation value exceeds the benchmark judgment level, constituting a trend reversal identification mark. This segment of the frame can be marked as a trend turning node, corresponding to the center of gravity offset slope reversal mark frame segment in the trend judgment step. This value is closely related to the intensity of the displacement change between frames and the relative amplitude difference. The larger the fluctuation amplitude evaluation value, the stronger the turning trend.

[0103] See also Figure 4 , the inertial movement module includes:

[0104] The center of gravity offset marking submodule calls the center of gravity offset slope reversal marking frame segment. For the center of gravity position point of each frame in the marked frame segment, it detects the vertical change trend between adjacent frames, calculates the offset slope value based on the vertical difference change rate of adjacent frames, and identifies the frame segment position where the slope changes from positive to negative or from negative to positive, thereby generating a center of gravity reversal frame sequence.

[0105] The frame images obtained should be a complete time series with a uniform time interval. For each frame, the bounding rectangle of the human body should be identified based on the image boundary information. The coordinates of the center of gravity of the frame are estimated by the center point of the rectangle and stored as the Y-axis vertical coordinate value corresponding to the frame. The vertical difference of the Y-axis coordinates of the center of gravity between adjacent frames is calculated in chronological order, and the inter-frame duration corresponding to the difference is recorded. After setting the time interval to a fixed value, the difference is used as the basis for trend judgment. On this basis, it is judged whether the difference of consecutive frames reverses direction, that is, when the Y-axis coordinates of a certain frame are reversed, the vertical difference is calculated. When the difference changes from positive to negative or from negative to positive, the frame is identified as the offset slope reversal point. At the same time, considering the existence of noise fluctuations, a minimum inter-frame interval needs to be set as a threshold. Frame segments with multiple consecutive reversals but an interval less than this threshold are eliminated to avoid misjudging slight jitter as a trend turning point. Combined with actual test data samples, the 12th, 27th, and 43rd frames can be marked as reversal node frames from continuous video frames, and their frame numbers can be summarized and organized to form a center of gravity reversal frame sequence. This sequence provides a time anchor point for subsequent key point change trends, and a center of gravity reversal frame sequence is obtained.

[0106] The key point extraction submodule calls the center of gravity inversion frame sequence to extract the vertical positioning values ​​of three key points in each frame: the pelvis, shoulders, and knees. It then generates a vertical positioning sequence for each key point in consecutive frames. The vertical position of the key point in each frame is determined by the direction of change in the vertical position of the key point, and a vertical direction sequence set is generated.

[0107] Perform image processing operations on each frame to obtain the vertical coordinates of the key points of the pelvis, shoulders and knees. In practical applications, known human posture recognition tools such as OpenPose can be selected to detect 17 or 25 regular key point positions from the image and filter the required key parts. For each key point in the image, its Y-axis coordinate is extracted and a time series is constructed. The position change trend of each key point in the continuous frame is identified, and the coordinate change values ​​of the two adjacent frames are converted into direction labels. If it moves upward, it is marked as "↑", if it moves downward, it is "↓", and if the position remains unchanged, it is "→". Then the direction label is set. The labels are combined into a complete direction sequence in the order of frame numbers, and three time-axis aligned direction labeling sequences are generated for the pelvis, shoulder, and knee respectively. The actual data sample sets the direction of the pelvis from frame 12 to frame 14 as [↑, ↑], the direction of the shoulder as [↑, →], and the direction of the knee as [↑, ↑] to form three direction trajectory sequences. In this process, it is necessary to ensure the image quality and the accuracy of key point detection to avoid direction misjudgment caused by recognition deviation. At the same time, the key point coordinate data extracted from the image frame and its direction transformation can be stored in a two-dimensional array respectively to facilitate subsequent calculation and use to generate a vertical direction sequence set.

[0108] The direction sequence screening submodule selects the direction identifiers of the three key points of the pelvis, shoulder, and knee within the same time period based on the vertical direction sequence set, determines whether they are in the same direction, records the continuous frame segments in the same direction state, and counts the frame length of the segment and the frequency of direction switching between the previous and subsequent same-direction segments. It records the total number of frames and the number of switching times for each segment of the same-direction movement, and generates the center of gravity inertia co-directional movement state.

[0109] It is necessary to extract and compare the direction marks of the three key points of pelvis, shoulder and knee in each frame to determine whether they are in the same direction state, that is, the directions of the three are "↑", "↓" or "→". If they are consistent, the frame is considered to be a co-directional frame. The continuous co-directional frames are grouped to form temporally continuous co-directional frame segments, and their start and end frame numbers are recorded. The number of frames contained in each co-directional frame segment is calculated. If the length of the segment exceeds the preset frame number threshold, if the number of consecutive frames is not less than 3 frames, it is considered to be a valid segment and saved. At the same time, each co-directional segment needs to be recorded during the traversal process. The direction type of the directional frame segment is determined, and the direction difference between two adjacent segments is judged. If there is a direction change, it is counted as a direction switch. The number of co-directional frame segments, the number of continuous frames, and the total number of direction switches are counted. Based on the example data, the 12th to 15th frames can be set as the "↑" segment, and the 16th to 20th frames can be set as the "↓" segment. Then two co-directional segments are formed and a direction switch is identified. The independence and continuity of the judgment logic of each segment must be ensured. The complete co-directional frame segment information and its switching record are constructed through accumulation to generate the center of gravity inertia co-directional push state.

[0110] See also Figure 5 , the phase judgment module includes:

[0111] The bone point vector extraction submodule extracts the continuous spatial displacement of three bone point pairs (hip and knee, knee and ankle, and shoulder and pelvis) in the vertical and anteroposterior directions based on the time frame segment marked by the co-directional movement of the center of gravity inertia. It obtains the motion vector sequence at each frame time point and calculates the unit time displacement of adjacent bone points in different directions to obtain the bone point direction vector sequence.

[0112] Extract the frame interval that meets the inertial propulsion characteristics from the time sequence identifier. This process is identified by setting the frame rate change threshold. When the displacement speed of three consecutive frames changes within 0.1m / s, it is recorded as an inertial common direction segment. Get all the time frame points in this segment, such as [1, 2, 3, 4, 5]. It is necessary to collect the three-dimensional spatial coordinate positions of the three groups of bone points of the hip and knee, knee and ankle, and shoulder and pelvis in each frame. Assume that in the first frame, the coordinates of the hip and knee group bone points are (1.0, 1.2, 0.8) and (1.2, 1.5, 1.0). By subtracting the x, y, and z coordinate components of adjacent bone points, the motion direction vector ( 0.2, 0.3, 0.2), and use this method to calculate the unit time displacement vector of the three groups of bone points in each direction frame by frame. The data sampling frequency is 60 frames / s, so the time interval is 1 / 60s. The vector velocity sequence is obtained by dividing the difference by the interval, such as 0.3 / 0.0167≈17.96m / s in the vertical direction and 0.2 / 0.0167≈11.98m / s in the front-back direction. The motion vectors of the three groups of bone points are recorded and organized into a time series in frame order. After sorting the data, six groups of vector sequences are formed by combining them with bone points according to direction. As the basis for subsequent analysis, the bone point direction vector sequence is obtained.

[0113] The angle change trend identification submodule calls the bone point direction vector sequence, calculates the angle value change between the direction vectors of the same group of bone points at adjacent time points, identifies the angle change direction, analyzes the angle change trend of each group of bone points in the full frame segment, and obtains the angle change trend parameter group;

[0114] The motion direction vectors of each group of bone points in adjacent frames are processed frame by frame, using the spatial angle formula:

[0115]

[0116] Calculate the angle between the direction vectors of the current frame and the next frame respectively. For example, if vector A = (1, 2, 2) and vector B = (2, 3, 1), their dot product is 1×2+2×3+2×1=10, and the modulus is:

[0117]

[0118] The angle is cos -1 (10 / (3×3.74))≈cos -1 (0.892)≈26.4°;

[0119] According to this method, the angle difference of each frame is calculated, and then a sequence is established for the angle change of each group of frames through the difference. The slope or derivative of the angle sequence is calculated as the angle change rate. If the rate is continuously positive, it means that the angle is increasing, otherwise it is decreasing. According to the angle change rate of each group of bone points at each frame time point, a trend array is formed to generate the angle change trend parameter group.

[0120] The cooperative fluctuation identification submodule analyzes the degree of movement synchronization and cooperative state based on the angle change trend parameter group and the angle change curve direction of the same frame time point between the bone points, and whether the change rate is in a similar range. The formula is:

[0121]

[0122] Calculate the synchronous angle synergy amplitude, call the segment continuity on the time frame axis, extract the segment boundary of the change direction, and obtain the center of gravity synergistic splitting fluctuation trend segment;

[0123] Among them, T s is the synchronization angle coordination amplitude, Represents the angle change value of the bone point group on the vertical axis in the oth frame, Represents the angle change value of the bone point group in the front and back axes in the oth frame, α o represents the difference in the rate of change of the angles of the three groups of bone points at the oth frame, β o represents the change rate of the angles of the three groups of bone points at the oth frame, γ o represents the consistency value of the angle change direction between the three groups of bone points at the oth frame, and M is the total number of time frames involved in the analysis;

[0124] The formula is beneficial in that it considers the changes in the vertical and fore-aft angles simultaneously, and introduces a joint penalty term for the angle change rate range and the consistency parameter. This allows the calculated synchronization angle synergy amplitude to reflect not only the amplitude of the change, but also the distribution of the change and the synergy characteristics, effectively establishing a measurement mechanism for the synchronization of spatial direction changes.

[0125] The formula is used to calculate the synchronization angle coordination amplitude T s Its rationality is reflected in the consistency of the spatial posture changes between the key bone points of the body in multiple dimensions. The formula introduces the vertical direction (Δθ v ) and the front-back direction (Δθ h ) describes the rate of change of the angle in different directions; the denominator contains the angle rate change values ​​in three directions (vertical, front-back, and comprehensive) (|α o -β o |+γ o) is used to measure the consistency between the directions of change. If the change rates in each direction are similar, the denominator becomes smaller and the coordination amplitude value increases, otherwise it decreases. By performing weighted averaging on the frames on the time axis, an overall evaluation of the motion change process is achieved. At the same time, the formula introduces a direction consistency identification value and an extreme difference penalty term to suppress noise interference caused by angle mutations, making the calculation results more stable. This design not only retains the direction amplitude information, but also comprehensively judges the consistency of spatial distribution. It is particularly suitable for detecting multi-person collaborative actions or motion patterns that require high internal coordination of the body. Through structured direction coordination change analysis, it can effectively identify whether the action is synchronous and consistent, and assist in anomaly detection and trend segmentation.

[0126] Change in angle to vertical axis Change in front and rear axle angle The difference in angle change rate α o , bone point angle change rate β o And the consistency identification value γ o The combined parameters are processed and the calculation is performed on each frame using the formula:

[0127]

[0128] in, and Obtained by the difference between adjacent frames in the angle sequence. For example, the vertical axis angle of frame 3 is 28.4° and that of frame 4 is 32.0°. The same applies to the front and back directions;

[0129] α o The difference between the maximum and minimum values ​​of the three groups of angle change rates within the frame, such as 0.28;

[0130] β o is the average value of the three groups of angle change rates, such as 0.17;

[0131] γ o is the consistency count value. If the three groups of angles change in the same direction, then γ o =3;

[0132] Substitute the above frame data into the following table:

[0133] Table 3 Co-fluctuation parameter values

[0134]

[0135]

[0136] As shown in Table 3, the parameter values ​​used for the formula calculation are listed;

[0137] Substitute the data in the table into the calculation to obtain the T value of each frame and calculate its average value:

[0138] Calculation of the first frame:

[0139]

[0140] And so on, we can summarize:

[0141]

[0142] The results show that the synchronization angle synergy amplitude is 3.96. The synchronization angle synergy amplitude indicates whether the changes of the skeleton in the vertical and anterior-posterior directions are synchronized and coordinated. The larger the value, the stronger the consistency and synchronous change characteristics of the bone point group in the spatial direction. The angles of each bone point in the current analysis segment show a certain synchronization trend in the vertical and anterior-posterior directions. The change intensity is close and the synergy direction is consistent, so a center of gravity synergistic splitting fluctuation trend segment is generated.

[0143] See also Figure 6 , the grade response module includes:

[0144] The time segmentation identification submodule splits the fluctuation trend segments based on the center of gravity, extracts the time frame segments of the differentiated fluctuation segments, distinguishes the time frame segments by frame sequence number, identifies the interruption points of the fluctuation continuity and divides the frames into segments, calculates the time span value through the change of the fluctuation characteristics between consecutive frames, and obtains the time frame segment interval;

[0145] The center of gravity data of the human body in a certain period of time is obtained. The data is recorded by the wearable inertial sensor and is expressed as a time-series center of gravity coordinate sequence in frames. It records the balance changes of the user during exercise and performs differentiated segment extraction on the fluctuation trend data. That is, the center of gravity changes in the entire recording interval are divided into time frame segments according to the frame number. During the operation, the data is first numbered frame by frame in the order of the frame number, and the coordinate change amplitudes between adjacent frames are compared. If the changes of several consecutive frames remain within a specific amplitude range, it is judged as a continuous fluctuation segment. If the coordinate jump value between a frame and the previous frame exceeds If the set interruption determination threshold is exceeded, it is determined that the point is a fluctuation continuity interruption point, and the timeline is divided into multiple continuous sub-segments. If the center of gravity changes continuously from the 15th frame to the 25th frame, this segment is identified as the first valid frame segment. The change value from the 40th frame to the 50th frame is continuously stable again, forming the second valid frame segment. After dividing the frame segments, the time span of each frame segment is calculated, that is, the difference between the start and end time of the frame segment. Frame segments with too short time spans will be excluded, and only segments with significant fluctuation characteristics are retained to obtain a set of time frame segment intervals containing several different time frame segments.

[0146] The risk level judgment submodule calls the time frame interval and, based on the two-dimensional parameter set consisting of the center of gravity fluctuation amplitude value and the fluctuation rate value, determines the belonging to the fluctuation interval corresponding to the mild, moderate, and severe fall risk levels, and obtains the risk level classification trend;

[0147] The center of gravity fluctuation characteristics within each segment are analyzed one by one. The center of gravity coordinates of the frames included in each segment are extracted, and the maximum and minimum coordinate values ​​within the segment are calculated to obtain the fluctuation amplitude of the frame segment. The fluctuation rate of each segment is also evaluated based on the position change between each frame and the frame rate. The fluctuation amplitude and fluctuation rate of each frame segment are integrated to form a two-dimensional parameter pair, which serves as a quantitative indicator of the intensity of the motion of the frame segment. The quantitative indicator is compared with the boundary range of the preset risk level. If the fluctuation amplitude is in the small range and the rate is slow, it can be classified as a mild risk level; if the fluctuation amplitude is in the middle range or the rate is close to the critical value, it is classified as a moderate risk level; if the fluctuation amplitude is severe and the rate is fast, it is classified as a severe risk level. In real-world scenarios, users' center of gravity changes little during daily walking, so most frames are classified as mild risk. If the center of gravity fluctuation increases when going up and down stairs or avoiding obstacles, moderate or severe risk segments appear. Each frame segment is assigned a risk level label. The frame segment risk level labels form an ordered sequence, which is the risk level classification trend.

[0148] The response parameter setting submodule classifies the risk level and arranges the corresponding relationships between the frame segments and the buzzer, flashing light, and notification instructions according to the risk level. It also maps the time axis according to the start and end time of the time frame segments, summarizes the time line and arranges the sequence to generate a set of center of gravity fall risk response time periods.

[0149] Entering the response parameter setting stage, it is necessary to match the corresponding warning parameter combination for each frame segment. The warning intensity corresponding to each level is preset. Mild risk is set to a low-frequency buzzer sound, a slow flashing light frequency and a slightly longer notification delay; moderate risk corresponds to a medium-frequency buzzer and a flashing of a synchronous frequency, as well as a moderate response delay; severe risk requires a high-frequency buzzer to sound rapidly, a fast-frequency light to flash, and a notification within the shortest delay. In actual applications, a home fall monitoring device is set. When a frame segment is identified as a moderate risk level, the buzzer will be set to sound twice per second, the light will flash once per second, and an alarm message will be pushed within two seconds. According to the start and end time of the time frame segment, the parameter setting value is bound to the corresponding time period to construct a complete timeline mapping structure. The response parameters of the frame segment and the time period information are integrated into an ordered sequence and merged and arranged in chronological order, which is the warning response set caused by the center of gravity change under each frame segment, forming a center of gravity fall risk response time period set.

[0150] The above are merely preferred embodiments of the present invention and do not limit the present invention in any other form. Any technician familiar with the profession may use the technical content disclosed above to change or modify it into an equivalent embodiment with equivalent changes and apply it to other fields. However, any simple modification, equivalent change and modification made to the above embodiment based on the technical essence of the present invention without departing from the content of the technical solution of the present invention still falls within the scope of protection of the technical solution of the present invention.

Claims

1. A fall warning system for the elderly based on motion trajectory analysis, characterized in that: The system comprises: The center of gravity recognition module obtains continuous image frames of elderly people in indoor activity spaces, extracts the vertical coordinates of the waist and feet key points in the image frames, performs consistency judgment on the direction of difference changes between adjacent frames, and generates continuous segments of center of gravity offset; The trend recognition module extracts the spatial position path of the elderly person's waist key points within three adjacent frames based on the continuous segment of the center of gravity offset, detects whether there is an initial reverse switch, and determines the fluctuation continuity of the center of gravity change trend before and after the reversal point, thereby generating a center of gravity offset slope reversal mark frame segment; The inertial shift module calls the center of gravity offset slope reversal mark frame segment, extracts the vertical positioning sequence of the pelvis, shoulder, and knee key points, determines whether there is a segment with consistent direction among the three points, and calculates the segment length and switching frequency to generate the center of gravity inertial co-directional shift state; The phase judgment module extracts the motion vector sequences of the three groups of bone point pairs in the vertical and anterior-posterior axis directions of the hip and knee, knee and ankle, and shoulder and pelvis according to the time frame segment marked by the co-directional movement state of the center of gravity inertia, analyzes the changes in the angles of the bone point direction vectors between adjacent frames, and generates a center of gravity collaborative splitting fluctuation trend segment.

2. The elderly fall warning system based on motion trajectory analysis according to claim 1 is characterized in that: The center of gravity offset continuous segment segment includes a height change amplitude sequence, a change direction consistency label, and a continuous frame segment number; the center of gravity offset slope reversal mark frame segment includes a reverse switching frame index, a trend fluctuation continuity label, and a waist path segment number; the center of gravity inertia co-directional movement state includes the co-directional segment length, the number of direction maintenance times, and the switching event frequency; the center of gravity collaborative split fluctuation trend segment includes an angle change trend sequence, a bone point collaborative state identifier, and an action deviation area number.

3. The elderly fall warning system based on motion trajectory analysis according to claim 1 is characterized in that: The center of gravity recognition module includes: The image coordinate extraction submodule obtains continuous image frames of the elderly in the indoor activity space, detects the key points of the waist and feet in the image frames, extracts the vertical coordinates of the key points of the waist and feet in the image frames, and generates a vertical coordinate sequence of the waist and feet; The key point difference calculation submodule sequentially extracts the coordinate height differences of the same key points between the differentiated frames based on the waist and foot vertical coordinate sequence, calculates the average vertical distance between the waist point and the foot points in the image frame, and obtains the waist and foot height difference sequence; The difference change judgment submodule determines whether the difference change direction between adjacent image frames is consistent based on the waist-foot height difference sequence. If the difference change direction in consecutive image frames is the same, the segments are connected, the change consistency score is calculated, and combined with the sequence continuity, a continuous segment of the center of gravity offset is generated.

4. The elderly fall warning system based on motion trajectory analysis according to claim 3 is characterized in that: The trend identification module includes: The path extraction submodule extracts the spatial positions of the elderly person's waist key points in three adjacent frames based on the center of gravity offset continuous segment, calculates the two-dimensional spatial path formed by the three points, obtains the speed direction value between every two frames, and obtains the speed change direction sequence; The reverse identification submodule calls the direction vector changes between adjacent frame segments in the speed change direction sequence, compares the angle value between the real-time direction vector and the previous direction vector to see whether a reverse switch occurs, determines whether there is an initial direction reversal, selects the frame segment where the angle changes from an acute angle to an obtuse angle as a reverse point segment, records the position of the direction switching critical point, and obtains the first reverse switching frame segment; The trend determination submodule adopts the first reverse switching frame segment to obtain the center of gravity position coordinates of two consecutive frames before and after the reversal point, identifies the center of gravity longitudinal axis offset difference sequence, makes a difference judgment on the sequence fluctuation direction, calculates the fluctuation amplitude evaluation value, and compares the fluctuation amplitude evaluation value with the center of gravity fluctuation judgment benchmark. If the judgment is valid, it is determined to be a trend reversal frame segment, and a center of gravity offset slope reversal mark frame segment is obtained.

5. The elderly fall warning system based on motion trajectory analysis according to claim 4 is characterized in that: The inertial shift module includes: The center of gravity offset marking submodule calls the center of gravity offset slope reversal marking frame segment, detects the vertical change trend between adjacent frames for the center of gravity position point of each frame in the marked frame segment, calculates the offset slope value based on the vertical difference change rate of adjacent frames, and identifies the frame segment position where the slope changes from positive to negative or from negative to positive, thereby generating a center of gravity reversal frame sequence; The key point extraction submodule calls the center of gravity inversion frame sequence to extract the vertical positioning values ​​of three key points of the pelvis, shoulder, and knee in each frame, generates a vertical positioning sequence in consecutive frames for the key points, determines the upward or downward trend based on the direction of change of the vertical position coordinates of the key points in the consecutive frames, and generates a vertical direction sequence set; The direction sequence screening submodule selects the direction identifiers of the three key points of the pelvis, shoulder, and knee within the same time period based on the vertical direction sequence set, determines whether they are in the same direction, records the frame segments of continuous same-direction state, and counts the frame length of the segment and the frequency of direction switching between the previous and subsequent same-direction segments. The total number of continuous frames and the number of switching times of each same-direction movement are recorded to generate the center of gravity inertia co-directional movement state.

6. The elderly fall warning system based on motion trajectory analysis according to claim 5 is characterized in that: The phase judgment module includes: The bone point vector extraction submodule extracts the continuous spatial displacement of the three bone point pairs of hip and knee, knee and ankle, and shoulder and pelvis in the vertical and anterior-posterior directions based on the time frame segment marked by the co-directional displacement state of the center of gravity inertia, obtains the motion vector sequence at each frame time point, and calculates the unit time displacement of adjacent bone points in different directions to obtain the bone point direction vector sequence; The angle change trend identification submodule calls the bone point direction vector sequence, calculates the angle value change between the bone point direction vectors of the same group at adjacent time points, identifies the angle change direction, analyzes the angle change trend of each group of bone points in the full frame segment, and obtains the angle change trend parameter group; The collaborative fluctuation identification submodule analyzes the degree of movement synchronization and the state of coordination based on the angle change trend parameter group and whether the directions of the angle change curves at the same frame time points between the bone points are consistent and whether the change rates are in a similar range. It calculates the coordination amplitude of the synchronization angle, calls the segment continuity on the time frame axis, extracts the segment boundaries of the change direction, and obtains the center of gravity collaborative splitting fluctuation trend segment.

7. The elderly fall warning system based on motion trajectory analysis according to claim 1 is characterized in that: The system further comprises a level response module: The level response module calls the center of gravity collaborative splitting fluctuation trend segment and sets it as mild, moderate, and severe fall risk levels respectively. The differentiated levels are triggered in the corresponding frame segments and time-mapped. The correspondence between the differentiated levels and the buzzer ringing frequency, flashing light frequency, and notification command delay is evaluated to generate a center of gravity fall risk response time period set. The center of gravity fall risk response period set includes a risk level label, a buzzer frequency value, a flashing light cycle code, and a notification delay parameter.

8. The elderly fall warning system based on motion trajectory analysis according to claim 7 is characterized in that: The level response module includes: The time segmentation identification submodule collaboratively splits the fluctuation trend segments based on the center of gravity, extracts the time frame segments of the differentiated fluctuation segments, distinguishes the time frame segments by frame sequence numbers, identifies the interruption points of the fluctuation continuity and divides the frames into segments, calculates the time span value by the change in the fluctuation characteristics between consecutive frames, and obtains the time frame segment interval; The risk level judgment submodule calls the time frame interval and, based on a two-dimensional parameter set consisting of a center of gravity fluctuation amplitude value and a fluctuation rate value, judges the belonging to the fluctuation interval corresponding to the mild, moderate, and severe fall risk levels, and obtains a risk level classification trend; The response parameter setting submodule combines and arranges the correspondence between the frame segments and the buzzer, flashing light, and notification instructions according to the risk level classification trend, maps the time axis according to the start and end time of the time frame segment, summarizes the time line and arranges the sequence to generate a center of gravity fall risk response time period set.

Images