A method and system for monitoring high-risk groups in a nursing home based on medical and nursing combination

Abstract

This invention relates to the field of smart elderly care monitoring technology, and discloses a monitoring method and system for high-risk groups in elderly care institutions based on integrated medical and elderly care. By modeling the process of returning to the room after an activity as a complete care task, walking, pausing, turning, and stopping at the door in video images are no longer understood as ordinary actions in isolation, but are continuously analyzed under the constraint of whether the elderly person successfully returns to their own room. Features such as target consistency, head and body orientation differences, approaching the elderly person's own room door without entering, and stopping at the door of a non-elderly person's room are introduced. Before the elderly person falls or mistakenly enters another room, the system can identify whether the process of returning to the room has deviated. It can also restore the disorientation behavior that is fragmented under the condition of crowd obstruction to the same continuous event, and more accurately distinguish between normal return to the room, short-term waiting, and true disorientation and wandering. This solves the problem that existing monitoring methods are difficult to continuously identify the deviation of the intention of high-risk elderly people returning to their rooms and the disorientation and wandering process in the scenario of returning to the room after an activity.

Description

Technical Field

[0001] This invention relates to the field of smart elderly care monitoring technology, and in particular to a method and system for monitoring high-risk groups in elderly care institutions based on the integration of medical and elderly care. Background Technology

[0002] In the daily operation of integrated medical and elderly care facilities, some high-risk residents, while still able to participate in group rehabilitation training, recreational activities, post-meal activities, and short-term social activities, also experience cognitive impairment, decreased spatial orientation, bradygait, weakened executive function, sequelae of stroke, Parkinson's syndrome, or weakness during post-operative recovery. The risks for these individuals do not always manifest directly as falls, but often gradually emerge during a routine that should be completed normally. For example, after group activities, they may deviate from their intended path, stop incorrectly, repeatedly turn back, search for their way around, or remain in their rooms for extended periods.

[0003] The corridor environment after an activity is typically characterized by dense crowds, frequent obstructions, obvious path forks, and individuals following each other. Some high-risk elderly individuals, after leaving the activity room, may initially follow the crowd ahead, but gradually lose their sense of direction after the crowd disperses or the reference point disappears. This can lead to behaviors such as passing by their own room without entering, hesitating in front of other people's rooms, pacing back and forth at forks in the road, and briefly lingering near the nursing station before continuing on the wrong path. Because these behaviors may not necessarily manifest as falls or obvious imbalances in the early stages, caregivers often struggle to detect them promptly through manual patrols alone. Traditional monitoring methods based on fall detection, boundary crossing detection, or prolonged static detection are also insufficient to accurately capture the complete process of deviation from the intended return to the room and the subsequent disorientation and wandering. In existing technologies, general behavior recognition methods typically only classify actions such as walking, standing, and stopping, lacking explicit constraints on the care target of returning to one's own room. Therefore, it is difficult to judge the risk from the angle of deviation from the target room. Ordinary target tracking methods are also prone to interruption under the condition of crowd obstruction after the event, which cuts the same disorientation process into multiple discontinuous small segments, reducing the effectiveness of continuous recognition.

[0004] Therefore, how to continuously identify and provide early warnings of the risk of high-risk individuals wandering around aimlessly during their return to their rooms after activities remains a technical problem that needs to be solved in the field of smart monitoring of integrated medical and elderly care institutions. Summary of the Invention

[0005] This invention provides a method and system for monitoring high-risk individuals in elderly care institutions based on the integration of medical and elderly care, in order to at least solve the problem that existing monitoring methods are unable to continuously identify the deviation of high-risk elderly individuals' intentions and their wandering or disoriented processes when returning to their rooms after activities.

[0006] To achieve the above objectives, this invention provides a method for monitoring high-risk populations in elderly care institutions based on integrated medical and elderly care, the method comprising the following steps: The system acquires video frame data, scene area configuration data, and room-specific data corresponding to the public activity areas and residential areas of elderly care institutions. Based on the video frame data, it performs image recognition and behavior monitoring to generate data for monitoring the return to room of high-risk individuals. Based on the data from the room return monitoring task, extract and form the room return behavior characteristic data of high-risk individuals during the process of returning to their own rooms after the event; Based on the room return behavior feature data, candidate disorientation fragment data of the high-risk object during the process of returning to their own room are identified and formed, and cross-interruption association processing is performed on the candidate disorientation fragment data to form continuous wandering chain data. Based on the continuous wandering chain data, the corresponding room data of the high-risk object, and the historical safe return path data, continuous risk identification processing is performed to generate the disorientation risk score data corresponding to the high-risk object. Based on the disorientation risk score data, an early warning judgment is performed, and when the judgment meets the preset conditions, a nursing prompt result containing the current area location, target room information, and path indication information is output.

[0007] Optionally, data for monitoring the return to room of high-risk individuals can be generated, specifically including: Acquire scene area configuration data for the activity room exit area, main corridor area, corridor branch area, area near the user's room door, area near the non-user's room door, and area near the nursing station, and establish a semantic topology relationship for returning to the room based on the scene area configuration data; At the end of the activity, human recognition and target tracking are performed on the video frame data corresponding to the public activity area and residential area of ​​the elderly care institution to identify human targets leaving the exit area of ​​the activity room and extract the identity representation information of the corresponding human targets. The identity information is matched with the data corresponding to the rooms of high-risk individuals to determine the corresponding high-risk individuals and their rooms. The time when the high-risk individuals leave the activity room exit area is determined as the start time of the return-to-room monitoring task. Based on the high-risk object, its room, and the start time, return-to-room monitoring task data is generated; wherein, the return-to-room monitoring task data includes at least the object identifier, the target room identifier, the task start time, and the scene area index relationship.

[0008] Optionally, based on the room return monitoring task data, extract and form room return behavior characteristic data of high-risk individuals during their return to their own rooms after the activity, specifically including: Continuous tracking is performed on the video frame data corresponding to the room return monitoring task data to extract the center position sequence of the high-risk object, and continuous trajectory data is formed based on the center position sequence; The movement direction and speed data of the high-risk object are calculated based on the changes in the center position at adjacent times to determine the target consistency coefficient. Head orientation data and body orientation data are extracted based on head key points and shoulder key points to form orientation search data; Based on the spatial relationship between the high-risk object and the area adjacent to the object's own door and the area adjacent to the non-object's door, data on the object's approach to the object's door without entering and data on the non-object's door stay are extracted to form door approach behavior data. Based on the continuous trajectory data, the target consistency coefficient, the orientation search data, and the door approach behavior data, return-to-room behavior feature data corresponding one-to-one with the high-risk objects are formed.

[0009] Optionally, based on the room return behavior characteristic data, candidate disorientation fragment data of the high-risk individual during the process of returning to their own room are identified and generated, specifically including: Within a sliding time window, the number of directional reversals, the degree of repeated arrivals in local areas, and the cumulative duration of low-speed dwelling for the high-risk objects are counted to form data on return behavior, local loop behavior, and abnormal dwelling behavior. The target consistency coefficient, orientation search data, backtracking behavior data, local loop behavior data, abnormal stay behavior data, data on approaching but not entering one's own room door, and data on staying at other people's room doors are fused together to form temporal disorientation candidate score data. Time intervals that continuously exceed a preset score threshold are identified as candidate misdirected segments, and candidate misdirected segment data are formed based on the start time, end time, dominant region, and segment feature summary of the candidate misdirected segments.

[0010] Optionally, cross-interruption association processing is performed on the candidate lost fragment data to form continuous wandering chain data, specifically including: For temporally adjacent or nearby candidate maze segments, extract the segment end position prediction result, segment start position result, object appearance representation result, target room deviation result, dominant region result, and candidate maze score result; Spatial continuity data is formed based on the predicted end position of the segment and the start position of the segment; appearance consistency data is formed based on the object appearance characterization results; target deviation continuity data is formed based on the target room deviation results; regional transfer rationality data is formed based on the dominant region results; and risk continuity data is formed based on the candidate disorientation score results. The spatial continuity data, appearance consistency data, target deviation continuity data, regional transfer rationality data, and risk continuity data are fused to form the cross-interruption correlation degree between candidate disorientation segments; When the cross-interruption correlation degree meets the preset correlation conditions, the corresponding candidate wandering segments are spliced ​​into the same continuous wandering chain, and continuous wandering chain data is formed.

[0011] Optionally, based on the continuous wandering chain data, the room data corresponding to the high-risk object, and the historical safe return path data, continuous risk identification processing is performed to generate disorientation risk score data corresponding to the high-risk object, specifically including: Based on the continuous loitering chain data, the following features are extracted: duration of failure to return to the room after leaving the activity room, cumulative amount of abnormal lingering at the door, cumulative number of backtracking, cumulative degree of target deviation, and cumulative degree of search behavior, forming chain-level summary feature data; Based on historical safe return-to-room path data, short-term waiting sample data, and nursing confirmation disorientation sample data, prototypes of normal return-to-room behavior, short-term waiting behavior, and disorientation wandering behavior were constructed respectively. The behavioral representation data at the current moment is matched with the normal return-to-room behavior prototype, the short-term waiting behavior prototype, and the disoriented wandering behavior prototype respectively to form the corresponding prototype matching coefficient data. Based on the prototype matching coefficient data, deviation correction is performed on the behavior representation data at the current moment, and the current corrected state data is formed by combining it with the state data at the previous moment. Misconception risk scores are generated based on the current correction status data and chain-level summary feature data.

[0012] Optionally, based on the current correction status data and chain-level summary feature data, a disorientation risk score is formed, specifically including: The previous moment's confusion risk score, the current correction status data, and the chain-level summary feature data are weighted and accumulated to form the current moment's confusion risk score; The disorientation risk score is compared with the individual risk threshold data of the corresponding high-risk object to form risk level data; When the risk level data reaches the preset alert level, the time-series evidence data corresponding to the current continuous wandering chain is retained.

[0013] Optionally, an early warning determination is performed based on the disorientation risk score data, and when the determination meets preset conditions, a nursing prompt result containing the current area location, target room information, and path indication information is output, specifically including: A tiered early warning system is generated based on disorientation risk scores and individual risk thresholds. When the graded early warning result reaches the preset level, video segment data of the time period before and after the alarm is extracted from the corresponding continuous wandering chain, and the current semantic region data, target room data and object center trajectory data are extracted. Path indication data is generated based on the object center trajectory data, and video clip data, current semantic region data, target room data and path indication data are combined to form nursing prompt results; The nursing prompt results are output to the nursing terminal, and the nursing feedback data returned by the corresponding nursing terminal is received.

[0014] Optional methods for monitoring high-risk groups in integrated medical and elderly care facilities also include: Based on the nursing feedback data and the trajectory data of this ward return monitoring task, the historical safe ward return path data is updated to form the updated historical safe ward return path data; Based on the false alarm results, effective early warning results, and high-risk results in the nursing feedback data, the individual risk threshold data is rolled over to form updated individual risk threshold data. When a new continuous wandering chain is formed again within a preset time range for the same high-risk object, the system determines whether it belongs to the same persistent disorientation event based on the temporal adjacency, path overlap, and target deviation consistency between the new and old continuous wandering chains, and inherits the previous warning status when it is determined to belong to the same persistent disorientation event.

[0015] Furthermore, to achieve the above objectives, the present invention also provides a high-risk population monitoring system for elderly care institutions based on integrated medical and elderly care, comprising: The acquisition module is used to acquire video frame data, scene area configuration data, and room corresponding data of high-risk individuals in the public activity areas and residential areas of the elderly care institution. Based on the video frame data, image recognition and behavior monitoring are performed to form the corresponding room return monitoring task data for high-risk individuals. The extraction module is used to extract and form the return-to-room behavior feature data of high-risk individuals during the process of returning to their own rooms after the event, based on the return-to-room monitoring task data; The identification module is used to identify and form candidate maze fragment data of the high-risk object during the process of returning to its own room based on the return behavior feature data, and to perform cross-interruption association processing on the candidate maze fragment data to form continuous wandering chain data. The forming module is used to perform continuous risk identification processing based on the continuous wandering chain data, the room corresponding data of the high-risk object, and the historical safe return path data, so as to form the disorientation risk score data corresponding to the high-risk object; The output module is used to perform early warning judgment based on the disorientation risk score data, and output a nursing prompt result containing the current area location, target room information and path indication information when the judgment meets the preset conditions.

[0016] The beneficial effects of this invention are as follows: It proposes a monitoring method and system for high-risk groups in elderly care institutions based on the integration of medical and elderly care. By modeling the process of returning to the room after an activity as a complete care task, walking, pausing, turning, and stopping at the door in video images are no longer understood as ordinary actions in isolation, but are continuously analyzed under the goal constraint of whether the elderly person has successfully returned to their own room. Features such as goal consistency, head and body orientation differences, approaching the elderly person's own room without entering, and stopping at the door of a non-elderly person's room are introduced. This allows for the identification of whether the return process has deviated before the elderly person falls or mistakenly enters another room. By using cross-interruption association to form a continuous wandering chain, the disoriented behaviors fragmented under the condition of crowd obstruction can be restored to the same continuous event. The continuous identification method based on behavior prototype matching and deviation correction can more accurately distinguish between normal return to the room, short-term waiting, and true disoriented wandering, thereby improving the practicality of the early warning results. Finally, the system outputs nursing prompts containing information on the current area location, target room, and path indication, and updates historical safe return paths and individual thresholds based on nursing feedback, so that the system can better meet the actual nursing needs of the institution after long-term operation. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating a method for monitoring high-risk populations in elderly care facilities based on integrated medical and elderly care, as described in an embodiment of the present invention. Figure 2 This is a schematic diagram of the structure of a high-risk population monitoring system for elderly care institutions based on the integration of medical and elderly care, according to an embodiment of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the 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 merely illustrative and not intended to limit the invention.

[0019] This invention provides a method for monitoring high-risk groups in elderly care institutions based on integrated medical and elderly care, referring to... Figure 1 As shown, it includes the following steps: S1: Acquire video frame data, scene area configuration data, and room corresponding data of high-risk individuals in the public activity areas and residential areas of the elderly care institution. Based on the video frame data, perform image recognition and behavior monitoring to form the corresponding room return monitoring task data for high-risk individuals.

[0020] Specifically, the system acquires scene area configuration data for the activity room exit area, main corridor area, corridor branching area, area near the user's room door, area near non-user's room door, and area near the nursing station, and establishes a semantic topology relationship for returning to the room based on the scene area configuration data. During the activity's end period, it performs human recognition and target tracking on video frame data corresponding to the public activity area and residential area of ​​the elderly care facility to identify human targets leaving the activity room exit area and extract the corresponding human target's identity representation information. It then matches this identity representation information with data corresponding to the rooms of high-risk individuals to determine the corresponding high-risk individuals and their rooms, and determines the time when the high-risk individuals leave the activity room exit area as the start time of the return-to-room monitoring task. Based on the high-risk individuals, their rooms, and the start time, it generates return-to-room monitoring task data. This return-to-room monitoring task data includes at least an object identifier, a target room identifier, a task start time, and a scene area index relationship.

[0021] In this embodiment of the invention, the video frame data is used to reflect the continuous image process of a high-risk individual leaving the activity room, entering the corridor, and continuing to return to their own room after the activity ends; the scene area configuration data is used to reflect the spatial distribution relationship of the activity room exit, main corridor, corridor branches, the area near the individual's room door, the area near the non-individual's room door, and the area near the nursing station; the high-risk individual's room correspondence data is used to reflect the one-to-one correspondence between the high-risk individual and their respective room. By uniformly organizing the above data, return-to-room monitoring task data oriented towards specific individuals, specific rooms, and specific return-to-room tasks can be formed, providing support for subsequent trajectory extraction, behavior recognition, and risk assessment.

[0022] It's easy to understand that if only ordinary surveillance video is acquired without establishing the relationship between the task's starting point and the target room, subsequent processing can only identify whether someone is walking or stopping in the video, but cannot identify whether the high-risk individual is deviating from their designated room. Therefore, in this step, the semantic topology relationship for returning to the room is first established based on the scene area configuration data. Specifically, the activity room exit area can be used as the starting area for the return-to-room task, the main corridor area as the continuous passage area, the corridor branching area as the direction-sensitive area, the area near the individual's room door as the target completion area, the area near non-the individual's room door as the error-prone area, and the area near the nursing station as a prompt area that facilitates manual intervention and observation. Furthermore, in one executable implementation, the above areas can be organized into a set of directed path relationships according to the actual corridor structure to ensure that when the subsequent system makes area jump judgments, it not only knows which area the target is currently in, but also whether the area conforms to the normal return-to-room process with the previous area and the subsequent target area.

[0023] In practical implementation, human recognition and target tracking can be performed on video frame data during the end of the activity. Human recognition here is not limited to a specific network structure; as long as it can stably separate individual human targets from the video frames and output target area information and identity representation information for subsequent continuous tracking, it is acceptable. Further, the identified target identity representation is matched with the data corresponding to the rooms of high-risk individuals to determine whether the object leaving the activity room exit area is a high-risk object and what its corresponding target room is. It should be noted that the return-to-room monitoring task described in this invention is only triggered when the object is identified as a high-risk object and its center position crosses the boundary of the activity room exit area. The resulting return-to-room monitoring task data includes at least the object identifier, target room identifier, task start time, and scene area index relationship.

[0024] In one specific implementation, a daytime activity area and its corresponding residential corridor in a senior care facility can be selected as the monitoring scenario. The activity room exit, main corridor, branching points, areas near each room door, and areas near the nursing station are pre-numbered. When a high-risk individual leaves through the activity room exit after an activity, the system reads their identity template and their room identifier, creating a return-to-room monitoring task record, for example, recorded as "Object A, Target Room R12, Task Start Time: 19:08:12, Task Starting Area: Activity Room Exit Area". This method ensures that subsequent processing always revolves around the correspondence between the individual and their room, preventing confusion caused by the actions of other elderly residents or caregivers in the footage.

[0025] S2: Based on the room return monitoring task data, extract and form the room return behavior feature data of high-risk individuals during the process of returning to their own rooms after the event.

[0026] Specifically, continuous tracking is performed on the video frame data corresponding to the room return monitoring task data to extract the center position sequence of the high-risk object, and continuous trajectory data is formed based on the center position sequence; the movement direction data and movement speed data of the high-risk object are calculated based on the center position changes at adjacent times to determine the target consistency coefficient; head orientation data and body orientation data are extracted based on head key points and shoulder key points to form orientation search data; based on the spatial relationship between the high-risk object and the adjacent area of ​​its own room door and the adjacent area of ​​other people's rooms door, data on approaching but not entering the own room door and data on staying at other people's rooms door are extracted to form room approach behavior data; based on the continuous trajectory data, the target consistency coefficient, the orientation search data, and the room approach behavior data, room return behavior feature data corresponding one-to-one with the high-risk object is formed.

[0027] In this embodiment of the invention, based on the room return monitoring task data formed in step S1, the continuous trajectory, orientation search, and door approach behavior of the high-risk object during its return to its own room after the activity are extracted to form room return behavior feature data. Specifically, the continuous trajectory data reflects the object's position change over time within the corridor space; the orientation search data reflects whether the object is walking while searching, stopping to search, or looking left and right at the door; and the door approach behavior data reflects whether the object has approached its own door, whether it has entered after approaching, and whether it lingers in front of a door other than its own. By organizing these three types of data together, a behavioral feature set can be formed, providing fine-grained input for subsequent candidate maze segment identification.

[0028] It's easy to understand that in the scenario of returning to one's room after an event, the risk of disorientation cannot be determined solely by speed, as some high-risk elderly individuals naturally walk slowly; similarly, the duration of pauses also fails to assess risk, as even short waits or yielding to others can cause pauses. Therefore, this step first extracts continuous trajectory data. Specifically, continuous tracking can be performed on the video frame sequence corresponding to the room return monitoring task to obtain the center position of the high-risk object at each moment, and then a sequence of center positions can be constructed from these continuous center positions. Furthermore, in one feasible implementation, the object's direction of movement can be obtained by subtracting the center positions from adjacent moments, and the object's speed can be obtained by considering the distance between adjacent positions and the time relationship between frames. It should be noted that the direction and speed of movement here are not final conclusions, but rather foundational data for subsequent target consistency, turnaround statistics, and abnormal dwell time statistics.

[0029] To determine whether an object's current movement is still directed towards its own room, this invention introduces a target consistency coefficient, which can be calculated using the following formula: ; in, The target consistency coefficient is represented by the coefficient of time t. This represents the actual direction of movement at time t; This represents the target direction vector pointing from the current center position to the position of the user's room door; Represents the magnitude of the actual movement direction vector; ε represents the magnitude of the target direction vector; ε represents a minimal constant to prevent the denominator from being zero. It is easy to understand that when... When the value is close to 1, it indicates that the object's current movement direction is relatively consistent with the direction of the user's door; when... When the value is close to 0, it means that the object's current movement direction is basically unrelated to the direction of the user's door; when... A negative value indicates that the object's current direction of movement has deviated from the direction of the door. This formula transforms the nursing semantic of whether the object is currently moving towards its own room into a computable quantity of directional consistency.

[0030] Building upon this, the present invention further extracts orientation search data. Specifically, head orientation and body orientation can be estimated based on head key points and shoulder key points respectively, and the difference between the two can be used to determine whether there is significant search behavior. The head-body orientation difference can be calculated using the following formula: ; in, This indicates the degree of difference in head and body orientation at time t; This indicates the head-facing angle at time t; The body orientation angle at time t is shown. It should be noted that the head orientation angle can be obtained from the direction information of the head key points, and the body orientation angle can be obtained from the shoulder key points or the direction of the shoulder-hip line. Furthermore, when... When the duration of multiple consecutive moments is large, it often indicates that although the object is still moving, its attention has shifted from returning to the room to searching left and right for the door; when the object stops in a certain area and A sustained increase in this value is more likely to indicate that the user is confirming a address or searching for direction. It should be noted that this value, along with the target consistency coefficient, constitutes a joint characteristic of directional deviation and search behavior.

[0031] Furthermore, this invention also extracts door approach behavior data. It's easy to understand that disorientation risk isn't just about not heading towards a room, but also about approaching a door but not whether it's one's own door, and whether one actually enters. Therefore, in practical applications, a proximity determination zone can be set outside each door's vicinity, and an entry determination line can be set at the door frame. When an object enters its own door's vicinity zone, the system begins accumulating its proximity and dwell time related to its own door. If the object does not cross its own door's entry determination line during the accumulation period, it is recorded as "approaching but not entering" data for its own door. Similarly, if the object enters a non-own door's vicinity zone and remains there, it is recorded as "staying at a non-own door" data. In one executable implementation, the "approaching but not entering" data for one's own door can be composed of both proximity level and duration of proximity without entry, while the "staying at a non-own door" data can be composed of both the distance from the target to the corresponding non-own door and the duration of stay. Through this design, the system can identify potential risks of an object lingering in front of the wrong door or reaching its own door but not entering before it mistakenly enters another's room.

[0032] In one specific implementation, after a high-risk individual leaves the activity room, they move along the main corridor towards their room. The system first continuously extracts their center position from video frames to form a trajectory; then it calculates the target consistency coefficient at several consecutive moments, identifying the initial segment of their journey... Maintaining a high level, but after approaching the bifurcation point of the corridor, The object gradually descended. Simultaneously, the system estimated the difference in head and body orientation using head and shoulder keypoints, revealing more pronounced left-right search movements at the bifurcation point. Subsequently, the object entered the area near its own room door but did not enter, continuing forward and pausing for several seconds in front of a door that was not its own. This process demonstrates that speed or pauses alone are insufficient to explain the issue; rather, the continuous trajectory, search orientation, and door approach behavior, when combined, clearly indicate abnormalities in its return-to-room process.

[0033] S3: Based on the room return behavior feature data, identify and form candidate disorientation fragment data of the high-risk object during the process of returning to its own room, and perform cross-interruption association processing on the candidate disorientation fragment data to form continuous wandering chain data.

[0034] Specifically, within a sliding time window, the number of directional reversals, the degree of repeated arrivals in local areas, and the cumulative duration of low-speed stays of the high-risk object are statistically analyzed to form backtracking behavior data, local looping behavior data, and abnormal stay behavior data. The target consistency coefficient, orientation search data, backtracking behavior data, local looping behavior data, abnormal stay behavior data, data on approaching but not entering one's own room door, and data on staying at other people's room doors are fused to form temporal disorientation candidate score data. Temporal intervals that continuously exceed a preset score threshold are identified as candidate disorientation segments, and candidate disorientation segment data are formed based on the start time, end time, dominant region, and segment feature summary of the candidate disorientation segments.

[0035] Based on this, for temporally adjacent or nearby candidate maze segments, the following data are extracted: segment end position prediction result, segment start position result, object appearance representation result, target room deviation result, dominant region result, and candidate maze score result. Spatial continuity data is formed based on the segment end position prediction result and segment start position result; appearance consistency data is formed based on the object appearance representation result; target deviation continuity data is formed based on the target room deviation result; regional transfer rationality data is formed based on the dominant region result; and risk continuity data is formed based on the candidate maze score result. The spatial continuity data, appearance consistency data, target deviation continuity data, regional transfer rationality data, and risk continuity data are fused to form the cross-interruption correlation degree between candidate maze segments. When the cross-interruption correlation degree meets the preset correlation conditions, the corresponding candidate maze segments are spliced ​​into the same continuous wandering chain, forming continuous wandering chain data.

[0036] In this embodiment of the invention, after extracting the room-returning behavior feature data in step S2, target deviation behavior, turning back behavior, local looping behavior, and abnormal dwelling behavior are identified based on the room-returning behavior feature data to form candidate disorientation segment data. Cross-interruption correlation processing is then performed on the candidate disorientation segment data to form continuous wandering chain data. The purpose of this step is to first filter out disorientation precursor segments worthy of further analysis from a continuous video stream containing a large number of normal room-returning behaviors, and then restore segments interrupted by crowd obstruction or short-term interruptions into the same continuous wandering chain. The advantages of this approach are twofold: firstly, it reduces the computational burden of directly performing deep judgments on all video frames; secondly, it ensures that continuous disorientation events are not interrupted by short-term obstruction.

[0037] It's easy to understand that a large number of normal waiting, yielding, stopping, and turning behaviors will occur simultaneously in the corridor after the event. Including all these ordinary behaviors in high-risk identification could easily lead to false alarms. Therefore, this step first counts the object's backtracking, partial looping, and abnormal stopping behaviors within a sliding time window. Backtracking behavior can be counted by comparing whether the angle between consecutive displacement vectors is close to reversing; partial looping behavior can be counted by observing whether the object returns to the same small area multiple times within a certain period; and abnormal stopping behavior can be identified by accumulating the duration of low speed or near-stationary conditions. Furthermore, in one executable implementation, the number of backtracking actions can be... Defined as the number of times the direction is significantly reversed within the current sliding time window, the degree of local loopback is... Defined as the degree of repetition and proximity of trajectory points within the current window to the current center position, and includes the cumulative amount of abnormal dwell times. Defined as the cumulative duration during which the speed is below a preset threshold within the current window.

[0038] After obtaining the above behavioral quantities, this invention integrates the target consistency coefficient, head and body orientation difference, number of backtracking attempts, local looping degree, cumulative abnormal dwell time, number of times the user approaches but does not enter their own room, and number of times the user dwells at a non-user's room as a candidate disorientation score. The candidate disorientation score can be calculated using the following formula: ; in, Represents the candidate misdirection score at time t; This represents the target consistency coefficient; Indicates the degree of difference in head and body orientation; Indicates the number of turnarounds; Indicates the degree of local loop closure; Indicates the cumulative amount of abnormal stays; This indicates that the person's room door is close but has not yet been entered; This indicates the number of times someone else stayed at the door; to This represents the weight of each feature in the candidate disorientation score. It's easy to understand that in the formula... This is used to unify the degree of directional deviation with other risk characteristics to the same direction of increased risk; the more the object's direction deviates from the person's door... The larger the value, the more obvious it becomes when searching for the object while walking around. The larger the value, the more pronounced the object's back-and-forth movement, partial looping, or lingering. , , The larger the value; when the object does not enter the room in front of the owner's door or lingers in front of a room that is not the owner's door. , The larger the value, the better. Therefore, this formula can fuse multiple weak anomalous precursors into a single candidate maze score that can change continuously over time.

[0039] Based on this, the system segments time intervals that continuously exceed a preset score threshold into candidate misdirection segments. It should be noted that, to avoid a large number of fragmented candidate segments caused by single-frame noise or instantaneous head turns, one executable implementation can set dual conditions: a continuous threshold exceeding time and a continuous fallback time. That is, a segment start point is formed only when the candidate misdirection score is continuously above the threshold for a certain duration, and a segment end point is formed only when the candidate misdirection score is continuously below the fallback threshold for a certain duration. This approach improves the stability of candidate segment boundaries. The obtained candidate misdirection segment data includes at least the start time, end time, dominant region, and segment summary, thus providing structured input for subsequent cross-interruption correlation.

[0040] Furthermore, a significant challenge in the scenario of returning to the room after an event lies in the fact that a single disorientation process is often briefly obstructed by others, carts, door frames, or corners, causing interruptions between candidate segments. If the system treats each interruption as the end of an event, a disorientation process lasting several minutes may be broken down into several unrelated small segments, leading to cumulative risk and failure. Therefore, this invention, after forming candidate disorientation segments, continues to perform cross-interruption association processing on temporally adjacent or near-adjacent candidate segments. This cross-interruption association processing simultaneously considers the spatial continuity, appearance consistency, target deviation continuity, reasonableness of area transfer, and risk continuity between segments.

[0041] To quantify the degree of cross-interrupt correlation between segments, this invention can introduce the following formula for calculating the degree of cross-interrupt correlation: ; in, Indicates the first The candidate disorientation fragment and the first Cross-interruption correlation between candidate misdirected segments; It indicates spatial continuity and is used to reflect whether the trend of the end position of the previous segment can reasonably predict the starting position of the next segment. Indicates visual consistency, used to reflect whether two segments belong to the same object; It indicates the continuity of the target deviation and is used to reflect whether the deviation state of the two segments relative to the person's room door continues; Indicates the rationality of regional transitions, reflecting whether the jump between the dominant areas of two segments conforms to the topological relationship between the corridor and the door. It indicates the continuity of risk and is used to reflect whether the degree of confusion between two consecutive segments has a continuous characteristic; to This represents the weights of each item. It should be noted that the aforementioned five types of continuity do not necessarily have to use the same unit of measurement; they can be normalized before calculation so that the final correlation degree can be directly used for unified threshold judgment.

[0042] Furthermore, in one executable implementation, the potential forward position trend of a candidate segment can be estimated by first estimating the displacement of the last few frames of the preceding candidate segment, and then compared with the center position of the first frame of the following candidate segment to establish spatial continuity; similarity can be calculated from the human appearance representation results of the preceding and following segments to establish appearance consistency; target deviation continuity can be extracted from the target consistency change trend and the association results of the user's door and non-user's door in the preceding and following segments; whether the preceding and following segments are still within a reasonable corridor range can be determined from the semantic regions to establish the rationality of region transfer; and the continuity of the change in the candidate disorientation scores of the preceding and following segments can be used to determine whether their risks have a successor relationship. When the cross-interruption correlation is higher than a preset threshold and the interruption duration does not exceed the tolerable duration, the system will splice the corresponding candidate segments into the same continuous wandering chain.

[0043] After forming a continuous wandering chain, the system further extracts chain-level summary features. These chain-level summary features include at least the total duration of not returning to the room after leaving the activity room, the cumulative amount of abnormal lingering at the door, the cumulative number of backtracking attempts, the cumulative degree of target deviation, and the cumulative degree of search behavior. It should be noted that these chain-level summary features no longer describe the local performance of a single frame, but rather the overall state of a complete and continuous event. For example, the cumulative amount of abnormal lingering at the door is not a single instance of lingering at a particular door, but rather the combined amount of not entering the user's own door and lingering at non-user's doors throughout the entire chain; similarly, the cumulative degree of target deviation is not a directional error at a single moment, but rather the comprehensive result of not effectively advancing towards the user's own door for an extended period throughout the entire chain.

[0044] In one specific implementation, a high-risk individual, after leaving the activity room, initially proceeds in the correct direction, then pauses briefly at a fork in the road and turns to search. They continue forward but miss their own room door, pausing for several seconds in front of another door. They are then obstructed by a nursing cart for approximately two seconds, before turning back at an adjacent location. The system first generates two candidate maze segments based on factors such as decreased target consistency, increased head and body orientation differences, failure to enter their own room door, and pausing at a door other than their own. Next, the system connects the two segments into a single continuous wandering chain based on spatial continuity, visual consistency, similar target deviation trends, and regional transition continuity, using cross-interruption association. Thus, this step not only filters out abnormal precursors from a continuous flow but also reconstructs the actual, ongoing maze process in complex corridor environments.

[0045] S4: Based on the continuous wandering chain data, the corresponding room data of the high-risk object, and the historical safe return path data, perform continuous risk identification processing to form the disorientation risk score data corresponding to the high-risk object.

[0046] Specifically, based on continuous wandering chain data, the duration of failure to return to the room after leaving the activity room, the cumulative amount of abnormal lingering at the door, the cumulative number of backtracking attempts, the cumulative degree of target deviation, and the cumulative degree of search behavior are extracted to form chain-level summary feature data. Based on historical safe return-to-room path data, short-term waiting sample data, and nursing confirmation of disorientation sample data, normal return-to-room behavior prototypes, short-term waiting behavior prototypes, and disorientation wandering behavior prototypes are constructed respectively. The behavioral representation data at the current moment is matched with the normal return-to-room behavior prototype, short-term waiting behavior prototype, and disorientation wandering behavior prototype to form corresponding prototype matching coefficient data. Based on the prototype matching coefficient data, deviation correction is performed on the behavioral representation data at the current moment, and combined with the state data at the previous moment to form the current corrected state data. Based on the current corrected state data and the chain-level summary feature data, disorientation risk score data is formed.

[0047] Furthermore, based on the current correction status data and chain-level summary feature data, a disorientation risk score is formed. Specifically, this includes: weighting and accumulating the disorientation risk score from the previous moment, the current correction status data, and the chain-level summary feature data to form the disorientation risk score at the current moment; comparing the disorientation risk score with the individual risk threshold data of the corresponding high-risk object to form risk level data; and retaining the temporal evidence data corresponding to the current continuous wandering chain when the risk level data reaches a preset alert level.

[0048] In this embodiment of the invention, after forming continuous wandering chain data, continuous risk identification processing is performed based on the continuous wandering chain data, the corresponding room data, and historical safe return-to-room path data to form disorientation risk score data for corresponding high-risk objects. The core of the continuous risk identification processing is that it does not directly make a one-time judgment on the strength of the anomaly in the current frame, but first matches the current behavior against three typical behavior reference modes: normal return to room, short-term waiting, and disorientation wandering. Then, it combines the state at the previous moment for deviation correction, and finally incorporates the corrected current state and the overall deterioration degree of the entire continuous wandering chain into the risk inertia accumulation. Through this processing, it is possible to more accurately distinguish between a temporary pause and continuous disorientation, avoiding falsely reporting ordinary waiting as high risk, and also avoiding repeatedly resetting continuous disorientation after breaking it up.

[0049] It's easy to understand that one of the biggest challenges in the scenario of returning to one's room after an event lies in the lack of a clear single-frame boundary between normal waiting, temporary hesitation, and genuine disorientation. To address this, this invention first constructs three behavioral prototypes in this step: a normal return-to-room behavior prototype, a short-term waiting behavior prototype, and a disoriented wandering behavior prototype. The normal return-to-room behavior prototype can be extracted from previously confirmed safe return paths and behavioral fragments; the short-term waiting behavior prototype can be extracted from common but non-risk-posing short pauses, yielding, and observation samples after an event ends; and the disoriented wandering behavior prototype can be summarized from samples of genuine disorientation events confirmed by nursing staff. It should be noted that the prototypes mentioned here are not literally limited to a specific model name; their essence is a reference representation of three typical behavioral patterns in the scenario.

[0050] Specifically, the current behavioral representation data is first matched with the three types of behavioral prototypes mentioned above to form the current matching coefficients for the three types of prototypes. The prototype matching coefficients can be calculated using the following formula: ; in, Represents the matching coefficient for the k-th type of behavior prototype at time t; Represents the current behavior representation vector at time t; Represents the prototype vector of the k-th behavior; Represents the similarity function; The temperature parameter is represented; the value of k ranges from 1 to 3, corresponding to the normal return-to-room behavior prototype, the short-term waiting behavior prototype, and the disoriented wandering behavior prototype, respectively. It's easy to understand that this formula allows the current behavioral representation to not be forcibly categorized into only one type, but to maintain varying degrees of similarity to all three prototypes simultaneously. For example, behavior at a certain moment may partially resemble normal return-to-room behavior, partially resemble short-term waiting behavior, and slightly resemble disoriented wandering behavior, thus more realistically reflecting the gradual process of an elderly person transitioning from a normal state to a risky state.

[0051] After obtaining the prototype matching coefficients, the system further performs deviation correction on the current behavioral representation based on these coefficients, and combines this with the state data from the previous time step to form the current corrected state. The current corrected state can be calculated using the following formula: ; in, Represents the current corrected state vector at time t; Let α represent the corrected state vector from the previous time step; α represent the state inertia coefficient, used to describe the degree of inheritance of the previous state from the current state; and β represent the current behavior representation retention coefficient, used to describe the direct contribution of the current real-time observation information to the corrected state. It should be noted that this equation obtains a reference state by weighting the three types of behavior prototypes according to the current matching coefficient. This reference state reflects the relative position of the current behavior among the three typical modes. Furthermore, by introducing the state vector from the previous time step... This avoids the system from drastically changing its judgment due to instantaneous changes in one or two frames, making it more suitable for recognizing gradual processes such as persistent disorientation.

[0052] In this embodiment of the invention, further, after forming the current correction state, the system does not immediately use it as the final risk. Instead, it continues to input the current correction state and the chain-level summary features of the continuous wandering chain into the risk inertia accumulation process to form the final disorientation risk score. The disorientation risk score can be calculated using the following formula: ; in, This represents the disorientation risk score at time t; This represents the disorientation risk score at the previous moment; μ represents the risk inertia coefficient. denoted by 'activation function'; w represents the output layer weight vector; b represents the bias term. Represents the current correction state vector; This indicates the cumulative number of abnormal stops at the entrance; Indicates the cumulative number of turnarounds; Indicates the cumulative degree of deviation from the target; , , This represents the weighting coefficient of the corresponding chain-level features. It's easy to understand that the first part of the formula is used to retain the historical risk of the same event, preventing the risk from being completely eliminated as soon as the normal direction is restored; the second part is used to introduce the immediate risk after deviation correction at the current moment; and the third part explicitly adds the cumulative features such as abnormal stops at the entrance, reversals, and target deviations in the entire continuous wandering chain, so that the risk score depends not only on the current behavior but also on the overall deterioration of the entire event.

[0053] Furthermore, in one feasible implementation, the cumulative amount of abnormal stops at the entrance... This can be accumulated from the user's approach to their own room door without entering, and from other users' lingering at their own room door; the number of times the user turns back is accumulated. It can be formed by the accumulation of candidate segments and significant reversal events within the entire wandering chain; the cumulative degree of target deviation. Can be within a period of time The cumulative statistics are generated. This process allows the system to assign a higher risk rating to objects that have repeatedly missed their own door and continue to linger in the wrong area, rather than making a judgment based solely on the action of the current frame.

[0054] In one specific implementation, a high-risk individual leaves the activity room exit after the activity ends, continuing towards their own room door for the initial part of the journey. At this point, the prototype matching result mainly favors the normal return-to-room behavior prototype, and the current correction state is relatively close to the normal state. Subsequently, the individual frequently searches left and right at the bifurcation point, passing their own room door but not entering, then lingering in front of a non-own room door and turning back multiple times. At this point, the prototype matching coefficient begins to gradually shift from the normal return-to-room prototype to the short-term waiting prototype and the disoriented wandering prototype, and the current correction state also gradually transitions from the normal state to the disoriented state. As the cumulative amount of abnormal lingering at the door, the cumulative number of turns, and the cumulative degree of target deviation increase, the disoriented risk score continues to rise, eventually reaching the medium-level warning threshold. Therefore, this step not only corrects the current behavior state but also effectively distinguishes between persistent disorientation and momentary hesitation using chain-level cumulative values.

[0055] S5: Based on the disorientation risk score data, perform an early warning judgment, and when the judgment meets the preset conditions, output a nursing prompt result containing the current area location, target room information and path indication information.

[0056] Specifically, a graded early warning result is formed based on disorientation risk score data and individual risk threshold data; when the graded early warning result reaches a preset level, video clip data of the time period before and after the alarm is extracted from the corresponding continuous wandering chain, and the current semantic region data, target room data, and object center trajectory data are extracted; path indication data is generated based on the object center trajectory data, and the video clip data, current semantic region data, target room data, and path indication data are combined to form a nursing prompt result; the nursing prompt result is output to the nursing terminal, and nursing feedback data returned by the corresponding nursing terminal is received.

[0057] Following this, based on the nursing feedback data and the trajectory data of this ward return monitoring task, the historical safe ward return path data is updated to form updated historical safe ward return path data; based on the false alarm results, effective warning results, and high-risk results in the nursing feedback data, the individual risk threshold data is rolled over to form updated individual risk threshold data; when the same high-risk object forms a new continuous wandering chain again within a preset time range, based on the temporal adjacency relationship, path overlap relationship, and target deviation consistency relationship between the new and old continuous wandering chains, it is determined whether they belong to the same persistent disorientation event, and if they are determined to belong to the same persistent disorientation event, the previous warning status is inherited.

[0058] In this embodiment of the invention, after generating a disorientation risk score, an early warning determination is performed based on the disorientation risk score data. When the determination meets preset conditions, a nursing prompt result containing the current area location, target room information, and path indication information is output. After the nursing prompt result is received by the nursing terminal and nursing feedback is returned, the historical safe return-to-room path and individual risk threshold are updated based on the nursing feedback. This step is used to transform the aforementioned continuous behavior recognition results into actual executable care actions, and enables the system to continuously self-correct through feedback to adapt to individual differences among different high-risk individuals.

[0059] It's easy to understand that what nursing staff truly need in their work is not a string of abstract scores, but rather specific nursing prompts, such as: whether to check, where the person is now, which is the target room, and how they just got there. Therefore, in one feasible implementation, several levels of individual risk thresholds can be pre-set for each high-risk individual, and the disorientation risk score can be compared with these thresholds to generate tiered early warning results, such as background record level, attention level, suggested nursing check level, and emergency intervention level. It's important to note that the thresholds do not have to be completely uniform; different individuals can use different threshold configurations based on their historical ward return habits, the number of previous disorientation events, and their nursing risk level, thereby improving individualized adaptation.

[0060] When the warning result reaches a preset level, the system extracts video clips from the corresponding continuous wandering chain, capturing the time periods before and after the alarm. Simultaneously, it extracts the current semantic region, target room identifier, and the object's central trajectory. Subsequently, a path diagram is generated based on the object's central trajectory. The video clip data, current semantic region data, target room data, and path diagram are combined to form a nursing prompt result, which is then pushed to the nursing terminal. The path diagram is designed to allow nursing staff to quickly understand which locations the object has passed through after leaving the activity room, where it is currently located, and whether it has missed its room door. This prompt result significantly reduces the nursing staff's judgment time and improves intervention efficiency.

[0061] After receiving nursing feedback, this invention further updates the historical safe return-to-room paths and individual risk thresholds. Specifically, if the nursing feedback indicates that the event was a false alarm, the corresponding level of individual risk threshold can be appropriately increased within a preset rolling window; if the nursing feedback indicates that the event was indeed a disorientation risk or had even developed into a more serious consequence, the corresponding level of individual risk threshold can be appropriately decreased, thereby improving the sensitivity of subsequent identification. Simultaneously, if the task is ultimately confirmed as a safe return to the room, the system can align the current return-to-room trajectory with existing historical safe return-to-room paths and update it at a fixed ratio to form a safe path baseline that more closely reflects the subject's current gait and return-to-room habits. It should be noted that updating historical safe return-to-room paths is not limited to a single algorithm. In practical applications, methods such as exponential sliding updates, key location point interpolation updates, or segmented path averaging updates can be used, as long as the historical baseline maintains long-term stability while slowly absorbing recent changes in the subject's habits.

[0062] Furthermore, in this embodiment of the invention, considering that the same high-risk object may generate multiple consecutive wandering chains with adjacent spatial locations and consistent target deviation trends within a short period of time, to avoid the nursing terminal receiving too many repeated alarms in the same persistent disorientation event, in one executable implementation, a continuous event inheritance judgment can be performed on the new and old consecutive wandering chains. If the newly formed consecutive wandering chain is adjacent to the previous alarmed consecutive wandering chain in time, overlaps in path, and is continuous in target deviation trend, then it can be determined that the two belong to the same persistent disorientation event, and the previous warning state is inherited, without re-initiating a completely independent new event. In this way, the interference of repeated alarms on nursing work can be further reduced.

[0063] In one specific implementation, a high-risk individual exhibits directional deviations and lingers outside their designated room after an activity ends. Once the system's calculated disorientation risk score reaches the medium-level alert threshold, it automatically pushes a notification to the nursing terminal, including a 10-second video clip before and after the alarm. The notification also displays the individual's current location as outside their designated room, the target room as R12, and a text message indicating that the path has already passed their designated room and continues forward, along with a corresponding path diagram. Nursing staff quickly go to the scene based on this notification and bring the individual back to their designated room. The terminal then marks the event as a valid warning and indicates that the individual has been manually guided back to their room. Upon receiving this feedback, the system classifies the event as a genuine disorientation event and slightly lowers the medium-level threshold for that individual in a scrolling window to issue a notification earlier next time.

[0064] Reference Figure 2 , Figure 2 This is a schematic diagram of the structure of a high-risk population monitoring system for elderly care institutions based on the integration of medical and elderly care, according to an embodiment of the present invention. Figure 2As shown, in an optional embodiment, the present invention also proposes a high-risk population monitoring system for elderly care institutions based on integrated medical and elderly care, comprising: The acquisition module 10 is used to acquire video frame data, scene area configuration data, and room corresponding data of high-risk people in the public activity area and residential area of ​​the elderly care institution, and to perform image recognition and behavior monitoring based on the video frame data to form the room return monitoring task data of the corresponding high-risk objects. Extraction module 20 is used to extract and form the return-to-room behavior feature data of high-risk individuals during the process of returning to their own rooms after the event, based on the return-to-room monitoring task data; The identification module 30 is used to identify and form candidate maze fragment data of the high-risk object during the process of returning to its own room based on the return behavior feature data, and to perform cross-interruption association processing on the candidate maze fragment data to form continuous wandering chain data. The forming module 40 is used to perform continuous risk identification processing based on the continuous wandering chain data, the room corresponding data of the high-risk object, and the historical safe return path data, so as to form the disorientation risk score data corresponding to the high-risk object. The output module 50 is used to perform early warning judgment based on the disorientation risk score data, and output a nursing prompt result containing the current area location, target room information and path indication information when the judgment meets the preset conditions.

[0065] Other embodiments or specific implementations of the present invention based on the monitoring system for high-risk groups in elderly care institutions that integrate medical and elderly care can refer to the above-mentioned method embodiments, and will not be repeated here.

[0066] It is understood that in the description of this specification, references to terms such as "one embodiment," "another embodiment," "other embodiments," or "first embodiment to Nth embodiment," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0067] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or system that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or system. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or system that includes that element.

[0068] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A method for monitoring high-risk groups in elderly care institutions based on integrated medical and elderly care, characterized in that, The method includes the following steps: The system acquires video frame data, scene area configuration data, and room-specific data corresponding to the public activity areas and residential areas of elderly care institutions. Based on the video frame data, it performs image recognition and behavior monitoring to generate data for monitoring the return to room of high-risk individuals. Based on the data from the room return monitoring task, extract and form the room return behavior characteristic data of high-risk individuals during the process of returning to their own rooms after the event; Based on the room return behavior feature data, candidate disorientation fragment data of the high-risk object during the process of returning to their own room are identified and formed, and cross-interruption association processing is performed on the candidate disorientation fragment data to form continuous wandering chain data. Based on the continuous wandering chain data, the corresponding room data of the high-risk object, and the historical safe return path data, continuous risk identification processing is performed to generate the disorientation risk score data corresponding to the high-risk object. Based on the disorientation risk score data, an early warning judgment is performed, and when the judgment meets the preset conditions, a nursing prompt result containing the current area location, target room information, and path indication information is output.

2. The method for monitoring high-risk populations in elderly care institutions based on integrated medical and elderly care as described in claim 1, characterized in that, Data for monitoring the return to room of high-risk individuals is generated, specifically including: Acquire scene area configuration data for the activity room exit area, main corridor area, corridor branch area, area near the user's room door, area near the non-user's room door, and area near the nursing station, and establish a semantic topology relationship for returning to the room based on the scene area configuration data; At the end of the activity, human recognition and target tracking are performed on the video frame data corresponding to the public activity area and residential area of ​​the elderly care institution to identify human targets leaving the exit area of ​​the activity room and extract the identity representation information of the corresponding human targets. The identity information is matched with the data corresponding to the rooms of high-risk individuals to determine the corresponding high-risk individuals and their rooms. The time when the high-risk individuals leave the activity room exit area is determined as the start time of the return-to-room monitoring task. Based on the high-risk object, its room, and the start time, return-to-room monitoring task data is generated; wherein, the return-to-room monitoring task data includes at least the object identifier, the target room identifier, the task start time, and the scene area index relationship.

3. The method for monitoring high-risk populations in elderly care institutions based on integrated medical and elderly care as described in claim 1, characterized in that, Based on the data from the room return monitoring task, the behavioral characteristics of high-risk individuals returning to their rooms after the activity are extracted and formed, specifically including: Continuous tracking is performed on the video frame data corresponding to the room return monitoring task data to extract the center position sequence of the high-risk object, and continuous trajectory data is formed based on the center position sequence; The movement direction and speed data of the high-risk object are calculated based on the changes in the center position at adjacent times to determine the target consistency coefficient. Head orientation data and body orientation data are extracted based on head key points and shoulder key points to form orientation search data; Based on the spatial relationship between the high-risk object and the area adjacent to the object's own door and the area adjacent to the non-object's door, data on the object's approach to the object's door without entering and data on the non-object's door stay are extracted to form door approach behavior data. Based on the continuous trajectory data, the target consistency coefficient, the orientation search data, and the door approach behavior data, return-to-room behavior feature data corresponding one-to-one with the high-risk objects are formed.

4. The method for monitoring high-risk populations in elderly care institutions based on integrated medical and elderly care as described in claim 3, characterized in that, Based on the aforementioned room-returning behavior characteristic data, candidate disorientation fragment data of the high-risk individual during the process of returning to their own room are identified and generated, specifically including: Within a sliding time window, the number of directional reversals, the degree of repeated arrivals in local areas, and the cumulative duration of low-speed dwelling for the high-risk objects are counted to form data on return behavior, local loop behavior, and abnormal dwelling behavior. The target consistency coefficient, orientation search data, backtracking behavior data, local loop behavior data, abnormal stay behavior data, data on approaching but not entering one's own room door, and data on staying at other people's room doors are fused together to form temporal disorientation candidate score data. Time intervals that continuously exceed a preset score threshold are identified as candidate misdirected segments, and candidate misdirected segment data are formed based on the start time, end time, dominant region, and segment feature summary of the candidate misdirected segments.

5. The method for monitoring high-risk populations in elderly care institutions based on integrated medical and elderly care as described in claim 4, characterized in that, Perform cross-interruption association processing on the candidate lost fragment data to form continuous wandering chain data, specifically including: For temporally adjacent or nearby candidate maze segments, extract the segment end position prediction result, segment start position result, object appearance representation result, target room deviation result, dominant region result, and candidate maze score result; Spatial continuity data is formed based on the predicted end position of the segment and the start position of the segment; appearance consistency data is formed based on the object appearance characterization results; target deviation continuity data is formed based on the target room deviation results; regional transfer rationality data is formed based on the dominant region results; and risk continuity data is formed based on the candidate disorientation score results. The spatial continuity data, appearance consistency data, target deviation continuity data, regional transfer rationality data, and risk continuity data are fused to form the cross-interruption correlation degree between candidate disorientation segments; When the cross-interruption correlation degree meets the preset correlation conditions, the corresponding candidate wandering segments are spliced ​​into the same continuous wandering chain, and continuous wandering chain data is formed.

6. The method for monitoring high-risk populations in elderly care institutions based on integrated medical and elderly care as described in claim 1, characterized in that, Based on the continuous wandering chain data, the room data corresponding to the high-risk object, and the historical safe return route data, continuous risk identification processing is performed to generate disorientation risk score data corresponding to the high-risk object, specifically including: Based on the continuous loitering chain data, the following features are extracted: duration of failure to return to the room after leaving the activity room, cumulative amount of abnormal lingering at the door, cumulative number of backtracking, cumulative degree of target deviation, and cumulative degree of search behavior, forming chain-level summary feature data; Based on historical safe return-to-room path data, short-term waiting sample data, and nursing confirmation disorientation sample data, prototypes of normal return-to-room behavior, short-term waiting behavior, and disorientation wandering behavior were constructed respectively. The behavioral representation data at the current moment is matched with the normal return-to-room behavior prototype, the short-term waiting behavior prototype, and the disoriented wandering behavior prototype respectively to form the corresponding prototype matching coefficient data. Based on the prototype matching coefficient data, deviation correction is performed on the behavior representation data at the current moment, and the current corrected state data is formed by combining it with the state data at the previous moment. Misconception risk scores are generated based on the current correction status data and chain-level summary feature data.

7. The method for monitoring high-risk populations in elderly care institutions based on integrated medical and elderly care as described in claim 6, characterized in that, Based on the current correction status data and chain-level summary feature data, a disorientation risk score is generated, specifically including: The previous moment's confusion risk score, the current correction status data, and the chain-level summary feature data are weighted and accumulated to form the current moment's confusion risk score; The disorientation risk score is compared with the individual risk threshold data of the corresponding high-risk object to form risk level data; When the risk level data reaches the preset alert level, the time-series evidence data corresponding to the current continuous wandering chain is retained.

8. The method for monitoring high-risk populations in elderly care institutions based on integrated medical and elderly care as described in claim 1, characterized in that, Based on the disorientation risk score data, an early warning judgment is performed, and when the judgment meets preset conditions, a nursing prompt result containing the current area location, target room information, and path indication information is output, specifically including: A tiered early warning system is generated based on disorientation risk scores and individual risk thresholds. When the graded early warning result reaches the preset level, video segment data of the time period before and after the alarm is extracted from the corresponding continuous wandering chain, and the current semantic region data, target room data and object center trajectory data are extracted. Path indication data is generated based on the object center trajectory data, and video clip data, current semantic region data, target room data and path indication data are combined to form nursing prompt results; The nursing prompt results are output to the nursing terminal, and the nursing feedback data returned by the corresponding nursing terminal is received.

9. The method for monitoring high-risk populations in elderly care institutions based on integrated medical and elderly care as described in claim 8, characterized in that, The method further includes: Based on the nursing feedback data and the trajectory data of this ward return monitoring task, the historical safe ward return path data is updated to form the updated historical safe ward return path data; Based on the false alarm results, effective early warning results, and high-risk results in the nursing feedback data, the individual risk threshold data is rolled over to form updated individual risk threshold data. When a new continuous wandering chain is formed again within a preset time range for the same high-risk object, the system determines whether it belongs to the same persistent disorientation event based on the temporal adjacency, path overlap, and target deviation consistency between the new and old continuous wandering chains, and inherits the previous warning status when it is determined to belong to the same persistent disorientation event.

10. A monitoring system for high-risk groups in elderly care institutions based on integrated medical and elderly care, characterized in that, The system includes: The acquisition module is used to acquire video frame data, scene area configuration data, and room corresponding data of high-risk individuals in the public activity areas and residential areas of the elderly care institution. Based on the video frame data, image recognition and behavior monitoring are performed to form the corresponding room return monitoring task data for high-risk individuals. The extraction module is used to extract and form the return-to-room behavior feature data of high-risk individuals during the process of returning to their own rooms after the event, based on the return-to-room monitoring task data; The identification module is used to identify and form candidate maze fragment data of the high-risk object during the process of returning to its own room based on the return behavior feature data, and to perform cross-interruption association processing on the candidate maze fragment data to form continuous wandering chain data. The forming module is used to perform continuous risk identification processing based on the continuous wandering chain data, the room corresponding data of the high-risk object, and the historical safe return path data, so as to form the disorientation risk score data corresponding to the high-risk object; The output module is used to perform early warning judgment based on the disorientation risk score data, and output a nursing prompt result containing the current area location, target room information and path indication information when the judgment meets the preset conditions.

Images