Edge computing neural network-based method for recognizing abnormal behavior of the elderly

Abstract

This invention relates to the fields of edge computing, computer vision, and intelligent health monitoring, specifically a method for identifying abnormal behavior in the elderly based on edge computing neural networks. The method includes: multimodal state recognition, constructing a fall confidence sequence, an emotion confidence sequence, a behavioral feature sequence, and a network delay sequence; network state weight redistribution, generating a second fusion weight combination; candidate execution configuration formation, configuring model priority, sampling frequency, and processing queue; edge inference execution and configuration evaluation, selecting the optimal weight combination; triggering local alarms, remote notifications, or local responses to network outages when the comprehensive risk score exceeds a threshold; and adaptive evolution based on historical behavior baselines constrained by the risk score.

Description

Technical Field

[0001] This invention relates to the fields of edge computing, computer vision, and intelligent health monitoring, specifically a method for identifying abnormal behavior in the elderly based on edge computing neural networks. Background Technology

[0002] In the current home-based monitoring environment for the elderly, video acquisition devices continuously generate a large amount of activity video frame data. This data usually requires multi-dimensional modal analysis involving falls, emotions, and daily trajectories.

[0003] To identify and alert on abnormal behavior in this video data, existing solutions generally employ a fixed edge inference architecture or a cloud-edge collaboration mechanism. This means that each neural network inference is executed sequentially according to a statically set resource allocation ratio and a fixed task queue. Although this solution has a certain comprehensive recognition capability when network communication is stable, it is highly dependent on a smooth network environment and lacks dynamic scheduling capabilities. When encountering network latency fluctuations or abnormal congestion, the fixed execution configuration on the edge side causes critical emergency hazard identification tasks and routine monitoring tasks to compete for computing resources. This results in prolonged processing time for critical identification tasks, delayed emergency alarm responses, and the solidified historical behavior baseline is easily interfered with by abnormal data, making it difficult to support highly reliable alarms and risk warnings in complex monitoring scenarios.

[0004] Therefore, how to dynamically allocate edge computing power to improve the timeliness and accuracy of abnormal behavior alarm output when communication networks fluctuate has become an urgent technical problem to be solved. Summary of the Invention

[0005] The purpose of this invention is to provide a method for identifying abnormal behavior in the elderly based on edge computing neural networks, and to solve the following technical problems:

[0006] It avoids delays in high-risk fall recognition results due to insufficient available computing power during network congestion or outages, and effectively prevents abnormal behavioral features from being incorrectly included in the normal behavioral baseline, thus preventing baseline contamination.

[0007] The objective of this invention can be achieved through the following technical solutions:

[0008] A method for recognizing abnormal behavior in the elderly based on edge computing neural networks operates on an edge computing node connecting a video acquisition device to a television or projector, including:

[0009] Multimodal state recognition: Obtain historical behavioral baselines that characterize the daily activity patterns of the elderly, call neural network models to process video and network state data, construct fall confidence, emotion confidence, behavioral features and network delay sequences, and determine the initial fusion weight set;

[0010] Network state weight reallocation: The initial fusion weight set is reallocated based on the network delay sequence to generate several second fusion weight combinations;

[0011] Candidate execution configuration formation: The combination of each second fusion weight is introduced into the edge inference cycle, the inference task is divided and the computing resources are allocated to form an execution configuration including model priority, sampling frequency and processing queue;

[0012] Edge inference execution: Apply the corresponding execution configuration to edge computing nodes to perform parallel inference and anomaly detection;

[0013] Configuration evaluation: Extract features from the reasoning process to obtain the corresponding configuration evaluation values;

[0014] Optimal configuration application: Based on the configuration evaluation value, the optimal second fusion weight combination is selected to generate a comprehensive risk score. When the judgment threshold is exceeded or triggered based on facial emotion recognition, an alarm response and remote care loop are triggered.

[0015] Baseline evolution adjustment: Adaptive evolution is performed based on the historical behavior baseline update rate constrained by the comprehensive risk score.

[0016] In one possible implementation, a historical behavioral baseline representing the daily activity patterns of the elderly is obtained. A pre-defined neural network model is then used to process the video data of the elderly's activities and network state data collected by the video acquisition device. This process constructs a fall confidence sequence, an emotion confidence sequence, a behavioral feature sequence, and a network delay sequence. Finally, an initial set of fusion weights corresponding to each multimodal state is determined, specifically as follows:

[0017] The system collects continuous video frames from video capture devices installed in the living areas of the elderly and forms an active video sequence in chronological order.

[0018] Human target detection and region cropping are performed on the activity video sequence to obtain a human image sequence corresponding to the human subject;

[0019] Perform posture change recognition and temporal motion extraction on human image sequences to obtain a fall confidence sequence representing the fall state;

[0020] Grayscale conversion and expression recognition are performed on the facial regions in human image sequences to obtain emotion confidence sequences representing emotional states;

[0021] Human contour extraction and trajectory association processing are performed on the activity video sequence to form a behavioral feature sequence under a continuous time window;

[0022] Collect network communication latency data between edge computing nodes and remote servers, and form a network latency sequence in chronological order;

[0023] Based on the basic attention levels corresponding to fall recognition, emotion recognition, and behavior monitoring, initial fusion weights are set for each recognition branch, resulting in an initial fusion weight set that corresponds one-to-one with each recognition branch.

[0024] In one possible implementation, the initial fusion weight set is reallocated based on the network delay sequence to generate several second fusion weight combinations, specifically:

[0025] Continuous sampling and statistical analysis of the network delay sequence are performed to obtain the network delay value corresponding to each sampling time.

[0026] Based on a preset network delay upper limit, the network delay values ​​corresponding to each sampling time are normalized to form a network state coefficient sequence that reflects the degree of network congestion.

[0027] The network state coefficient sequence is divided into intervals to determine the normal network interval, the fluctuating network interval, and the abnormal network interval.

[0028] Based on the network state coefficient sequence, the degree of computing power focus on the edge side is mapped to form a computing power focus sequence corresponding to each sampling time.

[0029] While keeping the correspondence between the recognition branches unchanged, the fall recognition weight, emotion recognition weight and behavior monitoring weight in the initial fusion weight set are adjusted incrementally or by decreasing, so as to form the weight adjustment results corresponding to different network state intervals.

[0030] By performing weight adjustment operations multiple times, several second fusion weight combinations with different degrees of tilt on the fall recognition side are obtained;

[0031] Establish a one-to-one correspondence between each second fusion weight combination and its corresponding network state interval for subsequent execution configuration.

[0032] In one possible implementation, the combinations of the second fusion weights are introduced into the edge inference cycle to divide the inference task and computing resources, forming an execution configuration that includes model priority, sampling frequency, and processing queue, specifically:

[0033] Edge computing nodes are configured with fixed scheduling cycles, and based on these fixed scheduling cycles, the continuous operation process is divided into several edge inference cycles.

[0034] Each second fusion weight combination is introduced into a single edge inference cycle as a priority constraint for each identification branch within that inference cycle;

[0035] Based on the corresponding second fusion weight combination, the fall detection priority processing period and the non-fall detection regular processing period are divided within a single edge inference cycle;

[0036] Based on the priority processing period for fall detection and the regular processing period for non-fall detection, the neural network processing resources and central processing resources of edge computing nodes are allocated to form a computing power resource quota corresponding to each detection branch;

[0037] Based on the computing power resource quota corresponding to each recognition branch, the sampling frequency of video frames, the execution frequency of emotion recognition, and the length of the behavior monitoring time window are adjusted to obtain the sampling frequency adjustment results;

[0038] The video frame task stream entering the edge computing node is sorted according to the priority and processing time of its respective recognition branch, forming a queue sorting result that includes a high-priority processing queue and a regular processing queue.

[0039] The computing power resource quota, sampling frequency adjustment results, and queue sorting results are combined to form an execution configuration that corresponds one-to-one with each second fusion weight combination.

[0040] In one possible implementation, the execution configurations corresponding to each second fusion weight combination are applied to edge computing nodes to perform multimodal parallel inference and abnormal behavior recognition, specifically:

[0041] The edge computing node is configured with video access links and basic recognition links. The video access links and basic recognition links are kept running continuously in the edge computing node, and a high-priority inference link for key anomaly identification is configured.

[0042] For each second fusion weight combination, during the operation of the edge computing node, the high-priority inference link is activated according to the execution configuration corresponding to the second fusion weight combination;

[0043] When the high-priority inference link is active, video frames within the corresponding time period will be prioritized for access to the fall recognition branch, and simultaneously accessed to the emotion recognition branch and behavior monitoring branch.

[0044] During the high-priority inference link activation period, the emotion recognition task and behavior monitoring task are executed at a reduced frequency or continuously, depending on the sampling frequency and time window length set in the execution configuration.

[0045] When the configured priority processing period ends, stop inputting new key recognition tasks into the high-priority inference link and restore subsequent video frames to the basic recognition link for continued processing;

[0046] The fall confidence results output by the fall recognition branch, the emotion confidence results output by the emotion recognition branch, and the behavioral feature results output by the behavior monitoring branch are synchronously aggregated to form a multimodal inference result corresponding to the execution configuration.

[0047] The multimodal inference results generated under each execution configuration are identified and output separately for subsequent configuration evaluation.

[0048] In one possible implementation, the inference process features under each second fusion weight combination are extracted, and the configuration evaluation value corresponding to each second fusion weight combination is obtained, specifically as follows:

[0049] For each applied second fusion weight combination, during the continuous operation of the edge computing node, the inference start time, inference end time, and anomaly identification output result under the corresponding execution configuration are recorded, and a single inference process is extracted.

[0050] Within each single inference process, the processing time between the fall recognition result input and the abnormal output of the video frame is statistically analyzed to form the emergency response latency feature.

[0051] Within each single inference process, the retention length of the emotion confidence result within the continuous time window and the degree of decay correlation between adjacent time steps are statistically analyzed to form the continuous emotion perception feature.

[0052] Within each single inference process, the magnitude of deviation of the statistical behavioral characteristic results from the historical behavioral baseline and the degree of fluctuation between adjacent windows form the stable behavioral deviation characteristics.

[0053] In each single inference process, the computing power occupancy ratio and task queue length of the edge computing node under each identification branch are recorded to form resource occupancy characteristics.

[0054] The emergency response delay characteristics, continuous emotion perception characteristics, behavioral deviation stability characteristics, and resource consumption characteristics are combined in a fixed order to form a configuration evaluation vector corresponding to each second fusion weight combination.

[0055] An evaluation function is constructed based on each feature in the configuration evaluation vector, and the configuration evaluation value corresponding to each second fusion weight combination is calculated.

[0056] In one possible implementation, the optimal second fusion weight combination is selected based on the configuration evaluation value, and a comprehensive risk score is generated under the optimal second fusion weight combination. When the comprehensive risk score exceeds the anomaly detection threshold, an abnormal behavior alarm response is triggered, specifically as follows:

[0057] The configuration evaluation values ​​corresponding to each second fusion weight combination are sorted according to the network state interval order to construct a sequence of weight combination-evaluation value correspondence.

[0058] In the weight combination-evaluation value correspondence sequence, identify the segment where the configuration evaluation value gradually converges with the increase of the fall recognition weight, and determine the second fusion weight combination corresponding to the segment as the candidate optimal weight combination;

[0059] The combination with the smallest total weight occupancy among the candidate optimal weight combinations is selected as the final optimal second fusion weight combination; the final optimal second fusion weight combination is configured to the edge computing node to perform weighted fusion on the fall confidence result, sentiment confidence result and behavioral feature result;

[0060] The emotional confidence results are subjected to time decay accumulation processing to form a time-weighted emotional risk quantity;

[0061] The deviation between the behavioral characteristic results and the historical behavioral baseline is normalized to form a behavioral deviation risk quantity;

[0062] The fall confidence result, time-weighted emotional risk and behavioral deviation risk are non-linearly mapped to generate a comprehensive risk score. When the comprehensive risk score exceeds the anomaly judgment threshold, an abnormal behavior alarm response is triggered.

[0063] Based on the emotion confidence results, facial emotion classification is performed. When an emotion belonging to a preset category that needs care or a positive emotion belonging to a preset category that needs sharing is identified, the remote care loop and alarm are triggered. Through the connection between the video acquisition device and the TV or projector, the user terminal establishes a video chat connection through a remote APP to conduct remote care.

[0064] In one possible implementation, the update rate of the historical behavioral baseline is constrained based on a comprehensive risk score, and adaptive evolution of the behavioral baseline is performed, specifically as follows:

[0065] Obtain the comprehensive risk score generated under the optimal second fusion weight combination, and establish a risk constraint relationship corresponding to the current behavioral feature results;

[0066] The historical behavior baseline and the current behavior feature results are aligned in the same dimension to form a baseline comparison vector for baseline update;

[0067] Based on the comprehensive risk score, the basic learning rate is reduced and mapped to obtain the dynamic learning rate corresponding to the current risk state.

[0068] When the comprehensive risk score is less than or equal to the anomaly detection threshold, the dynamic learning rate is increased so that the current behavioral feature results can be used to update the historical behavioral baseline.

[0069] When the comprehensive risk score is greater than the anomaly detection threshold, the dynamic learning rate is reduced or limited to zero, thereby blocking the writing of abnormal behavior features into the historical behavior baseline.

[0070] Based on the dynamic learning rate, a weighted update is performed on the historical behavioral baseline and the current behavioral feature results to form a new behavioral baseline;

[0071] The new behavioral baseline is written back to the behavioral monitoring branch for subsequent behavioral deviation identification and comprehensive risk score generation.

[0072] The beneficial effects of this invention are:

[0073] 1. This invention collects network latency sequences and reallocates the initial fusion weight set to generate several second fusion weight combinations. These combinations are then introduced into the edge inference cycle to divide the inference task and computing resources, forming an execution configuration that includes model priority, sampling frequency, and processing queue. This mechanism enables edge computing nodes to flexibly adjust the weight allocation and computing power quota among fall detection, emotion recognition, and behavior monitoring based on network congestion. During network fluctuations or anomalies, priority is given to ensuring computing resources and video access processing for fall detection tasks, effectively solving the computing power mismatch problem caused by fixed-ratio allocation in existing methods and ensuring the priority response time for high-risk abnormal behaviors.

[0074] 2. This invention extracts emergency response latency features, continuous emotion perception features, behavioral deviation stability features, and resource consumption features under the execution configuration to calculate the configuration evaluation value corresponding to each second fusion weight combination, and selects the optimal second fusion weight combination based on the evaluation value; it performs closed-loop optimization based on the actual operating characteristics of edge computing nodes under different configurations, which improves the response speed of key anomalies while effectively balancing the computing load to meet the preset feature retention rate, thus ensuring the comprehensiveness and accuracy of the comprehensive judgment of multimodal abnormal behavior.

[0075] 3. This invention proposes to constrain the update rate of historical behavioral baselines based on comprehensive risk scores, and to perform adaptive evolution of behavioral baselines. The system obtains a dynamic learning rate based on the current comprehensive risk score mapping. When a high-risk state is identified, the dynamic learning rate is automatically reduced or limited to zero, thereby directly blocking the writing of current abnormal behavioral features into the historical behavioral baseline. Blocking the writing of current abnormal behavioral features into the historical behavioral baseline ensures the accuracy and reliability of the data of the elderly's normal daily activities baseline and improves the benchmark accuracy of subsequent behavioral deviation judgments. Attached Figure Description

[0076] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0077] Figure 1 This is a flowchart illustrating the method for identifying abnormal behavior of the elderly based on edge computing neural networks in this application. Detailed Implementation

[0078] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0079] Please see Figure 1 A method for recognizing abnormal behavior in the elderly based on edge computing neural networks, running on an edge computing node connecting a video acquisition device and a television or projector, including:

[0080] Multimodal state recognition: Obtain historical behavioral baselines that characterize the daily activity patterns of the elderly, call neural network models to process video and network state data, construct fall confidence, emotion confidence, behavioral features and network delay sequences, and determine the initial fusion weight set;

[0081] Network state weight reallocation: The initial fusion weight set is reallocated based on the network delay sequence to generate several second fusion weight combinations;

[0082] Candidate execution configuration formation: The combination of each second fusion weight is introduced into the edge inference cycle, the inference task is divided and the computing resources are allocated to form an execution configuration including model priority, sampling frequency and processing queue;

[0083] Edge inference execution: Apply the corresponding execution configuration to edge computing nodes to perform parallel inference and anomaly detection;

[0084] Configuration evaluation: Extract features from the reasoning process to obtain the corresponding configuration evaluation values;

[0085] Optimal configuration application: Based on the configuration evaluation value, the optimal second fusion weight combination is selected to generate a comprehensive risk score. When the judgment threshold is exceeded or triggered based on facial emotion recognition, an alarm response and remote care loop are triggered.

[0086] Baseline evolution adjustment: Adaptive evolution is performed based on the historical behavior baseline update rate constrained by the comprehensive risk score;

[0087] A historical behavioral baseline representing the daily activity patterns of the elderly is obtained. A pre-defined neural network model is used to process the video data of the elderly's activities and network state data collected by the video acquisition device. Fall confidence sequences, emotion confidence sequences, behavioral feature sequences, and network delay sequences are constructed, and initial fusion weight sets corresponding to each multimodal state are determined, specifically as follows:

[0088] The system collects continuous video frames from a video acquisition device installed in the elderly's living area and forms an active video sequence in chronological order. Human target detection and region cropping are then performed on the active video sequence to obtain a sequence of human images corresponding to the human subject.

[0089] Pose change recognition and temporal motion extraction are performed on human image sequences to obtain a fall confidence sequence representing the fall state; grayscale conversion and expression recognition are performed on the face region in the human image sequence to obtain an emotion confidence sequence representing the emotional state.

[0090] Human contour extraction and trajectory association processing are performed on the activity video sequence to form a behavioral feature sequence under a continuous time window; network communication latency data between edge computing nodes and remote servers are collected and network latency sequence is formed in chronological order;

[0091] Based on the basic attention levels corresponding to fall recognition, emotion recognition, and behavior monitoring, initial fusion weights are set for each recognition branch to obtain an initial fusion weight set that corresponds one-to-one with each recognition branch.

[0092] The initial fusion weight set is redistributed based on the network delay sequence to generate several second fusion weight combinations. Specifically, the network delay sequence is continuously sampled and statistically analyzed to obtain the network delay value corresponding to each sampling time.

[0093] Based on a preset network delay upper limit, the network delay values ​​corresponding to each sampling time are normalized to form a network state coefficient sequence that reflects the degree of network congestion.

[0094] The network state coefficient sequence is divided into intervals to determine the normal network interval, the fluctuating network interval, and the abnormal network interval; based on the network state coefficient sequence, the degree of computing power focus on the edge side is mapped to form a computing power focus sequence corresponding to each sampling time.

[0095] While keeping the correspondence between the recognition branches unchanged, the fall recognition weight, emotion recognition weight and behavior monitoring weight in the initial fusion weight set are adjusted incrementally or by decreasing, so as to form the weight adjustment results corresponding to different network state intervals.

[0096] By performing weight adjustment operations multiple times, several second fusion weight combinations with different degrees of tilt on the fall recognition side are obtained; a one-to-one correspondence is established between each second fusion weight combination and the corresponding network state interval for subsequent configuration.

[0097] The system is equipped with a video acquisition device, a local sound and light alarm device, a television or projector, and an edge computing node. The edge computing node uses an edge computing chip with a neural network processor and communicates with a remote server.

[0098] The video acquisition device continuously sends video frames of the elderly person's living area to the edge computing node, which then performs fall recognition, emotion recognition, behavior monitoring, and comprehensive risk assessment locally.

[0099] When the network communication status changes, the edge computing node does not wait for the remote server to return the recognition result, but adjusts the weight and inference resources of each recognition branch based on the network latency collected locally;

[0100] In this embodiment, the historical behavior baseline is used to represent the behavioral characteristics of the elderly in their normal living state. This historical behavior baseline can be pre-saved in the local storage area by the edge computing node, or it can be saved in the edge computing node after being issued by the remote server during the initialization phase.

[0101] The edge computing node reads this historical behavior baseline during runtime and provides it to the behavior monitoring branch as reference data for subsequent deviation judgment;

[0102] After the continuous video frames output by the video acquisition device enter the edge computing node, they are first written into the video buffer according to the acquisition time. The edge computing node then forms an active video sequence based on the frame timestamps.

[0103] The human target detection process identifies the main human body region from the active video sequence and distributes the extracted human image sequence to different recognition branches;

[0104] The fall detection branch performs posture change recognition and temporal motion extraction on human image sequences to obtain fall confidence results for each time step; these results are written into the fall confidence cache in chronological order to form a fall confidence sequence.

[0105] The emotion recognition branch locates the face region from the human image sequence, performs grayscale processing on the face region, performs expression recognition, obtains the emotion confidence results corresponding to each time step, and writes them into the emotion confidence cache to form an emotion confidence sequence.

[0106] The behavior monitoring branch performs human contour extraction and trajectory association processing on the activity video sequence, generates behavior feature results within a continuous time window, and saves them as a behavior feature sequence according to the window order;

[0107] Edge computing nodes also periodically detect the communication latency between themselves and remote servers; each detected network communication latency is accompanied by the sampling time and written into the network status record area, forming a network latency sequence in chronological order;

[0108] When the weight redistribution logic is triggered, the edge computing node independently reads the network latency sequence and maintains the original recorded state of the fall confidence sequence, sentiment confidence sequence, and behavioral feature sequence unchanged.

[0109] In the initial state, the edge computing nodes set an initial fusion weight set for each recognition branch based on the basic attention level corresponding to fall recognition, emotion recognition, and behavior monitoring;

[0110] The weights in this set maintain a one-to-one correspondence with the recognition branches. For example, the fall recognition weight corresponds to the fall recognition branch, the emotion recognition weight corresponds to the emotion recognition branch, and the behavior monitoring weight corresponds to the behavior monitoring branch. This correspondence is not changed in the subsequent redistribution process to avoid branch mismatch after the weight adjustment.

[0111] In response to network anomalies or communication latency exceeding a preset threshold, the edge computing node triggers a network state weight redistribution mechanism. The edge computing node continuously samples and statistically analyzes the network latency sequence to obtain the network latency value at each sampling time, and performs normalization processing based on a preset network latency upper limit to form a network state coefficient sequence.

[0112] The network status coefficient can be obtained as follows: calculate the ratio of the current network latency value to the preset network latency upper limit, and limit the result exceeding the upper limit to within 1.0; the closer the coefficient is to 1.0, the closer the current network status is to a network abnormality or network outage.

[0113] Edge computing nodes divide the network into intervals based on the network state coefficient sequence to determine the normal network interval, the fluctuating network interval, and the abnormal network interval; the interval division results are written into the scheduling status record area for subsequent execution configuration formation process to read.

[0114] Edge computing nodes map the degree of computing power focus on the edge side based on the network state coefficient, forming a computing power focus sequence; the degree of computing power focus can be determined by adding 1 to the product of the adjustment coefficient and the network state coefficient; when the network state coefficient increases, the degree of computing power focus increases accordingly.

[0115] During weight redistribution, the branch correspondence between fall recognition, emotion recognition and behavior monitoring of edge computing nodes remains unchanged, and only the corresponding weights are adjusted incrementally or by decreasing weight.

[0116] Specifically, when the corresponding network state coefficient is greater than the preset threshold, the fall recognition weight increases with the increase of computing power focus, while the emotion recognition weight and behavior monitoring weight decrease with the increase of computing power focus; when the network state is normal, each weight is close to the initial fusion weight set.

[0117] The initial fusion weight parameters for fall recognition, emotion recognition, and behavior monitoring are set to 0.4, 0.3, and 0.3, respectively, and the computational power focus parameter corresponding to the network state is set to 1.5.

[0118] Edge computing nodes can multiply the fall detection weights by the level of focus to obtain a new weight after incremental adjustment, i.e. Then, according to the principle of overall weight normalization, the remaining 0.4 weight is proportionally distributed to the other two branches, so that the emotion recognition weight and behavior monitoring weight are reduced to 0.2 respectively, thereby generating the second fusion weight combination in the current environment.

[0119] Edge computing nodes generate several second fusion weight combinations containing different fall recognition weight allocation ratios by performing weight adjustments multiple times;

[0120] Each second fusion weight combination is associated with a normal network interval, a fluctuating network interval, or an abnormal network interval, and is written into the weight combination record area.

[0121] In the case of missing network latency samples, the edge computing node does not directly interrupt the identification process, but reads the network latency value in the previous valid sampling period as the replacement value for the current period and retains the replacement record in the network status recording area.

[0122] If network latency values ​​cannot be obtained for several consecutive cycles, the network is treated as an abnormal interval, so that subsequent resource allocation prioritizes the fall detection branch.

[0123] If no human body is detected in a certain video frame, the edge computing node will not write the frame into the human image sequence, but will continue to process subsequent frames; the behavior monitoring branch will form behavior feature results in the current window based on the obtained human body contour and trajectory records, and will not update the behavior feature sequence of the current window if it cannot form a valid result.

[0124] In the aforementioned residential monitoring scenario, when the remote server communication latency is lower than a preset latency threshold, the edge computing node maintains the normal multi-branch recognition ratio.

[0125] When the communication delay of the remote server is higher than the preset delay threshold, the network state coefficient increases accordingly. The edge computing node will write the second fusion weight combination corresponding to the network fluctuation interval or network abnormal interval into the scheduling state record area.

[0126] The subsequent candidate execution configuration formation process directly reads this record, without needing to re-analyze the original network latency data, thus keeping the video recognition process separate from the network scheduling process.

[0127] In this embodiment of the invention, each second fusion weight combination is introduced into the edge inference cycle to divide the inference task and computing resources, forming an execution configuration that includes model priority, sampling frequency and processing queue. Specifically, the edge computing node is configured with a fixed scheduling beat, and based on the fixed scheduling beat, the continuous running process is divided into several edge inference cycles; each second fusion weight combination is introduced into a single edge inference cycle as a priority constraint for each identification branch within that inference cycle.

[0128] Based on the corresponding second fusion weight combination, a fall recognition priority processing period and a non-fall recognition regular processing period are divided within a single edge inference cycle; based on the fall recognition priority processing period and the non-fall recognition regular processing period, the neural network processing resources and central processing resources of the edge computing nodes are allocated to form a computing power resource quota corresponding to each recognition branch;

[0129] Based on the computing power resource quota corresponding to each recognition branch, the sampling frequency of video frames, the execution frequency of emotion recognition, and the length of the behavior monitoring time window are adjusted to obtain the sampling frequency adjustment result; the video frame task flow entering the edge computing node is sorted according to the priority and processing time of its respective recognition branch to form a queue sorting result containing a high-priority processing queue and a regular processing queue.

[0130] The computing power resource quota, sampling frequency adjustment results and queue sorting results are combined to form an execution configuration that corresponds one-to-one with each second fusion weight combination;

[0131] The execution configurations corresponding to each second fusion weight combination are applied to the edge computing nodes to perform multimodal parallel inference and abnormal behavior recognition. Specifically, the edge computing nodes are configured with video access links and basic recognition links. The video access links and basic recognition links are kept running continuously in the edge computing nodes, and high-priority inference links for key anomaly recognition are configured.

[0132] For each second fusion weight combination, during the operation of the edge computing node, the high-priority inference link is activated according to the execution configuration corresponding to the second fusion weight combination;

[0133] When the high-priority inference link is active, video frames within the corresponding time period will be prioritized for access to the fall recognition branch, and simultaneously accessed to the emotion recognition branch and behavior monitoring branch.

[0134] During the high-priority inference link activation period, the emotion recognition task and behavior monitoring task are executed at a reduced frequency or continuously, depending on the sampling frequency and time window length set in the execution configuration.

[0135] When the configured priority processing period ends, stop inputting new key recognition tasks into the high-priority inference link and restore subsequent video frames to the basic recognition link for continued processing;

[0136] The fall confidence results output by the fall recognition branch, the emotion confidence results output by the emotion recognition branch, and the behavioral feature results output by the behavior monitoring branch are synchronously aggregated to form a multimodal inference result corresponding to the execution configuration. The multimodal inference results formed under each execution configuration are identified and output separately for subsequent configuration evaluation.

[0137] The inference process features under each second fusion weight combination are extracted, and the configuration evaluation value corresponding to each second fusion weight combination is obtained, specifically:

[0138] For each applied second fusion weight combination, during the continuous operation of the edge computing node, the inference start time, inference end time, and anomaly identification output result under the corresponding execution configuration are recorded, and a single inference process is extracted.

[0139] Within each single inference process, the processing time between the fall recognition result input and the abnormal output of the video frame is statistically analyzed to form the emergency response latency feature.

[0140] Within each single inference process, the retention length of the emotion confidence result within the continuous time window and the degree of decay correlation between adjacent time steps are statistically analyzed to form the emotion continuous perception feature; within each single inference process, the deviation of the behavioral feature result from the historical behavioral baseline and the degree of fluctuation between adjacent windows are statistically analyzed to form the behavioral deviation stability feature.

[0141] In each single inference process, the computing power occupancy ratio and task queuing length of the edge computing node under each identification branch are recorded to form resource occupancy features; the emergency response latency features, emotion continuous perception features, behavior deviation stability features and resource occupancy features are combined in a fixed order to form a configuration evaluation vector corresponding to each second fusion weight combination.

[0142] An evaluation function is constructed based on each feature in the configuration evaluation vector, and the configuration evaluation value corresponding to each second fusion weight combination is calculated.

[0143] The second fusion weight combination is introduced into the edge inference cycle, and the model priority, sampling frequency and processing queue configuration parameters are adjusted synchronously to prevent network anomalies from causing lag in the output of recognition results.

[0144] Edge computing nodes are configured with fixed scheduling cycles, and the continuous operation process is divided into multiple edge inference cycles according to these fixed scheduling cycles.

[0145] At the start of each edge inference cycle, the scheduler reads the second fusion weight combination corresponding to the current network state interval from the weight combination record area and writes the combination into the scheduling parameter record for the current cycle;

[0146] This record serves as a priority constraint for each identification branch within this cycle and remains valid until the end of this cycle, thus avoiding repeated changes in the queue order due to short-term fluctuations in network latency within the same cycle.

[0147] The scheduler divides the fall detection priority processing period and the non-fall detection regular processing period within a single edge inference cycle according to the relative size of the fall detection weights in the second fusion weight combination. The fall detection priority processing period is used to prioritize the execution of the inference task of the fall detection branch.

[0148] During the non-fall recognition routine processing period, routine tasks of the emotion recognition branch and behavior monitoring branch are performed; edge computing nodes allocate neural network processing resources and central processing resources accordingly, forming computing power resource quotas corresponding to each recognition branch; these quotas are written into the execution configuration for this cycle, and can be directly read during the inference process;

[0149] When forming sampling parameters, the edge computing nodes adjust the video frame sampling frequency, emotion recognition execution frequency, and behavior monitoring time window length according to the computing power resource quota of each recognition branch to obtain the sampling frequency adjustment result.

[0150] When the weight of the fall detection side is increased, the fall detection branch receives a higher processing priority and a higher video frame access frequency;

[0151] The frequency of emotion recognition tasks can be reduced according to the execution configuration, and the time window length of behavior monitoring tasks can be adjusted according to the execution configuration. The above adjustments do not change the video frame records in the original video buffer, but only change the task flow entering each recognition branch.

[0152] The video frame task flow entering the edge computing node is sorted according to the priority of the recognition branch to which it belongs and the current processing time period; during the priority processing time period of fall recognition, video frame tasks related to the fall recognition branch are pushed into the high priority processing queue.

[0153] Emotion recognition and behavior monitoring tasks are pushed into the regular processing queue or delayed according to the frequency reduction rules set in the execution configuration.

[0154] During the non-fall detection routine processing period, tasks in the routine processing queue continue to be scheduled and executed; the computing power resource quota, sampling frequency adjustment result, and queue sorting result together constitute the execution configuration corresponding to each second fusion weight combination, and are stored in the execution configuration record area;

[0155] During edge inference execution, the edge computing node maintains the continuous operation of the video access link and the basic recognition link; the video access link is used to receive video frames output by the video acquisition device.

[0156] The basic identification link is used to perform multi-branch identification in a conventional manner; in addition to the above links, the edge computing node is configured with a high-priority inference link for key anomaly identification; when the execution configuration corresponding to a certain second fusion weight combination is selected, the edge computing node activates the high-priority inference link according to the execution configuration;

[0157] When the high-priority inference link is active, video frames within the corresponding time period are prioritized for access to the fall recognition branch, and simultaneously accessed to the emotion recognition branch and behavior monitoring branch;

[0158] The emotion recognition and behavior monitoring tasks are not canceled, but rather subjected to down-frequency processing or continuous processing based on the sampling frequency and time window length set in the execution configuration;

[0159] Fall detection results can be output first, while emotional confidence results and behavioral feature results are still retained for subsequent comprehensive judgment;

[0160] When the configured priority processing period ends, the edge computing node stops inputting new key recognition tasks into the high-priority inference link, and subsequent video frames are restored to the basic recognition link for further processing.

[0161] The fall confidence results output by the fall recognition branch, the emotion confidence results output by the emotion recognition branch, and the behavioral feature results output by the behavior monitoring branch are written into the same inference result cache and aligned by timestamps or period identifiers.

[0162] The edge computing nodes generate multimodal inference results corresponding to the current execution configuration. The multimodal inference results generated under different execution configurations are respectively accompanied by the execution configuration identifier and the second fusion weight combination identifier, so that the subsequent configuration evaluation process can distinguish their sources.

[0163] If the high-priority processing queue experiences task backlog within a certain period, the edge computing nodes will prioritize retaining tasks that have entered the fall recognition branch, and reduce the execution frequency of emotion recognition tasks or extend the behavior monitoring time window according to the execution configuration.

[0164] If no effective face region is formed within a certain period, the emotion recognition branch will not output new emotion confidence results at that time step, and subsequent continuous emotion perception features will only be statistically analyzed based on the retained emotion confidence results.

[0165] If the behavior monitoring branch cannot form a complete trajectory association result within the current window, the behavior feature result will not participate in the deviation stability statistics of this window and will be processed in the next valid window.

[0166] The configuration evaluation process is used to compare the effects of different second fusion weight combinations in actual edge operation; for each applied second fusion weight combination, the edge computing node records the inference start time, inference end time and anomaly identification output results under the corresponding execution configuration during continuous operation, and extracts a single inference process from the operation record;

[0167] Within a single inference process, the edge computing nodes statistically analyze the processing time between the fall recognition result input and the abnormal output of the video frame, forming an emergency response latency feature; this feature reflects the response status of the fall recognition branch under the current execution configuration.

[0168] When forming emergency response delay characteristics, the difference between the preset maximum allowable processing delay limit and the actual processing time is obtained, and the difference is normalized to generate emergency response delay feature values. The emergency response delay feature values ​​are positively correlated with the difference.

[0169] Meanwhile, edge computing nodes statistically analyze the retention length of sentiment confidence results and the degree of decay correlation between adjacent time steps within a continuous time window. Based on the degree of decay correlation parameter, they generate continuous sentiment perception features. The value of continuous sentiment perception features is positively correlated with the system's continuous perception capability.

[0170] The edge computing node multiplies the sentiment confidence results output at each effective time step within a continuous time window by an exponentially decreasing decay coefficient configured according to the time step interval at the current moment, and sums the product results to obtain the decay correlation parameter.

[0171] Based on the attenuation correlation parameter, continuous emotion perception features are generated. The values ​​of continuous emotion perception features are positively correlated with the temporal distribution density and temporal proximity of emotion confidence results.

[0172] In terms of behavior monitoring, edge computing nodes statistically analyze the deviation of behavioral characteristics from historical baselines and the degree of fluctuation between adjacent windows. They obtain the difference between a preset upper limit for fluctuation tolerance and the weighted sum of deviation and fluctuation, generating a stable behavior deviation feature. The stable behavior deviation feature value is positively correlated with the stability parameter of the behavior feature. In terms of resources, edge computing nodes record and accumulate the computing power occupancy ratio and task queue length under each identification branch to form a resource occupancy feature. The value of the resource occupancy feature is positively correlated with the system computing power load.

[0173] The aforementioned emergency response delay characteristics, continuous emotion perception characteristics, behavioral deviation from stability characteristics, and resource consumption characteristics are combined in a fixed order to form a configuration evaluation vector;

[0174] Edge computing nodes assign positive weight coefficients to emergency response latency features, continuous emotion perception features, and behavior deviation stability features, and negative weight coefficients to resource consumption features. The features in the configuration evaluation vector are weighted and summed to obtain the configuration evaluation value corresponding to each second fusion weight combination.

[0175] The configuration evaluation value is written to the evaluation result record area and associated with the corresponding second fusion weight combination;

[0176] During the continuous monitoring of elderly people living alone, when the network enters a fluctuating range, the edge computing nodes apply a second fusion weight combination with different fall recognition weight ratios in several inference cycles and record the corresponding inference process features.

[0177] If a combination increases the priority of fall detection but causes the queue length of the regular processing queue to exceed the queue length threshold, the resource consumption feature will reflect its negative impact in the configuration evaluation value; if a combination can maintain a low processing time for fall detection and retain the necessary emotional and behavioral results, its configuration evaluation value will be increased accordingly; this evaluation result is used in the optimal configuration application process.

[0178] In this embodiment of the invention, the optimal second fusion weight combination is selected based on the configuration evaluation value, and a comprehensive risk score is generated under the optimal second fusion weight combination. When the comprehensive risk score exceeds the anomaly judgment threshold, an abnormal behavior alarm response is triggered. Specifically, the configuration evaluation values ​​corresponding to each second fusion weight combination are sorted according to the network state interval order to construct a weight combination-evaluation value correspondence sequence.

[0179] In the weight combination-evaluation value correspondence sequence, the segment where the configured evaluation value gradually converges with the increase of the fall recognition weight is identified, and the corresponding second fusion weight combination in the segment is determined as the candidate optimal weight combination; from the candidate optimal weight combinations, the combination with the smallest total weight of multiple branches is selected as the final optimal second fusion weight combination; the final optimal second fusion weight combination is configured to the edge computing node, and weighted fusion is performed on the fall confidence result, emotion confidence result and behavioral feature result;

[0180] The emotional confidence results are subjected to time decay accumulation processing to form a time-weighted emotional risk quantity; the deviation between the behavioral characteristic results and the historical behavioral baseline is normalized to form a behavioral deviation risk quantity.

[0181] The fall confidence result, time-weighted emotional risk and behavioral deviation risk are non-linearly mapped to generate a comprehensive risk score. When the comprehensive risk score exceeds the anomaly judgment threshold, an abnormal behavior alarm response is triggered.

[0182] Based on the emotion confidence results, facial emotion classification is performed. When an emotion belonging to a preset category that needs care or a preset category that needs to be shared is identified as a positive emotion, a remote care loop and alarm are triggered. Through the connection of the video capture device to a TV or projector, the user can establish a video chat connection through a remote APP to conduct remote care.

[0183] The update rate of historical behavior baselines is constrained based on comprehensive risk scores, and adaptive evolution of behavior baselines is performed. Specifically, the comprehensive risk score generated under the optimal second fusion weight combination is obtained, and a risk constraint relationship corresponding to the current behavior feature results is established. Historical behavior baselines and current behavior feature results are aligned in the same dimension to form a baseline comparison vector for baseline updates.

[0184] The basic learning rate is reduced and mapped based on the comprehensive risk score to obtain the dynamic learning rate corresponding to the current risk status. When the comprehensive risk score is less than or equal to the anomaly judgment threshold, the dynamic learning rate is increased so that the current behavioral feature results can participate in the historical behavior baseline update.

[0185] When the comprehensive risk score is greater than the anomaly detection threshold, the dynamic learning rate is reduced or limited to zero, thereby blocking the writing of abnormal behavior features into the historical behavior baseline.

[0186] Based on the dynamic learning rate, the historical behavioral baseline and the current behavioral feature results are weighted and updated to form a new behavioral baseline; the new behavioral baseline is written back to the behavioral monitoring branch for subsequent behavioral deviation identification and comprehensive risk score generation.

[0187] The optimal second fusion weight combination is selected based on the changing trend of the configuration evaluation value, and when the increase in the fall recognition weight causes the improvement of the configuration evaluation value to converge, the resource consumption tilt is restricted to ensure multimodal comprehensive judgment.

[0188] The edge computing node reads the configuration evaluation value corresponding to each second fusion weight combination from the evaluation result recording area, sorts them according to the network state interval order, and constructs a weight combination-evaluation value correspondence sequence;

[0189] In this sequence, each position corresponds to a second fusion weight combination and its configuration evaluation value, and retains the occupancy status of fall recognition weight, emotion recognition weight and behavior monitoring weight in the combination;

[0190] Edge computing nodes identify a segment in this sequence where the configuration evaluation value increases with the fall detection weight and the first derivative is less than the slope threshold. This segment indicates that while the configuration evaluation value can still improve when the fall detection weight is further increased, the improvement is already limited.

[0191] Since the weights in each second fusion weight combination often exhibit a discrete distribution, the system specifically decomposes the calculation of the first derivative into the difference quotient of the configuration evaluation values ​​between adjacent combinations in a structured manner.

[0192] The difference quotient parameter is calculated based on the ratio of the difference in configuration evaluation value corresponding to adjacent combinations to the difference in fall recognition weight; when the difference quotient parameter is less than the preset slope threshold, the edge computing node determines the corresponding sequence segment as a segment with limited improvement.

[0193] In the aforementioned change zone, the edge computing node determines the position where it first reaches a local maximum and the subsequent change is less than a preset threshold, and the second fusion weight combination corresponding to this position is determined as the candidate optimal weight combination.

[0194] If there are multiple combinations of second fusion weights that satisfy this change condition, then the one with the smallest total weight for emotion recognition and behavior monitoring is selected as the final optimal combination of second fusion weights.

[0195] The selected result is written to the currently effective configuration record area, and the subsequent comprehensive risk scoring process directly reads this final optimal second fusion weight combination;

[0196] Under the final optimal second fusion weight combination, the edge computing nodes perform weighted fusion of fall confidence results, sentiment confidence results, and behavioral feature results;

[0197] The fall confidence results are directly provided by the fall recognition branch; the emotion confidence results are first preserved in chronological order and then accumulated through time decay to form a time-weighted emotion risk quantity, with a corresponding gain weight distribution; the behavioral feature results are compared with the historical behavioral baseline, and the edge computing nodes calculate the degree of deviation between the two and perform normalization processing to form a behavioral deviation risk quantity;

[0198] The comprehensive risk score can be generated using a non-linear mapping method; for example, the fall confidence result, the time-weighted emotional risk, and the behavioral deviation risk are multiplied by the corresponding weights in the final optimal second fusion weight combination and then summed to obtain the total risk accumulation term.

[0199] Then, by constructing a nonlinear mapping function containing a negative exponential term, specifically the constant 1 minus the negative exponent of the natural constant for the total risk accumulation term, the formula is expressed as:

[0200]

[0201] in, For the total risk accumulation term, map it to a value range of... arrive A comprehensive risk score between them;

[0202] Based on the mapping formula, the comprehensive risk score exhibits a nonlinear asymptotic mapping response that tends towards 1 under the condition that the total risk accumulation term increases.

[0203] Through this process, when the fall confidence result is greater than the confidence threshold, or when both emotional risk and behavioral deviation risk increase simultaneously, the comprehensive risk score can increase non-linearly with the increase of input risk.

[0204] The edge computing node compares the comprehensive risk score with the anomaly detection threshold; when the comprehensive risk score exceeds the anomaly detection threshold, the edge computing node triggers an abnormal behavior alarm response.

[0205] Alarm responses can include local alarms, remote notifications, or local responses during network outages; they also include a remote care loop and alarm triggered by facial emotion recognition, specifically: when the emotion confidence result identifies emotions requiring care, such as loneliness, sadness, anger, and worry, or positive emotions requiring sharing, such as happiness and joy, a care instruction is triggered and an alarm prompt is executed; when the network is normal, the edge computing node can trigger a local audio-visual alarm and send a notification to the remote server or remote client; simultaneously, through the connection of the video capture device to a TV or projector, the associated user can access video chat via a remote APP to conduct remote care;

[0206] If the overall risk score does not exceed the anomaly determination threshold, the edge computing node retains the score and the corresponding inference result for use in the subsequent behavior baseline evolution process;

[0207] If there are no valid sentiment confidence results within the scoring period, the edge computing node only performs time decay accumulation processing on the retained sentiment confidence results; if there are no available sentiment confidence results in the current period, the time-weighted sentiment risk is not considered as a new source of risk.

[0208] If the behavioral characteristic results cannot be compared with the historical behavioral baseline in the same dimension, the behavioral deviation risk amount will not be updated in the current period, and the previous effective behavioral deviation risk amount will continue to be used or the next effective behavioral window will be waited for. The above processing is to avoid the interruption of the comprehensive risk scoring process due to the lack of data in a single branch.

[0209] Based on the comprehensive risk score constraint, a corresponding dynamic update rate is configured for the historical behavior baseline; the edge computing node obtains the comprehensive risk score generated under the final optimal second fusion weight combination, and establishes the risk constraint relationship between the score and the current behavior feature result;

[0210] This constraint is written to the baseline update cache along with the current behavioral feature result, enabling the baseline update process to obtain the risk level corresponding to the current behavioral feature result;

[0211] Edge computing nodes perform same-dimensional alignment processing on historical behavior baselines and current behavior feature results to form a baseline comparison vector for baseline updates;

[0212] If the dimensions of the historical behavior baseline and the current behavior feature result are consistent, then proceed directly to the learning rate calculation; if the dimensions cannot be aligned due to the invalidity of the current behavior window, then the historical behavior baseline will not be updated in this cycle, and the original historical behavior baseline will continue to be used for subsequent recognition.

[0213] The dynamic learning rate is obtained by reducing the base learning rate through the comprehensive risk score; the dynamic learning rate can be determined by multiplying the base learning rate by a risk deduction item, and the risk deduction item decreases as the comprehensive risk score increases;

[0214] When the comprehensive risk score is in the low-risk range, the dynamic learning rate is relatively increased, allowing the current behavioral feature results to participate in the historical behavioral baseline update; when the comprehensive risk score is in the high-risk range, the dynamic learning rate is reduced or limited to zero, thereby blocking the writing of abnormal behavioral features to the historical behavioral baseline.

[0215] Edge computing nodes perform weighted updates on historical behavior baselines and current behavior feature results based on dynamic learning rates to form new behavior baselines; the new behavior baselines are written back to the behavior monitoring branch and overwrite or update locally stored historical behavior baseline records;

[0216] When calculating the risk of behavioral deviation, the subsequent behavior monitoring branch will use this new behavior baseline as a reference; if the current comprehensive risk score is high and the dynamic learning rate is zero, the new behavior baseline will be consistent with the original historical behavior baseline, and the behavior monitoring branch will continue to use the original historical behavior baseline.

[0217] During the ongoing operation of residences for elderly people living alone, when the daily routines of the elderly change, the current behavioral characteristics corresponding to the low-risk score can be gradually incorporated into the historical behavioral baseline update, so that the behavioral monitoring branch can adapt to changes in normal activity patterns.

[0218] When a fall, prolonged bed rest, or other abnormal behavior leads to an increase in the overall risk score, the dynamic learning rate is reduced or limited to zero, and the current abnormal behavior characteristics are not written into the historical behavior baseline. The system maintains a baseline that does not contain abnormal behavior characteristic data and uses it as a reference standard for subsequent behavior deviation identification and overall risk scoring.

[0219] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A method for identifying abnormal behavior in the elderly based on edge computing neural networks, characterized in that, Edge computing nodes operating on the connection between video capture devices and televisions or projectors include: Multimodal state recognition: Obtain historical behavioral baselines that characterize the daily activity patterns of the elderly, call neural network models to process video and network state data, construct fall confidence, emotion confidence, behavioral features and network delay sequences, and determine the initial fusion weight set; Network state weight reallocation: The initial fusion weight set is reallocated based on the network delay sequence to generate several second fusion weight combinations; Candidate execution configuration formation: The combination of each second fusion weight is introduced into the edge inference cycle, the inference task is divided and the computing resources are allocated to form an execution configuration including model priority, sampling frequency and processing queue; Edge inference execution: Apply the corresponding execution configuration to edge computing nodes to perform parallel inference and anomaly detection; Configuration evaluation: Extract features from the reasoning process to obtain the corresponding configuration evaluation values; Optimal configuration application: Based on the configuration evaluation value, the optimal second fusion weight combination is selected to generate a comprehensive risk score. When the judgment threshold is exceeded or triggered based on facial emotion recognition, an alarm response and remote care loop are triggered. Baseline evolution adjustment: Adaptive evolution is performed based on the historical behavior baseline update rate constrained by the comprehensive risk score.

2. The method for identifying abnormal behavior in the elderly based on edge computing neural networks according to claim 1, characterized in that, A historical behavioral baseline representing the daily activity patterns of the elderly is obtained. A preset neural network model is then used to process the video data of the elderly's activities and network state data collected by the video acquisition device. This process constructs a fall confidence sequence, an emotion confidence sequence, a behavioral feature sequence, and a network delay sequence. Finally, an initial fusion weight set corresponding to each multimodal state is determined, specifically as follows: The system collects continuous video frames from video capture devices installed in the living areas of the elderly and forms an active video sequence in chronological order. Human target detection and region cropping are performed on the activity video sequence to obtain a human image sequence corresponding to the human subject; The human image sequence is subjected to posture change recognition and temporal motion extraction to obtain a fall confidence sequence representing the fall state; Grayscale conversion and expression recognition processing are performed on the facial regions in the human image sequence to obtain an emotion confidence sequence representing the emotional state; Human contour extraction and trajectory association processing are performed on the activity video sequence to form a behavioral feature sequence under a continuous time window; Collect network communication latency data between edge computing nodes and remote servers, and form a network latency sequence in chronological order; Based on the basic attention levels corresponding to fall recognition, emotion recognition, and behavior monitoring, initial fusion weights are set for each recognition branch, resulting in an initial fusion weight set that corresponds one-to-one with each recognition branch.

3. The method for identifying abnormal behavior in the elderly based on edge computing neural networks according to claim 1, characterized in that, The initial fusion weight set is reallocated based on the network delay sequence to generate several second fusion weight combinations, specifically: The network delay sequence is continuously sampled and statistically analyzed to obtain the network delay value corresponding to each sampling time. Based on a preset network delay upper limit, the network delay values ​​corresponding to each sampling time are normalized to form a network state coefficient sequence that reflects the degree of network congestion. The network state coefficient sequence is divided into intervals to determine the normal network interval, the fluctuating network interval, and the abnormal network interval. Based on the network state coefficient sequence, the degree of computing power focus on the edge side is mapped to form a computing power focus sequence corresponding to each sampling time. While keeping the correspondence between the recognition branches unchanged, the fall recognition weight, emotion recognition weight and behavior monitoring weight in the initial fusion weight set are adjusted incrementally or by decreasing, so as to form the weight adjustment results corresponding to different network state intervals. By performing the weight adjustment operation multiple times, several second fusion weight combinations with different degrees of tilt on the fall recognition side are obtained; Establish a one-to-one correspondence between each second fusion weight combination and its corresponding network state interval for subsequent execution configuration.

4. The method for identifying abnormal behavior in the elderly based on edge computing neural networks according to claim 1, characterized in that, The second fusion weights are combined and introduced into the edge inference cycle to divide the inference task and computing resources, forming an execution configuration that includes model priority, sampling frequency, and processing queue, specifically: The edge computing node is configured with a fixed scheduling cycle, and based on the fixed scheduling cycle, the continuous operation process is divided into several edge inference cycles; Each second fusion weight combination is introduced into a single edge inference cycle as a priority constraint for each identification branch within that inference cycle; Based on the corresponding second fusion weight combination, the fall detection priority processing period and the non-fall detection regular processing period are divided within a single edge inference cycle; Based on the priority processing period for fall detection and the regular processing period for non-fall detection, the neural network processing resources and central processing resources of the edge computing nodes are allocated to form a computing power resource quota corresponding to each detection branch; Based on the computing power resource quota corresponding to each recognition branch, the sampling frequency of video frames, the execution frequency of emotion recognition, and the length of the behavior monitoring time window are adjusted to obtain the sampling frequency adjustment results; The video frame task stream entering the edge computing node is sorted according to the priority and processing time of its respective recognition branch, forming a queue sorting result that includes a high-priority processing queue and a regular processing queue. The computing power resource quota, sampling frequency adjustment result, and queue sorting result are combined to form an execution configuration that corresponds one-to-one with each second fusion weight combination.

5. The method for identifying abnormal behavior in the elderly based on edge computing neural networks according to claim 1, characterized in that, The execution configurations corresponding to each second fusion weight combination are applied to the edge computing nodes to perform multimodal parallel inference and abnormal behavior recognition, specifically as follows: The edge computing node is configured with a video access link and a basic identification link. The video access link and the basic identification link are kept running continuously in the edge computing node, and a high-priority inference link for key anomaly identification is configured. For each second fusion weight combination, during the operation of the edge computing node, the high-priority inference link is activated according to the execution configuration corresponding to the second fusion weight combination; When the high-priority inference link is active, video frames within the corresponding time period will be prioritized for access to the fall recognition branch, and simultaneously accessed to the emotion recognition branch and behavior monitoring branch. During the high-priority inference link activation period, the emotion recognition task and behavior monitoring task are executed at a reduced frequency or continuously, depending on the sampling frequency and time window length set in the execution configuration. When the configured priority processing period ends, stop inputting new key recognition tasks into the high-priority inference link and restore subsequent video frames to the basic recognition link for continued processing; The fall confidence results output by the fall recognition branch, the emotion confidence results output by the emotion recognition branch, and the behavioral feature results output by the behavior monitoring branch are synchronously aggregated to form a multimodal inference result corresponding to the execution configuration. The multimodal inference results generated under each execution configuration are identified and output separately for subsequent configuration evaluation.

6. The method for identifying abnormal behavior in the elderly based on edge computing neural networks according to claim 1, characterized in that, The inference process features under each second fusion weight combination are extracted, and the configuration evaluation value corresponding to each second fusion weight combination is obtained, specifically: For each applied second fusion weight combination, during the continuous operation of the edge computing node, the inference start time, inference end time, and anomaly identification output result under the corresponding execution configuration are recorded, and a single inference process is extracted. Within each single inference process, the processing time between the fall recognition result input and the abnormal output of the video frame is statistically analyzed to form the emergency response delay feature; Within each single inference process, the retention length of the emotion confidence result within the continuous time window and the degree of decay correlation between adjacent time steps are statistically analyzed to form the continuous emotion perception feature. Within each single inference process, the magnitude of deviation of the statistical behavioral characteristic results from the historical behavioral baseline and the degree of fluctuation between adjacent windows form the stable behavioral deviation characteristics. In each single inference process, the computing power occupancy ratio and task queue length of the edge computing node under each identification branch are recorded to form resource occupancy characteristics. The emergency response delay characteristics, continuous emotion perception characteristics, behavioral deviation stability characteristics, and resource consumption characteristics are combined in a fixed order to form a configuration evaluation vector corresponding to each second fusion weight combination. An evaluation function is constructed based on each feature in the configuration evaluation vector, and the configuration evaluation value corresponding to each second fusion weight combination is calculated.

7. The method for identifying abnormal behavior in the elderly based on edge computing neural networks according to claim 1, characterized in that, The optimal second fusion weight combination is selected based on the configuration evaluation value, and a comprehensive risk score is generated under the optimal second fusion weight combination. When the comprehensive risk score exceeds the anomaly detection threshold, an abnormal behavior alarm response is triggered, specifically as follows: The configuration evaluation values ​​corresponding to each second fusion weight combination are sorted according to the network state interval order to construct a sequence of weight combination-evaluation value correspondence. In the weight combination-evaluation value correspondence sequence, identify the segment where the configuration evaluation value gradually converges with the increase of the fall recognition weight, and determine the second fusion weight combination corresponding to the segment as the candidate optimal weight combination; The combination with the smallest total weight occupancy among the candidate optimal weight combinations is selected as the final optimal second fusion weight combination; the final optimal second fusion weight combination is configured to the edge computing node to perform weighted fusion on the fall confidence result, sentiment confidence result and behavioral feature result; The emotional confidence results are subjected to time decay accumulation processing to form a time-weighted emotional risk quantity; The deviation between the behavioral characteristic results and the historical behavioral baseline is normalized to form a behavioral deviation risk quantity. The fall confidence result, time-weighted emotional risk and behavioral deviation risk are nonlinearly mapped to generate a comprehensive risk score. When the comprehensive risk score exceeds the anomaly judgment threshold, an abnormal behavior alarm response is triggered. Based on the emotion confidence results, facial emotion classification is performed. When an emotion belonging to a preset category that needs care or a positive emotion belonging to a preset category that needs sharing is identified, the remote care loop and alarm are triggered. Through the connection between the video acquisition device and the TV or projector, the user terminal establishes a video chat connection through a remote APP to conduct remote care.

8. The method for identifying abnormal behavior in the elderly based on edge computing neural networks according to claim 1, characterized in that, The update rate of the historical behavioral baseline is constrained based on the comprehensive risk score, and the adaptive evolution of the behavioral baseline is performed, specifically as follows: Obtain the comprehensive risk score generated under the optimal second fusion weight combination, and establish a risk constraint relationship corresponding to the current behavioral feature results; Historical behavior baselines and current behavior feature results are aligned in the same dimension to form a baseline comparison vector for baseline updates; Based on the comprehensive risk score, the basic learning rate is reduced and mapped to obtain the dynamic learning rate corresponding to the current risk state. When the comprehensive risk score is less than or equal to the anomaly determination threshold, the dynamic learning rate is increased so that the current behavioral feature results can participate in the historical behavioral baseline update. When the comprehensive risk score is greater than the anomaly determination threshold, the dynamic learning rate is reduced or limited to zero, thereby blocking the writing of abnormal behavior features to the historical behavior baseline. Based on the dynamic learning rate, a weighted update is performed on the historical behavioral baseline and the current behavioral feature results to form a new behavioral baseline; The new behavioral baseline is written back to the behavioral monitoring branch for subsequent behavioral deviation identification and comprehensive risk score generation.

Images