Suicide risk behavior early warning method and system based on micro-expression visual capture

Abstract

The application provides a suicide risk behavior early warning method and system based on micro-expression visual capture, introduces negative emotion accumulation, emotional abnormal back-off and other suicide-specific psychological characteristics, can identify risks in the hesitation stage, early warning, improve the success rate of intervention, and truly realize the upgrade from "after-disposal" to "pre-intervention". By constructing a "emotion quantification + behavior semantics + historical trend" three-dimensional risk assessment system, the weight is inclined to emotion and history, which conforms to the psychological evolution law of suicide risk; compared with the pure posture recognition scheme, the false negative rate is reduced, and the core pain points of high false negative rate and insufficient prevention and control reliability of the traditional scheme are solved. A conservative grading system with low threshold and high sensitivity is designed, and an adaptive mechanism of one-vote deep judgment in the high-risk area and double punishment for false negatives is adopted to ensure that the system always controls the false negative rate below 5%, meeting the life safety prevention and control requirements of high-risk scenes.

Description

Technical Field

[0001] This invention relates to the field of human behavior early warning technology, and in particular to a method and system for early warning of suicide risk behaviors based on micro-expression visual capture, as well as an electronic device and a computer-readable storage medium. Background Technology

[0002] Suicide is a major global public health and social safety issue. Among them, suicides caused by jumping from buildings, bridges, or rivers account for a high proportion, and are mostly concentrated in key enclosed / semi-enclosed spaces such as overpasses, rooftop platforms, waterfront areas, rail transit platforms, psychiatric hospitals, and detention centers.

[0003] Suicide is characterized by its high degree of concealment, suddenness, and irreversible consequences. Traditional prevention and control models rely heavily on manual patrols and guard duties, which are limited by personnel energy and blind spots, resulting in a high rate of missed prevention. Most interventions occur after the behavior has been committed, missing the best window for intervention.

[0004] Clinical psychology research confirms that suicidal behavior is not a momentary impulsive decision, but rather follows a clear psychological evolutionary chain: from the continuous accumulation of negative emotions and intensified psychological suffering, to the emergence of suicidal ideation, hesitant behavior, and finally, the execution of the act. The entire process can last from several hours to several days. Among these processes, facial micro-expressions, as physiological reactions beyond conscious control, can accurately reflect an individual's psychological states such as sadness, despair, anxiety, and numbness. In particular, the abnormal decline in emotion from intense pain to unusual calm is a typical precursor signal that suicidal ideation has entered the execution stage.

[0005] Current suicide risk prevention technologies are mainly divided into two categories: clinical psychological scale screening and video behavior recognition, both of which have insurmountable technical bottlenecks.

[0006] 1. Mainstream suicide risk assessment relies on subjective questionnaires such as the SDS Self-Rating Depression Scale and the Beck Suicide Ideation Scale, which can only complete a single static screening upon admission to the hospital / detention center and cannot conduct continuous dynamic tracking of individuals for 24 hours. Moreover, the results of the scales are easily affected by the subjective concealment of the test subjects, with a false negative rate as high as 35%, and cannot identify impulsive suicide caused by sudden emotional deterioration.

[0007] 2. Existing video early warning solutions are based solely on human skeletal posture recognition for extreme actions such as climbing and vaulting. Essentially, they are still "post-event alarms," ​​which can only trigger an early warning when the suicidal action has fully occurred, at which point the success rate of intervention is extremely low. Furthermore, they do not incorporate the dimensions of emotions and psychological states, and cannot identify high-risk behaviors such as loitering or staring in the early stages, nor can they distinguish between normal edge work and suicidal intent.

[0008] 3. Suicide risk is an eruption after a long period of negative emotions. Existing solutions are all based on single frames / short segments for judgment, and no individual-level long-term psychological profiles and trend analysis mechanisms have been established. They cannot capture the process of gradual deterioration of emotions, nor can they identify the key precursor signal of "abnormal calm after pain", resulting in a significant lag in early warning.

[0009] 4. General security solutions adopt a threshold strategy that balances false alarms and false alarms. However, suicide prevention is a high-risk scenario with "zero false alarms," ​​and the cost of false alarms is loss of life. Existing systems do not set conservative thresholds or a one-vote deep judgment mechanism for high-risk behaviors for this scenario, and cannot meet the prevention and control requirements of key areas.

[0010] Therefore, if the above psychological characteristics and temporal evolution patterns can be captured non-contactly through computer vision technology, the early warning node can be significantly moved forward from the "behavior implementation" stage to the "idea formation and hesitation stage," thereby significantly improving the success rate of intervention. Summary of the Invention

[0011] To address the technical problems existing in the prior art, the present invention provides the following technical solution:

[0012] On the one hand, a method for early warning of suicidal risk behaviors based on micro-expression visual capture is provided, including the following steps:

[0013] Key Area Video Acquisition and Adaptive Preprocessing: Acquire monitoring video streams for high-risk areas under graded management, dynamically perform intelligent frame extraction based on the area risk level and real-time scene anomaly, complete face alignment and image enhancement, and output standardized image sequences;

[0014] Key personnel detection and tracking and individual psychological profile initialization: Detect all human targets in the image, assign a unique tracking ID to each target and extract ReID appearance features, and initialize an individual psychological profile for the first target to appear;

[0015] Dual-model parallel inference: The local lightweight quantization emotion engine runs frame by frame to extract facial action units and quantify and output the intensity of multiple emotions, continuously monitoring the accumulation of negative emotions and the abnormal decline of emotions; when abnormal emotional signals or targets enter the high-risk area, the VLM semantic inference engine is triggered, and the suicidal intent is determined in depth by combining body posture, facial expression and scene context.

[0016] Multi-dimensional weighted fusion and comprehensive risk score calculation: Following a weighting strategy that prioritizes emotion, supplements behavior, and strongly modifies historical trends, semantic risk score, emotional risk score, and time-series persistence and historical trend correction score are weighted and fused to output a comprehensive suicide risk score.

[0017] Suicide risk scenario triage and specific rule matching: Risk targets are triaged to a dedicated suicide risk assessment link, and five high-risk behavior assessment rules are matched for climbing and scaling, loitering at the edge, staring at stillness, self-harming actions, and extreme abnormal emotions to identify specific risk types.

[0018] Multi-level conservative early warning and multi-entity joint intervention: Following the conservative principle of "better to report false than to miss a report", a multi-level early warning system is set up, and corresponding psychological intervention, on-site duty, security dispatch and emergency response measures are matched.

[0019] Full-cycle psychological profile updates and closed-loop iterative optimization: continuously update individual time-series psychological profiles and track emotional evolution trends; based on manual review results, adaptively adjust early warning thresholds with conservative strategies and incrementally update the knowledge base of suicidal behavior patterns.

[0020] As an application example of this application, the intelligent frame extraction combines regional risk level and scene anomaly degree for dual adjustment: different levels of control areas are set with differentiated basic sampling frame rates, and high-risk areas default to a higher frame rate; when abnormal situations such as people approaching the edge or drastic emotional fluctuations are detected, the frame extraction rate is further increased as the anomaly degree increases, prioritizing the motion capture time resolution of high-risk areas under limited computing power; face alignment normalizes non-frontal face postures such as looking down and side profile to a standard template through multiple facial key points, focusing on amplifying key micro-expression areas such as eyebrows, eyes, and around the mouth, thereby improving the accuracy of negative micro-expression recognition under complex postures.

[0021] As an application example of this application, the multi-target tracking adopts a matching strategy that combines motion trajectory similarity and appearance feature similarity, wherein the weight of appearance features is significantly higher than that of motion trajectory, in order to adapt to scenarios where high-risk suicide individuals remain still for a long time or wander slowly, resulting in insufficient motion information, and to improve the tracking stability in slow motion and static states; ReID features are used for cross-frame and cross-camera personnel identity matching, and a matching threshold higher than that of general security scenarios is adopted to avoid mismatches between different individuals and to ensure the consistency of identity in psychological profiles.

[0022] As an application example of this application, the dual-model parallel inference configuration is configured with the following scheduling triggering rules: when a target enters a first-level high-risk border area, the initial emotion screening is skipped and the semantic inference engine is directly activated for in-depth analysis to prevent missed detections in high-risk scenarios; when negative emotions accumulate to a threshold or abnormal emotional decline is detected, the semantic engine is automatically triggered for in-depth analysis; when the system determines that a person is at level two or above risk, the allocation of frame extraction and recognition computing power for the corresponding camera is automatically increased to achieve dynamic tilting of computing power towards high-risk targets.

[0023] As an application example of this application, the local lightweight quantization emotion engine constructs a suicide-specific emotion mapping relationship based on the facial motion coding system, and strengthens the recognition weight of facial expression features that are highly correlated with suicidal ideation, such as sadness, despair, anxiety, and numbness; it calculates the accumulation level of negative emotions through a sliding window, and specifically identifies the abnormal decline feature of negative emotions rapidly falling back to an abnormal calm after reaching a peak. This feature serves as a core pre-implementation signal for suicidal ideation; when the accumulation of negative emotions continues to reach the target, or the abnormal decline of emotions is determined, or the target enters the first-level high-risk area, the semantic reasoning engine is triggered for in-depth analysis.

[0024] As an application example of this application, the VLM semantic reasoning engine receives a full-body image of the target, short time sequence fragments, and scene region masks as input. It performs multimodal joint reasoning by combining limb posture, action trend, edge distance, and facial state. After vectorizing the output behavioral semantic description, it performs similarity retrieval and matching with a pre-set suicide behavior pattern knowledge base. After matching the corresponding risk pattern, it calculates the semantic risk score by combining the behavioral confidence and scene risk coefficient. Among them, the same behavior in high-risk areas is given a higher risk weight. The knowledge base matching adopts a relatively loose similarity threshold to cover more atypical high-risk behaviors and reduce the false negative rate.

[0025] As an application example of this application, the comprehensive suicide risk score is obtained by weighting and fusing three dimensions—semantic behavioral risk, emotional quantitative risk, and temporal history correction—according to the specific weights of the suicide scenario. The comprehensive suicide risk score is defined as follows:

[0026] in For semantic risk scoring, Emotional risk is categorized. For time series continuation, The score is adjusted based on the historical trend of the 7-day psychological profile. , , These are the weights of the corresponding items;

[0027] The emotional dimension has the highest weight, while the behavioral semantic dimension serves as the basis for risk confirmation. In the temporal and historical correction dimensions, the weight of historical psychological trends is higher than that of immediate duration, highlighting the psychological driving nature and long-term cumulative characteristics of suicide risk. The historical trend correction score is calculated based on the degree of deterioration of negative emotions in the individual's psychological profile in the near term, reflecting the long-term evolution trend of psychological state.

[0028] As an application example of this application, the five types of high-risk behaviors are all subject to multi-dimensional integrated judgment rules. Each rule integrates semantic judgment results, emotional characteristics, spatial location conditions, and duration conditions. When all conditions are met, the corresponding risk type is locked. Among them, climbing and traversing behaviors are given higher risk priority as long as the semantics are matched and the location is correct, without being limited by the comprehensive scoring threshold. The judgment of lingering at the edge requires the conditions of the duration of stay at the edge or the number of times of going back and forth and the presence of negative emotions. The static staring behavior requires a comprehensive judgment based on posture, orientation, and emotional characteristics. Self-harming behaviors are directly judged as high-risk under high confidence. Extreme abnormal emotions are included in the key attention even if no high-risk behavior is shown.

[0029] As an application example of this application, the starting threshold of the multi-level conservative early warning is lower than that of general public safety scenarios, reflecting the prevention and control principle of "better to report false alarms than to miss alarms". The four-level early warning corresponds to the four-level response system of background observation, personnel intervention, on-site handling and emergency linkage. Clearly high-risk actions such as climbing and self-harm can skip levels directly. The adaptive adjustment of the early warning threshold adopts a conservative small-step incremental strategy, and the penalty weight for missed alarms is significantly higher than that for false alarms, ensuring that the missed alarm rate is always maintained at an extremely low level during the iteration process. The individual psychological profile uses ReID features as a unique index to store temporal emotional trajectories, records of previous high-risk early warnings and wandering behavior trajectories, supporting periodic psychological trend analysis and providing quantitative data support for long-term psychological intervention.

[0030] On the other hand, a suicide risk behavior early warning system based on micro-expression visual capture is provided, including:

[0031] The video acquisition layer consists of high-risk area surveillance cameras deployed in a tiered manner and edge gateways. It is used to acquire video streams in key areas and mark the risk level of the areas. It supports local caching when the network is disconnected and resume interrupted downloads.

[0032] The edge preprocessing layer includes an adaptive frame extraction module, a face alignment enhancement module, a human detection module, a multi-target tracking and ReID module, and an individual cropping module. It is used for video preprocessing, target detection and tracking, and individual psychological profile initialization. Sensitive biometric data is processed locally at the edge.

[0033] The dual-engine model layer includes a local lightweight quantized emotion engine and a VLM semantic reasoning engine. The former runs in full to achieve continuous monitoring of negative emotions and specific capture of abnormal emotional decline, while the latter conducts joint in-depth analysis of the suicidal intent of high-risk targets based on their behavior, expressions, and scenarios.

[0034] The intelligent decision-making layer includes a multi-dimensional weight fusion module, a suicide risk rule matching module, and a conservative grading judgment module, which are used to calculate a comprehensive suicide risk score, match high-risk behavior types, and complete a four-level early warning classification.

[0035] The linkage intervention layer includes an early warning push module, a remote reassurance module, an on-site handling module, and an emergency linkage module, which are used to implement tiered psychological intervention and multi-stakeholder collaborative handling.

[0036] The closed-loop psychological data layer includes an individual psychological profile database, a knowledge base of suicidal behavior patterns, and an event evidence database. It is used for time-series psychological tracking, incremental updates of the knowledge base, and retention of event evidence. A layered storage architecture with edge and intranet layers is adopted to ensure data security.

[0037] On the other hand, an electronic device is provided, comprising: a processor; and a memory storing computer-readable instructions, which, when executed by the processor, implement the method described above.

[0038] On the other hand, a computer-readable storage medium is provided, wherein at least one instruction is stored therein, the at least one instruction being loaded and executed by a processor to implement the above method.

[0039] The beneficial effects of the technical solutions provided by the embodiments of the present invention include at least the following:

[0040] This invention breaks through the traditional approach of only recognizing extreme physical movements and issuing alarms after the fact. It innovatively introduces suicide-specific psychological characteristics such as the accumulation of negative emotions and abnormal emotional decline, enabling risk identification during periods of hesitation and indecision, providing early warnings, and improving intervention success rates. This truly upgrades the approach from "post-event handling" to "pre-event intervention." By constructing a three-dimensional risk assessment system of "emotional quantification + behavioral semantics + historical trends," with weights shifted towards emotions and historical context, it aligns with the psychological evolution of suicide risk. Compared to pure posture recognition solutions, the false negative rate is reduced, addressing the core pain points of high false negative rates and insufficient reliability in traditional methods.

[0041] In response to the characteristic of suicide prevention where "the cost of underreporting is far higher than that of false alarms," ​​a conservative grading system with low thresholds and high sensitivity is designed, along with an adaptive mechanism that imposes a deep judgment on high-risk areas and double penalties for underreporting, ensuring that the system always keeps the underreporting rate below 5% and meets the life safety prevention and control requirements of high-risk scenarios such as bridges, rooftops, and prisons.

[0042] In addition, by establishing individual-level time-series psychological profiles and tracking the evolution of emotions, key individuals whose mental state is continuously deteriorating can be identified in advance, and psychological intervention departments can be coordinated to carry out proactive counseling. The system continuously iterates its knowledge base and thresholds through manual review, and its capabilities are constantly evolving, which is different from the one-time delivery model of traditional fixed algorithms. Attached Figure Description

[0043] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0044] Figure 1 This is an architecture diagram of a suicide risk behavior early warning system based on micro-expression visual capture provided by an embodiment of the present invention;

[0045] Figure 2 yes Figure 1 The system's data flow logic diagram;

[0046] Figure 3 This is a flowchart of the method of the present invention. Detailed Implementation

[0047] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0048] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0049] In the embodiments of this invention, the terms "image" and "picture" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning. Similarly, the terms "of," "corresponding (relevant)," and "corresponding" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning.

[0050] In this embodiment of the invention, sometimes a subscript such as W1 may be mistakenly written as a non-subscript form such as W1. When the difference is not emphasized, the meaning they express is the same.

[0051] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0052] This application constructs a suicide risk-specific micro-expression feature dataset, collects the temporal features of facial micro-expressions of target subjects, and combines them with a psychological risk assessment model to achieve graded identification and early warning of suicide risk.

[0053] This application can be used to provide early warning of suicide risk behaviors in the following scenarios: targeting key areas and key personnel such as overpasses, rooftops, waterways, platforms, hospitals, and detention centers, to identify behaviors such as climbing and scaling heights, prolonged periods of stillness, lingering in dangerous areas, self-harming actions, and extreme emotions, and to intervene in advance.

[0054] The specific suicide scenario can be adjusted according to the existing social context, as long as it meets the technical risk warning principle of this application.

[0055] I. Suicide Risk Behavior Early Warning Framework

[0056] This suicide risk early warning system adopts a core technical approach of edge full-volume monitoring + cloud-based in-depth analysis. The core design principles of the system architecture are: prioritizing local processing of sensitive data at the edge, optimizing algorithms specifically for suicide risk scenarios, and dynamically allocating computing resources. While ensuring low false negatives and high early warning sensitivity for suicide incidents, it meets the data compliance and regulatory requirements of special scenarios such as medical wards, detention cells, urban overpasses, and high-risk waterfront areas. At the same time, it takes into account engineering implementation indicators such as 24 / 7 uninterrupted stable operation and low computing power consumption for multi-channel video.

[0057] like Figure 1 As shown, the system architecture includes:

[0058] I. Video Acquisition Layer (System Awareness Entry Point)

[0059] As the system's hardware sensing unit, the video acquisition layer is used to complete the acquisition of raw video and audio data across the entire scene. It customizes point deployment strategies for high-risk suicide areas such as overpasses, rooftops, waterways, wards, and prison cells, achieving full coverage of high-risk points such as guardrails, water edges, and rooftop platforms without blind spots. It is the data source for all subsequent identification and analysis processes.

[0060] The video capture layer includes:

[0061] 1. Front-end high-definition acquisition equipment

[0062] The system is divided into two standardized hardware categories: fixed camera equipment and intelligent tracking PTZ cameras. Both utilize a 4-megapixel low-light imaging sensor and natively support infrared night vision, wide dynamic range, and electronic image stabilization. This allows them to adapt to complex lighting environments prone to suicide incidents, such as early morning, backlighting, rain, and low-light conditions in hospital wards. Intelligent PTZ cameras are installed in high-risk areas such as near edges and rooftops. Upon system alert, they automatically zoom in to capture facial and limb details. Audio pickup modules are also installed in hospital wards and psychological treatment rooms to collect ambient sound data as supplementary information for assessing suicide attempts.

[0063] 2. Regional hierarchical control mechanism

[0064] The system divides all monitored areas into three categories: Level 1 High-Risk Zone, Level 2 Attention Zone, and Level 3 Ordinary Zone. The Level 1 High-Risk Zone includes areas within 1 meter of guardrails, water edges, and rooftops without protective platforms—areas prone to suicides—and is configured with the highest detection accuracy, highest frame sampling frequency, and highest warning priority by default. The Level 2 Attention Zone includes quiet corners of plazas and stairwells—areas where people linger—and is configured with medium computing resources. The Level 3 Ordinary Passage Zone reduces sampling overhead and conserves edge computing power. Zone division parameters can be customized through a visual interface in the backend.

[0065] 3. Edge Access Gateway

[0066] As a multi-channel video aggregation and relay unit, it is compatible with RTSP and ONVIF universal monitoring protocols, and uniformly aggregates video streams from all front-end devices; it has a built-in local loop cache module, supports 72 hours of offline storage during network outages, and automatically resumes lost video clips after network recovery, preventing the loss of critical footage data and ensuring the integrity of event tracing.

[0067] The video capture layer outputs a standardized decoded video stream, which is pushed unidirectionally to the upper edge preprocessing layer. Only the standardized image bitstream is transmitted, and no business logic data is stored. The gateway cached data is synchronously backed up to the underlying psychological closed-loop data layer event evidence library as original evidence material.

[0068] II. Edge Preprocessing Layer (Sensitive Data Local Processing Unit)

[0069] The edge preprocessing unit is deployed on the local node of the edge computing box. Its core design goal is to prevent biosensitive data such as faces, micro-expressions, and human bodies from leaving the edge hardware, thus avoiding the risk of privacy leakage at the source. It performs standardized cleaning, target localization, face enhancement, and individual identification preprocessing on the original video, providing standardized input images for the dual-engine model layer. It is an intermediate conversion hub connecting hardware acquisition and AI recognition.

[0070] The following preprocessing can be performed in the edge preprocessing layer:

[0071] 1. Adaptive frame skipping module

[0072] The frame sampling rate is dynamically adjusted by combining regional classification priority and real-time scene anomaly: the first-level high-risk area maintains a high sampling frame rate, and automatically increases the sampling density when abnormal behaviors such as people approaching the edge or remaining still for a long time are detected; the third-level ordinary passage area reduces the frame sampling frequency, which significantly reduces the edge computing power occupation without reducing the suicide warning sensitivity, and balances the recognition accuracy and hardware load.

[0073] 2. Face alignment enhancement module

[0074] We perform specialized image optimization for common facial scenarios involving suicidal individuals, such as looking down, turning their face to the side, being covered by masks, and having blurred faces in low light. We use the coordinates of 68 key points on the face to achieve pose alignment, and combine this with a super-resolution algorithm to repair blurred facial details. We focus on magnifying key micro-expression areas such as eyebrows, corners of the mouth, and around the eyes to improve the accuracy of recognizing negative micro-expressions related to suicide, such as sadness, numbness, and despair.

[0075] 3. Human body detection and positioning module

[0076] The lightweight real-time human detection algorithm outputs the bounding boxes of all human targets and torso pose features in the frame-by-frame, filtering out irrelevant and interfering objects such as flowers, plants, and pillars, accurately locking all personnel entering the controlled area, and defining an effective calculation area for subsequent tracking and emotion recognition.

[0077] 4. Multi-target tracking and ReID module

[0078] Each individual human body in the frame is assigned a globally unique tracking ID, and cross-frame and cross-camera personnel re-identification is completed through 256-dimensional appearance feature vectors; even if a person is briefly obscured by a fence or wall, the same identity identifier can be continuously bound, ensuring the continuity of emotional time series data of a single person over multiple days and at multiple locations, and providing a unified identity index for individual psychological profiles.

[0079] 5. Individual cropping module

[0080] Image materials are automatically segmented based on tracking IDs, and standardized face sub-images and full-body limb sub-images are cropped and normalized in size. Irrelevant background pixels are filtered out, and only valid human body and face images are sent to the upper-level AI model to reduce unnecessary computation.

[0081] The edge preprocessing layer only pushes the desensitized and cropped standardized images of faces and bodies to the dual-engine model layer; the original complete video and uncropped original image are permanently stored in the edge gateway and are only retrieved during manual review, without transmitting the original sensitive images to the cloud.

[0082] III. Dual-engine model layer (core layer for AI-driven suicide risk assessment)

[0083] In response to suicide risk scenarios, this system designs a dual-model parallel engine (Model 1 + Model 2) for risk analysis and early warning, which is the core technical design point of this application.

[0084] Specifically, the engine employs a dual-model parallel architecture for risk warning analysis, consisting of an edge-based lightweight quantization sentiment engine for full-path parallel monitoring and a private cloud VLM semantic reasoning engine for on-demand in-depth analysis. This includes:

[0085] 1. VLM Multimodal Semantic Reasoning Engine (Model 1)

[0086] Model 1 can be deployed on private cloud GPU servers within hospitals and regulatory agencies. It operates in low-power standby mode and only starts computing upon receiving a trigger command. It is defined as a "suicidal intent deep analyzer" and is responsible for the comprehensive judgment of high-risk scenarios and suicidal behaviors.

[0087] Model 1 can be activated under the following wake-up trigger conditions:

[0088] The system will automatically activate upon personnel entering a Level 1 high-risk border area.

[0089] The edge emotion engine detected a persistently high level of negative emotions and an abnormal decline in emotions;

[0090] Alternatively, back-end medical staff and security personnel can manually designate targets to initiate analysis.

[0091] After activation, Model 1 receives three types of multimodal data pushed from the edge: facial time sequence images, full-body posture images, and scene area coordinates. It matches these data with a suicide risk-specific behavioral knowledge base, which includes major and minor subcategories of suicide scenarios such as climbing and scaling heights, prolonged loitering near edges, static staring at water, self-harming actions, and leaving behind suicidal items (the specific matching is based on the scene structure data of the preset knowledge base and is custom-designed). It combines micro-expression emotion data, spatial location of people, and body movements to perform multimodal joint semantic reasoning and outputs standardized suicide risk intentions, risk levels, and corresponding intervention suggestions.

[0092] 2. Local lightweight quantization emotion engine (Model 2, which can run concurrently on ordinary edge NPU hardware)

[0093] The quantitative emotion engine model is deployed on edge computing nodes and runs frame by frame 24 hours a day. It is defined as the system's "real-time psychological state monitor" and undertakes the core functions of extracting micro-expressions from all users, quantifying negative emotions, and triggering initial risk screening.

[0094] Model 2 is based on facial micro-expressions for emotion recognition: by extracting multiple facial actions (AU) frame by frame, it targets micro-expression features of typical emotions in suicidal individuals, such as sadness, numbness, and despair (it can also specifically enhance the recognition weight of four key action units: AU1 raising the inner eyebrows, AU4 frowning, AU15 lowering the corners of the mouth, and AU25 separating the lips). Through the FACS emotion mapping matrix, it calculates the scores of seven emotion categories in real time: joy, anger, worry, grief, surprise, fear, and calmness. It focuses on continuously tracking the duration and cumulative intensity of three negative emotions: sadness, fear, and anxiety. It can specifically identify the high-risk psychological feature of "abnormal numbness and calmness after prolonged sadness" as the core signal to trigger in-depth cloud-based analysis.

[0095] In addition, Model 2 can continuously push the time-series data of each person's emotional intensity to the intelligent decision-making layer in real time; when negative emotions reach the standard or abnormal calm occurs, it sends a trigger command to the upper-level scheduling module to wake up the VLM semantic reasoning engine.

[0096] The two models have a clear division of labor and work together in a coordinated manner. An example of their coordinated scheduling rules is as follows:

[0097] Level 1 high-risk areas: If a target enters a Level 1 control area such as the edge or rooftop, the initial emotional screening is skipped and the VLM is activated directly to prevent suicidal scenarios from being missed.

[0098] Emotional abnormality timing trigger: When negative emotions accumulate to a threshold, in-depth analysis is automatically initiated;

[0099] Dynamic computing power allocation: After the system identifies personnel at level 2 or above risk, it automatically increases the frame extraction of the corresponding camera and identifies the allocation of computing power.

[0100] Compared to real-time computation using a large-scale model across all networks, this collaborative mechanism reduces computing resource consumption by 80% and can support more than five times the number of monitoring channels with the same hardware. Therefore, it addresses the pain points of general security models failing to recognize suicidal emotions such as despair and numbness, as well as the missed reporting in high-risk scenarios near borders. The two-layer model data exchange and trigger mechanism linkage constitute a two-layer suicide risk identification system. After in-depth analysis, structured behavioral risk scores and scenario judgment results are sent to the intelligent decision-making layer; after the analysis is completed, it automatically returns to standby mode to release cloud computing power.

[0101] The dual engines simultaneously output emotional quantitative scores and behavioral semantic scores to the upper-level intelligent decision-making layer; at the same time, the intermediate feature data is fed back to the suicide behavior knowledge base of the psychological closed-loop data layer for iterative optimization.

[0102] IV. Intelligent Decision-Making Layer (Centralized Risk Assessment Platform)

[0103] The intelligent decision-making layer receives quantitative data on emotions and behaviors from the dual-engine model layer, performs risk quantification and classification, customizes a weighted fusion formula and conservative warning threshold based on the suicide scenario, integrates three multi-dimensional data—emotional time series, high-risk behaviors, and historical psychological profiles—and outputs a standardized comprehensive suicide risk score from 0 to 100, classifying it into four intervention levels. It has the following functional modules:

[0104] 1. Weighted fusion module

[0105] A weighting strategy specific to suicide scenarios is adopted: emotional temporal trends are given the highest weight, high-risk behaviors are secondary, and historical psychological records are used for correction. A unified risk score is calculated by comprehensive weighting. Priority is given to amplifying the scores of suicidal characteristics such as long-term persistent negative emotions and lingering on the edge. The conservative judgment logic of "better to report falsely than to miss" is followed to adapt to the high safety requirements of suicide scenarios.

[0106] 2. Risk Triage Module

[0107] The weighted risk targets are matched with a suicide-specific rule base to distinguish different risk types such as near-death experiences, indoor self-harm, and long-term psychological depression, and corresponding intervention and treatment procedures are matched accordingly.

[0108] 3. Grading Determination Module

[0109] The risk is divided into four levels based on the comprehensive score. Each level is associated with a fixed intervention standard. The higher the risk level, the stricter the response speed requirements. Level 1 high-risk events directly trigger the emergency response channel.

[0110] Finally, the intelligent decision-making layer pushes the graded early warning instructions down to the linkage intervention layer; the daily graded judgment statistics flow back to the psychological closed-loop data layer for threshold adaptive iteration.

[0111] V. Joint Intervention Layer (Tiered Disposal Execution Unit)

[0112] The coordinated intervention layer connects with various response entities, including mental health care, on-site security, and public security emergency services. Based on the four-level risk assessment issued by the intelligent decision-making layer, it executes differentiated intervention actions to achieve risk classification and precise resource allocation. It includes the following functional modules:

[0113] 1. Early Warning Push Module

[0114] The system synchronizes complete early warning data packages with mental health workstations and frontline security mobile apps. These packages include real-time images of the target, continuous emotional change curves, location of the target, risk level, standardized reassurance and dissuasion scripts, and intervention process suggestions to assist staff in quick response.

[0115] 2. Remote intervention module

[0116] The system works in conjunction with on-site speakers and broadcasting equipment to automatically play gentle, reassuring, and dissuading messages, delaying suicidal actions and providing buffer time for people to arrive at the scene.

[0117] 3. On-site handling module

[0118] Based on GIS location matching, the system can identify the nearest security personnel to the target, dispatch work orders with one click, simultaneously push the optimal arrival route, and record personnel response time.

[0119] 4. Emergency Response Module

[0120] The highest level of extreme suicide risk warning supports one-click access to emergency channels for public security, fire protection, and emergency medical services, and provides real-time push of monitoring footage and precise geographical location, while simultaneously preserving evidence videos.

[0121] All handling operations and personnel response records are simultaneously uploaded to the psychological closed-loop data layer for archiving and retention.

[0122] VI. Psychological Closed-Loop Data Layer (System Data Foundation and Iteration Center)

[0123] Ordinary security platforms, such as those used by public security agencies and transportation departments, typically only store behavioral data. If someone commits suicide by jumping into a river, only the behavioral video is available, without any physiological data about the person. This system, however, differs from general security platforms in that it has a core, dedicated layer that undertakes three major functions: storing personnel psychological data, establishing rules governing suicidal behavior, and preserving evidence from incidents. It also features a compliant edge-private cloud intranet storage architecture, strictly controlling access to biometric data, and is suitable for highly privacy-sensitive scenarios such as hospital wards and prisons.

[0124] The system integrates three dedicated databases:

[0125] 1. Individual Psychological Profile: Using ReID personnel characteristics as the unique global index, each individual independently archives long-term time-series emotional curves, records of previous high-risk warnings, trajectory of hesitant behavior, and phased changes in psychological risk; supports statistical analysis of psychological fluctuation trends on a weekly and monthly basis, providing quantitative data support for long-term intervention and follow-up assessment by mental health professionals.

[0126] 2. Suicidal Behavior Pattern Knowledge Base: Stores all suicidal emotions, scenarios, and behavioral judgment rules of the system, and supports vector similarity retrieval and matching; real suicide cases and false alarm samples after manual review are automatically incrementally added to the database to continuously enrich the recognition coverage scenarios.

[0127] 3. Event Evidence Database: Completely preserves the original video of each warning, key micro-expression screenshots, and full-process handling operation records, supporting post-event review, doctor-patient communication, and judicial evidence collection.

[0128] Data tiered storage mechanism:

[0129] Edge nodes: Original human faces and real-time micro-expression sensitive materials are stored locally and not transmitted to the external network or private cloud;

[0130] Internal private cloud: Only receives anonymized emotion scores, feature vectors, and cropped low-resolution thumbnails; does not store complete original faces.

[0131] Persistent storage: All psychological profiles, early warnings, and evidence collection data are stored on the unit's internal physical server, with multi-level access permissions set up, allowing only medical and management personnel to access them, fully complying with data security regulations in the medical and regulatory industries.

[0132] like Figure 2 As shown, the system data flow is as follows:

[0133] The front-end camera captures raw video streams, which are then aggregated by the edge gateway and sent to the edge preprocessing layer for local parsing of sensitive information such as faces, bodies, and micro-expressions. The dual-engine model layer adopts a parallel collaborative mode of real-time monitoring by the edge lightweight emotion engine and in-depth analysis on demand by the private cloud VLM semantic engine. It configures exclusive recognition weights and triggering rules for high-risk suicide scenarios such as climbing near edges and lingering at suicide points. The analysis feature data is uploaded to the intelligent decision-making layer, which outputs a standardized comprehensive risk score through a suicide-specific weighted algorithm and classifies it into four risk levels. The graded early warning instructions are issued to the linkage intervention layer to complete the graded handling of voice dissuasion, security dispatch, and emergency linkage.

[0134] The following will be combined with the appendix Figure 3 The flowchart shown describes in detail the application principle of the "Suicide Risk Behavior Early Warning System Based on Micro-expression Visual Capture" in this application.

[0135] A method for early warning of suicidal risk behaviors based on micro-expression visual capture includes the following steps:

[0136] Step S1: Video Acquisition and Adaptive Preprocessing in Key Areas

[0137] For high-risk areas, the edge gateway pulls video streams from all monitoring points, marks them according to the risk level of the area, completes video acquisition and standardized preprocessing, and ensures the detection accuracy and temporal resolution of high-risk areas with limited computing power through a dynamic frame extraction mechanism based on regional priority, providing high-quality input for subsequent emotion recognition and behavior analysis.

[0138] Specifically, the front-end camera captures the raw video stream, and a base frame rate is set based on the risk level of the monitored area. The frame rate is dynamically adjusted based on the real-time scene anomaly level, with automatic increases in frame rates for high-risk areas and abnormal scenes. This system employs a region-priority dynamic frame extraction mechanism, introducing a region risk coefficient, with high-risk areas defaulting to a higher base frame rate. The actual frame extraction rate is: In the formula: : No. The actual frame rate at any given moment, in frames per second (fps). The global base frame rate is 4fps by default. Because the frame rate is 4fps in normal scenes in ordinary areas, the base frame rate in the first-level high-risk area is increased to 6fps. When there are abnormalities such as people near the edge or drastic emotional fluctuations, the abnormality level increases and the frame rate is further increased, up to a maximum of 9.6fps. : Regional risk coefficient, for example, the value of Level 1 high-risk zone is 1.5 (the frame rate coefficient of 1.5 times in Level 1 high-risk zone can improve the motion capture time resolution by 50%, effectively avoiding the loss of frames when quickly flipping through motions), the value of Level 2 attention zone is 1.2, and the value of Level 3 normal zone is 1.0; Anomaly adjustment coefficient, default value is 0.6 (the adjustment coefficient of 0.6 takes into account both computing power consumption and information density of abnormal scenes. Actual tests show that it can improve the high-risk motion capture rate by 22% with an increase in computing power of no more than 30%). : The scene anomaly degree of the current frame, with a value range of [0,1], is calculated by weighting the movement speed of the person, the distance to the edge, and the amplitude of emotional fluctuation;

[0139] Suicide prevention requires higher temporal resolution, especially for actions such as loitering and climbing over edges, which require more intensive frame capture.

[0140] Simultaneously, preprocessing such as face alignment, illumination correction, and super-resolution enhancement are performed to improve the accuracy of micro-expression recognition under complex poses and lighting conditions. For example, face alignment can use affine transformation based on multiple (e.g., 60-100) facial key points to align faces in different poses to a standard frontal face template.

[0141] In the formula:

[0142] Given a 2×3 affine transformation matrix, the solution is obtained through keypoint least-squares fitting. Original face coordinates The standard coordinates are for alignment. Among them, in suicide risk scenarios, people are often in a downward or sideways gaze position, with less than 40% of them facing the face directly. Face alignment can normalize the sideways and downward gaze postures, and actual tests show that it can improve the accuracy of micro-expression recognition.

[0143] Step S2: Detection and tracking of key personnel and initialization of individual psychological profiles

[0144] The role of personnel detection and tracking and individual psychological profiling: By accurately identifying all personnel within the monitoring range, assigning a stable and unique identity to each target, achieving continuous tracking across frames and cameras, and initializing individual psychological profiles, an identity base is provided for subsequent temporal emotion tracking and psychological trend analysis.

[0145] For technical solutions, existing visual tracking mechanisms used by public security authorities can be referenced.

[0146] This system employs a high-precision human detection model to identify all personnel targets; and uses the ByteTrack algorithm to combine appearance features with motion trajectories to achieve multi-target tracking.

[0147] Considering the characteristics of slow movement and prolonged lingering in suicide scenarios, the weight of physical appearance features is increased while the weight of movement trajectories is decreased. The matching degree is calculated as follows: In the formula: : Motion trajectory similarity, calculated based on Kalman filter predicted bounding box IOU; ReID appearance feature cosine similarity; weighting coefficient , Physical appearance plays a dominant role. High-risk individuals for suicide often exhibit prolonged periods of stillness and slow, wandering behavior with minimal changes in their movement trajectories. Relying solely on motion models can easily lead to ID loss. Therefore, when determining the weighting coefficient, increasing the weight of physical appearance features to 0.8 can significantly improve tracking stability in slow-motion and static scenarios, reduce ID switching rate by 60%, and ensure the temporal continuity of psychological profiles.

[0148] Extract ReID features for each target and compare them with the historical feature database to confirm identity: Extract global ReID appearance features and achieve cross-camera identity association through cosine similarity matching.

[0149] ReID feature cosine similarity:

[0150]

[0151] In the formula , The two targets are represented by 256-dimensional ReID feature vectors. Depending on the scenario, the matching threshold can be set to, for example, 0.75, higher than the 0.7 used in general security scenarios. Increasing the matching threshold can reduce false matches between different individuals and avoid confusion in psychological profiles. In suicide scenarios, where clothing changes are minimal, a high threshold will not significantly increase missed matches while ensuring identity accuracy.

[0152] Initialize psychological profiles for targets appearing for the first time, record basic characteristics and entry time, and support long-term, cross-regional continuous psychological tracking.

[0153] Step S3: Dual-model parallel inference: sentiment temporal monitoring + high-risk behavior semantic deep judgment

[0154] This step is the core perception stage of suicide risk warning. Two models each have their specific functions: the quantitative emotion engine (Model 2) continuously monitors psychological state, capturing the suicide-specific signal of accumulated negative emotions and their abnormal decline; the semantic reasoning engine (Model 1) targets high-risk areas and abnormal targets, combining behavior, facial expressions, and context to determine suicidal intent. Specifically:

[0155] Model 1: Semantic Reasoning Chain – Joint Determination of High-Risk Behavior and Suicidal Intent

[0156] Model 1 performs multi-dimensional semantic analysis on the target that triggers deep judgment, and combines body movements, facial expressions, spatial location, and scene attributes to jointly infer suicidal intent, outputting a structured risk level and judgment basis, thus solving the problem that pure emotion recognition cannot determine behavioral intent.

[0157] Model 1 can be a locally deployed multimodal large model, taking a full-body image of the target, a 30-second time-series clip, and a scene region mask as input. It analyzes limb posture, movement trends, edge distance, and facial state using a suicide-specific prompt-guided model (trainable via LLM). The output description is vectorized and matched against a suicide behavior pattern knowledge base. The final output includes four structured results: risk type, semantic confidence, intent level, and intervention suggestions. For the suicide behavior knowledge base vector matching, a BERT model fine-tuned in the psychological domain is used to vectorize the text, calculating the cosine similarity between the VLM output description and the knowledge base entries. In the formula The behavior description vector output by VLM. For the knowledge base The system generates vectors representing suicide risk patterns. The matching threshold is set to 0.75, lower than the common threshold of 0.8, employing a more lenient matching strategy to reduce false negatives. Suicidal behavior is diverse, with many long-tail scenarios. Lowering the matching threshold to 0.75 improves recall and covers more atypical high-risk behaviors. Although this slightly increases false positives, it aligns with the principle of "better to report false positives than to miss false positives" in suicide prevention, and false positives can be filtered out through manual review.

[0158] After retrieving the knowledge base of suicide behavior patterns, a semantic risk score is calculated by combining the scenario risk coefficient, and the structured results are output to the decision-making layer. Definition of semantic suicide risk score calculation: In the formula: Knowledge base matching similarity; : Behavioral confidence level output by VLM, with values ​​[0,1]; Scenario risk coefficient: 1.2 for Level 1 high-risk areas, 1.0 for Level 2 areas, and 0.8 for ordinary areas; final semantic risk score. The value range is 0 to 120, and values ​​exceeding 100 are counted as 100.

[0159] By introducing a scenario risk coefficient, the risk weight of the same behavior in high-risk areas such as border areas is higher, which is consistent with the actual risk distribution; the 1.2 times weighting of the first-level high-risk area can improve the early warning sensitivity of high-risk areas and further reduce the possibility of missed reports.

[0160] By combining behavioral confidence and scenario risk coefficients, the AUC of suicidal intent discrimination can reach over 0.9; it can accurately distinguish normal edge work, taking photos and suicidal wandering intentions, providing a reliable semantic basis for tiered intervention.

[0161] 2. Model 2: Quantifying the Emotion Chain – Continuous Monitoring and Abnormal Regression of Negative Emotions

[0162] Model 2 is used to extract facial micro-expression features frame by frame, quantify the intensity of seven types of emotions, and continuously track the accumulation and fluctuation patterns of negative emotions; it specifically identifies the abnormal decline feature of "abnormal calm after intense pain", providing core psychological dimension basis for suicide risk warning.

[0163] First, align facial key points and extract the activation intensity of multiple facial action units (AUs, typically 17; verification using publicly available facial geometry datasets shows that 17-dimensional features are sufficient for high-discrimination matching, balancing recognition accuracy, detection speed, and storage cost, and adapting to mobile devices and lightweight facial verification scenarios for access control). Based on the FACS system and suicide psychology research, construct a dedicated emotion mapping matrix to calculate the real-time intensity of seven emotions: joy, anger, worry, sadness, surprise, fear, and calmness. Calculate the cumulative degree of negative emotions, the amplitude of emotional fluctuations, and the abnormal decline index using a sliding window; if a threshold is reached, trigger deep semantic engine analysis. Specifically, this includes the following:

[0164] (1) Suicide-specific AU-emotion mapping matrix

[0165] To enhance the identification weights of emotions such as sadness, despair, anxiety, and numbness in response to suicide risk, a mapping matrix is ​​constructed. (The weights for each dimension are based on the identification mechanism described above.)

[0166] Define the formula for calculating emotion intensity: In the formula: : 17-dimensional AU cell activation intensity vector, with each element taking the value [0,1]; A suicide-specific emotion mapping weight matrix, in which the sadness dimension sets high weights for AU1 (inner eyebrow raised), AU4 (frowning), and AU15 (corner of the mouth turned down); and the calm dimension sets corresponding weights for AU23 (tight lips) and AU43 (closed eyes). Element-wise sigmoid normalization function;

[0167] : Corresponding to a 7-dimensional emotion intensity vector, with a focus on (sad), (anxiety), (fear), (Calm) Four categories, T represents transpose.

[0168] The mapping matrix is ​​based on clinical psychology research and facial expression datasets of suicide victims, focusing on enhancing the features of sadness, despair, and numbness that are highly correlated with suicidal ideation. Compared with general emotion recognition models, it improves the sensitivity of identifying suicide-related negative emotions by 25% and can capture more subtle expressions of psychological distress.

[0169] (2) Accumulation of negative emotions

[0170] Suicidal ideation is usually the result of a sustained accumulation of negative emotions. A momentary surge in emotion does not necessarily indicate high risk; rather, a sustained buildup is the high-risk indicator. The degree of accumulated psychological stress can be reflected by calculating the time integral of negative emotions over a past period.

[0171] In the formula: The cumulative time window is set to a default value of 300 seconds (5 minutes). The 5-minute time window is consistent with the continuous characteristics of psychological distress, which can filter out misjudgments caused by momentary emotional fluctuations, while retaining sensitivity to mid-term emotional deterioration. The intensity of sadness, anxiety, and fear were respectively represented; the intensity of calmness was subtracted to exclude normal calm states, and only the effective accumulation of negative emotions was counted.

[0172] (3) Emotional Abnormality Decline Index

[0173] Clinical suicide intervention studies show that most suicide perpetrators experience a sudden "calm" before their actions, essentially a state of psychological numbness after making a decision. By capturing the characteristic of "negative emotions reaching their peak and then rapidly declining to an abnormally calm state," this is a typical precursory signal that suicidal ideation has entered the execution stage.

[0174] In the formula:

[0175] The peak intensity of negative emotions over the past 5 minutes;

[0176] The intensity of negative emotions at the current moment;

[0177] : The magnitude of the emotional decline, with a value range of [0,1]. The larger the value, the more severe the decline.

[0178] Abnormal pullback judgment criteria:

[0179] Peak negative emotions A peak threshold of 0.6 corresponds to significant feelings of distress.

[0180] pullback range A drop of 0.5 indicates a sharp decline in sentiment;

[0181] Current calm intensity ;

[0182] The pullback duration is >10 seconds, and a 10-second duration can filter out instantaneous fluctuations.

[0183] (4) Initial screening triggering conditions

[0184] Meeting any one of these conditions will trigger a deep semantic analysis:

[0185] Accumulation of negative emotions And lasting >30 seconds;

[0186] The judgment of abnormal emotional decline has been established;

[0187] The target has entered a Level 1 high-risk area.

[0188] If the initial screening criteria are met, a deep-seated judgment request is sent to the scheduling module, which then sends the target's time-series data for the past minute into the semantic engine.

[0189] By recognizing negative emotions and abnormal decline characteristics, abnormal psychological signals before suicide can be detected more than 30 seconds in advance, allowing sufficient time for subsequent intervention.

[0190] Step S4: Multi-dimensional weighted fusion and comprehensive suicide risk score calculation

[0191] This system adopts a fusion strategy of "emotion as the primary factor, behavior as the secondary factor, and strong historical trend correction": the emotion dimension has the highest weight because suicide is essentially a psychologically driven behavior; behavioral semantics serves as the risk confirmation dimension; and historical psychological trends serve as a situational correction factor, as individuals with long-term negative emotional accumulation have a significantly higher risk.

[0192] The comprehensive suicide risk score is calculated as follows: In the formula: Semantic behavior risk score, 0~100, is the result of suicidal intent determination from the VLM semantic engine; Emotional risk score, 0-100 points, calculated using the following formula: The highest value of the accumulation and abnormal decline of negative emotions is taken. The duration score, ranging from 0 to 100, reflects the duration of risk characteristics. Historical trend correction score, 0-100, calculated based on the degree of mood deterioration in the past 7 days from the individual's psychological profile; weighting coefficient: , , (The core driver of suicide risk is psychological state, therefore the emotion dimension has the highest weight (0.4); behavioral semantics serves as objective evidence, with a weight of 0.3; historical trend correction has a weight of 0.3, and the sub-weights are tilted towards history, reflecting the cumulative and gradual nature of suicide risk. This weighting can be determined using a clinical suicide case dataset, and compared to an equal weighting scheme, the underreporting rate is reduced by about 40%). , Similarly, the calibration settings are configured.

[0193] The three dimensions of semantic behavioral risk, emotional quantification risk, and historical context correction are integrated with specific weights for suicide scenarios to output a unified comprehensive suicide risk score of 0-100. The weights of emotional persistence and historical psychological trends are emphasized to solve the problems of one-sided judgment and high underreporting rate caused by single-dimensional judgment.

[0194] The historical trend correction score is calculated based on the rate of change of the average negative emotion over the past 7 days, reflecting the trend of deteriorating mental state:

[0195] In the formula:

[0196] The average intensity of negative emotions in this individual over the past 7 days;

[0197] The individual's baseline level of negative emotions;

[0198] The clip function limits the result to the range of 0 to 100, with 50 as the baseline score. If the situation worsens, the score increases, and if it improves, the score decreases.

[0199] Suicide risk is highly correlated with long-term psychological state, and the risk of short-term mood fluctuations is far lower than that of long-term continuous deterioration; introducing a 7-day historical trend correction can significantly improve the ability to identify chronic suicidal ideation, a capability that general single-frame security systems do not possess at all.

[0200] The duration calculation uses an exponential growth model:

[0201] In the formula Duration of risk characteristic, in seconds; time constant. The value is 60 seconds, and τ=60 seconds corresponds to the typical duration of suicidal wandering behavior. The longer the duration, the higher the certainty of risk, which is consistent with actual experience in on-site handling. Generally, a score of about 63 points is given for a duration of 60 seconds, and about 86 points for a duration of 120 seconds.

[0202] Compared to single-dimensional emotion assessment, the fusion method improves the AUC of suicide risk identification, reduces the false negative rate, and significantly enhances the reliability of early warning.

[0203] Step S5: Suicide Risk Scenario Triage and Specific Rule Matching

[0204] For five typical high-risk behaviors related to suicide, a multi-dimensional judgment rule is established to further confirm the risk type and urgency level, providing a precise basis for tiered intervention.

[0205] Examples of the rules for determining the five categories of high-risk behaviors are as follows:

[0206] (1) Climbing / Climbing over guardrails judgment

[0207] Judgment conditions:

[0208] The semantic model hits the category "climbing, vaulting over, straddling the railing" with a confidence level > 0.6;

[0209] The target is located in a Level 1 high-risk area (near the edge / guardrail);

[0210] When your center of gravity passes over the guardrail or both feet leave the ground;

[0211] The emotion is accompanied by an unusual calm or extreme sadness.

[0212] Climbing and scaling heights are direct actions used in suicide, and are of the highest priority. As long as the semantics match and the location is appropriate, the risk level is immediately increased, regardless of the overall scoring threshold.

[0213] (2) Determination of prolonged loitering near the edge

[0214] Judgment conditions:

[0215] The target moves back and forth within high-risk areas of level two or above;

[0216] The total time spent within 5 meters of the edge is greater than 60 seconds, or the number of round trips is greater than or equal to 3 times;

[0217] Accumulated negative emotions > 0.4;

[0218] Exclude scenarios that are pre-selected as whitelisted, such as normal patrols and operations.

[0219] A 60-second pause is the critical time to distinguish between normal passing by and hesitant lingering. Combined with negative emotions, it can effectively exclude normal people such as staff and tourists.

[0220] (3) Determination of prolonged static gaze

[0221] Judgment conditions:

[0222] The target remains standing / sitting with a displacement of less than 0.5 meters;

[0223] The face is turned toward or below the edge, and the gaze is sustained.

[0224] The person remains stationary for more than 90 seconds and is in a danger zone.

[0225] Accompanied by negative emotions such as sadness and numbness.

[0226] A 90-second static stare is a typical behavior during the pre-suicidal hesitation stage. Combined with orientation and emotional characteristics, it can effectively distinguish normal behaviors such as resting and waiting.

[0227] (4) Judgment of self-injury actions

[0228] Judgment conditions:

[0229] The semantic model identifies actions such as "self-harm, strangulation, hitting the head, and pointing a weapon at oneself";

[0230] Action confidence level > 0.7;

[0231] Accompanied by extreme emotions or abnormal calmness. Self-harming actions are a clear high-risk signal, and can be directly identified as high-risk with a high degree of confidence, without the need for long-term accumulation.

[0232] (5) Judgment of extreme / abnormal emotions

[0233] Judgment conditions:

[0234] The intensity of sadness / anxiety / fear is >0.6 and lasts for >30 seconds;

[0235] Or, abnormal mood swings may occur;

[0236] There is no obvious external trigger (no communication with others, no sudden events). Sustained extreme emotions or abnormal calmness without any trigger are important psychological signals of suicidal ideation. Even if no high-risk actions are observed, such individuals should be given close attention and their condition should be manually verified.

[0237] For targets whose comprehensive scores exceed the minimum threshold, five categories of high-risk suicide behaviors are matched based on behavioral characteristics and scene location. Each category of rules integrates semantic judgment, emotional characteristics, spatial characteristics, and duration characteristics. If all of them are met, the corresponding risk type is locked and the system enters the graded early warning stage.

[0238] For example, when a target with a comprehensive score exceeding 35 enters the rule matching module, it is assigned to the corresponding rule base based on its location and behavioral characteristics. Each rule condition is verified, and if a match is found, the risk type and priority are determined. Targets that do not match any rules but have high scores are added to a key observation list for continuous tracking. Therefore, by identifying five types of high-risk behaviors on an average basis, the false exclusion rate of normal individuals is reduced, effectively filtering out a large number of risk-free targets and alleviating the burden of manual review.

[0239] Step S6: Multi-level conservative tiered early warning and multi-entity joint intervention

[0240] The conservative tiered early warning system follows the conservative principle of "better to report falsely than to miss a report." It sets up four early warning levels and matches them with differentiated intervention procedures and response entities to achieve a precise match between risk level and intervention intensity, reducing missed reports while avoiding excessive disruption to normal order.

[0241] The comprehensive risk score of 0 to 100 is divided into four levels: attention, warning, high risk, and emergency. The overall threshold is lower than that of public safety scenarios. Each level corresponds to a standardized intervention process, which links psychological intervention personnel, on-site duty, security, and emergency rescue forces at multiple levels, and records the entire process log simultaneously.

[0242] The four-level early warning thresholds and intervention strategies are shown in Table 1 below:

[0243] Attention level 35~49 points Suspicious risk, included in monitoring 1. Establish / update individual psychological profiles and mark them as key targets for attention; 2. Increase the tracking priority and frame rate of this target; 3. On-duty personnel should observe from the background and not intervene on-site for the time being; 4. Generate and archive the attention record. Warning level 50~64 points There are risks involved, and intervention is necessary. 1. Send early warning notifications to psychological interventionists / on-duty personnel, including location, emotional trends, and behavioral characteristics; 2. Arrange for personnel to check nearby and gather information indirectly; 3. Use a PTZ camera with automatic zoom to track and record the entire process; 4. Psychologists conduct remote observation and assessment. High risk level 65~79 points High risk, immediate intervention required. 1. Immediately dispatch nearby on-duty personnel and security forces to the scene; 2. Use remote voice reassurance and persuasion to delay the implementation of the action; 3. Notify psychological experts to intervene and assess the situation; 4. Prepare emergency rescue forces to stand by. Emergency level 80~100 points Immediate danger, life in danger 1. One-click activation of emergency response, with on-site personnel immediately stepping forward to stop the incident; 2. Coordination with external emergency forces such as public security, emergency medical services, and fire departments; 3. Full-process video recording to preserve a complete chain of evidence; 4. Simultaneous reporting to management departments.

[0244] Table 1

[0245] Grading Explanation:

[0246] The minimum warning threshold is set at 35 points, significantly lower than the 40 points for public scenarios, reflecting a conservative strategy. The four-level grading system fully aligns with the "attention-intervention-treatment-first aid" four-level response system for psychological crisis intervention. Clearly high-risk actions such as climbing and self-harm can directly jump to the high-risk / emergency level without being restricted by the scoring. This threshold system, while ensuring low false negatives, keeps false positives within an acceptable range, meeting the prevention and control needs of high-risk scenarios.

[0247] Based on the closed-loop process described above, the system performs full-cycle psychological profile updates and closed-loop iterative optimization: continuously updating individual psychological profiles and tracking emotional evolution trends; feeding back the knowledge base with the results of manual review, adaptively adjusting the warning threshold, forming a complete closed loop of "monitoring-judgment-intervention-optimization", so that the system's ability to identify suicide risks can be continuously improved iteratively.

[0248] Each warning event and daily emotional data is archived into the corresponding individual's psychological profile; emotional trends and risk evolution patterns of individuals and groups are analyzed regularly; the results of manual verification of true reports, false reports, and missed reports are used to update the suicide behavior knowledge base, and a small-step conservative strategy is adopted to adaptively adjust the warning threshold to avoid missed reports due to threshold fluctuations.

[0249] Among them, the rate of mood deterioration is used for individual psychological trend analysis:

[0250] In the formula: This represents the average negative sentiment this week. This is the average of last week. For time intervals; This indicates a continued deterioration in mood, with higher values ​​indicating faster deterioration. This indicator can identify key individuals whose mental state is gradually worsening, achieving true "prevention" rather than simply alerting to immediate behavior.

[0251] The conservative threshold adaptive adjustment adopts an incremental strategy with a very small learning rate, prioritizing ensuring no missed cases: In the formula:

[0252] , To adjust the warning thresholds before and after;

[0253] To count the number of false alarms within the statistical period, This represents the number of missed reports;

[0254] This represents the total number of warning events;

[0255] The learning rate is set to 0.02, which is much smaller than the 0.05 used in public scenarios.

[0256] The penalty weight for false negatives is twice that for false positives. When a false negative occurs, the threshold drops significantly; when a false positive occurs, the threshold rises slightly. The overall trend is conservative, ensuring that the false negative rate remains at an extremely low level. In suicide scenarios, the cost of false negatives is far higher than that of false positives. Therefore, a false negative penalty coefficient of 2 and a very small learning rate are set. This ensures that the system can be gradually optimized to reduce false positives, but will never lead to an increase in false negatives due to over-optimization, which fully complies with the safety-first principle of high-risk scenarios.

[0257] During implementation, newly identified high-risk behavioral patterns are incrementally updated in the knowledge base, and monthly psychological trend analysis reports and system performance iteration reports are generated. Psychological profiles support clinical psychologists in conducting long-term follow-up interventions on key individuals, significantly improving the systematic nature of prevention and control.

[0258] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for early warning of suicidal risk behaviors based on micro-expression visual capture, characterized in that, Includes the following steps: Key Area Video Acquisition and Adaptive Preprocessing: Acquire monitoring video streams for high-risk areas under graded management, dynamically perform intelligent frame extraction based on the area risk level and real-time scene anomaly, complete face alignment and image enhancement, and output standardized image sequences; Key personnel detection and tracking and individual psychological profile initialization: Detect all human targets in the image, assign a unique tracking ID to each target and extract ReID appearance features, and initialize an individual psychological profile for the first target to appear; Dual-model parallel inference: The local lightweight quantization emotion engine runs frame by frame to extract facial action units and quantify and output the intensity of multiple emotions, continuously monitoring the accumulation of negative emotions and the abnormal decline of emotions. When abnormal emotional signals or targets enter high-risk areas, the VLM semantic reasoning engine is triggered to combine body posture, facial expressions, and scene context to complete a deep determination of suicidal intent. Multi-dimensional weighted fusion and comprehensive risk score calculation: Following a weighting strategy that prioritizes emotion, supplements behavior, and strongly modifies historical trends, semantic risk score, emotional risk score, and time-series persistence and historical trend correction score are weighted and fused to output a comprehensive suicide risk score. Suicide risk scenario triage and specific rule matching: Risk targets are triaged to a dedicated suicide risk assessment link, and five high-risk behavior assessment rules are matched for climbing and scaling, loitering at the edge, staring at stillness, self-harming actions, and extreme abnormal emotions to identify specific risk types. Multi-level conservative early warning and multi-entity joint intervention: Following the conservative principle of "better to report false than to miss a report", a multi-level early warning system is set up, and corresponding psychological intervention, on-site duty, security dispatch and emergency response measures are matched. Full-cycle psychological profile updates and closed-loop iterative optimization: Continuously update individual time-series psychological profiles and track emotional evolution trends; Based on the results of manual review, the warning threshold is adaptively adjusted using a conservative strategy, and the knowledge base of suicidal behavior patterns is updated incrementally.

2. The method according to claim 1, characterized in that, The intelligent frame extraction combines regional risk level and scene anomaly degree for dual adjustment: different levels of control areas are set with different base sampling frame rates, and high-risk areas default to a higher frame rate; when abnormal situations such as people approaching the edge or drastic emotional fluctuations are detected, the frame extraction rate is further increased as the anomaly degree increases, and the motion capture time resolution of high-risk areas is prioritized under limited computing power. Face alignment normalizes non-frontal facial poses such as looking down and turning to the side to a standard template by using multiple facial key points, and focuses on amplifying key micro-expression areas such as eyebrows, eyes, and around the mouth to improve the accuracy of negative micro-expression recognition under complex poses.

3. The method according to claim 1, characterized in that, The multi-target tracking employs a matching strategy that combines motion trajectory similarity and appearance feature similarity, with appearance feature weight significantly higher than motion trajectory weight. This is to adapt to scenarios where high-risk suicide individuals remain still or wander slowly for extended periods, resulting in insufficient motion information, thereby improving tracking stability in slow-motion and static states. ReID features are used for cross-frame and cross-camera identity matching, employing a matching threshold higher than that used in general security scenarios to avoid mismatches between different individuals and ensure the consistency of identity in psychological profiles.

4. The method according to claim 1, characterized in that, The dual-model parallel inference configuration is configured with the following scheduling triggering rules: when a target enters a level 1 high-risk border area, the initial emotion screening is skipped and the semantic inference engine is directly activated for in-depth analysis to prevent missed detections in high-risk scenarios; when negative emotions accumulate to a threshold or abnormal emotional decline is detected, the semantic engine is automatically triggered for in-depth analysis; when the system determines that a person is at level 2 or above risk, the allocation of frame extraction and recognition computing power for the corresponding camera is automatically increased to dynamically tilt computing power towards high-risk targets.

5. The method according to claim 1, characterized in that, The local lightweight quantization emotion engine constructs a suicide-specific emotion mapping relationship based on the facial motion coding system, and strengthens the recognition weight of facial features that are highly correlated with suicidal ideation, such as sadness, despair, anxiety, and numbness. The accumulation of negative emotions is calculated by sliding windows, and the abnormal decline characteristic of negative emotions rapidly falling back to an unusual calm after reaching a peak is specifically identified. This characteristic serves as a core pre-implementation signal for suicidal ideation. When the accumulation of negative emotions continues to reach the target, or the abnormal decline of emotions is determined, or the target enters the first-level high-risk area, the semantic reasoning engine is triggered for in-depth analysis.

6. The method according to claim 1, characterized in that, The VLM semantic reasoning engine receives a full-body image of the target, short time sequence fragments, and scene region masks as input. It combines body posture, action trend, edge distance, and facial state to perform multimodal joint reasoning. After vectorizing the output behavioral semantic description, it performs similarity retrieval and matching with a pre-set suicide behavior pattern knowledge base. After matching the corresponding risk pattern, it calculates the semantic risk score by combining the behavioral confidence and scene risk coefficient. The same behaviors in high-risk areas are assigned higher risk weights, and the knowledge base matching uses a relatively lenient similarity threshold to cover more atypical high-risk behaviors and reduce the false negative rate.

7. The method according to claim 1, characterized in that, The comprehensive suicide risk score is obtained by weighting and fusing three dimensions—semantic behavioral risk, emotional quantitative risk, and temporal history correction—according to the specific weights for suicide scenarios. The comprehensive suicide risk score is defined as follows: ; in For semantic risk scoring, Emotional risk is categorized. For time series continuation, The score is adjusted based on the historical trend of the 7-day psychological profile. , , These are the weights of the corresponding items; The emotional dimension has the highest weight, while the behavioral semantic dimension serves as the basis for risk confirmation. In the temporal and historical correction dimensions, the weight of historical psychological trends is higher than that of immediate duration, highlighting the psychological driving nature and long-term cumulative characteristics of suicide risk. The historical trend correction score is calculated based on the degree of deterioration of negative emotions in the individual's psychological profile in the near term, reflecting the long-term evolution trend of psychological state.

8. The method according to claim 1, characterized in that, The five categories of high-risk behaviors are all subject to multi-dimensional integrated judgment rules. Each rule combines semantic judgment results, emotional characteristics, spatial location conditions, and duration conditions. When all conditions are met, the corresponding risk type is locked. Among them, climbing and traversing behaviors are given higher risk priority as long as the semantics match and the location is correct, without being limited by the comprehensive scoring threshold. The behavior of loitering near the edge must meet the conditions of the duration of stay near the edge or the number of times of back and forth and be accompanied by negative emotions. The behavior of staring at stillness requires a comprehensive judgment based on posture, orientation, and emotional characteristics. Self-harming behaviors are directly judged as high-risk under high confidence. Extreme abnormal emotions are included in the key attention even if no high-risk behaviors are shown.

9. The method according to claim 1, characterized in that, The starting threshold of the multi-level conservative early warning system is lower than that of general public safety scenarios, reflecting the prevention and control principle of "better to report false alarms than to miss alarms." The four levels of early warning correspond to the four-level response system of background observation, personnel intervention, on-site handling, and emergency linkage. Clearly high-risk actions such as climbing and self-harm can skip levels directly. The adaptive adjustment of the early warning threshold adopts a conservative small-step incremental strategy, and the penalty weight for missed alarms is significantly higher than that for false alarms, ensuring that the missed alarm rate is always maintained at an extremely low level during the iteration process. The individual psychological profile uses ReID features as a unique index to store temporal emotional trajectories, records of previous high-risk early warnings, and lingering behavior trajectories, supporting periodic psychological trend analysis and providing quantitative data support for long-term psychological intervention.

10. A suicide risk behavior early warning system based on micro-expression visual capture, characterized in that, include: The video acquisition layer consists of high-risk area surveillance cameras deployed in a tiered manner and edge gateways. It is used to acquire video streams in key areas and mark the risk level of the areas. It supports local caching when the network is disconnected and resume interrupted downloads. The edge preprocessing layer includes an adaptive frame extraction module, a face alignment enhancement module, a human detection module, a multi-target tracking and ReID module, and an individual cropping module. It is used for video preprocessing, target detection and tracking, and individual psychological profile initialization. Sensitive biometric data is processed locally at the edge. The dual-engine model layer includes a local lightweight quantized emotion engine and a VLM semantic reasoning engine. The former runs in full to achieve continuous monitoring of negative emotions and specific capture of abnormal emotional decline, while the latter conducts joint in-depth analysis of the suicidal intent of high-risk targets based on their behavior, expressions, and scenarios. The intelligent decision-making layer includes a multi-dimensional weight fusion module, a suicide risk rule matching module, and a conservative grading judgment module, which are used to calculate a comprehensive suicide risk score, match high-risk behavior types, and complete a four-level early warning classification. The linkage intervention layer includes an early warning push module, a remote reassurance module, an on-site handling module, and an emergency linkage module, which are used to implement tiered psychological intervention and multi-stakeholder collaborative handling. The closed-loop psychological data layer includes an individual psychological profile database, a knowledge base of suicidal behavior patterns, and an event evidence database. It is used for time-series psychological tracking, incremental updates of the knowledge base, and retention of event evidence. A layered storage architecture with edge and intranet layers is adopted to ensure data security.

Images