Computer vision-based psychological health assessment system for higher vocational college students

Abstract

The application belongs to the technical field of computers, and specifically discloses a higher vocational student psychological health evaluation system based on computer vision. The system comprises a scene adaptation collection module, a multi-granularity behavior analysis engine, a psychological state quantitative modeling module and a dynamic early warning and feedback module. By collecting anonymized video streams in multiple natural scenes, and hierarchically analyzing micro-physiological, medium-action and macro-social behavior characteristics, the psychological state is quantitatively evaluated and graded early warning and feedback are realized, so as to improve the objectivity, initiative and timeliness of psychological health evaluation.

Description

Technical Field

[0001] This invention belongs to the field of computer technology, specifically relating to a computer vision-based mental health assessment system for vocational college students. Background Technology

[0002] Mental health assessment is an important interdisciplinary field integrating psychology and information technology, aiming to quantify and diagnose an individual's psychological state using scientific methods. With the development of artificial intelligence technology, intelligent assessment systems based on multimodal data have become a research hotspot in this field. Computer vision technology, due to its non-contact and non-invasive characteristics, shows potential in behavior and emotion recognition.

[0003] Mental health assessments targeting specific student groups are a key area of ​​research in educational psychology and campus safety management. Vocational college students, as a specific group of young people, are affected by multiple factors such as academic pressure and career planning. Traditional scale assessments or interviews have inherent limitations, including high subjectivity, limited coverage, and difficulty in routine implementation. Utilizing computer vision technology to automatically analyze student behavior and provide objective evidence for early warning and intervention has become a fundamental goal of this technological approach.

[0004] Existing technologies typically rely on collecting visual information such as facial expressions and postures from students in controlled, static environments for assessment. This approach lacks ecological validity; the collected behavioral data is obtained in a "performative" state where students are aware they are being observed, making it difficult to trigger their natural stress responses and emotional expressions in real learning and life scenarios. This leads to a discrepancy between the assessment results and the students' true psychological state. Existing systems often focus on single-modal visual feature analysis, failing to fully consider the diversity of activity scenarios and the complexity of behavioral patterns among vocational college students, and lacking sufficient sensitivity to identify subtle, persistent abnormal behavioral patterns. Existing assessment systems suffer from issues with the accuracy and timeliness of early warnings, failing to meet the needs for routine and precise screening and monitoring of the mental health of vocational college students. Summary of the Invention

[0005] The purpose of this invention is to provide a computer vision-based mental health assessment system for vocational college students, in order to solve the technical problems of low ecological validity, insufficient identification of natural stress behaviors, and insufficient sensitivity to subtle and persistent abnormal behavior patterns in complex scenarios caused by data collection in controlled static environments in existing technologies.

[0006] To achieve the above objectives, this invention provides a computer vision-based mental health assessment system for vocational college students, comprising: The scene-adaptive acquisition module is used to continuously collect anonymized video stream data from students in multiple preset uncontrolled natural scenes on campus. A multi-granularity behavior analysis engine, connected to the scene adaptation acquisition module, is used to perform layered analysis on the received anonymized feature video stream and extract multi-layered behavior features; the multi-granularity behavior analysis engine includes a micro-physiological behavior analysis sub-module, a meso-action behavior analysis sub-module, and a macro-social behavior analysis sub-module. The psychological state quantification modeling module is connected to the multi-granularity behavior analysis engine and is used to integrate multi-level behavioral features and map them to the potential psychological state dimension through a machine learning model; the psychological state quantification modeling module includes a feature fusion layer and a state inference layer. The dynamic early warning and feedback module, connected to the psychological state quantitative modeling module, is used to execute tiered early warnings and personalized feedback based on the output of the quantitative model and historical trends. The dynamic early warning and feedback module has a built-in three-level early warning rule base. After an early warning is triggered, the dynamic early warning and feedback module automatically generates a structured assessment report and pushes it to the authorized mental health teacher work platform through an encrypted interface. The dynamic early warning and feedback module is also connected to the campus digital terminal, used to push self-care tips and online psychological resource links to students with their informed consent.

[0007] Preferably, the scene adaptation acquisition module is deployed with multiple distributed visual sensing nodes, each of which integrates a wide-angle optical lens and an infrared supplementary light unit. The visual sensing node has a built-in edge computing unit for real-time anonymization of the acquired raw video stream data. The real-time anonymization process includes face region detection and blurring to generate an anonymized feature video stream that retains only the human skeleton joint coordinate sequence and global behavioral contour. All anonymized video streams are transmitted to the central data server via the campus-dedicated network.

[0008] Preferably, the microscopic physiological behavior analysis submodule is used to analyze non-contact physiological indicators from the human head region contour sequence in the anonymized feature video stream; the analysis process includes: calculating the pixel brightness change sequence of a specified area of ​​the face region using optical flow; performing bandpass filtering on the brightness change sequence, with the filter passband set to 0.8 Hz to 3 Hz; performing spectral analysis on the filtered signal, extracting its dominant frequency component as the estimated heart rate value, and calculating the energy proportion of the signal in the respiratory frequency band as a respiratory rhythm stability index; simultaneously, by tracking key points of the eye contour, calculating the blink frequency and average duration of a single blink per unit time; The mesoscopic action behavior analysis submodule is used to identify and quantify predefined action units and behavior patterns from the human skeleton joint coordinate sequence of anonymized feature video streams. The mesoscopic action behavior analysis submodule has a built-in action recognition model based on temporal convolutional networks. For each identified action unit, its timestamp, duration, and frequency of occurrence in a specific scenario are recorded. By analyzing the transition probability between consecutive action units, a Markov model of an individual's behavior sequence within a specific time period is constructed. The macro-social behavior analysis submodule is used to analyze the spatial relationships and interaction patterns among multiple individuals in a public scene video stream. The analysis process includes: determining the position and motion trajectory of all individuals in the video frame through target detection and tracking algorithms; calculating the real-time Euclidean distance between any two individuals, and determining that the two are in a proximity state when the real-time Euclidean distance is continuously less than a preset threshold of 1.5 meters; statistically analyzing the average number of companions in a proximity state of the target individual within a unit observation period as its social contact density index; quantifying its directional behavior by analyzing the change in the angle between the direction of the line connecting the target individual and its nearest companion and its own motion direction; and statistically analyzing the proportion of time the target individual spends at the edge of the group in a group activity scene as a negative indicator of its social participation.

[0009] Preferably, the feature fusion layer is used to receive time-series feature vectors from micro, meso, and macro sub-modules, and perform moving average and standardization processing with minutes as the time window; the feature fusion layer uses an attention mechanism to assign weights to different feature dimensions under different scenarios; The state inference layer comprises multiple parallel deep neural network models, each corresponding to a psychological state dimension, including emotional valence, emotional arousal, social motivation, and behavioral vitality. Each neural network model takes a weighted and fused feature vector as input and outputs a continuous score between 0 and 100. The state inference layer also includes a long short-term memory network to analyze the changing trends of the scores of each psychological state dimension over time scales of several days or even several weeks.

[0010] Preferably, the optical flow calculation in the microscopic physiological behavior analysis submodule specifically adopts the Lucas-Cannard optical flow algorithm. The Lucas-Cannard optical flow algorithm delineates multiple regions of interest with a size of 16 pixels by 16 pixels in the facial region, calculates the pixel displacement vector field of these regions between consecutive video frames, performs principal component analysis on the vector field, extracts the basis vectors representing periodic motion, and projects them onto the color change channel.

[0011] Preferably, the pre-trained temporal convolutional network in the mesoscopic action behavior analysis submodule has a structure comprising 8 one-dimensional convolutional layers and 3 fully connected layers, with convolutional kernel sizes of 7, 5, and 3, respectively; the input of the temporal convolutional network is a sequence of three-dimensional coordinates of 33 skeleton joints extracted from the past 64 frames of video.

[0012] Preferably, the attention mechanism in the psychological state quantification modeling module has a weight generation network that is a two-layer fully connected network. The fully connected network takes the current scene encoding and the original feature vector as input and outputs a weight vector with the same dimension as the original feature vector. After the weight vector is normalized by a soft maximization function, it is multiplied element-wise with the original feature vector.

[0013] Preferably, the three-level early warning rule base in the dynamic early warning and feedback module includes: The first level of alert is triggered when the score of any psychological state dimension is less than 20% of the individual's historical baseline for three consecutive days, or when the score on a single day is less than 30 percentage points of the group norm. The second level of advice is an interview-level warning, triggered when two or more related dimensions simultaneously reach the attention-level threshold, or when a clear social avoidance trend pattern is detected. The third level of emergency intervention warning is triggered by the detection of extreme behavioral decline combined with abnormal physiological indicators, or the appearance of a predefined risky behavior pattern in a scenario. The individual historical baseline value is determined by statistically analyzing the moving average and standard deviation of the student's scores for each psychological state dimension in the same scenario and time period over the past 30 days; the individual historical baseline value is updated every 7 days.

[0014] Preferably, the macro-social behavior analysis submodule is further used to introduce a group interaction modeling mechanism based on graph neural networks; the group interaction modeling mechanism regards all individuals in the video frame as graph nodes, and the proximity relationship and interaction frequency between individuals as edge weights to construct a dynamic social graph; a graph convolutional network is used to embed the dynamic social graph to classify the role of individuals in the group.

[0015] Preferably, the early warning rule base of the dynamic early warning and feedback module is further extended with early warning rules based on changes in group roles; when it is detected that a student has changed from a core leader role to an isolated observer role within two consecutive weeks, a second-level suggested interview-level early warning is triggered.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention overcomes the limitations of traditional assessments, which are restricted to static and controlled environments, by deploying a scenario-adaptive data collection module to continuously collect anonymized behavioral data in multiple uncontrolled natural scenarios. Students' behaviors and physiological reactions in an unaware, natural state have greater ecological validity. The system captures their true stress responses and emotional patterns, improving the correlation and credibility between assessment data and actual psychological states, thus laying a solid data foundation for assessment.

[0017] 2. This invention, by constructing a multi-granularity behavior analysis engine, achieves comprehensive and multi-layered extraction of behavioral features, ranging from microscopic physiological signals and mesoscopic specific actions to macroscopic social interactions. This hierarchical analysis framework not only enhances the ability to capture hidden indicators such as heart rate variability and subtle stereotyped movements, but also, through macroscopic social behavior analysis, places the assessment of an individual's psychological state within their real interpersonal relationship context. This qualitatively improves the sensitivity and specificity in identifying persistent and situational abnormal behaviors such as social avoidance and decreased participation, enabling earlier and more accurate detection of potential risks.

[0018] 3. This invention maps high-dimensional heterogeneous behavioral characteristics to interpretable psychological state dimensions through a psychological state quantification modeling module, and utilizes a dynamic early warning and feedback module to achieve hierarchical responses based on data and rules. This system not only provides instantaneous state quantification but also tracks long-term trends, achieving a leap from static assessment to dynamic monitoring. The closed-loop design of the early warning mechanism and personalized feedback transforms technological discoveries into actionable intervention clues, serving the early warning and precise intervention needs of campus mental health work, and enhancing the initiative, scientific rigor, and timeliness of mental health management. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the overall technical solution architecture of the present invention; Figure 2 This is a schematic diagram of the core principle framework of the multi-granularity behavior parsing engine in this invention; Figure 3 This is a flowchart illustrating the logical flow of scene adaptation data collection and anonymization processing in this invention. Figure 4 This is a schematic diagram of the interaction and data flow between the quantitative modeling of psychological states and dynamic early warning in this invention. Detailed Implementation

[0020] Example 1: The overall technical architecture of the computer vision-based mental health assessment system for vocational college students proposed in this invention is as follows: Figure 1As shown, it includes a scene adaptation and acquisition module, a multi-granularity behavior analysis engine, a psychological state quantitative modeling module, and a dynamic early warning and feedback module. The modules are connected and logically coordinated through a dedicated campus network to form a complete closed loop from natural behavior perception, multi-level feature extraction, psychological state mapping to graded intervention response.

[0021] The scene-adaptive acquisition module is deployed in multiple pre-defined uncontrolled natural scenes within the campus, including classroom learning scenes, library study scenes, cafeteria dining scenes, playground activity scenes, and dormitory public area scenes. Each scene is equipped with a distributed visual sensing node, which integrates a wide-angle optical lens and an infrared illumination unit to ensure stable acquisition of video stream data under conditions of strong daylight, low nightlight, and even complete absence of visible light. The wide-angle optical lens has a field of view of no less than 120 degrees to cover as large a public activity area as possible; the infrared illumination unit uses an 850nm wavelength near-infrared light source with a maximum illumination distance of 8 meters and features automatic power adjustment, dynamically adjusting the output brightness according to the ambient light intensity to avoid image quality degradation due to overexposure or underexposure.

[0022] All visual sensing nodes have built-in edge computing units, which are equipped with lightweight neural network inference engines for real-time anonymization processing of video streams locally. The processing flow is as follows: Figure 3 As shown, the system first performs face region detection on the original video frames using a face localization algorithm based on a single-stage object detector, achieving a localization accuracy of over 95%. Then, Gaussian blurring is applied to the detected face regions with a blur kernel size of 31 pixels and a standard deviation of 15, ensuring that identifiable facial features cannot be recovered. The system further extracts the human skeleton joint coordinate sequence and global behavioral contour. Skeletal joint extraction employs a multi-person pose estimation algorithm based on heatmap regression, outputting the two-dimensional coordinates of 33 key points, including the head, shoulders, elbows, wrists, hips, knees, and ankles. The global behavioral contour is generated using background subtraction and foreground segmentation algorithms, preserving the overall human motion shape but not containing any identity information. The final anonymized feature video stream only contains the aforementioned skeleton coordinate sequence and contour mask. The original video data is immediately discarded after edge node processing and is not stored or uploaded in any form, thus strictly protecting student privacy and security.

[0023] All anonymized video streams are transmitted to the central data server via a dedicated encrypted campus network. This network employs virtual LAN isolation technology, is physically isolated from other campus systems, and uses end-to-end transport layer security protocols to ensure data integrity and confidentiality during transmission. The central data server utilizes a distributed storage architecture, sharding data according to timestamps and scene types, supporting high-concurrency read / write operations and millisecond-level retrieval, providing data support for subsequent multi-granularity behavior analysis.

[0024] The multi-granularity behavior analysis engine connects to a central data server to perform hierarchical analysis on the received anonymized feature video stream, extracting multi-level features ranging from microscopic physiological behavior and mesoscopic motor behavior to macroscopic social behavior. The overall framework of this multi-granularity behavior analysis engine is as follows: Figure 2 As shown, it consists of a micro-physiological behavior analysis submodule, a meso-action behavior analysis submodule, and a macro-social behavior analysis submodule. The three submodules run in parallel and each independently processes behavioral signals in its corresponding dimension.

[0025] The microscopic physiological behavior analysis submodule focuses on extracting non-contact physiological indicators from the human head region contour sequence in anonymized feature video streams. Since the face in the original video is blurred, this submodule instead utilizes the periodic micro-movements and local brightness variations of the head contour to indirectly infer physiological states. This submodule first delineates multiple 16×16 pixel regions of interest within the head contour region. These regions are evenly distributed in areas with rich blood flow and thin skin, such as the forehead, cheeks, and jaw. Then, the Lucas-Candard optical flow algorithm is used to calculate the pixel displacement vector field of these regions between consecutive video frames. This Lucas-Candard optical flow algorithm assumes constant optical flow within a local window and achieves displacement estimation by solving the following system of equations: ; , and These represent the gradients of the image in the x, y, and time directions, respectively. and The optical flow vector components to be solved are shown. Principal component analysis is performed on the optical flow vector fields of all regions of interest to extract the basis vectors representing periodic motion, which are then projected onto the color change channel to enhance the signal-to-noise ratio of the subtle periodic changes in skin color caused by subcutaneous blood flow. The enhanced brightness change sequence is then bandpass filtered from 0.8 Hz to 3 Hz, covering typical human heart and respiratory rates. The filtered signal is then subjected to Fast Fourier Transform (FFT) for spectral analysis, and the frequency corresponding to the dominant peak is taken as the estimated heart rate value, in beats per minute. Simultaneously, the energy proportion of the signal in the 0.2 to 0.5 Hz frequency band is calculated as an indicator of respiratory rhythm stability. A lower respiratory rhythm stability index indicates more irregular breathing, which may reflect anxiety or tension.

[0026] This microscopic physiological behavior analysis submodule calculates blinking behavior parameters by tracking key points of the eye contour. The system defines an effective blink as an event with an eyelid closure duration between 100 and 500 milliseconds. The number of effective blinks per minute is counted using a sliding window to obtain the blink frequency; simultaneously, the duration of each blink is recorded and its average value is calculated. Research shows that individuals in anxious states exhibit increased blink frequency and shortened blink duration; therefore, these two parameters are considered important physiological indicators of emotional arousal.

[0027] The mesoscopic action behavior analysis submodule is used to identify and quantify specific action units and behavioral patterns from the human skeleton joint coordinate sequence of anonymized feature video streams. This submodule incorporates a pre-trained temporal convolutional network model, whose input is a 3D coordinate sequence (x, y, z) of 33 skeleton joints extracted from the past 64 frames of video. The z-coordinate is calculated from the 2D coordinates using a monocular depth estimation algorithm. The network structure consists of 8 one-dimensional convolutional layers and 3 fully connected layers, with kernel sizes of 7, 5, and 3, respectively, used to capture long-term, medium-term, and short-term action dependencies. This temporal convolutional network model is trained on a large-scale dataset containing more than 100,000 hours of labeled video and can accurately identify 52 predefined action units, including leaning forward while sitting, frequent changes in posture, crossed arms, head down, wandering gaze, and rapid pacing.

[0028] For each identified action unit, the system records its timestamp, duration, and frequency in a specific scenario. For example, in a library study scenario, if "leaning forward" lasts for more than 5 minutes and is accompanied by "wandering gaze," it may indicate inattention or excessive cognitive load. In a cafeteria scenario, if "crossed arms" occurs frequently and co-occurs with "looking down," it may reflect defensiveness or social avoidance tendencies. Furthermore, this mesoscopic action behavior analysis submodule constructs a Markov model of an individual's behavioral sequence over a specific time period. All action units are treated as discrete states in a state space, and the number of transitions between any two states is counted to construct a transition probability matrix. By calculating the entropy value of this transition probability matrix, the regularity of an individual's behavioral patterns can be quantified: the lower the entropy value, the more stereotyped and repetitive the behavior, commonly seen in individuals with depressive or autistic tendencies; the higher the entropy value, the more random and unpredictable the behavior, potentially related to impulse control disorders.

[0029] The macro-social behavior analysis submodule analyzes the spatial relationships and interaction patterns among multiple individuals in public scenes. First, it continuously tracks all individuals in a video frame using a multi-target tracking algorithm (such as DeepSORT), obtaining the center coordinates and velocity vector of each individual in each frame. It then calculates the real-time Euclidean distance between any two individuals. When this real-time Euclidean distance remains less than a threshold of 1.5 meters for more than 5 seconds, the system determines that the two are in a "proximity" state; this threshold references the upper limit of intimate distance in human social interactions. The macro-social behavior analysis submodule then calculates the average number of companions in a "proximity" state for each target individual within a unit observation period, using this as an indicator of their social contact density. A higher index indicates higher social activity.

[0030] The system quantifies social tendency by analyzing the change in the angle between the direction of the line connecting the target individual to their nearest companion and the target individual's own direction of movement. If the angle is consistently less than 45 degrees and the target individual moves towards their companion, it is considered an active approach behavior; if the angle is greater than 135 degrees and the target individual moves away from their companion, it is considered an avoidance behavior. The system records the ratio of approach to avoidance behaviors per unit of time as a positive indicator of social motivation. Simultaneously, in group activities, the system identifies group boundaries using clustering algorithms and calculates the percentage of time the target individual spends at the group edge (i.e., at a distance greater than 80% of the group's centroid), serving as a negative indicator of social participation. A higher percentage of this time indicates a greater tendency for the individual to isolate themselves, a significant behavioral marker of social withdrawal.

[0031] The psychological state quantification modeling module receives all output from the multi-granularity behavior analysis engine, integrates temporal feature vectors with more than 120 dimensions, and maps them to interpretable psychological state dimensions through a machine learning model. The interaction relationships and data flow of this psychological state quantification modeling module are as follows: Figure 4 As shown, it includes a feature fusion layer and a state inference layer.

[0032] The feature fusion layer first performs time alignment and standardization on all input features. All features are averaged over a 1-minute time window with a 10-second step size to smooth out instantaneous noise and preserve trend information. Subsequently, a dynamic attention mechanism is used to assign weights to different feature dimensions for different scenarios. This dynamic attention mechanism's weight generation network is a two-layer fully connected network. The input is a composite vector formed by concatenating the current scene encoding and the original feature vector, and the output is a weight vector with the same dimensions as the original feature vector. This weight vector is normalized using a soft maximization function and then multiplied element-wise with the original feature vector to achieve dynamic feature weighting. For example, in a self-study scenario, attention weights are more likely to be assigned to features such as blink frequency, posture stability, and gaze wandering frequency; in a playground activity scenario, weights are more likely to be assigned to features such as social contact density, the proportion of time spent at the edge of a group, and the coefficient of variation of movement speed.

[0033] The state inference layer comprises four parallel deep neural network models, corresponding to the dimensions of emotional valence, emotional arousal, social motivation, and behavioral vitality, respectively. Each deep neural network model is a three-layer fully connected network with 256, 128, and 64 neurons in the hidden layers, respectively, and uses modified linear units as the activation function. The deep neural network models take a weighted and fused 120-dimensional feature vector as input and output a continuous score between 0 and 100. This continuous score represents the instantaneous state intensity of an individual in the behavioral vitality dimension; for example, a higher emotional valence score indicates a more positive emotion, while a lower behavioral vitality score indicates a weaker willingness to participate. All deep neural network models are trained end-to-end on a large-scale student behavioral-psychological paired dataset calibrated using professional psychological assessment tools (such as PHQ-9, GAD-7, and SCL-90), with training samples covering more than 5000 vocational college students. This ensures that the output of the deep neural network models achieves a Pearson correlation coefficient of over 0.75 with the gold standard assessment results, demonstrating statistical significance.

[0034] The state inference layer also includes a long short-term memory network (LSTM) to analyze the trends in scores across various psychological state dimensions over time scales of several days or even weeks. This LTM network takes the daily average scores for each dimension as input sequences and outputs trend features, including continuously declining scores, abnormally increased volatility, and periodic missing values. These trend patterns are encoded as high-order risk features and directly input into the dynamic early warning and feedback module.

[0035] The dynamic early warning and feedback module executes a three-tiered early warning and personalized feedback based on the output and historical trends of the psychological state quantitative modeling module. This module has a built-in rule base. The first level of warning is the attention level, triggered by the following conditions: any psychological state dimension score is less than 20% of the student's personal historical baseline value for three consecutive days, or a single day's score is less than 30 percentage points below the norm for the same grade and gender group. The personal historical baseline value is determined by statistically analyzing the moving average and standard deviation of the student's scores for each dimension over the past 30 days in the same scenario and time period. The baseline value is updated every 7 days to adapt to the student's natural growth and seasonal changes.

[0036] The second-level warning is the recommended interview level, triggered by: two or more related dimensions simultaneously reaching the attention level threshold, or the long short-term memory network detecting a clear social avoidance trend pattern.

[0037] The third level of warning is the emergency intervention level. The triggering conditions are: the behavioral vitality dimension score drops sharply to below 20 in a single day, accompanied by a heart rate variability index that is less than the normal range, or a predefined risky behavior pattern is detected in open scenarios such as playgrounds.

[0038] Once an alert is triggered, the system automatically generates a structured assessment report, which includes the dimensions that triggered the alert, specific behavioral evidence, historical change curves, and preliminary interpretation suggestions based on clinical guidelines. This structured assessment report is pushed to the authorized mental health teacher work platform through an encrypted interface that meets the requirements of Level 3 National Information Security Protection System, and is accessible only to full-time teachers holding digital certificates.

[0039] Meanwhile, with the student's informed consent, the system can push non-labeled, positive self-care tips to the student's campus digital terminal, such as "You seem a little tired lately, why not try our 5-minute mindfulness audio" or "The library has launched a new stress management micro-course, welcome to try it," along with links to optional online psychological resources. All push content is reviewed by psychology experts to avoid stigmatization or anxiety and to ensure constructive and supportive feedback.

[0040] In summary, this embodiment deploys a distributed visual sensing network in a real campus environment, combines it with edge computing to collect behavioral data while protecting privacy, and constructs a mental health assessment and intervention system with high ecological validity, strong sensitivity, and timely response through multi-granularity analysis, dynamic modeling, and intelligent early warning. This addresses the shortcomings of existing technologies in capturing natural behavior, identifying subtle anomalies, and tracking long-term trends.

[0041] Example 2: Building upon Example 1, this example enhances the macro-social behavior analysis submodule by introducing a group interaction modeling mechanism based on graph neural networks to improve the accuracy of individual role identification in complex group dynamics. Specifically, in playground activity scenarios or large gatherings, traditional Euclidean distance and orientation angle analysis are insufficient to characterize hierarchical relationships and influence flow within multi-person groups. Therefore, this example treats all individuals in a video frame as graph nodes, using proximity relationships and interaction frequency as edge weights to construct a dynamic social graph. This graph is updated every 5 seconds, and node features include individual movement speed, acceleration, orientation angle, and historical social contact density; edge weights are determined by the product of the reciprocal of the proximity distance and the co-occurrence duration.

[0042] A graph convolutional network (GCNN) is employed for embedding learning on a dynamic social graph. This GCNN consists of three layers of graph convolution operations, each aggregating information from neighboring nodes to update the representation of the central node. The final output node embedding vectors are used to classify an individual's role within the group, such as "core leader," "marginal follower," and "isolated observer." Research indicates that students who consistently occupy the "isolated observer" role have a higher risk of depression than those in other roles. This role label is used as a new output dimension in the macro-social behavior analysis submodule and input into the psychological state quantification modeling module, further improving the granularity and accuracy of the social motivation dimension assessment.

[0043] This embodiment expands the rule base of the dynamic early warning and feedback module, adding early warning rules based on changes in group roles. For example, if a student changes from a "core leader" role to an "isolated observer" within two consecutive weeks, even if their scores in each psychological state dimension do not reach the traditional thresholds, the system will still trigger a second-level early warning, indicating a possible sharp decline in social functioning due to a significant life event. This enhanced mechanism allows the system to focus not only on the individual's absolute state but also on their relative position changes within the social network, thereby achieving earlier and more contextualized risk identification.

[0044] Through the above enhancements, this embodiment improves the ability to assess the psychological state of individuals in complex social scenarios while maintaining the original system architecture. It is applicable to highly interactive educational scenarios such as club activities and practical training courses in higher vocational colleges, further strengthening the ecological validity and practical value of the system.

[0045] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0046] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A computer vision-based mental health assessment system for vocational college students, characterized in that, include: The scene-adaptive acquisition module is used to continuously collect anonymized video stream data from students in multiple preset uncontrolled natural scenes on campus. A multi-granularity behavior parsing engine, connected to the scene adaptation acquisition module, is used to perform layered parsing on the received anonymized feature video stream and extract multi-layered behavior features. The multi-granularity behavior analysis engine includes a micro-physiological behavior analysis sub-module, a meso-action behavior analysis sub-module, and a macro-social behavior analysis sub-module. The psychological state quantification modeling module is connected to the multi-granularity behavior analysis engine and is used to integrate multi-level behavioral features and map them to potential psychological state dimensions through machine learning models. The psychological state quantification modeling module includes a feature fusion layer and a state inference layer; The dynamic early warning and feedback module is connected to the psychological state quantitative modeling module and is used to perform graded early warning and personalized feedback based on the output of the quantitative model and historical trends. The dynamic early warning and feedback module has a built-in three-level early warning rule base; After an alert is triggered, the dynamic early warning and feedback module automatically generates a structured assessment report and pushes it to the authorized mental health teacher's work platform through an encrypted interface. The dynamic early warning and feedback module is also connected to the campus digital terminal to push self-care tips and online psychological resource links to students with their informed consent.

2. The computer vision-based mental health assessment system for vocational college students according to claim 1, characterized in that, The scene adaptation acquisition module is equipped with multiple distributed visual sensing nodes, each of which integrates a wide-angle optical lens and an infrared supplementary light unit. The visual sensing node has a built-in edge computing unit for real-time anonymization of the acquired raw video stream data. The real-time anonymization process includes face region detection and blurring to generate an anonymized feature video stream that retains only the human skeleton joint coordinate sequence and global behavioral contour. All anonymized video streams are transmitted to the central data server via the campus-dedicated network.

3. The computer vision-based mental health assessment system for vocational college students according to claim 2, characterized in that, The microscopic physiological behavior analysis submodule is used to analyze non-contact physiological indicators from the human head region contour sequence in an anonymized feature video stream. The analysis process includes: calculating the pixel brightness change sequence of a specified area of ​​the face region using optical flow; performing bandpass filtering on the brightness change sequence, with the filter passband set to 0.8 Hz to 3 Hz; performing spectral analysis on the filtered signal, extracting its dominant frequency component as the estimated heart rate value, and calculating the energy proportion of the signal in the respiratory frequency band as a respiratory rhythm stability indicator; simultaneously, by tracking key points of the eye contour, calculating the blink frequency and average duration of a single blink per unit time. The mesoscopic action behavior analysis submodule is used to identify and quantify predefined action units and behavior patterns from the human skeleton joint coordinate sequence of anonymized feature video streams. The mesoscopic action behavior analysis submodule has a built-in action recognition model based on temporal convolutional networks. For each identified action unit, its timestamp, duration, and frequency of occurrence in a specific scenario are recorded. By analyzing the transition probability between consecutive action units, a Markov model of an individual's behavior sequence within a specific time period is constructed. The macro-social behavior analysis submodule is used to analyze the spatial relationships and interaction patterns among multiple individuals in a public scene video stream. The analysis process includes: determining the position and motion trajectory of all individuals in the video frame through target detection and tracking algorithms; calculating the real-time Euclidean distance between any two individuals, and determining that the two are in a proximity state when the real-time Euclidean distance is continuously less than a preset threshold of 1.5 meters; statistically analyzing the average number of companions in a proximity state of the target individual within a unit observation period as its social contact density index; quantifying its directional behavior by analyzing the change in the angle between the direction of the line connecting the target individual and its nearest companion and its own motion direction; and statistically analyzing the proportion of time the target individual spends at the edge of the group in a group activity scene as a negative indicator of its social participation.

4. The computer vision-based mental health assessment system for vocational college students according to claim 3, characterized in that, The feature fusion layer is used to receive time-series feature vectors from micro, meso, and macro sub-modules, and perform moving average and standardization processing with minutes as the time window; the feature fusion layer uses an attention mechanism to assign weights to different feature dimensions under different scenarios; The state inference layer comprises multiple parallel deep neural network models, each corresponding to a psychological state dimension, including emotional valence, emotional arousal, social motivation, and behavioral vitality. Each neural network model takes a weighted and fused feature vector as input and outputs a continuous score between 0 and 100. The state inference layer also includes a long short-term memory network to analyze the changing trends of the scores of each psychological state dimension over time scales of several days or even several weeks.

5. The computer vision-based mental health assessment system for vocational college students according to claim 4, characterized in that, The optical flow calculation in the microscopic physiological behavior analysis submodule specifically adopts the Lucas-Cannard optical flow algorithm. The Lucas-Cannard optical flow algorithm delineates multiple regions of interest with a size of 16 pixels × 16 pixels in the facial region. By calculating the pixel displacement vector field of these regions between consecutive video frames, and then performing principal component analysis on the vector field, the basis vectors representing periodic motion are extracted and projected onto the color change channel.

6. The computer vision-based mental health assessment system for vocational college students according to claim 5, characterized in that, The pre-trained temporal convolutional network in the mesoscopic action behavior analysis submodule contains 8 one-dimensional convolutional layers and 3 fully connected layers, with convolutional kernel sizes of 7, 5, and 3, respectively. The input of the temporal convolutional network is the three-dimensional coordinate sequence of 33 skeleton joints extracted from the past 64 frames of video.

7. The computer vision-based mental health assessment system for vocational college students according to claim 6, characterized in that, The attention mechanism in the psychological state quantification modeling module has a weight generation network that is a two-layer fully connected network. The fully connected network takes the current scene encoding and the original feature vector as input and outputs a weight vector with the same dimension as the original feature vector. After the weight vector is normalized by a soft maximization function, it is multiplied element-wise with the original feature vector.

8. The computer vision-based mental health assessment system for vocational college students according to claim 7, characterized in that, The three-level early warning rule base in the dynamic early warning and feedback module includes: The first level of alert is triggered when the score of any psychological state dimension is less than 20% of the individual's historical baseline for three consecutive days, or when the score on a single day is less than 30 percentage points of the group norm. The second level of advice is an interview-level warning, triggered when two or more related dimensions simultaneously reach the attention-level threshold, or when a clear social avoidance trend pattern is detected. The third level of emergency intervention warning is triggered by the detection of extreme behavioral decline combined with abnormal physiological indicators, or the appearance of a predefined risky behavior pattern in a scenario. The individual historical baseline value is determined by statistically analyzing the moving average and standard deviation of the student's scores for each psychological state dimension in the same scenario and time period over the past 30 days; the individual historical baseline value is updated every 7 days.

9. The computer vision-based mental health assessment system for vocational college students according to claim 8, characterized in that, The macro-social behavior analysis submodule is also used to introduce a group interaction modeling mechanism based on graph neural networks. The group interaction modeling mechanism treats all individuals in the video frame as graph nodes, and uses the proximity relationship and interaction frequency between individuals as edge weights to construct a dynamic social graph. A graph convolutional network is used to embed the dynamic social graph to classify the role of individuals in the group.

10. The computer vision-based mental health assessment system for vocational college students according to claim 9, characterized in that, The early warning rule base of the dynamic early warning and feedback module is also extended with early warning rules based on changes in group roles; when it is detected that a student has changed from a core leader role to an isolated observer role within two consecutive weeks, a second-level suggested interview-level early warning is triggered.

Images