Cognitive risk detection method based on browser camera eye movement estimation and quality control

Abstract

The application discloses a kind of based on browser camera eye movement estimation and quality control cognitive risk detection method, belong to computer vision and digital medical field.It includes: complete phased visual cognitive screening task chain and collect facial video stream;Screen gaze point coordinates of official task section are obtained based on facial video stream;Each screen gaze point coordinate is resampled to uniform time axis;When the quality control result of official task section is passed, frame-by-frame timing feature containing basic space-time feature and event feature is constructed based on screen gaze point coordinate, and task-level summary feature is extracted for each official task section;Frame-by-frame timing feature and each task-level summary feature are input into risk inference model, and cognitive risk assessment result is obtained.The present application converts calibration section sample into conversation-specific screen drop point projection, realizes stable screen gaze point estimation under uncontrolled conditions;Explicit quality control is completed before scoring, and finally the cognitive risk probability is output after probability calibration.

Description

Technical Field

[0001] This invention belongs to the field of computer vision and digital medical technology, specifically relating to a cognitive risk detection method based on browser camera eye-tracking estimation and quality control. Background Technology

[0002] Early detection of cognitive decline and related risk states typically relies on paper-and-pencil scales, specialist interviews, or specialized eye-tracking devices. These methods are either highly dependent on venues, personnel, and equipment, or have high data collection costs, making it difficult to cover low-friction scenarios such as community screening, primary care triage, waiting room staffing, and home follow-up.

[0003] In existing camera-based remote eye-tracking solutions, many methods upload the raw video to a server or directly proceed to the scoring process after obtaining the gaze point. The screen gaze point, which is the point where the user's gaze falls on the screen, plays a crucial role in subsequent eye-tracking event reconstruction, task-level feature extraction, and cognitive risk inference. However, these solutions are prone to problems such as unstable gaze points, calibration failures, and poor cross-session comparability under conditions of browser frame rate fluctuations, differences in terminal computing power, asynchronous network uploads, head pose drift, partial occlusion, and changes in the geometric relationships of different device placements.

[0004] Furthermore, many existing cognitive screening programs conflate "whether a score is possible" with "whether a risk exists." Low-quality samples are either silently discarded or forcibly fed into the model for scoring, making it difficult to distinguish between "the subject is indeed low-risk" and "this sample is uninterpretable" in real-world deployment scenarios. This also fails to provide a clear basis for determining whether recalibration, retesting, or manual review is necessary.

[0005] Furthermore, most general browser camera eye-tracking estimation methods focus on general gaze regression, lacking an integrated design for the compliance of calibration anchor points within the screening task chain, task stage switching, quality control gating, and result threshold governance. Relying solely on a single static calibration and directly trusting the fitted result can easily lead to problems such as underfitting near the calibration point, accumulated task stage drift, and unstable final screening probability in deployment scenarios.

[0006] Therefore, there is an urgent need for a cognitive risk detection method built around real browser deployment scenarios. Summary of the Invention

[0007] To address the aforementioned problems in existing technologies, this invention provides a cognitive risk detection method based on browser camera eye-tracking estimation and quality control. This method acquires usable facial key point sequences without uploading the original video, transforms calibrated segment samples into session-specific screen landing point projections, and maintains the estimation usability from facial key points to screen landing points even under uncontrolled conditions such as unstable browser frame rates, asynchronous network uploads, head pose drift, and partial occlusion. Explicit quality control is performed before scoring, and a probability-calibrated subject-level cognitive risk probability is output based on a fixed-order visual cognitive screening task chain.

[0008] To achieve the above objectives, the technical solution of the present invention is as follows:

[0009] A cognitive risk detection method based on browser camera eye-tracking estimation and quality control includes:

[0010] S1. The browser loads the phased visual cognition screening task chain and calls the browser's front-facing camera to capture the subject's facial video stream during the execution of the phased visual cognition screening task chain; the phased visual cognition screening task chain includes various visual cognition task tests executed sequentially in the calibration segment and the formal task segment.

[0011] S2. Obtain the screen gaze coordinates of the formal task segment based on the facial video stream;

[0012] S3. Resample the coordinates of each screen gaze point to a unified time axis;

[0013] S4. Calculate the quality indicators of the formal task segment and determine the quality control results of the formal task segment.

[0014] S5. When the quality control result of the formal task segment is passed, construct frame-by-frame temporal features containing basic spatiotemporal features and event features based on the resampled screen gaze point coordinates, and extract task-level summary features for each formal task segment.

[0015] S6. Input the frame-by-frame temporal features and the task-level aggregated features into the risk inference model to obtain the cognitive risk probability, and output the final cognitive risk screening result based on the cognitive risk probability.

[0016] Furthermore, in step S2, the screen gaze coordinates of the formal task segment are obtained based on the following method:

[0017] S2.1 The browser extracts the coordinates of facial key points frame by frame from the facial video stream. The set of facial key point coordinates detected in a single frame forms a facial key point sample. The task identifier of the task segment to which each facial key point sample belongs is recorded. Several facial key point samples sorted by timestamp are obtained and uploaded to the processing end.

[0018] S2.2 The processing end extracts facial key point samples of the calibration segment based on the task identifier, and constructs a calibration support set based on the facial key point samples of the calibration segment;

[0019] S2.3. Solve the initial session projection matrix based on the calibration support set, and iteratively refine the initial session projection matrix to obtain the final session projection matrix;

[0020] S2.4. Based on the quality indicators of the calibration support set, determine the quality control results of the calibration segment;

[0021] S2.5 In response to the quality control result being passed, for any facial key point sample within the formal task segment, the final session projection matrix is ​​used to perform mapping calculations to obtain the screen gaze point coordinates corresponding to the facial key point sample.

[0022] Furthermore, in step S2.2, the calibration support set is constructed based on the following method:

[0023] Identify the time window of each calibration anchor point in the calibration segment, that is, the time period presented by each calibration anchor point;

[0024] The facial key point samples of the calibration segment are divided into the corresponding anchor point time windows based on the timestamp;

[0025] Within each anchor point time window, invalid frames are removed, stable head poses are screened, and blinking frames are filtered for facial key point samples to obtain valid samples.

[0026] The valid samples within each anchor time window are aggregated into a calibration support set, which is represented as follows:

[0027]

[0028] in, To support the sample size; Indicates the calibration support set within the first A set of facial key point coordinates for each supporting sample; Indicates the first The anchor point's real screen reference coordinates corresponding to the anchor point's time window where each supporting sample is located; This represents the confidence weight.

[0029] Furthermore, in step S2.3, based on the calibrated support set, the initial session projection matrix is ​​solved using regularized least squares with projection prior constraints:

[0030]

[0031] in, This is the initial session projection matrix. The projective prior obtained from offline training; The regularization coefficient is used. Represents the square of the Frobenius norm; Indicates the calibration support set within the first Extended features of each supporting sample ; The output of the shared geometric representation network represents the first... Geometric descriptive features of each supporting sample , Represents a shared geometric representation network. Indicates the first A normalized set of key points is obtained by performing translation and scale normalization on the coordinates of all facial key points within a support sample; This is a local eye geometry vector constructed based on normalized keypoints; This represents the head pose vector.

[0032] Furthermore, the final session projection matrix is ​​obtained based on the following method:

[0033] Using the initial session projection matrix as the initial value, the calibration support set is iteratively refined using a recursive least squares algorithm:

[0034] Let the first Round The inverse covariance matrix of each supporting sample before update is: Then the recursive update formula is:

[0035]

[0036]

[0037]

[0038]

[0039] in, For the first Round absorption Gain vector for each support sample; and The first Round Each session projection matrix supports both the pre- and post-update samples; For the first Round The two-dimensional reprojection residuals of each supporting sample, i.e., the real screen reference coordinates. Compared with current forecasts The difference between them; and The first Round Each support sample inverse covariance matrix before and after update;

[0040] The forgetting factor of the recursive least squares algorithm is set to 1, and the anchor point residual is calculated after each round:

[0041]

[0042] in, The length of the screen diagonal; Indicates the first All rounds completed A session projection matrix supporting sample absorption;

[0043] If the decrease in the residual of the anchor point between two adjacent rounds is less than a preset threshold, the update is stopped in advance, and the final session projection matrix is ​​obtained. Otherwise, continue with the next round of iterations to refine until the maximum number of iterations is reached.

[0044] Furthermore, the quality metrics of the calibration support set include the percentage of valid calibration frames. , effective coverage of calibration anchor points Calibration residuals and calibration head posture stability index At least one of them; if and only if Greater than the first threshold and Greater than the second threshold and Less than the third threshold and If the result is less than the fourth threshold, the quality control result is considered passed; otherwise, it is considered failed.

[0045] Furthermore, in step S4, the quality indicators of the formal task segment include the percentage of effective frames for key points in the formal task. Task coverage percentage of abnormal interruptions Formal mission head posture stability indicators and the proportion of abnormal coordinates At least one of them; if and only if Greater than the fifth threshold and Greater than the sixth threshold and Less than the seventh threshold and Less than the eighth threshold and If the result is less than the ninth threshold, the quality control result is considered passed; otherwise, it is considered failed.

[0046] Furthermore, the frame-by-frame temporal features for each moment on the unified timeline are constructed as follows:

[0047]

[0048]

[0049]

[0050]

[0051] in, Represents the basic spatiotemporal characteristics; Indicates characteristics that indicate the type of eye movement event; This indicates the start and end positions and length encoding of the current eye-tracking event segment; To unify the timeline The task identifier of the task segment to which each time grid point belongs; , Represents the x and y coordinates of the screen gaze point; , and They represent the first The horizontal velocity, vertical velocity, and velocity magnitude at each time grid point , , , This is the resampling interval; and Indicates the first Screen gaze coordinates at each time grid point The coordinates relative to the normalized width and height of the screen display area; , and They represent the first The gaze, saccade, and smoothing tracking event indicators for each time grid point are obtained based on the continuous screen gaze point coordinates and velocity magnitude sequence on a unified time grid. If it belongs to the corresponding eye movement event, it is set to 1; otherwise, it is set to 0. , , and These represent the normalized start position, end position, normalized position from the start position, and normalized position from the end position of the current eye-tracking event segment, respectively. and These represent the sine and cosine period codes for the length of the current eye-tracking event segment, respectively.

[0052] Furthermore, the risk reasoning model is a hierarchical temporal model, including a frame-by-frame feature mapping layer, a task type embedding layer, a temporal coding layer, a task-level attention pooling layer, and a subject-level fusion layer;

[0053] frame-by-frame temporal features By task identifier The feature sequence is divided into corresponding task segments, the first segment... Within the first task segment The features of each time grid point are denoted as follows: ; In addition to In addition to the basic spatiotemporal features, eye-movement event type indication features, and eye-movement event segment location coding features, let them be denoted as The frame-by-frame feature mapping layer uses The input is processed by a linear or multi-layer perceptron to obtain a frame-by-frame feature map representation. ;

[0054] The task type embedding layer is based on In Generate task embedding and will With the frame-by-frame feature mapping representation By adding or concatenating the elements and then performing linear projection, a frame-level hiding representation is obtained. ;

[0055] The temporal coding layer uses frame-level hidden representation sequences within the same task stimulus fragment. As input, a temporal encoder is used to model temporal dependencies within the same task segment, gaze-saccade transitions, and cross-event contexts, outputting a frame-level context representation. ,in Indicates the first The effective time grid length after resampling of each task segment;

[0056] The task-level attention pooling layer is based on the first... Frame-level context representation of each task segment Using this as input, calculate the weight of the risk evidence for that task segment at each time grid point. And aggregate to obtain the task segment time sequence representation. :

[0057]

[0058]

[0059] in, This represents a learnable query vector, a task-type conditional query, or an equivalent query representation generated by a gated attention network. Represents the learnable weight matrix;

[0060] The subject-level fusion layer receives data according to a fixed task sequence. ,in , They represent the first The temporal representation of each task segment, task-level aggregated features, and missing masks are obtained. The subject-level fusion layer uses cross-task attention, gated weighted averaging, or multilayer perceptron for fusion to obtain the subject-level representation. and output the original risk score. and original risk probability ;

[0061] The original risk probability The calibrated risk probability is obtained through parameterized probability calibration. ; calibrated risk probability With demographic vector By merging the results, the final screening probability is obtained. :

[0062]

[0063]

[0064] in, , , These are standardized age, gender, and education level codes, respectively. ~ For demographic parameters;

[0065] Final screening probability With configurable referral thresholds When comparing, If the risk is high, then there is cognitive risk; otherwise, it is low risk.

[0066] Furthermore, the formal task segment includes the following segments executed sequentially: visual pairing comparison encoding segment, cue-guided visual search demonstration segment, formal cue-guided visual search segment, delayed visual pairing comparison extraction segment, free viewing segment, and visual motion exploration task VisMET segment. The task-level summary features of each formal task segment are discriminative task-level temporal features.

[0067] The beneficial effects of this invention are:

[0068] (1) The present invention completes the extraction of facial key points on the browser side, and the processing end receives facial key point samples instead of the original facial video, which reduces the bandwidth pressure and privacy risks caused by video upload and is easy to deploy in scenarios such as home, community and waiting room.

[0069] (2) The present invention uses calibration segments to form a session projection matrix and absorbs the current session residuals through iterative refinement, so that the estimation of screen gaze points under different screen sizes, camera positions, subject facial differences and slight head posture changes is more stable.

[0070] (3) The present invention sets quality control rules in both the calibration stage and the formal task stage. It first determines whether the current data collection is usable, and then enters the risk reasoning stage. It can distinguish between "data collection is unreadable" and "the cognitive risk is low", thereby reducing misjudgment caused by low-quality data.

[0071] (4) The present invention resamples the gaze points of different task segments to a unified time axis and constructs frame-by-frame temporal features and task-level summary features according to the task identifier, which makes it easy to use eye-tracking evidence in visual pairing comparison, visual search, free viewing and dynamic visual exploration in a unified manner.

[0072] (5) This invention combines frame-by-frame temporal features, task-level summary features and demographic information, and outputs cognitive risk probability and screening results after probability calibration. The results are clear and easy to configure referral thresholds according to different application scenarios.

[0073] (6) This invention does not rely on a dedicated eye tracker, nor does it require complicated offline operations. The screening and testing process can be completed in a regular terminal browser, which is conducive to expanding the screening coverage and providing a basis for recalibration, retesting or manual review. Attached Figure Description

[0074] Figure 1 This is a flowchart of a cognitive risk detection method based on browser camera eye-tracking estimation and quality control in an embodiment of the present invention. Detailed Implementation

[0075] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0076] like Figure 1 As shown, this embodiment provides a cognitive risk detection method based on browser camera eye-tracking estimation and quality control. This method collects eye-tracking data from subjects using the browser's front-facing camera and combines it with a phased visual cognitive screening task chain design to achieve early screening for the risk of cognitive decline. Specifically, it includes the following steps:

[0077] S1. The browser loads the phased visual cognition screening task chain and calls the browser's front-facing camera to capture the subject's facial video stream during the execution of the phased visual cognition screening task chain; the phased visual cognition screening task chain includes various visual cognition task tests executed sequentially in the calibration segment and the formal task segment.

[0078] S2. Obtain the screen gaze coordinates of the formal task segment based on the facial video stream;

[0079] S3. Resample the coordinates of each screen gaze point to a unified time axis;

[0080] S4. Calculate the quality indicators of the formal task segment and determine the quality control results of the formal task segment.

[0081] S5. When the quality control result of the formal task segment is passed, construct frame-by-frame temporal features containing basic spatiotemporal features and event features based on the resampled screen gaze point coordinates, and extract task-level summary features for each formal task segment.

[0082] S6. Input the frame-by-frame temporal features and the task-level aggregated features into the risk inference model to obtain the cognitive risk probability, and output the final cognitive risk screening result based on the cognitive risk probability.

[0083] In one embodiment, in step S1, the screening page is first launched in the user terminal's standard web browser to perform camera permission request, light check, face centering guidance and viewing distance estimation until the subject's face is detected to be within the preset viewfinder and the eyes are clearly visible.

[0084] A phased visual cognition screening task chain with a unified timeline is loaded on the browser, guiding participants to complete various visual cognition tasks sequentially. During this process, the browser uses the front-facing camera to capture facial video streams of the participants. The phased visual cognition screening task chain includes a calibration segment, a visual pairing comparison encoding segment, a cue-guided visual search demonstration segment, a formal cue-guided visual search segment, a delayed visual pairing comparison extraction segment, a free viewing segment, and a VisMET (Vismokinesis-Motion Exploration) segment, which are executed sequentially. The visual pairing comparison encoding segment, the cue-guided visual search demonstration segment, the formal cue-guided visual search segment, the delayed visual pairing comparison extraction segment, the free viewing segment, and the VisMET segment constitute the formal task segment.

[0085] The test task for the calibration segment is as follows: the user looks at several calibration anchor points presented on the screen in a fixed or random order (such as a nine-point calibration segment, where the nine calibration anchor points are the center point, top left point, top center point, top right point, left center point, right center point, bottom left point, bottom center point, and bottom right point). The number, screen reference coordinates, and start and end times of each calibration anchor point are pre-written into a unified timeline.

[0086] The test task of the visual pairing comparison coding segment is as follows: several familiar stimulus images or objects are presented on the screen in a preset order, and a gaze cross is set before each stimulus is presented. The subject naturally watches and encodes the stimulus content according to the prompts on the page. The trial number, AOI boundary and presentation start and end time of each stimulus are written into a unified time axis for subsequent novelty preference and recognition-related feature calculation. Among them, AOI (area of ​​interest) refers to a closed area that is predefined by stimulus metadata and mapped to the screen coordinate system. It can be a rectangle, polygon or other area where the gaze point can be determined.

[0087] The test task content of the visual search demonstration segment with prompts is as follows: the screen first presents a target prompt image or target prompt symbol, then presents a gaze cross and a small number of demonstration search scenarios, guiding the subject to understand the operational rules of finding the target object in the scenario according to the prompts; this segment is used for practice, task comprehension confirmation and quality auxiliary assessment, and is not used as a separate diagnostic basis for formal screening conclusions.

[0088] The test task for the formal prompt-guided visual search segment is as follows: In multiple formal trials, a target prompt image, a gaze cross, and a search scene are presented sequentially. Subjects search for the target object in the scene based on the prompts. The system records the target AOI, target corner AOI, non-target corner AOI, and the start and end times of the trials based on stimulus metadata. This data is used to calculate features such as target discovery delay, initial target gaze latency, path efficiency, percentage of incorrect corner dwells, and number of corner revisits.

[0089] The test task of the delayed visual pairing comparison extraction segment is as follows: The screen presents pairing trials consisting of familiar stimuli and novel stimuli, and the subject freely views each pairing stimulus; The system calculates the novelty preference ratio, the initial novelty fixation latency, the number of AOI switching, the number of revisits, and the left-right offset characteristics based on the familiar stimulus AOI and the novel stimulus AOI to reflect the eye movement evidence related to delayed recognition. Among them, the familiar stimulus AOI is the region of interest surrounding the encoded or previously seen stimuli, and the novel stimulus AOI is the region of interest surrounding the newly appearing stimuli in the same pairing trial.

[0090] The test task for the free viewing segment is as follows: the screen presents natural scenes, complex images, or other visual stimuli without a clear search target, and the subject views them naturally according to the prompts; the system extracts exploration features such as scanning path length, gaze dispersion, region of interest coverage, center offset, transfer entropy, and scene exploration uniformity based on preset AOI or screen partitions.

[0091] The VisMET segment of the visual motion exploration task involves the following: a moving target or dynamic scene with a predetermined trajectory, speed, or direction change is presented on the screen, and the subject naturally follows or watches the motion stimulus according to the prompts; the system extracts dynamic visual exploration features such as following continuity, speed consistency, and phase shift based on the spatiotemporal relationship between the target trajectory and the gaze trajectory.

[0092] In one embodiment, in step S2, the screen gaze coordinates of the formal task segment are obtained based on the following method:

[0093] S2.1 The browser needs to continuously collect facial key point samples of the subject throughout the entire task chain execution process, and record the task identifier of the task segment to which each facial key point sample belongs, so as to align the eye-tracking data with the specific visual cognition task stage.

[0094] S2.2 Extract facial key point samples from the calibration segment based on the task identifier and further construct the calibration support set.

[0095] S2.3. Solve the initial session projection matrix based on the calibration support set, and iteratively refine the initial session projection matrix to obtain the final session projection matrix.

[0096] S2.4. Based on the quality indicators of the calibration support set, determine the quality control results of the calibration segment.

[0097] S2.5 In response to the quality control result being passed, for any facial key point sample in the formal task segment, the final session projection matrix is ​​used to perform mapping calculation to obtain the screen gaze point coordinates corresponding to the facial key point sample.

[0098] The browser-side uses computer vision algorithms (such as facial landmark detection models) to extract facial landmark coordinates frame by frame. These facial landmarks include landmarks in the periorbital region and common anchor points. The periorbital landmarks include the inner and outer corners of the eyes, the center of the iris, and the upper and lower eyelids. The common anchor points include the bridge of the nose, the tip of the nose, the area between the eyebrows, the forehead, and the chin. For example, each frame extracts 468 two-dimensional or three-dimensional coordinate data points for facial landmarks, including those at the corners of the eyes, the center of the iris, and the bridge of the nose.

[0099] In one embodiment, in step S2.2, the calibration support set is constructed based on the following method:

[0100] S2.2.1 Identify the anchor time window corresponding to each calibration anchor point, that is, the time period presented by each calibration anchor point. This time period is determined according to the start and end time of each calibration anchor point presented on the screen.

[0101] S2.2.2. Divide the facial key point samples of the calibration segment into the corresponding anchor point time window according to the timestamp.

[0102] S2.2.3. Within each anchor point time window, invalid frame removal, stable head posture screening, and blinking frame filtering are performed on the facial keypoint samples to obtain valid samples. Invalid frame removal aims to exclude abnormal data caused by the subject's own actions or environmental factors. For example, when blinking is detected (e.g., the distance between eyelid keypoints is less than a preset threshold), facial occlusion (e.g., keypoint confidence is lower than a preset threshold), or a missing face is detected, the frame is marked as invalid and removed. Head posture stability screening is used to exclude frames where the subject's head is shaking violently. For example, the rate of change of head yaw angle, pitch angle, and roll angle between consecutive frames is calculated. If the rate of change exceeds a preset threshold, it is judged as unstable head posture and removed. Through the above screening, noisy data can be effectively removed to ensure the data quality used for subsequent calculations.

[0103] S2.2.4. Aggregate the valid samples within each anchor point time window into a calibration support set, which is represented as:

[0104]

[0105] in, To support the sample size; Indicates the calibration support set within the first A set of facial key point coordinates for each supporting sample; Indicates the first The actual screen reference coordinates of the anchor point corresponding to the anchor point time window of each supporting sample, that is, the screen reference coordinates of the calibrated anchor point that the user should be looking at at that moment. The confidence weight is represented by a weighted average of the overall confidence of facial key points, the visibility of periorbital key points, and head pose stability. If continuous confidence is not currently output, the confidence weight of the selected valid support samples will be uniformly set to [value missing]. .

[0106] By constructing a calibration support set, the original noisy acquisition data is transformed into a high-quality sample set with supervised information, laying the foundation for the accurate solution of the projection matrix and high-precision screen gaze point estimation.

[0107] In one embodiment, the core of step S2.3 lies in establishing a session-adaptive projection relationship. Traditional static calibration methods often rely on a general projection model, which is difficult to adapt to the differences in facial geometry among different subjects and the variations in device placement. This embodiment dynamically solves for the initial projection matrix shared by the same session based on the calibration support set of the current session (i.e., the complete session including both calibration and non-calibration sections of the formal test tasks), and further introduces an iterative refinement mechanism to overcome the problem of static calibration failure. Specifically, it includes:

[0108] S2.3.1: Construct normalized keypoints, local eye geometry vectors, and head pose vectors based on the calibration support set;

[0109] S2.3.2: Encode the normalized keypoints, local eye geometry vectors, and head pose vectors to obtain geometric description features, and further construct extended features;

[0110] S2.3.3: Introducing projection priors, based on the calibration support set, extended features, and projection priors, the sum of reprojection error and regularization term is minimized to obtain the initial session projection matrix;

[0111] S2.3.4: Using the initial session projection matrix as the initial value, perform multiple rounds of iterative optimization on the calibration support set to absorb the residuals and obtain the final session projection matrix.

[0112] In one embodiment, the normalized keypoints are constructed as follows:

[0113] For any support sample in the calibration support set Inner Coordinates of key points Perform translation and scale normalization to obtain normalized key points. :

[0114]

[0115] in, A facial reference center; Scale factor; To prevent stable terms with a denominator of zero; Take the midpoint between the outer corners of the left and right eyes. Take the Euclidean distance between the outer corners of the left and right eyes, or Take the midpoint between the centers of the left and right eyes. Take the Euclidean distance between the centers of the left and right eyes.

[0116] In one embodiment, the local eye geometry vector is constructed based on normalized keypoints, specifically as follows:

[0117] Let the centers of the left and right irises be respectively and The coordinates of the inner and outer corners of the left eye are respectively , The coordinates of the inner and outer corners of the right eye are respectively , The key points of the upper and lower eyelids of the left eye are respectively , The key points of the upper and lower eyelids of the right eye are respectively , Common anchor point set Includes key points such as the bridge of the nose, the tip of the nose, and the area between the eyebrows; local eye geometry vectors. Concatenate the following subvectors in a fixed order: Relative to each , Two-dimensional offset, Relative to each , Two-dimensional offset; to , Normalized distance, to , The normalized distance; the opening and closing of the left and right eyelids, which are the ratios of the distance between the upper and lower eyelids to the distance between the inner and outer corners of the eyes; , to The normalized distances of all common anchor points; and the differences and ratios of the corresponding quantities for the left and right eyes (i.e., the same parameters corresponding to the left and right eyes).

[0118] In one embodiment, step S2.3.2, which involves encoding the normalized keypoints, local eye geometry vectors, and head pose vectors to obtain geometric description features, and further constructing extended features, specifically includes:

[0119]

[0120] in, Indicates the first The normalized keypoint set of each supporting sample; Represents the head pose vector; , and These represent the yaw angle, pitch angle, and roll angle of the head relative to the camera, respectively. The relative depth surrogate quantity derived from interocular distance or facial scale; This represents a shared geometric representation network (which may employ an equivariant graphical neural network EGNN, a lightweight multilayer perceptron, or other equivalent structures) or a parameterized feature projector; Indicates by The output of the first Geometric description features of each supporting sample;

[0121] Based on the geometric description features Build extended features .

[0122] In one embodiment, in step S2.3.3, by introducing a projection prior, the objective function can be expressed as:

[0123]

[0124] in, This is the initial session projection matrix. The projective prior obtained from offline training; The regularization coefficient is used. denoted as the squared Frobenius norm. The objective function consists of two terms: the first is the reprojection error, which measures how well the projection matrix fits the current session data; the second is the regularization term, which constrains the projection matrix not to deviate too far from the general prior.

[0125] In one embodiment, to improve the conformity of the projection results to the calibration anchor points, the shared geometric representation network is trained during the offline training phase. Introducing calibration anchor point fidelity loss :

[0126]

[0127] in, The session projection matrix is ​​obtained from the current support set; For the first Extended features of each supporting sample; For the first The anchor point screen reference coordinates corresponding to the anchor point time window of each supporting sample. The calibration anchor point fidelity loss and the query sample prediction loss are jointly optimized, so that the encoder not only pursues the reduction of overall error, but also explicitly reduces the reprojection error on the calibration anchor point.

[0128] In one embodiment, in step S2.3.4, using the initial session projection matrix as the initial value, multiple rounds of iterative optimization are performed on the calibration support set to absorb residuals, resulting in the final session projection matrix; specifically:

[0129] Let the first Round The inverse covariance matrix of each supporting sample before update is: Then the recursive update formula is:

[0130]

[0131]

[0132]

[0133]

[0134] in, For the first Round absorption Gain vector for each support sample; and The first Round Each session projection matrix supports both the pre- and post-update samples; For the first Round The two-dimensional reprojection residuals of each supporting sample, i.e., the real screen reference coordinates. Compared with current forecasts The difference between them; and The first Round Each support sample inverse covariance matrix before and after update; For the first Extended features of each supporting sample; during the first round of initialization , , To initialize the regularization coefficients, It is an identity matrix; subsequent rounds use the last value of the previous round as the initial value.

[0135] Using the recursive formula described above, each iteration fine-tunes the projection matrix based on the residuals of the current supporting samples. The multi-round iteration process essentially allows the projection matrix to continuously absorb residuals, gradually approaching the optimal solution. This iterative refinement mechanism significantly improves the observability of the projection results to the calibration anchor points, ensuring that the projected screen gaze point coordinates are as close as possible to the true calibration anchor point coordinates. This solves the problem of static calibration failure and improves the accuracy of screen gaze point prediction.

[0136] In one embodiment, the forgetting factor of the recursive least squares algorithm is set to 1 to maintain accurate recursive absorption of the current calibrated support set; the anchor point residual is calculated after each round:

[0137]

[0138] in, The length of the screen diagonal; Indicates the first All rounds completed A session projection matrix supporting sample absorption;

[0139] When the decrease in residual at two adjacent anchor points satisfies: If the update is stopped early, the final session projection matrix will be obtained. This matrix is ​​used to calculate screen gaze points in all subsequent non-calibration stages; otherwise, continue to the next round of iteration refinement until the maximum number of iterations is reached; the number of iterations is set according to actual needs, for example, 3 to 5 rounds.

[0140] In one embodiment, in step S2.4, the quality indicators of the calibration segment include the percentage of valid calibration frames. , effective coverage of calibration anchor points Calibration residuals and calibration head posture stability index At least one of the following; the calculation formula is as follows:

[0141]

[0142]

[0143]

[0144]

[0145] in, The number of frames that are removed from the calibration segment if valid facial landmarks are detected and are not occluded, blink abnormally, or have insufficient confidence. This represents the total number of frames collected in the calibration segment. The number of calibration anchor points required to reach a preset minimum number of valid frame samples within the calibration segment; To determine the total number of anchor points; , indicating the first The predicted screen gaze coordinates are obtained from the final session projection matrix of the support samples; , and These represent the standard deviations of the head yaw angle, pitch angle, and roll angle for each valid frame sample within the calibration segment, respectively. Indicates the interocular distance scale factor of each valid frame sample within the calibration segment. The coefficient of variation.

[0146] If and only if the percentage of valid frames is calibrated Greater than the first threshold and the effective coverage of the calibrated anchor points Greater than the second threshold and the calibration residual Less than the third threshold and calibrated head posture stability index If the result is below the fourth threshold, the quality control result is considered passed; otherwise, it is considered failed and recalibration is required. If the number of recalibrations exceeds the maximum allowed number of recalibrations... If the calibration pass conditions are still not met, the data acquisition will be output as unreadable.

[0147] This embodiment constructs a logical rule for determining whether a result passes or fails by setting threshold values ​​for each indicator. This multi-dimensional quality control mechanism can effectively eliminate low-quality data caused by environmental interference or low subject cooperation, ensuring the reliability of subsequent screen fixation point estimation.

[0148] In one embodiment, step S2.5 applies the final session projection matrix to each valid facial keypoint sample within the non-calibration segment, i.e., the formal task segment. Calculate its screen gaze coordinates :

[0149]

[0150] in, This represents the extended features obtained from the effective facial keypoint samples in the t-th frame of the non-calibrated segment using the same method as in step S2.3, i.e. , , , The first The normalized keypoint set, local eye geometry vector, and head pose vector of the effective facial keypoint samples in the frame;

[0151] The predicted screen gaze coordinates are further rewritten as .

[0152] In one embodiment, in step S3, the resampling frequency is 60 Hz; resampling and smoothing are performed only within the same task segment to avoid signal leakage across task segments. Preferably, Savitzky-Golay smoothing, exponential moving average, or a combination of both are used within the segment to reduce high-frequency noise caused by browser sampling jitter.

[0153] In one embodiment, in step S4, the quality indicators of the formal task segment include the percentage of effective frames for key points in the formal task. Task coverage percentage of abnormal interruptions Formal mission head posture stability indicators and the proportion of abnormal coordinates At least one of them; the formula is as follows:

[0154]

[0155]

[0156]

[0157]

[0158]

[0159] in, This refers to the number of frames in the formal task segment that are detected as valid facial landmarks and are not occluded, have abnormal blinking, or are removed by rules of insufficient confidence. This represents the total number of frames acquired in the formal task segment. This represents the number of time grid points that passed the validity screening and were successfully resampled within the formal task segment. This represents the theoretical number of time grid points obtained based on the total duration of the formal task segment and the target resampling frequency. The duration of a given consecutive missing or invalid interval. To determine the minimum duration threshold for abnormal interruptions, For all Exceed The number of theoretical time grid points covered by the anomalous discontinuities; The number of time grid points for coordinates that are out of bounds, jump coordinates that exceed the speed limit, or screen gaze point coordinates with a confidence level below a threshold.

[0160] If and only if the percentage of valid frames at key points of the formal task is... Greater than the fifth threshold and task coverage The proportion of abnormally discontinuous events exceeding the sixth threshold Less than the seventh threshold and the head pose stability index for formal tasks Less than the eighth threshold and the proportion of abnormal coordinates If the result is less than the ninth threshold, the quality control result is considered passed; otherwise, it is considered failed and the test must be repeated in step S1.

[0161] In one embodiment, in step S5, frame-by-frame temporal features are constructed for each moment on a unified time axis as follows:

[0162]

[0163]

[0164]

[0165]

[0166] in, It represents basic spatiotemporal features such as task identifier, screen coordinates, and velocity; Indicates characteristics that indicate the type of eye movement event; This indicates the start and end positions and length encoding of the current eye-tracking event segment; To unify the timeline The task identifier of the task segment to which each time grid point belongs; , Represents the x and y coordinates of the screen gaze point; , and They represent the first The horizontal velocity, vertical velocity, and velocity magnitude at each time grid point , , , This is the resampling interval; and Indicates the first Screen gaze coordinates at each time grid point The coordinates relative to the normalized width and height of the screen display area; , and They represent the first The gaze, saccade, and smooth tracking event indicators for each time grid point are assigned values ​​based on the event reconstruction results; if the event belongs to the corresponding eye-tracking event, the value is 1, otherwise it is 0. , , and Let these represent the normalized start position, end position, normalized distance from the start position, and normalized distance from the end position of the current eye-tracking event segment, respectively. The relative positions of each time grid point within its respective eye-tracking event segment are calculated; and The sine and cosine period codes, representing the length of the current eye-tracking event segment, are obtained by substituting the duration of the current event segment or the length of the time grid into a preset period code function.

[0167] In one embodiment, event reconstruction can transform the original coordinate sequence into a sequence of eye-tracking events with semantic information, providing standardized input data for subsequent cognitive risk feature extraction. Event reconstruction is based on continuous screen gaze coordinates on a unified temporal grid. and its velocity modulus sequence The eye movement events are divided into fixation, saccades, and smooth tracking segments. This embodiment reconstructs eye movement events based on a velocity threshold determination rule. Specifically:

[0168] When the velocity magnitude of the continuous time grid interval Less than the gaze velocity threshold Furthermore, the spatial dispersion within this interval is less than the maximum permissible spatial dispersion threshold. Furthermore, the duration of this interval is not shorter than the shortest fixation duration. When this occurs, it is marked as a gaze event;

[0169] When the velocity magnitude of the continuous time grid interval Greater than the saccadic speed threshold Or the displacement of the gaze center between two adjacent gaze events is greater than the adjacent gaze center displacement threshold. Meanwhile, the duration of this interval falls within the allowed duration range of a saccadic event. If so, it is marked as a scanning event;

[0170] When the distance between the coordinates of the screen gaze point and the reference coordinates of the dynamic target trajectory is less than a preset threshold within a continuous time grid interval. Furthermore, the angle between the direction of gaze velocity and the direction of target motion is less than a preset threshold. And the duration is not less than the shortest duration of the smoothed tracking event. If it is, then it is marked as a smooth tracking event.

[0171] In one embodiment, for each formal task segment, discriminative task-level temporal features, i.e., task-level summary features, are extracted, as follows:

[0172] In the visual pairing comparison encoding segment and the delayed visual pairing comparison extraction segment, the system calculates the novelty preference ratio, the initial novelty fixation latency, the number of AOI switches, the number of revisits, and the left-right bias feature based on the familiar stimulus AOI and the novel stimulus AOI. Specifically, the initial novelty fixation latency is the time from the start of stimulus presentation to the first entry into the novel stimulus AOI and the formation of a fixation event; the number of AOI switches is the number of times the AOI to which adjacent fixation events belong changes; the number of revisits is the number of times a certain AOI is re-entered after leaving it; the left-right bias feature is the ratio of the difference in dwell time between the left and right AOIs to the total dwell time; the novelty preference ratio... Defined as:

[0173]

[0174] in, The total fixation duration of an AOI (Aspect of Interest) falling on a novel stimulus within the same paired trial; The total fixation duration when the familiar stimulus AOI falls within the same paired trial; To prevent stable terms with a denominator of zero.

[0175] In the demonstration and formal visual search sections guided by cues, the target AOI refers to the AOI surrounding the cued target object in the search scene; the four-corner candidate AOIs refer to candidate AOIs predefined in the four corner regions according to the search scene layout, where the region corresponding to the cued target is the target corner AOI, and the remaining regions are non-target corner AOIs. The target AOI, target corner AOI, and non-target corner AOIs are all determined by the target object number, screen position, and region boundaries in the stimulus metadata. The system calculates the target discovery latency, initial target fixation latency, path efficiency, erroneous corner dwell rate, corner revisit count, and fixation dispersion based on the target AOI and the four-corner candidate AOIs. Specifically, the target discovery latency is the time from the start of the search scene to the first entry of the gaze point into the target AOI; the first target gaze latency is the time from the start of the search scene to the formation of the first gaze event within the target AOI; the erroneous corner dwell ratio is the proportion of the gaze dwell time in non-target corner AOIs to the total gaze dwell time in the four candidate corner AOIs; the corner revisit count is the number of times a corner AOI is re-entered after leaving any corner AOI; gaze dispersion can be represented by the variance of the gaze centers or the average pairwise distance within the same trial; the path efficiency... Defined as:

[0176]

[0177] in, This indicates the shortest theoretical path length to the target, which can be taken as the straight-line distance from the starting gaze point to the center of the target AOI. This represents the actual scanning path length, which can be the sum of the distances between adjacent gaze centers within the search scene; To prevent stable terms with a denominator of zero.

[0178] In the free-viewing segment, features such as scan path length, gaze dispersion, region of interest coverage, center bias, transition entropy, and scene exploration uniformity are extracted. Specifically, scan path length is the sum of distances between adjacent gaze centers; gaze dispersion is the variance or average distance of a gaze center relative to its mean center; the preset AOIs in the free-viewing segment refer to the scene semantic regions pre-annotated in the stimulus metadata, or regions obtained by dividing the display area according to rules such as fixed grids or center-periphery partitioning; therefore, the total number of preset AOIs is the number of accessible regions of interest defined in the metadata for this free-viewing stimulus. Region of interest coverage is the ratio of the number of accessed AOIs to the total number of preset AOIs corresponding to this free-viewing stimulus; center bias is the ratio of gaze dwell time in the central region to the total gaze dwell time; transition entropy is the entropy calculated based on the AOI transition probability matrix; and scene exploration uniformity is the normalized entropy of the dwell time distribution of each AOI.

[0179] In the VisMET segment of the visual motion exploration task, features such as follow continuity, velocity consistency, and phase shift to motion stimuli were extracted. Follow continuity is defined as the percentage of time the gaze point remains within the neighborhood of the moving target; velocity consistency is the stability of the correlation or ratio between the gaze velocity and the target velocity; and phase shift is the time lag when the cross-correlation between the gaze trajectory and the target trajectory is at its maximum.

[0180] In one embodiment, the risk reasoning model is a hierarchical temporal model, including a frame-by-frame feature mapping layer, a task type embedding layer, a temporal coding layer, a task-level attention pooling layer, and a subject-level fusion layer.

[0181] frame-by-frame temporal features By task identifier The feature sequence is divided into corresponding task segments, the first segment... Within the first task segment The features of each time grid point are denoted as follows: ; In addition to In addition to the basic spatiotemporal features, eye-movement event type indication features, and eye-movement event segment location encoding, a total of 16 features are denoted as follows: The frame-by-frame feature mapping layer uses As input, it passes through a linear or multi-layer perceptron. After processing, a frame-by-frame feature map representation is obtained. :

[0182]

[0183] The task type embedding layer is based on In Generate task embedding and will With the frame-by-frame feature mapping representation By adding or concatenating the elements and then performing linear projection, a frame-level hiding representation is obtained. :

[0184]

[0185] or

[0186] The temporal coding layer uses frame-level hidden representation sequences within the same task stimulus fragment. As input, a Transformer encoder, a gated recurrent network, or an equivalent temporal encoder is used to model the temporal dependencies within the same task segment, gaze-saccade transformation, and cross-event context. The output is a frame-level context representation. :

[0187]

[0188] in, Indicates the first The effective time grid length after resampling of each task segment Indicates the first time before timing coding Within the first task segment Hidden representation of each time grid point Represents the first time sequence after coding Within the first task segment Context representation of each time grid point.

[0189] The task-level attention pooling layer is based on the first... Frame-level context representation of each task segment Using this as input, calculate the weight of the risk evidence for that task segment at each time grid point. And aggregate to obtain the task segment time sequence representation. :

[0190]

[0191]

[0192] in, This represents a learnable query vector, a task-type conditional query, or an equivalent query representation generated by a gated attention network. Represents the learnable weight matrix. Normalization is performed within the same task segment; thus, the model can highlight more discriminative time segments within the same task segment, such as the initial target fixation latency, long-term error angle dwell time, and abnormal free-view concentration.

[0193] The subject-level fusion layer receives data according to a fixed task sequence. ,in , They represent the first Temporal representation of each task segment, task-level summary features, and missing mask. This is used to mark whether the task segment is missing, whether it is deemed invalid by quality control rules, and whether the corresponding features are available. The subject-level fusion layer can use cross-task attention, gated weighted averaging, or multilayer perceptron for fusion to obtain the subject-level representation. and output the original risk score. and original risk probability :

[0194]

[0195]

[0196]

[0197] The original risk probability The calibrated risk probability is further obtained through parameterized probability calibration. Preferably, logic calibration or beta calibration is used:

[0198] Logic calibration:

[0199] beta calibration:

[0200] Among them, logic calibration , and in beta calibration , , All of these are calibration parameters fitted on the validation set.

[0201] Furthermore, the calibrated risk probability With demographic vector By merging the results, the final screening probability is obtained. :

[0202]

[0203] Demographic integration:

[0204] in, , , These are standardized age, gender, and education level codes, respectively. ~ These are demographic parameters.

[0205] Final screening probability With configurable referral thresholds When comparing, If the risk is high, then there is cognitive risk; otherwise, it is low risk.

[0206] Finally, it should be noted that the above embodiments are intended to illustrate the technical solutions of the present invention and do not constitute any limitation on the present invention. Those skilled in the art should fully understand that modifications to the technical solutions described in the foregoing embodiments or equivalent substitutions for any part or all of the technical features are entirely feasible. Such modifications or substitutions, as long as they do not depart from the scope of protection defined by the claims of the present invention, should be considered reasonable extensions of the present invention.

Claims

1. A cognitive risk detection method based on browser camera eye-tracking estimation and quality control, characterized in that, include: S1. The browser loads the phased visual cognition screening task chain and calls the browser's front-facing camera to capture the subject's facial video stream during the execution of the phased visual cognition screening task chain; the phased visual cognition screening task chain includes various visual cognition task tests executed sequentially in the calibration segment and the formal task segment. S2. Obtain the screen gaze coordinates of the formal task segment based on the facial video stream; S3. Resample the coordinates of each screen gaze point to a unified time axis; S4. Calculate the quality indicators of the formal task segment and determine the quality control results of the formal task segment. S5. When the quality control result of the formal task segment is passed, construct frame-by-frame temporal features containing basic spatiotemporal features and event features based on the resampled screen gaze point coordinates, and extract task-level summary features for each formal task segment. S6. Input the frame-by-frame temporal features and the task-level aggregated features into the risk inference model to obtain the cognitive risk probability, and output the final cognitive risk screening result based on the cognitive risk probability.

2. The cognitive risk detection method based on browser camera eye-tracking estimation and quality control according to claim 1, characterized in that, In step S2, the screen gaze coordinates of the formal task segment are obtained based on the following method: S2.1 The browser extracts the coordinates of facial key points frame by frame from the facial video stream. The set of facial key point coordinates detected in a single frame forms a facial key point sample. The task identifier of the task segment to which each facial key point sample belongs is recorded. Several facial key point samples sorted by timestamp are obtained and uploaded to the processing end. S2.2 The processing end extracts facial key point samples of the calibration segment based on the task identifier, and constructs a calibration support set based on the facial key point samples of the calibration segment; S2.

3. Solve the initial session projection matrix based on the calibration support set, and iteratively refine the initial session projection matrix to obtain the final session projection matrix; S2.

4. Based on the quality indicators of the calibration support set, determine the quality control results of the calibration segment; S2.5 In response to the quality control result being passed, for any facial key point sample within the formal task segment, the final session projection matrix is ​​used to perform mapping calculations to obtain the screen gaze point coordinates corresponding to the facial key point sample.

3. The cognitive risk detection method based on browser camera eye-tracking estimation and quality control according to claim 2, characterized in that, In step S2.2, the calibration support set is constructed based on the following method: Identify the time window of each calibration anchor point in the calibration segment, that is, the time period presented by each calibration anchor point; The facial key point samples of the calibration segment are divided into the corresponding anchor point time windows based on the timestamp; Within each anchor point time window, invalid frames are removed, stable head poses are screened, and blinking frames are filtered for facial key point samples to obtain valid samples. The valid samples within each anchor time window are aggregated into a calibration support set, which is represented as follows: in, To support the sample size; Indicates the calibration support set within the first A set of facial key point coordinates for each supporting sample; Indicates the first The anchor point's real screen reference coordinates corresponding to the anchor point's time window where each supporting sample is located; This represents the confidence weight.

4. The cognitive risk detection method based on browser camera eye-tracking estimation and quality control according to claim 3, characterized in that, In step S2.3, based on the calibrated support set, the initial session projection matrix is ​​solved using regularized least squares with projection prior constraints: in, This is the initial session projection matrix. The projective prior obtained from offline training; The regularization coefficient is used. Represents the square of the Frobenius norm; Indicates the calibration support set within the first Extended features of each supporting sample ; The output of the shared geometric representation network represents the first... Geometric descriptive features of each supporting sample , Represents a shared geometric representation network. Indicates the first A normalized set of key points is obtained by performing translation and scale normalization on the coordinates of all facial key points within a support sample; This is a local eye geometry vector constructed based on normalized keypoints; This represents the head pose vector.

5. The cognitive risk detection method based on browser camera eye-tracking estimation and quality control according to claim 4, characterized in that, The final session projection matrix is ​​obtained based on the following method: Using the initial session projection matrix as the initial value, the calibration support set is iteratively refined using a recursive least squares algorithm: Let the first Round The inverse covariance matrix of each supporting sample before update is: Then the recursive update formula is: in, For the first Round absorption Gain vector for each support sample; and The first Round Each session projection matrix supports both the pre- and post-update samples; For the first Round The two-dimensional reprojection residuals of each supporting sample, i.e., the real screen reference coordinates. Compared with current forecasts The difference between them; and The first Round Each support sample inverse covariance matrix before and after update; The forgetting factor of the recursive least squares algorithm is set to 1, and the anchor point residual is calculated after each round: in, The length of the screen diagonal; Indicates the first All rounds completed A session projection matrix supporting sample absorption; If the decrease in the residual of the anchor point between two adjacent rounds is less than a preset threshold, the update is stopped in advance, and the final session projection matrix is ​​obtained. Otherwise, continue with the next round of iterations to refine until the maximum number of iterations is reached.

6. The cognitive risk detection method based on browser camera eye-tracking estimation and quality control according to claim 5, characterized in that, The quality metrics for the calibration support set include the percentage of valid calibration frames. , effective coverage of calibration anchor points Calibration residuals and calibration head posture stability index At least one of them; if and only if Greater than the first threshold and Greater than the second threshold and Less than the third threshold and If the result is less than the fourth threshold, the quality control result is considered passed; otherwise, it is considered failed.

7. The cognitive risk detection method based on browser camera eye-tracking estimation and quality control according to claim 1, characterized in that, In step S4, the quality indicators of the formal task segment include the percentage of effective frames for key points in the formal task. Task coverage percentage of abnormal interruptions Formal mission head posture stability indicators and the proportion of abnormal coordinates At least one of them; if and only if Greater than the fifth threshold and Greater than the sixth threshold and Less than the seventh threshold and Less than the eighth threshold and If the result is less than the ninth threshold, the quality control result is considered passed; otherwise, it is considered failed.

8. The cognitive risk detection method based on browser camera eye-tracking estimation and quality control according to claim 5, characterized in that, The frame-by-frame temporal features for each moment on a unified timeline are constructed as follows: in, Represents the basic spatiotemporal characteristics; Indicates characteristics that indicate the type of eye movement event; This indicates the start and end positions and length encoding of the current eye-tracking event segment; To unify the timeline The task identifier of the task segment to which each time grid point belongs; , Represents the x and y coordinates of the screen gaze point; , and They represent the first The horizontal velocity, vertical velocity, and velocity magnitude at each time grid point , , , This is the resampling interval; and Indicates the first Screen gaze coordinates at each time grid point The coordinates relative to the normalized width and height of the screen display area; , and They represent the first The gaze, saccade, and smoothing tracking event indicators for each time grid point are obtained based on the continuous screen gaze point coordinates and velocity magnitude sequence on a unified time grid. If it belongs to the corresponding eye movement event, it is set to 1; otherwise, it is set to 0. , , and These represent the normalized start position, end position, normalized position from the start position, and normalized position from the end position of the current eye-tracking event segment, respectively. and These represent the sine and cosine period codes for the length of the current eye-tracking event segment, respectively.

9. The cognitive risk detection method based on browser camera eye-tracking estimation and quality control according to claim 8, characterized in that, The risk reasoning model is a hierarchical temporal model, including a frame-by-frame feature mapping layer, a task type embedding layer, a temporal coding layer, a task-level attention pooling layer, and a subject-level fusion layer. frame-by-frame temporal features By task identifier The feature sequence is divided into corresponding task segments, the first segment... Within the first task segment The features of each time grid point are denoted as follows: ; In addition to In addition to the basic spatiotemporal features, eye-movement event type indication features, and eye-movement event segment location coding features, let them be denoted as The frame-by-frame feature mapping layer uses The input is processed by a linear or multi-layer perceptron to obtain a frame-by-frame feature map representation. ; The task type embedding layer is based on In Generate task embedding and will With the frame-by-frame feature mapping representation By adding or concatenating the elements and then performing linear projection, a frame-level hiding representation is obtained. ; The temporal coding layer uses frame-level hidden representation sequences within the same task stimulus fragment. As input, a temporal encoder is used to model temporal dependencies within the same task segment, gaze-saccade transitions, and cross-event contexts, outputting a frame-level context representation. ,in Indicates the first The effective time grid length after resampling of each task segment; The task-level attention pooling layer is based on the first... Frame-level context representation of each task segment Using this as input, calculate the weight of the risk evidence for that task segment at each time grid point. And aggregate to obtain the task segment time sequence representation. : in, This represents a learnable query vector, a task-type conditional query, or an equivalent query representation generated by a gated attention network. Represents the learnable weight matrix; The subject-level fusion layer receives data according to a fixed task sequence. ,in , They represent the first The temporal representation of each task segment, task-level aggregated features, and missing masks are obtained. The subject-level fusion layer uses cross-task attention, gated weighted averaging, or multilayer perceptron for fusion to obtain the subject-level representation. and output the original risk score. and original risk probability ; The original risk probability The calibrated risk probability is obtained through parameterized probability calibration. ; calibrated risk probability With demographic vector By merging the results, the final screening probability is obtained. : in, , , These are standardized age, gender, and education level codes, respectively. ~ For demographic parameters; Final screening probability With configurable referral thresholds When comparing, If the risk is high, then there is cognitive risk; otherwise, it is low risk.

10. The cognitive risk detection method based on browser camera eye-tracking estimation and quality control according to claim 1, characterized in that, The formal task segment includes the following segments executed sequentially: visual pairing comparison encoding segment, cue-guided visual search demonstration segment, formal cue-guided visual search segment, delayed visual pairing comparison extraction segment, free viewing segment, and visual motion exploration task VisMET segment. The task-level summary features of each formal task segment are discriminative task-level temporal features.

Images