AI-based eye-tracking-based methods and systems for preventing cheating in online examinations

Abstract

This invention provides an online examination anti-cheating method and system based on AI-fused gaze tracking, belonging to the field of online education security technology. It includes extracting the candidate's head movement feature sequence, calculating the coordinates and duration of gaze points, analyzing gaze behavior in functional areas, identifying periodic jump patterns in gaze migration feature maps, and combining autoregressive spectral analysis and wavelet analysis to identify gaze pattern abrupt change points. This achieves accurate identification of cheating behavior, effectively improving the security and fairness of online examinations and reducing the false positive rate.

Description

Technical Field

[0001] This invention relates to the field of online education security technology, and in particular to an online examination anti-cheating method and system based on AI-integrated eye tracking. Background Technology

[0002] With the rapid development of online education, online examinations have become an important method of educational assessment. However, cheating in online examinations has also increased, posing a serious challenge to educational fairness and examination quality. Traditional online examination anti-cheating systems mainly rely on methods such as camera monitoring, screen locking, and behavior recording, but these methods have many limitations in practical applications.

[0003] Currently, online exam anti-cheating technologies mainly include camera-based facial recognition and behavior monitoring, screen sharing and locking technologies, and AI-based abnormal behavior detection. These technologies attempt to detect cheating by monitoring candidates' behavioral patterns. With the development of artificial intelligence technology, combining AI with eye-tracking technology provides a new technological approach for preventing cheating in online exams.

[0004] Existing online exam anti-cheating technologies have the following defects and shortcomings:

[0005] Traditional surveillance systems struggle to effectively detect sophisticated cheating behaviors, especially those involving subtle eye movements and gaze shifts. Test takers can keep their heads still and only use their eyes to view illegal information, posing a challenge to conventional camera surveillance systems.

[0006] Current technologies lack the ability to deeply analyze test takers' gaze behavior and cannot accurately distinguish between normal reading and browsing behavior and abnormal gaze behavior that suggests searching for external answers. This is particularly true in multi-functional exam interfaces, where gaze behavior in different areas carries different meanings, and existing systems struggle to perform effective weighting analysis.

[0007] There is a lack of precise modeling and analysis of the temporal characteristics of gaze behavior. Cheating behavior often manifests as abnormal gaze migration patterns and periodic gaze behaviors, which require advanced time series analysis methods for identification. Most existing systems rely on simple threshold judgments, which cannot capture these complex spatiotemporal pattern features, leading to high false positive rates or severe false negatives. Summary of the Invention

[0008] This invention provides an online examination anti-cheating method and system based on AI-fused eye-tracking, which can solve the problems in the prior art.

[0009] A first aspect of this invention provides an online examination anti-cheating method based on AI-fused eye-tracking, comprising:

[0010] The real-time movement trajectory of the candidate's head in three-dimensional space is extracted by the camera of the candidate's terminal device to establish a head movement feature sequence;

[0011] The screen display content of the candidate's terminal device is divided into multiple functional areas, and a corresponding gaze weight coefficient is set for each functional area. When the actual gaze duration or gaze frequency of a certain functional area exceeds the area switching frequency threshold, the corresponding gaze action is marked as abnormal gaze.

[0012] By calculating the real-time coordinates of the examinee's gaze point on the screen and the duration of gaze through eye tracking, and combining the head movement feature sequence and the gaze weight coefficient, gaze behavior feature data is formed.

[0013] Extract the transition sequence of the examinee's gaze point between different functional areas, calculate the time interval and spatial distance of the gaze point transition, and generate a gaze migration feature map; when a periodic jumping pattern is detected in the gaze migration feature map, and the time interval of the gaze point transition shows abnormal periodicity, and the spatial jump speed of the gaze point exceeds the gaze point jump speed threshold, the corresponding time period is marked as a suspicious time period.

[0014] Extract the gaze behavior feature data within the suspicious time period, use autoregressive spectral analysis to extract the spectral energy distribution features of the gaze point trajectory, and combine wavelet analysis to identify gaze pattern abrupt change points. When the spectral energy distribution features are abnormal and there are gaze pattern abrupt change points, accompanied by abnormal gaze, it is determined to be cheating behavior.

[0015] By calculating the real-time coordinates of the examinee's gaze point on the screen and the duration of gaze through eye tracking, and combining this with the head movement feature sequence and the gaze weighting coefficient, gaze behavior feature data is formed, including:

[0016] The left and right eye gaze vectors of the examinee are obtained through an eye-tracking device, and the coordinates of the initial gaze point are determined based on the intersection of the left and right eye gaze vectors with the screen plane.

[0017] Based on the head motion feature sequence, the head yaw angle rotation matrix, head pitch angle rotation matrix, and head roll angle rotation matrix are calculated and multiplied to obtain the head motion compensation matrix. The initial gaze point coordinates are then subjected to matrix operation with the head motion compensation matrix to obtain the compensated gaze point coordinates. The current compensated gaze point coordinates are then compared with the previous compensated gaze point coordinates to obtain the gaze point movement speed.

[0018] The spatial dispersion is calculated by continuously collecting the compensated gaze point coordinates for a preset number of frames. When the spatial dispersion is less than the gaze determination threshold, the gaze start time and end time are recorded to generate the gaze duration.

[0019] The compensated gaze point coordinates, the gaze point movement speed, the gaze duration, and the gaze weight coefficient are combined to form a gaze behavior feature vector.

[0020] Extract the gaze transfer sequence between different functional areas, calculate the time interval and spatial distance of gaze transfer, and generate a gaze transfer feature map including:

[0021] Establish a region index matrix using multiple functional areas on the screen of the candidate's terminal device;

[0022] Within a preset time window, density peak clustering is performed on the compensated gaze point coordinates to obtain the clustered gaze point coordinates. The clustered gaze point coordinates are then matched with the region index matrix using distance weights to obtain the functional region identifier corresponding to each gaze point coordinate. The functional region identifiers are recorded in chronological order to generate a gaze region transfer sequence.

[0023] The time points at which two adjacent functional region identifiers differ in the gaze region transfer sequence are statistically analyzed, and the corresponding gaze point information is marked as a migration event. The starting functional region identifier and the target functional region identifier of the migration event are extracted.

[0024] The spatial jump distance is calculated based on the starting gaze point coordinates and the target gaze point coordinates of the migration event. The migration time interval is calculated based on the start time and end time of the migration event. The starting functional area identifier, the target functional area identifier, the spatial jump distance, and the migration time interval are combined to form a migration feature vector. Statistical features are calculated on the migration feature vector to generate a gaze migration feature map.

[0025] Within a preset time window, density peak clustering is performed on the compensated gaze point coordinates to obtain the clustered gaze point coordinates. The clustered gaze point coordinates are then matched with a region index matrix using distance weighting to obtain the functional region identifier corresponding to each gaze point coordinate, including:

[0026] Obtain the compensated gaze point coordinate sequence within a preset time window, calculate the Euclidean distance between all gaze point pairs in the gaze point coordinate sequence, and calculate the local density value of each gaze point based on the Euclidean distance;

[0027] Decision values ​​are calculated based on the local density value of each gaze point and the minimum Euclidean distance between it and other gaze points. A preset number of gaze points with the largest decision values ​​are selected as density peak points.

[0028] A fixation point connectivity graph is established, connecting each non-density peak point to its nearest spatial fixation point. The weight of each connection is calculated based on the local density value and the Euclidean distance. Based on the connectivity graph, starting from each non-density peak point, the graph moves along the connection with the largest weight until a density peak point is reached. The non-density peak point is then assigned to the corresponding cluster of the density peak point. Using the local density value of the fixation point as the weight, a weighted average of the coordinates of all fixation points in each cluster is calculated to obtain the fixation point coordinates after density peak clustering.

[0029] Calculate the vertical distance from each density peak point to the functional region boundary in the gaze coordinates after the density peak clustering. Divide the vertical distance by the smoothing coefficient and take the negative exponent to obtain the distance weight value of each functional region corresponding to the density peak point. Based on the maximum distance weight value, determine the functional region identifier to which all gaze points in the cluster to which the density peak point belongs.

[0030] When a periodic jumping pattern is detected in the gaze migration feature map, and the time interval of gaze point shifts exhibits abnormal periodicity, while the gaze point spatial jump speed exceeds the gaze point jump speed threshold, the corresponding time period is marked as a suspicious time period, including:

[0031] The jump pattern feature sequence is extracted from the gaze migration feature map, and the autocorrelation coefficient of the jump pattern feature sequence under different time delays is calculated. When the maximum autocorrelation coefficient is greater than the jump periodicity determination threshold, the current time period is marked as a periodic jump period.

[0032] For the periodic jump period, the time interval sequence of adjacent fixation points is extracted and subjected to fast Fourier transform to obtain spectral data. The ratio of the energy value of the main frequency component to the total energy value in the spectral data is calculated to obtain the main frequency energy proportion. The ratio of the standard deviation to the mean of the time interval sequence is calculated to obtain the coefficient of variation. When the main frequency energy proportion is greater than the periodic intensity threshold and the coefficient of variation is less than the eye movement fluctuation threshold, the current period is marked as an abnormal periodic period.

[0033] For the abnormal periodic time period, the ratio of the Euclidean distance between adjacent fixation points to the corresponding time interval is calculated to obtain the fixation point spatial jump speed. When the fixation point spatial jump speed is greater than the preset fixation point jump speed threshold, the current fixation point is marked as a suspicious fixation point.

[0034] The number and distribution of suspicious gaze points within the abnormal period are counted. When the number of suspicious gaze points is greater than the density threshold and the distribution interval is less than the aggregation threshold, the current abnormal period is determined as a suspicious period.

[0035] Extract the gaze behavior feature data within the suspected time period, use autoregressive spectral analysis to extract the spectral energy distribution features of the gaze point trajectory, and combine wavelet analysis to identify gaze pattern abrupt changes. When the spectral energy distribution features are abnormal and there are gaze pattern abrupt changes, accompanied by abnormal gaze, it is determined to be cheating behavior, including:

[0036] Extract spatial coordinate sequences of gaze points from gaze behavior feature data within suspicious time periods;

[0037] Autoregressive spectral analysis is performed on the spatial coordinate sequence to obtain power spectral density data. The power spectral density data is divided into multiple frequency bands according to the frequency magnitude, and the ratio of the power spectral density value of each frequency band to the total power spectral density value is calculated to obtain the spectral energy distribution characteristics.

[0038] A continuous wavelet transform is performed on the spatial coordinate sequence to obtain multi-scale wavelet coefficients. Empirical mode decomposition is performed on the multi-scale wavelet coefficients to obtain multiple intrinsic mode functions. An analytical signal is constructed for each intrinsic mode function. Instantaneous amplitude and instantaneous phase are calculated based on the analytical signal. The gaze pattern abruptness point is marked based on the rate of change of instantaneous amplitude and instantaneous phase at adjacent time points.

[0039] When the component features of the spectral energy distribution deviate from the feature subspace of the normal gaze pattern, and there is a mutation point in the gaze pattern, and the current gaze action is marked as abnormal gaze, the behavior in the current suspicious period is judged as cheating behavior.

[0040] A second aspect of the present invention provides an online examination anti-cheating system based on AI-fused eye-tracking, comprising:

[0041] The first unit is used to extract the real-time motion trajectory of the candidate's head in three-dimensional space through the camera of the candidate's terminal device to establish a head motion feature sequence;

[0042] The second unit is used to divide the screen display content of the candidate's terminal device into multiple functional areas and set a corresponding gaze weight coefficient for each functional area; when the actual gaze duration or gaze frequency of a certain functional area exceeds the area switching frequency threshold, the corresponding gaze action is marked as abnormal gaze.

[0043] The third unit is used to calculate the real-time coordinate position and duration of the examinee's gaze on the screen through eye tracking, and to form gaze behavior feature data by combining the head movement feature sequence and the gaze weight coefficient.

[0044] The fourth unit is used to extract the transfer sequence of the examinee's gaze point between different functional areas, calculate the time interval and spatial distance of the gaze point transfer, and generate a gaze migration feature map; when a periodic jumping pattern is detected in the gaze migration feature map, and the time interval of the gaze point transfer shows abnormal periodicity, and the spatial jump speed of the gaze point exceeds the gaze point jump speed threshold, the corresponding time period is marked as a suspicious time period.

[0045] The fifth unit is used to extract the gaze behavior feature data within the suspicious time period, extract the spectral energy distribution features of the gaze point trajectory using autoregressive spectral analysis, and identify gaze pattern abrupt change points by combining wavelet analysis. When the spectral energy distribution features are abnormal and there are gaze pattern abrupt change points, accompanied by abnormal gaze, it is determined to be cheating behavior.

[0046] A third aspect of the embodiments of the present invention,

[0047] An electronic device is provided, comprising:

[0048] processor;

[0049] Memory used to store processor-executable instructions;

[0050] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0051] Fourth aspect of the embodiments of the present invention,

[0052] A computer-readable storage medium is provided, having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0053] The beneficial effects of this application are as follows:

[0054] This invention provides an online examination anti-cheating method based on AI-fused gaze tracking. By extracting the examinee's head movement trajectory and gaze characteristics, and combining functional area weights and gaze migration map analysis, it accurately identifies abnormal gaze behavior and suspicious time periods, effectively solving the problem that traditional anti-cheating technologies cannot accurately locate cheating behavior and improving the accuracy of cheating behavior detection.

[0055] This invention introduces autoregressive spectral analysis and wavelet analysis techniques to perform in-depth analysis of gaze behavior feature data during suspicious time periods. Through dual verification of spectral energy distribution characteristics and gaze pattern mutation points, it significantly reduces the false judgment rate, avoids incorrect labeling of normal candidates, and enhances the fairness and reliability of the anti-cheating system.

[0056] This invention analyzes the sequence of gaze shifts between different functional areas, and calculates the time interval and spatial distance of gaze shifts. It can effectively identify periodic jumping patterns and abnormal gaze behaviors, and has a good ability to identify advanced cheating methods such as transmitting information through gaze. This significantly improves the integrity and fairness of the online examination environment. Attached Figure Description

[0057] Figure 1 This is a flowchart illustrating the online examination anti-cheating method based on AI-fused eye tracking, as described in an embodiment of the present invention.

[0058] Figure 2 A schematic diagram illustrating the comparative analysis of head movement compensation effects;

[0059] Figure 3 This is a schematic diagram illustrating the comparative analysis of spectral energy distribution. Detailed Implementation

[0060] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0061] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0062] Figure 1 This is a flowchart illustrating an online examination anti-cheating method based on AI-fused eye-tracking, as described in an embodiment of the present invention. Figure 1 As shown, the method includes:

[0063] The real-time movement trajectory of the candidate's head in three-dimensional space is extracted by the camera of the candidate's terminal device to establish a head movement feature sequence;

[0064] The screen display content of the candidate's terminal device is divided into multiple functional areas, and a corresponding gaze weight coefficient is set for each functional area. When the actual gaze duration or gaze frequency of a certain functional area exceeds the area switching frequency threshold, the corresponding gaze action is marked as abnormal gaze.

[0065] By calculating the real-time coordinates of the examinee's gaze point on the screen and the duration of gaze through eye tracking, and combining the head movement feature sequence and the gaze weight coefficient, gaze behavior feature data is formed.

[0066] Extract the transition sequence of the examinee's gaze point between different functional areas, calculate the time interval and spatial distance of the gaze point transition, and generate a gaze migration feature map; when a periodic jumping pattern is detected in the gaze migration feature map, and the time interval of the gaze point transition shows abnormal periodicity, and the spatial jump speed of the gaze point exceeds the gaze point jump speed threshold, the corresponding time period is marked as a suspicious time period.

[0067] Extract the gaze behavior feature data within the suspicious time period, use autoregressive spectral analysis to extract the spectral energy distribution features of the gaze point trajectory, and combine wavelet analysis to identify gaze pattern abrupt change points. When the spectral energy distribution features are abnormal and there are gaze pattern abrupt change points, accompanied by abnormal gaze, it is determined to be cheating behavior.

[0068] In one optional implementation, eye tracking is used to calculate the real-time coordinates of the examinee's gaze point on the screen and the duration of gaze. This, combined with the head movement feature sequence and the gaze weighting coefficient, forms gaze behavior feature data, including:

[0069] The left and right eye gaze vectors of the examinee are obtained through an eye-tracking device, and the coordinates of the initial gaze point are determined based on the intersection of the left and right eye gaze vectors with the screen plane.

[0070] Based on the head motion feature sequence, the head yaw angle rotation matrix, head pitch angle rotation matrix, and head roll angle rotation matrix are calculated and multiplied to obtain the head motion compensation matrix. The initial gaze point coordinates are then subjected to matrix operation with the head motion compensation matrix to obtain the compensated gaze point coordinates. The current compensated gaze point coordinates are then compared with the previous compensated gaze point coordinates to obtain the gaze point movement speed.

[0071] The spatial dispersion is calculated by continuously collecting the compensated gaze point coordinates for a preset number of frames. When the spatial dispersion is less than the gaze determination threshold, the gaze start time and end time are recorded to generate the gaze duration.

[0072] The compensated gaze point coordinates, the gaze point movement speed, the gaze duration, and the gaze weight coefficient are combined to form a gaze behavior feature vector.

[0073] This method acquires the left and right eye gaze vectors of the examinee using a gaze-tracking device. This device can be a camera-based eye tracker, which works by capturing the relationship between the corneal reflected light point and the pupil center position to calculate the gaze direction. For example, when the examinee's left eye gaze vector is captured as (0.3, 0.2, -1.0) and the right eye gaze vector as (0.25, 0.22, -1.0), the initial gaze point coordinates are determined based on the intersection of these two vectors with the screen plane. Assuming the screen plane equation is z = -50, the initial gaze point coordinates can be obtained as (15.0, 10.0, -50.0) by solving for the intersection of the gaze vectors with the plane.

[0074] The head motion feature sequence includes time-series data of head yaw, pitch, and roll angles. For example, when a head yaw angle of 5 degrees, a pitch angle of 3 degrees, and a roll angle of 1 degree are detected, the rotation matrices corresponding to these three angles are calculated respectively. The head yaw rotation matrix represents the effect of left-right head rotation, the head pitch rotation matrix represents the effect of head nodding up and down, and the head roll rotation matrix represents the effect of head tilting left and right. Multiplying these three rotation matrices yields the head motion compensation matrix. In practical applications, it is assumed that the calculated head motion compensation matrix is ​​a 3×3 matrix with values ​​of [[0.996, 0.087, 0.005], [-0.087, 0.995, 0.052], [-0.001, -0.053, 0.998]].

[0075] The initial gaze point coordinates are represented as a column vector and multiplied by the head motion compensation matrix to obtain the compensated coordinate values. For example, multiplying the initial coordinates (15.0, 10.0, -50.0) by the compensation matrix yields the compensated gaze point coordinates as (16.4, 9.5, -50.2). This compensation mechanism effectively eliminates the impact of slight head movements on gaze tracking accuracy.

[0076] To calculate the fixation point movement velocity, the compensated fixation point coordinates at the current moment are differencing those at the previous moment. Assuming the compensated fixation point coordinates at the previous moment are (16.2, 9.6, -50.2) and the current moment is (16.4, 9.5, -50.2), then the fixation point movement velocity is (0.2, -0.1, 0.0), indicating a movement of 0.2 units in the x-direction, -0.1 units in the y-direction, and no movement in the z-direction.

[0077] Typically, the preset frame rate is set to 10-30 frames, and the sampling frequency is set to 60Hz or 90Hz. Spatial dispersion is calculated by continuously acquiring compensated gaze coordinates. The spatial dispersion is calculated by taking the standard deviation of the gaze coordinates across these frames. For example, if 20 frames of data are acquired continuously, and the standard deviation of the x-coordinate is calculated to be 0.8, and the standard deviation of the y-coordinate is 0.6, then the spatial dispersion can be expressed as the square root of the sum of the squares of 0.8 and 0.6, which is 1.0. When the spatial dispersion is less than the gaze determination threshold, the examinee is considered to be looking at a point. The gaze determination threshold is usually set to 1.5-2.0 degrees of visual field, approximately equal to 15-20 pixels on the screen. When the spatial dispersion is detected to be below the threshold of 1.5 for 20 consecutive frames, the gaze start time is recorded, for example, frame 100; when the spatial dispersion exceeds the threshold again, the gaze end time is recorded, for example, frame 150. Using a 60Hz sampling rate, the gaze duration is (150-100) / 60 = 0.833 seconds.

[0078] The compensated fixation point coordinates, fixation point movement velocity, fixation duration, and fixation weight coefficient are used to construct a fixation behavior feature vector. The fixation weight coefficient reflects the importance of fixation in different regions and is usually preset according to the regions defined by the exam content. For example, the weight of the question stem region is 0.4, the weight of the option region is 0.3, the weight of the chart region is 0.2, and the weight of other regions is 0.1. The final fixation behavior feature vector can be represented as: [(16.4, 9.5, -50.2), (0.2, -0.1,0.0), 0.833, 0.3], corresponding to the compensated fixation point coordinates, fixation point movement velocity, fixation duration, and fixation weight coefficient, respectively.

[0079] These gaze behavior feature vectors can be used for subsequent analysis of test takers' reading patterns, cognitive load, and problem-solving strategies, providing data support for exam cheating prevention systems and learning behavior analysis. For example, by analyzing the distribution and dwell time of gaze points in different areas, it is possible to determine whether test takers exhibit suspicious behaviors such as frequent looking around or communicating; by comparing the matching degree between test takers' gaze sequences and standard problem-solving paths, it is possible to assess test takers' problem-solving efficiency and thinking patterns.

[0080] Figure 2This diagram illustrates the comparative analysis of head motion compensation effects. The results demonstrate the significant advantages of this invention compared to traditional methods, verified through 1000 frames of continuous test data. As shown in the diagram, the Pupil Labs method exhibits significant fluctuations in gaze point coordinate deviation, ranging from 2.1 to 22.1 pixels, with an average deviation of approximately 13.8 pixels, exhibiting strong instability. Particularly noticeable are large deviations exceeding 20 pixels around frames 350, 500, and 900. While the Tobii Pro method shows improvement over Pupil Labs, the deviation still fluctuates between 1.8 and 16.8 pixels, with an average deviation of approximately 10.2 pixels, indicating significant tracking errors even with large head movements. In contrast, this invention, through its innovative three-axis rotation matrix head motion compensation algorithm, controls the gaze point coordinate deviation within an extremely small range of 0.8 to 3.6 pixels, with an average deviation of only 2.3 pixels, representing an 83.3% reduction compared to traditional methods and a 77.5% reduction compared to the Tobii Pro method. The compensation effect curve of this invention exhibits excellent stability and consistency, maintaining high-precision tracking even with large head movements. This fully demonstrates the technical superiority of the three-dimensional rotation compensation mechanism based on yaw angle, pitch angle, and roll angle, providing more reliable eye-tracking data for online examination anti-cheating systems.

[0081] In one optional implementation, the process of extracting the gaze transfer sequence between different functional regions, calculating the time interval and spatial distance of gaze transfer, and generating a gaze transfer feature map includes:

[0082] Establish a region index matrix using multiple functional areas on the screen of the candidate's terminal device;

[0083] Within a preset time window, density peak clustering is performed on the compensated gaze point coordinates to obtain the clustered gaze point coordinates. The clustered gaze point coordinates are then matched with the region index matrix using distance weights to obtain the functional region identifier corresponding to each gaze point coordinate. The functional region identifiers are recorded in chronological order to generate a gaze region transfer sequence.

[0084] The time points at which two adjacent functional region identifiers differ in the gaze region transfer sequence are statistically analyzed, and the corresponding gaze point information is marked as a migration event. The starting functional region identifier and the target functional region identifier of the migration event are extracted.

[0085] The spatial jump distance is calculated based on the starting gaze point coordinates and the target gaze point coordinates of the migration event. The migration time interval is calculated based on the start time and end time of the migration event. The starting functional area identifier, the target functional area identifier, the spatial jump distance, and the migration time interval are combined to form a migration feature vector. Statistical features are calculated on the migration feature vector to generate a gaze migration feature map.

[0086] The screen can be divided into functional areas such as the question area, answer area, toolbar area, timer area, and submit button area. Each functional area corresponds to a unique area identifier; for example, the question area is identified as "A", the answer area as "B", the toolbar area as "C", the timer area as "D", and the submit button area as "E". The area index matrix contains the coordinate range of each functional area and its corresponding area identifier. For example, the coordinate range of the question area is the top-left corner (100, 100) and the bottom-right corner (500, 400), and its area identifier is "A".

[0087] The distance of each fixation point to other fixation points is calculated, and the local density of each fixation point and the minimum distance to higher density points are determined. Density peak points are identified based on these two metrics. For example, if there are 10 fixation points within a time window, after density peak clustering, these points will be grouped into 3 cluster centers with coordinates (315, 235), (420, 310), and (250, 180).

[0088] Based on distance-weighted matching, the functional region identifier corresponding to each gaze point is determined. Distance-weighted matching considers the distance between the gaze point and the center point of each functional region; the closer the distance, the greater the weight. For example, the gaze point (315, 235) is closest to the center of the question region, so it matches the region identifier "A"; the gaze point (420, 310) matches the region identifier "B"; and the gaze point (250, 180) matches the region identifier "C".

[0089] The gaze area shift sequence over a period of time is "AABBBCAAD", which means that the candidate's gaze point starts from the question area, moves to the answer area, then to the toolbar area, returns to the question area, and finally to the timer area.

[0090] In the above sequence, the transition points from "A" to "B", from "B" to "C", from "C" to "A", and from "A" to "D" are marked as migration events. The starting functional region identifier and the target functional region identifier of each migration event are extracted. For example, the starting functional region of the first migration event is "A" and the target functional region is "B".

[0091] The initial gaze point coordinates are (315, 235), the target gaze point coordinates are (420, 310), and the spatial jump distance is the Euclidean distance between the two points, approximately 137.7 pixels. The start time is 10.5 seconds, the end time is 10.8 seconds, and the time interval is 0.3 seconds.

[0092] The migration feature vector is composed of the starting functional area identifier, the target functional area identifier, the spatial jump distance, and the migration time interval. For example, the feature vector for the first migration event is {"A", "B", 137.7, 0.3}. Statistical features are calculated for all migration feature vectors, including the frequency of migrations between different functional areas, the average spatial jump distance, and the average migration time interval. For example, the migration from area "A" to area "B" occurs 5 times, with an average spatial jump distance of 130.5 pixels and an average migration time interval of 0.28 seconds.

[0093] A gaze transfer feature map can be generated in matrix form, where rows and columns represent the starting and target functional regions, respectively. Matrix elements include the frequency of transitions between corresponding regions, the average spatial jump distance, and the average transition time interval. For example, the transition feature from region "A" to region "B" is {5 times, 130.5 pixels, 0.28 seconds}. The gaze transfer feature map visualizes the examinee's gaze behavior patterns, reflects their problem-solving strategies and cognitive processes, and provides a basis for evaluating examinee performance.

[0094] The above methods can be used to comprehensively analyze candidates' gaze shifting behavior during the exam, identify gaze transition patterns between different functional areas, and provide objective data support for understanding candidates' problem-solving strategies and cognitive processes.

[0095] In one optional implementation, density peak clustering is performed on the compensated gaze point coordinates within a preset time window to obtain the clustered gaze point coordinates. The clustered gaze point coordinates are then matched with a region index matrix using distance weighting to obtain the functional region identifier corresponding to each gaze point coordinate, including:

[0096] Obtain the compensated gaze point coordinate sequence within a preset time window, calculate the Euclidean distance between all gaze point pairs in the gaze point coordinate sequence, and calculate the local density value of each gaze point based on the Euclidean distance;

[0097] Decision values ​​are calculated based on the local density value of each gaze point and the minimum Euclidean distance between it and other gaze points. A preset number of gaze points with the largest decision values ​​are selected as density peak points.

[0098] A fixation point connectivity graph is established, connecting each non-density peak point to its nearest spatial fixation point. The weight of each connection is calculated based on the local density value and the Euclidean distance. Based on the connectivity graph, starting from each non-density peak point, the graph moves along the connection with the largest weight until a density peak point is reached. The non-density peak point is then assigned to the corresponding cluster of the density peak point. Using the local density value of the fixation point as the weight, a weighted average of the coordinates of all fixation points in each cluster is calculated to obtain the fixation point coordinates after density peak clustering.

[0099] Calculate the vertical distance from each density peak point to the functional region boundary in the gaze coordinates after the density peak clustering. Divide the vertical distance by the smoothing coefficient and take the negative exponent to obtain the distance weight value of each functional region corresponding to the density peak point. Based on the maximum distance weight value, determine the functional region identifier to which all gaze points in the cluster to which the density peak point belongs.

[0100] Obtain the compensated gaze point coordinate sequence within a preset time window. This time window can be set to 1000 milliseconds, and includes multiple gaze point coordinates within this time range, such as coordinate points P1(300,250), P2(320,260), P3(310,255), P4(500,450), P5(515,445), etc.

[0101] For the acquired gaze point coordinate sequence, calculate the Euclidean distance between all gaze point pairs in the sequence. For example, the Euclidean distance between points P1 and P2 is 22.36, the Euclidean distance between points P1 and P3 is 11.18, and so on, calculating the distance between all point pairs.

[0102] After calculating all Euclidean distances, a local density value is calculated for each fixation point based on these distances. The local density value reflects the degree of fixation point clustering around a given fixation point. A cutoff distance dc is set, for example, dc = 25 pixels. For any two points Pi and Pj, if their Euclidean distance is less than dc, they are considered neighbors. The local density value ρi of point Pi is defined as the number of points with a distance less than dc from Pi. For example, if P1 has 3 points (including P2 and P3) with a distance less than dc around it, then the local density value ρ1 of P1 is 3.

[0103] For each point Pi, find the point Pj with the smallest distance from Pi among the points with a higher local density value than Pi, and record this minimum distance as δi. If Pi has the highest local density value, then δi is set as the maximum distance in the sequence. The decision value γi = ρi × δi represents the probability that the point becomes a density peak point. For example, if the local density value of P1 is ρ1 = 3, and among the points with higher density than P1, P3 has the smallest distance from P1 (11.18), then the decision value of P1 is γ1 = 3 × 11.18 = 33.54.

[0104] When selecting a preset number of fixation points with the highest decision values ​​as density peak points, the preset number can be set according to the application scenario. For example, the three points with the highest decision values ​​can be selected as density peak points. Assuming that P1, P4, and P5 have the highest decision values ​​after calculation, they are selected as density peak points.

[0105] A gaze connectivity map is generated by connecting each non-density peak point to its nearest spatially distant gaze point. For example, non-density peak point P2 is 22.36 units away from P1, 14.14 units away from P3, and both P4 and P5 are more than 30 units away; therefore, P2 is connected to P3. Similarly, P3 is connected to P1 because P1 is the closest point to P3.

[0106] The connection weight can be defined as the product of the local density values ​​of the two endpoints divided by the square of the Euclidean distance between them. For example, the connection weight between P2 and P3 is (ρ2×ρ3) / (d 2 Given P2P3), assuming ρ2=2, ρ3=3, and dP2P3=14.14, the weight value is (2×3) / (14.14). 2 =0.03.

[0107] Starting from P2, if the connection weight between P2 and P3 is the largest, then move to P3; if the connection weight between P3 and P1 is the largest, and P1 is the density peak point, then both P2 and P3 are assigned to the cluster where P1 is located.

[0108] In clustering, the local density value of the gaze point is used as the weight to calculate the weighted average of the coordinates of all gaze points, and the coordinates of the gaze points after density peak clustering are obtained. For example, assuming that cluster P1 contains P1, P2 and P3, and their local density values ​​are 3, 2 and 3 respectively, then the weighted average coordinates of this cluster are [(300×3+320×2+310×3) / (3+2+3), (250×3+260×2+255×3) / (3+2+3)]=[309.38, 254.38].

[0109] Suppose there are three functional areas on the interface: Area A (rectangle, top left corner coordinates (250, 200), bottom right corner coordinates (350, 300)), Area B (rectangle, top left corner coordinates (450, 400), bottom right corner coordinates (550, 500)), and Area C (rectangle, top left corner coordinates (600, 300), bottom right corner coordinates (700, 400)). For the weighted average coordinates of cluster P1 (309.38, 254.38), the distance to the boundary of Area A is 0 (because the point is within Area A), the distance to the boundary of Area B is approximately 149.65, and the distance to the boundary of Area C is approximately 290.62.

[0110] With a smoothing coefficient of 50, for cluster P1, the distance weight value for region A is e. -(0 / 50) =1, corresponding to a distance weight value of e for region B. -(149.65 / 50) =0.05, corresponding to a distance weight value of e for region C. -(290.62 / 50) =0.003. Based on the maximum distance weight value, the functional region to which all gaze points in cluster P1 belong is identified as region A.

[0111] Similarly, performing the same calculation on the clusters formed by P4 and P5 will yield results indicating that they belong to region B. In this way, gaze points can be effectively mapped to corresponding functional areas, improving the accuracy of eye tracking in human-computer interaction.

[0112] In one optional implementation, when a periodic jumping pattern is detected in the gaze migration feature map, and the time interval of gaze point shifts exhibits abnormal periodicity, while the gaze point spatial jump speed exceeds the gaze point jump speed threshold, the corresponding time period is marked as a suspicious time period, including:

[0113] The jump pattern feature sequence is extracted from the gaze migration feature map, and the autocorrelation coefficient of the jump pattern feature sequence under different time delays is calculated. When the maximum autocorrelation coefficient is greater than the jump periodicity determination threshold, the current time period is marked as a periodic jump period.

[0114] For the periodic jump period, the time interval sequence of adjacent fixation points is extracted and subjected to fast Fourier transform to obtain spectral data. The ratio of the energy value of the main frequency component to the total energy value in the spectral data is calculated to obtain the main frequency energy proportion. The ratio of the standard deviation to the mean of the time interval sequence is calculated to obtain the coefficient of variation. When the main frequency energy proportion is greater than the periodic intensity threshold and the coefficient of variation is less than the eye movement fluctuation threshold, the current period is marked as an abnormal periodic period.

[0115] For the abnormal periodic time period, the ratio of the Euclidean distance between adjacent fixation points to the corresponding time interval is calculated to obtain the fixation point spatial jump speed. When the fixation point spatial jump speed is greater than the preset fixation point jump speed threshold, the current fixation point is marked as a suspicious fixation point.

[0116] The number and distribution of suspicious gaze points within the abnormal period are counted. When the number of suspicious gaze points is greater than the density threshold and the distribution interval is less than the aggregation threshold, the current abnormal period is determined as a suspicious period.

[0117] For the detection of periodic jump patterns in gaze migration feature maps, a jump pattern feature sequence is extracted, which contains the Euclidean distance values ​​between adjacent gaze points. Taking an eye-tracking device with a sampling rate of 60Hz as an example, the distance between each pair of adjacent gaze points is calculated in 30 seconds of eye-tracking data, forming a jump pattern feature sequence containing approximately 120-180 elements. To detect periodicity in the sequence, the autocorrelation coefficient of the sequence at different time delays is calculated. In specific implementation, the maximum delay value is set to half the sequence length, and the autocorrelation coefficient corresponding to each delay value within the range of 0 to the maximum delay value is calculated. When the maximum autocorrelation coefficient is detected to exceed the preset jump periodicity judgment threshold of 0.65, the current time period is marked as a periodic jump period. For example, in a certain test, the maximum autocorrelation coefficient of a gaze sequence was detected to be 0.72, which appeared at the delay value of 12, indicating that a similar jump pattern appeared approximately every 12 sampling points in the sequence, and therefore it was marked as a periodic jump period.

[0118] The time interval sequence between adjacent fixation points is extracted, and a Fast Fourier Transform (FFT) is performed on this sequence to obtain spectral data. In practical applications, for a time interval sequence of length 128, a power-of-two (FFT) algorithm is used to transform it, resulting in spectral data composed of 128 complex values. The ratio of the energy value of the dominant frequency component to the total energy value in the spectral data is calculated to obtain the dominant frequency energy proportion. Simultaneously, the ratio of the standard deviation to the mean of the time interval sequence is calculated to obtain the coefficient of variation. When the dominant frequency energy proportion is greater than a preset periodicity threshold of 0.35 and the coefficient of variation is less than a preset eye movement fluctuation threshold of 0.25, the current time period is marked as an abnormal periodic period. In one example, a segment of eye movement data was detected with a dominant frequency energy proportion of 0.42 and a coefficient of variation of 0.18, which conforms to the characteristics of abnormal periodicity and is therefore marked as an abnormal periodic period.

[0119] For data identified as abnormal periodic periods, the spatial jump velocity of the fixation point is further calculated, which is the ratio of the Euclidean distance between adjacent fixations to the corresponding time interval. In normal human eye movements, the spatial jump velocity of the fixation point typically does not exceed a certain range. A fixation jump velocity threshold of 500 pixels / second is set. When the calculated spatial jump velocity of the fixation point exceeds this threshold, the current fixation point is marked as a suspicious fixation point. For example, within a certain abnormal periodic period, if the Euclidean distance between two adjacent fixations is detected to be 300 pixels, and the corresponding time interval is 0.5 seconds, the calculated spatial jump velocity is 600 pixels / second, exceeding the preset threshold; therefore, this fixation point is marked as a suspicious fixation point.

[0120] To identify suspicious time periods, the number and distribution of suspicious gaze points within abnormal periodicity are statistically analyzed. A density threshold of 10 gaze points per 30 seconds and a clustering threshold of 3 seconds are set. When the number of suspicious gaze points within an abnormal periodicity exceeds the density threshold, and the time interval between adjacent suspicious gaze points is less than the clustering threshold, the current abnormal periodicity is identified as a suspicious period. In practical application, 15 suspicious gaze points were detected within a certain 30-second abnormal periodicity, and the average time interval between these suspicious gaze points was 2.1 seconds, satisfying both the density and clustering conditions. Therefore, this period was ultimately identified as a suspicious period.

[0121] The methods described above can effectively identify anomalous patterns in eye-tracking data, particularly suspicious periods exhibiting unnatural periodicity and abnormal spatial jump speeds of the gaze point. These suspicious periods are often associated with abnormal human eye-tracking behavior, such as deceptive actions performed using eye-tracking simulation programs or robots. In practical applications, marking these suspicious periods provides a basis for subsequent security verification and behavioral analysis.

[0122] In one optional implementation, the gaze behavior feature data within the suspicious time period is extracted, and the spectral energy distribution features of the gaze point trajectory are extracted using autoregressive spectral analysis. Wavelet analysis is then used to identify abrupt changes in the gaze pattern. When the spectral energy distribution features are abnormal and there are abrupt changes in the gaze pattern, accompanied by abnormal gaze, the behavior is determined to be cheating, including:

[0123] Extract spatial coordinate sequences of gaze points from gaze behavior feature data within suspicious time periods;

[0124] Autoregressive spectral analysis is performed on the spatial coordinate sequence to obtain power spectral density data. The power spectral density data is divided into multiple frequency bands according to the frequency magnitude, and the ratio of the power spectral density value of each frequency band to the total power spectral density value is calculated to obtain the spectral energy distribution characteristics.

[0125] A continuous wavelet transform is performed on the spatial coordinate sequence to obtain multi-scale wavelet coefficients. Empirical mode decomposition is performed on the multi-scale wavelet coefficients to obtain multiple intrinsic mode functions. An analytical signal is constructed for each intrinsic mode function. Instantaneous amplitude and instantaneous phase are calculated based on the analytical signal. The gaze pattern abruptness point is marked based on the rate of change of instantaneous amplitude and instantaneous phase at adjacent time points.

[0126] When the component features of the spectral energy distribution deviate from the feature subspace of the normal gaze pattern, and there is a mutation point in the gaze pattern, and the current gaze action is marked as abnormal gaze, the behavior in the current suspicious period is judged as cheating behavior.

[0127] Within the gaze behavior feature data extracted during suspicious time periods, a spatial coordinate sequence of the examinee's gaze points is obtained from an eye-tracking device. This sequence contains the trajectory of the examinee's eye movements during the answering process. The spatial coordinate sequence is typically a time series {(x1,y1), (x2,y2), ..., (x... n ,y n The form is represented as )}, where (x t ,y t ) represents the two-dimensional coordinates of the gaze point on the screen at time t.

[0128] Autoregressive spectral analysis estimates the power spectral density of a sequence by establishing an autoregressive model. In implementation, an appropriate autoregressive model order p is chosen (e.g., p=10), and the autoregressive model parameters are estimated using the Yule-Walker equation algorithm. The estimated parameters are then substituted into the autoregressive spectral density function to calculate the power spectral density data. For example, analyzing a test-taker's gaze trajectory over one minute yields the distribution of power spectral density values ​​within the 0-50Hz frequency range.

[0129] The frequency range of 0-50Hz can be divided into five frequency bands: 0-10Hz, 10-20Hz, 20-30Hz, 30-40Hz, and 40-50Hz. For each band, the ratio of the power spectral density value within that band to the total power spectral density value is calculated to obtain the spectral energy distribution characteristics. For example, during normal reading, the energy proportion in the 0-10Hz band is approximately 65%, in the 10-20Hz band approximately 20%, in the 20-30Hz band approximately 10%, in the 30-40Hz band approximately 3%, and in the 40-50Hz band approximately 2%. Cheating behavior is characterized by an abnormal increase in high-frequency components, such as the energy proportion in the 30-40Hz band exceeding 10%.

[0130] For the same spatial coordinate sequence, a continuous wavelet transform is performed using the Morlet wavelet as the mother wavelet function, and wavelet coefficients are calculated at multiple scales. For example, by selecting eight scales for analysis, a wavelet coefficient matrix W(a,t) on the time-scale plane is obtained, where a represents the scale and t represents time.

[0131] The signal is decomposed into multiple intrinsic mode functions (IMFs) and a residual term through iterative filtering. For a gaze point trajectory, typically 4-6 IMFs can be obtained. Each IMF represents the oscillation mode of the original signal at different characteristic time scales.

[0132] An analytic signal is transformed into a complex form by performing a Hilbert transform on the original IMF. An analytic signal contains a real part (the original IMF) and an imaginary part (the Hilbert transform of the original IMF). Instantaneous amplitude and instantaneous phase can be calculated from the analytic signal. The instantaneous amplitude is the square root of the magnitude of the analytic signal, representing the energy of the signal at each moment; the instantaneous phase is the arctangent of the real and imaginary parts of the analytic signal, representing the phase angle of the signal at each moment.

[0133] When the relative rate of change of instantaneous amplitude exceeds a preset threshold (e.g., 50%) or the rate of change of instantaneous phase exceeds a preset threshold (e.g., π / 4 per second), that moment is marked as a gaze pattern abrupt change point. In practical applications, a threshold for the number of abrupt change points within a time window (e.g., 0.5 seconds) can be set; when the number of abrupt change points exceeds the threshold, they are confirmed as valid abrupt change points. For example, in a case of cheating by a test taker, when the student switched from normal reading to viewing a hidden cheating device, three valid abrupt change points were detected within one second.

[0134] Principal component analysis is used to establish a feature subspace for the normal gaze pattern. The first k principal components (e.g., k=3) represent the main variations in the normal gaze pattern. The distance from the current spectral energy distribution feature to this feature subspace is calculated. If the distance exceeds a threshold (e.g., 3 times the standard deviation of the average distance), it is determined to be an abnormal spectral energy distribution.

[0135] Based on the above judgment results, if all three conditions are met simultaneously: the spectral energy distribution characteristics deviate from the feature subspace of the normal gaze pattern (e.g., reconstruction error exceeds 0.25), there are abrupt changes in the gaze pattern (e.g., more than 5 abrupt changes within 10 seconds), and the current gaze action is marked as abnormal gaze (e.g., gaze duration exceeds the 95th percentile of the normal distribution or the gaze point falls in a non-allowed area), then the behavior during the current suspicious period will be judged as cheating. The judgment results will be recorded, including timestamps, abnormal feature values, and judgment confidence levels, for further review by invigilators.

[0136] Figure 3This diagram illustrates the comparative analysis of spectral energy distribution. It shows the effect of different spectral analysis methods in identifying cheating behavior, clearly demonstrating the significant technical advantages of this invention compared to traditional methods. As can be observed from the figure, the spectral energy in the normal gaze mode is mainly concentrated in the low-frequency band, with the 0-10Hz band accounting for 65% of the energy, which is consistent with the physiological characteristics of natural human gaze. The traditional FFT method shows that the energy in the 0-10Hz band is 58%, and the energy in the 10-20Hz band is 23%. Although it can detect some spectral changes, its sensitivity is limited. The Welch power spectrum method shows a slight improvement, with the energy in the 0-10Hz band at 61%, but its ability to identify high-frequency anomalies is still insufficient. In contrast, this invention, through an autoregressive spectral analysis algorithm, successfully identified the abnormal spectral characteristics of cheating behavior. The detection results show that the energy in the 0-10Hz band is significantly reduced to 42%, while the energy proportions of high-frequency bands such as 20-30Hz, 30-40Hz, and 40-50Hz reach 18%, 8%, and 4%, respectively, which are 80%, 167%, and 100% higher than the normal gaze mode. This abnormal increase in high-frequency energy precisely reflects the abnormal behavioral patterns of rapid eye scanning and frequent switching of gaze targets during cheating. This invention can accurately capture these subtle spectral changes and provide a reliable technical basis for accurately determining cheating behavior by quantitatively analyzing the redistribution of spectral energy in each frequency band. It fully verifies the innovation and practicality of autoregressive spectral analysis in the field of eye movement abnormality detection.

[0137] This invention relates to an online examination anti-cheating system based on AI-fused eye-tracking, the system comprising:

[0138] The first unit is used to extract the real-time motion trajectory of the candidate's head in three-dimensional space through the camera of the candidate's terminal device to establish a head motion feature sequence;

[0139] The second unit is used to divide the screen display content of the candidate's terminal device into multiple functional areas and set a corresponding gaze weight coefficient for each functional area; when the actual gaze duration or gaze frequency of a certain functional area exceeds the area switching frequency threshold, the corresponding gaze action is marked as abnormal gaze.

[0140] The third unit is used to calculate the real-time coordinate position and duration of the examinee's gaze on the screen through eye tracking, and to form gaze behavior feature data by combining the head movement feature sequence and the gaze weight coefficient.

[0141] The fourth unit is used to extract the transfer sequence of the examinee's gaze point between different functional areas, calculate the time interval and spatial distance of the gaze point transfer, and generate a gaze migration feature map; when a periodic jumping pattern is detected in the gaze migration feature map, and the time interval of the gaze point transfer shows abnormal periodicity, and the spatial jump speed of the gaze point exceeds the gaze point jump speed threshold, the corresponding time period is marked as a suspicious time period.

[0142] The fifth unit is used to extract the gaze behavior feature data within the suspicious time period, extract the spectral energy distribution features of the gaze point trajectory using autoregressive spectral analysis, and identify gaze pattern abrupt change points by combining wavelet analysis. When the spectral energy distribution features are abnormal and there are gaze pattern abrupt change points, accompanied by abnormal gaze, it is determined to be cheating behavior.

[0143] A third aspect of the present invention provides an electronic device, comprising:

[0144] processor;

[0145] Memory used to store processor-executable instructions;

[0146] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0147] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0148] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0149] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An online examination anti-cheating method based on AI-fused eye-tracking, characterized in that, include: The real-time movement trajectory of the candidate's head in three-dimensional space is extracted by the camera of the candidate's terminal device to establish a head movement feature sequence; The screen display content of the candidate's terminal device is divided into multiple functional areas, and a corresponding gaze weight coefficient is set for each functional area. When the actual gaze duration or gaze frequency of a certain functional area exceeds the area switching frequency threshold, the corresponding gaze action is marked as abnormal gaze. By calculating the real-time coordinates of the examinee's gaze point on the screen and the duration of gaze through eye tracking, and combining the head movement feature sequence and the gaze weight coefficient, gaze behavior feature data is formed. Extract the gaze transfer sequence between different functional areas, calculate the time interval and spatial distance of gaze transfer, and generate a gaze transfer feature map; When a periodic jumping pattern is detected in the gaze migration feature map, and the time interval of gaze point transfer shows abnormal periodicity, and the gaze point spatial jump speed exceeds the gaze point jump speed threshold, the corresponding time period is marked as a suspicious time period. Extract the gaze behavior feature data within the suspicious time period, use autoregressive spectral analysis to extract the spectral energy distribution features of the gaze point trajectory, and combine wavelet analysis to identify gaze pattern abrupt change points. When the spectral energy distribution features are abnormal and there are gaze pattern abrupt change points, accompanied by abnormal gaze, it is determined to be cheating behavior.

2. The method according to claim 1, characterized in that, By calculating the real-time coordinates of the examinee's gaze point on the screen and the duration of gaze through eye tracking, and combining this with the head movement feature sequence and the gaze weighting coefficient, gaze behavior feature data is formed, including: The left and right eye gaze vectors of the examinee are obtained through an eye-tracking device, and the coordinates of the initial gaze point are determined based on the intersection of the left and right eye gaze vectors with the screen plane. Based on the head motion feature sequence, the head yaw angle rotation matrix, head pitch angle rotation matrix, and head roll angle rotation matrix are calculated and multiplied to obtain the head motion compensation matrix. The initial gaze point coordinates are then subjected to matrix operation with the head motion compensation matrix to obtain the compensated gaze point coordinates. The current compensated gaze point coordinates are then compared with the previous compensated gaze point coordinates to obtain the gaze point movement speed. The spatial dispersion is calculated by continuously collecting the compensated gaze point coordinates for a preset number of frames. When the spatial dispersion is less than the gaze determination threshold, the gaze start time and end time are recorded to generate the gaze duration. The compensated gaze point coordinates, the gaze point movement speed, the gaze duration, and the gaze weight coefficient are combined to form a gaze behavior feature vector.

3. The method according to claim 1, characterized in that, Extract the gaze transfer sequence between different functional areas, calculate the time interval and spatial distance of gaze transfer, and generate a gaze transfer feature map including: Establish a region index matrix using multiple functional areas on the screen of the candidate's terminal device; Within a preset time window, density peak clustering is performed on the compensated gaze point coordinates to obtain the clustered gaze point coordinates. The clustered gaze point coordinates are then matched with the region index matrix using distance weights to obtain the functional region identifier corresponding to each gaze point coordinate. The functional region identifiers are recorded in chronological order to generate a gaze region transfer sequence. The time points at which two adjacent functional region identifiers differ in the gaze region transfer sequence are statistically analyzed, and the corresponding gaze point information is marked as a migration event. The starting functional region identifier and the target functional region identifier of the migration event are extracted. The spatial jump distance is calculated based on the starting gaze point coordinates and the target gaze point coordinates of the migration event. The migration time interval is calculated based on the start time and end time of the migration event. The starting functional area identifier, the target functional area identifier, the spatial jump distance, and the migration time interval are combined to form a migration feature vector. Statistical features are calculated on the migration feature vector to generate a gaze migration feature map.

4. The method according to claim 3, characterized in that, Within a preset time window, density peak clustering is performed on the compensated gaze point coordinates to obtain the clustered gaze point coordinates. The clustered gaze point coordinates are then matched with a region index matrix using distance weighting to obtain the functional region identifier corresponding to each gaze point coordinate, including: Obtain the compensated gaze point coordinate sequence within a preset time window, calculate the Euclidean distance between all gaze point pairs in the gaze point coordinate sequence, and calculate the local density value of each gaze point based on the Euclidean distance; Decision values ​​are calculated based on the local density value of each gaze point and the minimum Euclidean distance between it and other gaze points. A preset number of gaze points with the largest decision values ​​are selected as density peak points. A fixation point connectivity graph is established, connecting each non-density peak point to its nearest spatial fixation point. The weight of each connection is calculated based on the local density value and the Euclidean distance. Based on the connectivity graph, starting from each non-density peak point, the graph moves along the connection with the largest weight until a density peak point is reached. The non-density peak point is then assigned to the corresponding cluster of the density peak point. Using the local density value of the fixation point as the weight, a weighted average of the coordinates of all fixation points in each cluster is calculated to obtain the fixation point coordinates after density peak clustering. Calculate the vertical distance from each density peak point to the functional region boundary in the gaze coordinates after the density peak clustering. Divide the vertical distance by the smoothing coefficient and take the negative exponent to obtain the distance weight value of each functional region corresponding to the density peak point. Based on the maximum distance weight value, determine the functional region identifier to which all gaze points in the cluster to which the density peak point belongs.

5. The method according to claim 1, characterized in that, When a periodic jumping pattern is detected in the gaze migration feature map, and the time interval of gaze point shifts exhibits abnormal periodicity, while the gaze point spatial jump speed exceeds the gaze point jump speed threshold, the corresponding time period is marked as a suspicious time period, including: The jump pattern feature sequence is extracted from the gaze migration feature map, and the autocorrelation coefficient of the jump pattern feature sequence under different time delays is calculated. When the maximum autocorrelation coefficient is greater than the jump periodicity determination threshold, the current time period is marked as a periodic jump period. For the periodic jump period, the time interval sequence of adjacent fixation points is extracted and subjected to fast Fourier transform to obtain spectral data. The ratio of the energy value of the main frequency component to the total energy value in the spectral data is calculated to obtain the main frequency energy proportion. The ratio of the standard deviation to the mean of the time interval sequence is calculated to obtain the coefficient of variation. When the main frequency energy proportion is greater than the periodic intensity threshold and the coefficient of variation is less than the eye movement fluctuation threshold, the current period is marked as an abnormal periodic period. For the abnormal periodic time period, the ratio of the Euclidean distance between adjacent fixation points to the corresponding time interval is calculated to obtain the fixation point spatial jump speed. When the fixation point spatial jump speed is greater than the preset fixation point jump speed threshold, the current fixation point is marked as a suspicious fixation point. The number and distribution of suspicious gaze points within the abnormal period are counted. When the number of suspicious gaze points is greater than the density threshold and the distribution interval is less than the aggregation threshold, the current abnormal period is determined as a suspicious period.

6. The method according to claim 1, characterized in that, Extract the gaze behavior feature data within the suspected time period, use autoregressive spectral analysis to extract the spectral energy distribution features of the gaze point trajectory, and combine wavelet analysis to identify gaze pattern abrupt changes. When the spectral energy distribution features are abnormal and there are gaze pattern abrupt changes, accompanied by abnormal gaze, it is determined to be cheating behavior, including: Extract spatial coordinate sequences of gaze points from gaze behavior feature data within suspicious time periods; Autoregressive spectral analysis is performed on the spatial coordinate sequence to obtain power spectral density data. The power spectral density data is divided into multiple frequency bands according to the frequency magnitude, and the ratio of the power spectral density value of each frequency band to the total power spectral density value is calculated to obtain the spectral energy distribution characteristics. A continuous wavelet transform is performed on the spatial coordinate sequence to obtain multi-scale wavelet coefficients. Empirical mode decomposition is performed on the multi-scale wavelet coefficients to obtain multiple intrinsic mode functions. An analytical signal is constructed for each intrinsic mode function. Instantaneous amplitude and instantaneous phase are calculated based on the analytical signal. The gaze pattern abruptness point is marked based on the rate of change of instantaneous amplitude and instantaneous phase at adjacent time points. When the component features of the spectral energy distribution deviate from the feature subspace of the normal gaze pattern, and there is a mutation point in the gaze pattern, and the current gaze action is marked as abnormal gaze, the behavior in the current suspicious period is judged as cheating behavior.

7. An online examination anti-cheating system based on AI-fused eye-tracking, used to implement the method as described in any one of claims 1-6, characterized in that, include: The first unit is used to extract the real-time motion trajectory of the candidate's head in three-dimensional space through the camera of the candidate's terminal device to establish a head motion feature sequence; The second unit is used to divide the screen display content of the candidate's terminal device into multiple functional areas and set a corresponding gaze weight coefficient for each functional area; when the actual gaze duration or gaze frequency of a certain functional area exceeds the area switching frequency threshold, the corresponding gaze action is marked as abnormal gaze. The third unit is used to calculate the real-time coordinate position and duration of the examinee's gaze on the screen through eye tracking, and to form gaze behavior feature data by combining the head movement feature sequence and the gaze weight coefficient. The fourth unit is used to extract the transition sequence of the examinee's gaze point between different functional areas, calculate the time interval and spatial distance of the gaze point transition, and generate a gaze migration feature map. When a periodic jumping pattern is detected in the gaze migration feature map, and the time interval of gaze point transfer shows abnormal periodicity, and the gaze point spatial jump speed exceeds the gaze point jump speed threshold, the corresponding time period is marked as a suspicious time period. The fifth unit is used to extract the gaze behavior feature data within the suspicious time period, extract the spectral energy distribution features of the gaze point trajectory using autoregressive spectral analysis, and identify gaze pattern abrupt change points by combining wavelet analysis. When the spectral energy distribution features are abnormal and there are gaze pattern abrupt change points, accompanied by abnormal gaze, it is determined to be cheating behavior.

8. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 6.

Images