Intelligent examination monitoring system

Abstract

The invention discloses an intelligent examination monitoring system, which relates to the technical field of examination monitoring, and constructs a video information basis with higher time precision by introducing an image time sequence sampling and high frame reconstruction mechanism, so that tiny behavior disturbance can be accurately captured and identified, and the accuracy of examination monitoring is improved. The problem that an existing system is slow to perceive subtle, short-time and high-concealment behaviors is solved. The method does not depend on single action or local feature judgment any more, but synchronously extracts three types of micro-action information of the head angle track Pot, the shoulder disturbance rate Pbs and the eye jump amplitude Pdm from the image, so that the recognition result has higher behavior coverage and linkage coupling expression ability, and the overall judgment precision of the abnormal behavior is improved. Through layered construction of a disturbance coupling index RZH and a behavior mutation index psi, point-like judgment is upgraded to section trend evaluation on the behavior recognition level, so that the system can recognize an abnormal behavior mode formed by continuous perturbation instead of abnormal triggering of a single action.

Description

Technical Field

[0001] The present invention relates to the technical field of examination monitoring, and in particular to an intelligent examination monitoring system. Background Art

[0002] In exam monitoring scenarios, surveillance systems are typically deployed in confined spaces such as schools, training institutions, and vocational examination centers. Using multiple cameras and audio capture devices, they continuously analyze and assess the risk of test takers' posture, language, interactions, and other behaviors during the exam. In practical applications, the "specific things" the system focuses on are detecting abnormal test takers' behavior during the exam, particularly "non-intrusive" behaviors that don't rely on leaving their seats, connecting devices, or other hard cheating methods, such as frequent glances, shoulder shaking, and frequent eye wandering.

[0003] In existing applications, traditional intelligent exam monitoring systems often rely on coarse-grained analysis based on rules, such as static area intrusion detection, voice intensity assessment, or blind-spot camera monitoring, to address these "non-displacement abnormal behaviors." These methods are effective for identifying large-scale, overt behaviors, but suffer from insufficient recognition and high false positive rates when it comes to micro-perturbations that are "low-amplitude, short-duration, and highly coordinated." For example, a slight head tilt or subtle movement by a test-taker is often insufficient to trigger pre-set rules, preventing the system from identifying potential risks early on. These systems generally lack the ability to fusedly perceive temporal behavioral trajectories, making them unable to effectively analyze and model the linked patterns of "micro-head movements, eye movements, and shoulder perturbations," which in turn affects overall recognition depth and accuracy.

[0004] The root cause of these shortcomings lies in the fact that most existing systems treat behavior recognition as a "static rule-triggering" process, lacking continuous modeling of dynamic behavioral characteristics and multi-parameter coupled analysis mechanisms. During exams, these flaws can have several serious consequences: First, candidates with genuine aberrant intent may communicate information through subtle, irregular, yet coordinated movements without being recognized by the system. Second, due to the system's lack of fine-tuning of its tolerance range for micro-perturbations, some normal stressful behaviors may be misclassified as abnormal, causing psychological stress for candidates, increasing false positive rates, and even compromising exam fairness. Summary of the Invention

[0005] In view of the deficiencies of the prior art, the present invention provides an intelligent examination monitoring system, which solves the problems mentioned in the background technology.

[0006] To achieve the above objectives, the present invention is implemented through the following technical solutions: an intelligent test monitoring system, including an image time series sampling and high-frame reconstruction module, a multi-parameter micro-motion trajectory extraction module, a disturbance coupling behavior recognition module, a behavior jump window recognition module, a behavior abnormality level determination module, and a dynamic warning and behavior feedback module;

[0007] The image time series sampling and high-frame reconstruction module collects the video stream during the examination through a camera and performs high-frame rate interpolation and reconstruction to obtain a high-precision image sequence I(t);

[0008] The multi-parameter micro-motion trajectory extraction module extracts features from the high-precision image sequence I(t) and fits it into a temporal feature set PF;

[0009] The disturbance coupling behavior identification module performs time series fitting on the acquired time series feature set PF and constructs the disturbance coupling index RZH;

[0010] The behavior jump window identification module analyzes the disturbance mutation degree of the disturbance coupling index RZH within the time sliding window T, and performs abnormal mutation detection on the disturbance coupling index RZH to obtain the behavior mutation index Ψ;

[0011] The behavior abnormality level determination module compares the obtained behavior mutation index Ψ with the multi-level threshold, classifies the behavior risk level, and obtains the classification label L;

[0012] The dynamic warning and behavior feedback module executes a differentiated strategy based on the classification label L, behavior mutation index Ψ and disturbance coupling index RZH.

[0013] Preferably, the image temporal sampling and high frame reconstruction module includes a motion vector estimation unit and a temporal image interpolation unit;

[0014] The motion vector estimation unit collects the video frame sequence in the video stream through the camera and extracts the motion vector field M between adjacent frames. The motion vector field M represents the spatial displacement trend of each pixel point (x, y) between consecutive frames.

[0015] The expression formula of the motion vector field M is as follows:

[0016] ;

[0017] Where ∂ represents the sign of the partial derivative, vi represents the brightness value of a pixel in the i-th frame image, uxi represents the horizontal motion velocity component of the pixel, that is, the speed of the optical flow in the x-direction, and vyi represents the vertical motion velocity component of the pixel, that is, the speed of the optical flow in the y-direction.

[0018] The temporal image interpolation unit is used to perform temporal interpolation processing on the video frame sequence based on the motion vector field M, insert multiple intermediate frames between adjacent frames, and obtain a high-precision image sequence I(t);

[0019] The pixel value of the intermediate frame is obtained by the following formula:

[0020] ;

[0021] Where vi(j)(x, y) represents the j-th interpolated frame between the i-th frame and the i+1-th frame, vi+1 represents the brightness value of a pixel in the i+1-th frame image, and λj represents the interpolation coefficient, which is the position factor of the j-th frame.

[0022] Preferably, the multi-parameter micro-motion trajectory extraction module includes a structure posture tracking unit and a dynamic fitting unit;

[0023] The structural posture tracking unit performs dynamic recognition of human body structural areas based on the high-precision image sequence I(t) and extracts the examinee's posture points, including the head angle trajectory Pot, the shoulder disturbance rate Pbs, and the gaze jump amplitude Pdm.

[0024] The head angle trajectory Pot is obtained by the following formula:

[0025] ;

[0026] Where, (xno, yno) represents the coordinates of the nose tip point, (xne, yne) represents the coordinates of the neck point, and arctan represents the inverse tangent function;

[0027] The shoulder disturbance rate Pbs is obtained by the following formula:

[0028] ;

[0029] Where d represents the integral sign, SL represents the left shoulder vector, and SR represents the right shoulder vector;

[0030] The left shoulder vector SL is obtained by the following formula:

[0031] ;

[0032] Where, (xL, yL) represents the coordinates of the left shoulder point;

[0033] The right shoulder vector SR is obtained by the following formula:

[0034] ;

[0035] Where, (xR, yR) represents the coordinates of the right shoulder point;

[0036] The gaze transition amplitude Pdm is obtained by the following formula:

[0037] ;

[0038] Where (gx(t), gy(t)) represents the mapping coordinates of the gaze projection direction at time t on the image plane, and (gx(t-1), gy(t-1)) represents the mapping coordinates of the gaze projection direction at time t-1 on the image plane.

[0039] Preferably, the dynamic fitting unit fits the acquired head angle trajectory Pot, shoulder disturbance rate Pbs, and gaze jump amplitude Pdm to obtain an initial feature set PC, and removes image noise fluctuations through a dynamic smoothing fitting algorithm to obtain a temporal feature set PF;

[0040] The time series feature set PF is obtained by the following formula:

[0041] ;

[0042] Where PCo(ti) represents the oth item of data in the initial feature set PC at time ti, W represents the sliding window size, and PFo(t) represents the oth item of data in the time series feature set PF at time t.

[0043] Preferably, the disturbance coupling behavior identification module includes a disturbance parameter derivative extraction unit and a nonlinear disturbance fusion index construction unit;

[0044] The disturbance parameter derivative extraction unit extracts features from the temporal feature set PF, including the head angle change rate Vpo, shoulder disturbance acceleration Apb, and gaze jump change rate Vpd;

[0045] The head angle change rate Vpo is obtained by taking the first derivative of the head angle change rate Vpo at time t;

[0046] The shoulder disturbance acceleration Apb is obtained by taking the second derivative of the shoulder disturbance rate Pbs at time t;

[0047] The gaze jump change rate Vpd is obtained by taking the first-order derivative of the gaze jump amplitude Pdm at time t;

[0048] The nonlinear perturbation fusion index construction unit nonlinearly combines the acquired head angle change rate Vpo, shoulder perturbation acceleration Apb and gaze jump change rate Vpd to obtain the perturbation coupling index RZH;

[0049] The disturbance coupling index RZH is obtained by the following formula:

[0050] ;

[0051] Where, represent the preset weight values ​​of the head angle change rate Vpo, shoulder disturbance acceleration Apb and gaze jump change rate Vpd, respectively, and .

[0052] Preferably, the behavior jump window recognition module includes a sliding window disturbance statistics unit and an exponential behavior recognition unit;

[0053] The sliding window disturbance statistics unit performs statistical processing on the disturbance coupling index RZH within the sliding time window, and dynamically calculates the average value μRZH and the fluctuation amplitude ΔRZH of the disturbance coupling index RZH;

[0054] The average value μRZH is obtained by the following formula:

[0055] ;

[0056] Where T represents the time sliding window size, RZH(ti) represents the disturbance coupling index at time ti;

[0057] The fluctuation amplitude ΔRZH is obtained by the following formula:

[0058] ;

[0059] Where RZH(t) represents the perturbation coupling index at time t, and μRZH(t) represents the average value at time t.

[0060] Preferably, the exponential behavior recognition unit performs weighted integration on the disturbance coupling index RZH within the sliding window to obtain the behavior mutation index Ψ;

[0061] The behavioral mutation index Ψ is obtained by the following formula:

[0062] ;

[0063] Where Ψ(t) represents the behavior mutation index at time t, e represents a constant, γ represents the exponential enhancement coefficient, and d represents the integral sign.

[0064] Preferably, the behavior abnormality level determination module includes a multi-level threshold dynamic generation unit and a multi-level behavior level classification unit;

[0065] The multi-level threshold dynamic generation unit generates real-time behavior judgment thresholds based on the characteristics of the behavior mutation index Ψ and dynamic adjustment factors: the exponential threshold THr1 indicates a mild warning threshold; the mutation threshold THr2 indicates a severe abnormality judgment. An exponential function structure is introduced to control the threshold response sensitivity, realizing a threshold setting strategy that is adaptive to individual behavior fluctuations.

[0066] The index threshold THr1 is obtained by the following formula:

[0067] ;

[0068] Where μΨ represents the mean of the mutation index, η1 represents the exponential adjustment coefficient, σΨ represents the standard deviation of the mutation index, C1 represents the exponential factor, E represents the exponential attenuation coefficient, and CVΨ represents the coefficient of variation, which is obtained by the ratio of the standard deviation σΨ of the mutation index to the mean μΨ of the mutation index;

[0069] The mutation threshold THr2 is obtained by the following formula:

[0070] ;

[0071] In the formula, η2 represents the mutation adjustment coefficient, C2 represents the mutation factor, ln represents the logarithmic function, G represents, D Ψ It represents the mutation amplitude difference, which is obtained by the difference between the maximum value of the mutation index Ψ and the minimum value of the mutation index Ψ. G represents the amplitude sensitivity factor.

[0072] Preferably, the multi-level behavior grade classification unit compares the obtained index threshold THr1 and mutation threshold THr2 with the behavior mutation index Ψ in real time, classifies the behavior risk level, and obtains the classification label L;

[0073] Risk level classification is obtained by matching in the following ways:

[0074] When the behavior mutation index Ψ ≤ the index threshold THr1, it indicates the first risk level and normal state;

[0075] When the index threshold THr1 < behavior mutation index Ψ ≤ mutation threshold THr2, it indicates the second risk level, suspicious state;

[0076] When the mutation threshold THr2 is less than the behavior mutation index Ψ, it indicates the third risk level, an abnormal state;

[0077] The classification label L is obtained by matching in the following way;

[0078] When the risk level is at the first risk level, the classification label L = 0, indicating a normal behavior state, and the system only records no response;

[0079] When the risk level is at the second risk level, the classification label L=1, indicating a suspicious state, and the system issues a soft warning.

[0080] When the risk level is at the third risk level, the classification label L=2, indicating a serious abnormality, the system triggers an alarm, video marking and manual review.

[0081] Preferably, the dynamic warning and behavior feedback module uses the behavior label L as the main trigger condition, and combines the behavior mutation index Ψ and the disturbance coupling index RZH to obtain the behavior response intensity RS and implement a differentiated response strategy;

[0082] The behavioral response intensity RS is obtained by the following formula:

[0083] ;

[0084] Where H1 represents the classification response coefficient, H2 represents the behavior mutation weight coefficient, H3 represents the coupling index response coefficient, and L(t) represents the classification label at time t. represents the indicator function, which is 1 if the condition is met and 0 otherwise, and θRZH represents the abnormal threshold of coupling strength;

[0085] Differentiated response strategies are matched and obtained in the following ways:

[0086] When the behavioral response intensity RS=0, no strategy is executed;

[0087] When the behavioral response intensity RS = 1, the differentiated response strategy is behavioral recording;

[0088] When the behavioral response intensity RS=2, the differentiated response strategy is video marking and soft prompting;

[0089] When the behavioral response intensity RS ≥ 3, the differentiated response strategies are alarm, report generation, and video locking.

[0090] The present invention provides an intelligent examination monitoring system, which has the following beneficial effects:

[0091] (1) When the system is running, the extracted motion vector field is used to interpolate and generate inter-frame images through the temporal image interpolation unit, obtaining a high-precision image sequence I(t) with a higher temporal resolution than the original acquisition. This processing method avoids the "behavior blur" or "action frame skipping" problems caused by the sampling rate limit of traditional systems, and ensures that small disturbances are accurately restored from the source. It is particularly suitable for identifying slight, short-term, and local abnormal movements. Through the synergy of the two sub-units of motion estimation and image interpolation, the system can not only complete the data of the current video stream, but also build a basic image layer for high-precision behavior recognition without changing the camera equipment. This mechanism provides data layer support for subsequent modules and constitutes the key foundation of the entire intelligent monitoring system.

[0092] (2) Through the structural posture tracking unit, key posture points such as head angle, shoulder disturbance and gaze projection are simultaneously extracted from the image sequence to construct a multi-dimensional disturbance information structure covering the face-upper body-visual direction. Compared with the traditional method of using only head posture or regional motion analysis, this module significantly enhances the system's ability to capture small abnormal movements by connecting multiple source trajectories in parallel. The system uses vector construction to continuously describe the shoulder state, and the spatial displacement of the left and right shoulder points relative to the neck point is used to form the left and right shoulder vectors, and the shoulder disturbance rate is calculated based on this. This spatial structured method improves the system's perception of non-rigid movements and makes up for the defects of previous systems that ignore or roughly process shoulder behavior.

[0093] (3) By modeling parameters such as the disturbance change rate, acceleration, and fluctuation intensity, the system can not only identify whether an anomaly exists at the moment, but also determine whether the anomaly is increasing, stabilizing, or weakening, shifting the monitoring logic from static judgment to dynamic process modeling. Nonlinear disturbance fusion logic enables the coordinated recognition of multiple minor behaviors, especially when the behavior amplitude is not high but the coupling is strong. It can accurately identify potential risks and solve the key pain point of traditional systems where "a single action is difficult to trigger an early warning." The exponential recognition mechanism and sliding window integral structure can quickly respond to disturbance surges, thereby triggering an early warning as soon as cheating behavior begins to trend, avoiding serious consequences and improving the system's ability to balance processing delays and misjudgment windows.

[0094] (4) By combining the classification label L with the behavioral response intensity RS, the system can record the behavioral level, response strategy, and risk factor of each abnormal event, thereby building a full-cycle behavioral data chain, providing a complete basis for subsequent manual verification, report archiving, and system learning. Not only does it complete the process from data extraction to risk identification, but it also realizes a closed-loop early warning control chain of "identification-classification-response-traceability", breaking through the limitation of traditional systems that can only "identify but cannot coordinate and process", greatly improving the actual control value of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0095] Figure 1 This is a block diagram of the intelligent examination monitoring system of the present invention;

[0096] Figure 2 This is a schematic diagram of the behavior mutation index process of the present invention;

[0097] Figure 3 This is a bar graph showing the distribution of the disturbance coupling index of the present invention. DETAILED DESCRIPTION

[0098] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0099] Example 1

[0100] The present invention provides an intelligent examination monitoring system, please refer to Figures 1 to 3 , including image time series sampling and high frame reconstruction module, multi-parameter micro-motion trajectory extraction module, disturbance coupling behavior recognition module, behavior jump window recognition module, behavior abnormality level judgment module and dynamic warning and behavior feedback module;

[0101] The image time series sampling and high-frame reconstruction module collects the video stream during the examination through a camera and performs high-frame rate interpolation and reconstruction to obtain a high-precision image sequence I(t);

[0102] The multi-parameter micro-motion trajectory extraction module extracts features from the high-precision image sequence I(t) and fits it into a temporal feature set PF;

[0103] The disturbance coupling behavior identification module performs time series fitting on the acquired time series feature set PF and constructs the disturbance coupling index RZH;

[0104] The behavior jump window identification module analyzes the disturbance mutation degree of the disturbance coupling index RZH within the time sliding window T, and performs abnormal mutation detection on the disturbance coupling index RZH to obtain the behavior mutation index Ψ;

[0105] The behavior abnormality level determination module compares the obtained behavior mutation index Ψ with the multi-level threshold, classifies the behavior risk level, and obtains the classification label L;

[0106] The dynamic warning and behavior feedback module executes a differentiated strategy based on the classification label L, behavior mutation index Ψ and disturbance coupling index RZH.

[0107] This embodiment introduces a mechanism for time-series image sampling and high-frame reconstruction to build a video information foundation with higher temporal precision, enabling accurate capture and identification of subtle behavioral perturbations. This addresses the existing system's inability to perceive subtle, short-lived, and highly covert behaviors. Rather than relying on single actions or local features, the system simultaneously extracts three types of micro-movement information from the image: head angle trajectory (Pot), shoulder perturbation rate (Pbs), and gaze jump amplitude (Pdm). This then constructs a unified time series feature set, resulting in recognition results with greater behavioral coverage and coupled expression capabilities, improving the overall accuracy of abnormal behavior detection.

[0108] Through the hierarchical construction of the disturbance coupling index RZH and the behavior mutation index Ψ, the present invention upgrades the behavior recognition level from "point judgment" to "segment trend assessment", so that the system can identify abnormal behavior patterns formed by continuous micro-perturbations, rather than relying solely on abnormal triggers of single actions. The system no longer uses static thresholds in the process of determining the abnormal level of behavior, but dynamically generates judgment criteria based on the actual disturbance distribution, and links the system feedback module through multi-level classification labels, thereby improving the adaptability to individual differences and enhancing the controllability and interpretability of the system. The dynamic warning module integrates the abnormal behavior time period, peak index and response level to build an integrated mechanism of behavior labeling, video marking, alarm triggering and report generation, supports post-event tracing and system learning updates, and enhances the closed-loop intelligent capabilities of the system.

[0109] This system is particularly suitable for identifying "non-disturbing cheating" behaviors, such as peeking at other people's test papers and communicating through micro-movements. It solves the problem of traditional systems being unable to identify such micro-perturbations and fills the monitoring blind spots in practical applications. Through time series fusion and coupling modeling, the system can accurately distinguish between "normal stress actions" and "abnormal behavior combinations," avoiding misjudging a candidate's natural head turns and blinks as abnormalities, thereby improving the credibility and practicality of the monitoring system. Utilizing the behavioral mutation index and window recognition mechanism, the system can identify the accumulation of potential risky behaviors, rather than relying solely on instantaneous action judgments. This system has stronger trend recognition capabilities, facilitating early warning and proactive intervention.

[0110] This system not only meets the needs of effectively identifying violations during exams, but also ensures a reasonable space for normal behavior of candidates, providing a high-precision, low-interference intelligent monitoring solution that meets the needs of "fair, safe, and accurate" monitoring in the modern intelligent education environment.

[0111] Example 2

[0112] This embodiment is explained in Example 1, please refer to Figure 2 ,Specifically: the image temporal sampling and high frame reconstruction module includes a ,motion vector estimation unit and a temporal image interpolation unit;

[0113] The motion vector estimation unit collects the video frame sequence in the video stream through the camera and extracts the motion vector field M between adjacent frames. The motion vector field M represents the spatial displacement trend of each pixel point (x, y) between consecutive frames.

[0114] The expression formula of the motion vector field M is as follows:

[0115] ;

[0116] Where ∂ represents the sign of the partial derivative, vi represents the brightness value of a pixel in the i-th frame image, uxi represents the horizontal motion velocity component of the pixel, and vyi represents the vertical motion velocity component of the pixel;

[0117] The temporal image interpolation unit is used to perform temporal interpolation processing on the video frame sequence based on the motion vector field M, insert multiple intermediate frames between adjacent frames, and obtain a high-precision image sequence I(t);

[0118] The pixel value of the intermediate frame is obtained by the following formula:

[0119] ;

[0120] Where vi(j)(x, y) represents the j-th interpolated frame between the i-th frame and the i+1-th frame, vi+1 represents the brightness value of a pixel in the i+1-th frame image, and λj represents the interpolation coefficient.

[0121] This embodiment incorporates a motion vector estimation unit during the image sampling phase. By analyzing the spatial displacement trends of each pixel between adjacent frames, a complete motion field representation structure is established. This improvement, based on the underlying image geometry, provides a continuous and realistic motion description foundation for subsequent image reconstruction and behavior extraction, rather than relying on static interpolation between frames.

[0122] This system uses a temporal image interpolation unit to interpolate the extracted motion vector field to generate interframe images, resulting in a high-precision image sequence I(t) with a higher temporal resolution than the original acquisition. This processing method avoids the "behavior blur" or "motion frame skipping" problems caused by sampling rate limitations in traditional systems, ensuring that small disturbances are accurately restored at the source. It is particularly suitable for identifying minor, short-term, and localized abnormal movements. Through the synergistic effect of the two sub-units of motion estimation and image interpolation, the system not only completes the data of the current video stream but also constructs a basic image layer for high-precision behavior recognition without replacing the camera equipment. This mechanism provides data layer support for subsequent modules and constitutes the key foundation of the entire intelligent monitoring system.

[0123] By constructing high-frame-rate image sequences, the system can accurately capture subtle movements during exams, such as slight head tilts, momentary jitters, and localized micro-movements, providing clear, continuous spatiotemporal cues for subsequent action recognition and risk assessment. Traditional monitoring systems, due to their fixed sampling frequency, are subject to the risk of inter-frame behavior jumps. This embodiment uses interpolation and padding to increase the effective frame rate to a higher range, significantly reducing the probability of false positives and missed detections. This approach is particularly suitable for capturing fleeting cheating behaviors during exams.

[0124] Example 3

[0125] This embodiment is explained in Example 2, please refer to Figure 1 ,Specifically: the multi-parameter micro-motion trajectory extraction module includes a ,structure posture tracking unit and a dynamic fitting unit;

[0126] The structural posture tracking unit performs dynamic recognition of human body structural areas based on the high-precision image sequence I(t) and extracts the examinee's posture points, including the head angle trajectory Pot, the shoulder disturbance rate Pbs, and the gaze jump amplitude Pdm.

[0127] The head angle trajectory Pot is obtained by the following formula:

[0128] ;

[0129] Where, (xno, yno) represents the coordinates of the nose tip point, (xne, yne) represents the coordinates of the neck point, and arctan represents the inverse tangent function;

[0130] The shoulder disturbance rate Pbs is obtained by the following formula:

[0131] ;

[0132] Where d represents the integral sign, SL represents the left shoulder vector, and SR represents the right shoulder vector;

[0133] The left shoulder vector SL is obtained by the following formula:

[0134] ;

[0135] Where, (xL, yL) represents the coordinates of the left shoulder point;

[0136] The right shoulder vector SR is obtained by the following formula:

[0137] ;

[0138] Where, (xR, yR) represents the coordinates of the right shoulder point;

[0139] The gaze transition amplitude Pdm is obtained by the following formula:

[0140] ;

[0141] Where (gx(t), gy(t)) represents the mapping coordinates of the gaze projection direction at time t on the image plane, and (gx(t-1), gy(t-1)) represents the mapping coordinates of the gaze projection direction at time t-1 on the image plane.

[0142] The dynamic fitting unit fits the acquired head angle trajectory Pot, shoulder disturbance rate Pbs, and gaze jump amplitude Pdm to obtain the initial feature set PC. It then uses a dynamic smoothing fitting algorithm to remove image noise fluctuations and obtain the temporal feature set PF.

[0143] The time series feature set PF is obtained by the following formula:

[0144] ;

[0145] Where PCo(ti) represents the oth item of data in the initial feature set PC at time ti, W represents the sliding window size, and PFo(t) represents the oth item of data in the time series feature set PF at time t.

[0146] This embodiment uses a structural posture tracking unit to simultaneously extract key posture points such as head angle, shoulder disturbance and gaze projection from the image sequence, and constructs a multi-dimensional disturbance information structure covering the face-upper body-visual direction. Compared with the traditional method of using only head posture or regional motion analysis, this module significantly enhances the system's ability to capture small abnormal movements by connecting multiple source trajectories in parallel. The system uses a vector construction method to continuously describe the shoulder state, and the spatial displacement of the left and right shoulder points relative to the neck point constitutes the left and right shoulder vectors, and uses this to calculate the shoulder disturbance rate. This spatial structuring method improves the system's perception of non-rigid movements and makes up for the defects of previous systems that ignore or roughly handle shoulder behavior.

[0147] The dynamic fitting unit not only collects preliminary data on head angle, gaze changes, and shoulder disturbances, but also introduces a sliding window smoothing mechanism to effectively suppress the interference of image noise and recognition jitter on the results, thereby generating a stable feature trajectory that can be used for subsequent coupling analysis and mutation identification. This time series modeling approach enables the system to not only detect "immediate anomalies" but also perceive "behavioral trends."

[0148] By integrating and extracting three key movements—head, shoulder, and gaze—this module significantly improves the system's ability to identify covert behaviors such as peeking at others' exam papers, subtle body signal transmission, and eye contact, thereby establishing a true "multi-parameter micro-perturbation recognition architecture." The introduction of a sliding window dynamic smoothing mechanism allows temporal features to retain behavioral trends while eliminating sudden fluctuations in values ​​caused by interference sources such as image jitter, lighting fluctuations, and key point recognition errors, thereby enhancing system robustness and reducing false positives.

[0149] Example 4

[0150] This embodiment is explained in Example 3, please refer to Figure 1 and Figure 3,Specifically: the disturbance coupling behavior identification module includes a ,disturbance parameter derivative extraction unit and a ,nonlinear disturbance fusion index construction unit;

[0151] The disturbance parameter derivative extraction unit extracts features from the temporal feature set PF, including the head angle change rate Vpo, shoulder disturbance acceleration Apb, and gaze jump change rate Vpd;

[0152] The head angle change rate Vpo is obtained by taking the first derivative of the head angle change rate Vpo at time t;

[0153] The shoulder disturbance acceleration Apb is obtained by taking the second derivative of the shoulder disturbance rate Pbs at time t;

[0154] The gaze jump change rate Vpd is obtained by taking the first-order derivative of the gaze jump amplitude Pdm at time t;

[0155] The nonlinear perturbation fusion index construction unit nonlinearly combines the acquired head angle change rate Vpo, shoulder perturbation acceleration Apb and gaze jump change rate Vpd to obtain the perturbation coupling index RZH;

[0156] The disturbance coupling index RZH is obtained by the following formula:

[0157] ;

[0158] Where, represent the preset weight values ​​of the head angle change rate Vpo, shoulder disturbance acceleration Apb and gaze jump change rate Vpd, respectively, and .

[0159] Specific examples:

[0160] Table 1: Nonlinear perturbation fusion index calculation table;

[0161] Group Number Head angle change rate Vpo Shoulder disturbance acceleration Apb Gaze jump change rate Vpd Perturbation coupling index RZH Number One 0.6 0.4 0.5 0.512 Number Two 0.3 0.6 0.7 0.533 Number Three 0.8 0.2 0.4 0.556 Number Four 0.5 0.7 0.6 0.601 Number Five 0.7 0.5 0.3 0.553

[0162] The behavior jump window recognition module includes a sliding window disturbance statistics unit and an exponential behavior recognition unit;

[0163] The sliding window disturbance statistics unit performs statistical processing on the disturbance coupling index RZH within the sliding time window, and dynamically calculates the average value μRZH and the fluctuation amplitude ΔRZH of the disturbance coupling index RZH;

[0164] The average value μRZH is obtained by the following formula:

[0165] ;

[0166] Where T represents the time sliding window size, RZH(ti) represents the disturbance coupling index at time ti;

[0167] The fluctuation amplitude ΔRZH is obtained by the following formula:

[0168] ;

[0169] Where RZH(t) represents the perturbation coupling index at time t, and μRZH(t) represents the average value at time t.

[0170] The exponential behavior recognition unit performs weighted integration of the disturbance coupling index RZH within the sliding window to obtain the behavior mutation index Ψ;

[0171] The behavioral mutation index Ψ is obtained by the following formula:

[0172] ;

[0173] Where Ψ(t) represents the behavior mutation index at time t, e represents a constant, γ represents the exponential enhancement coefficient, and d represents the integral sign.

[0174] This implementation significantly enhances the ability to perceive behavioral trends by analyzing the time derivative features of head angle, shoulder disturbance, and gaze jumps, including the head angle change rate Vpo, shoulder disturbance acceleration Apb, and gaze jump change rate Vpd. Unlike traditional systems that only extract static or instantaneous values, this module can identify the intensification and development trends of the examinee's movements, enabling the system to achieve "pre-emptive perception and dynamic response" behavioral judgment capabilities.

[0175] By nonlinearly fusing multiple perturbation derivative parameters, a single "perturbation coupling index" is formed to express the synergy of system micro-perturbations. This allows previously dispersed multi-dimensional posture data to reflect the risk patterns of cross-channel linkage behavior in fusion modeling, significantly enhancing the depth of recognition of complex cheating actions. The sliding window perturbation statistics unit dynamically models the system's behavioral performance within the current time period, obtaining the current stable level and fluctuation amplitude of the perturbation. This analysis mechanism can identify behavioral mutations such as sudden increases in perturbations in a short period of time and abnormal deviations from the coupling index. It provides a dual reference of steady-state baseline and perturbation deviation for behavioral risk assessment, enhancing the system's sensitivity to transient and accumulated anomalies.

[0176] The exponential behavior recognition unit performs a weighted integration of the disturbance index within a sliding window and uses an exponential enhancement structure to strengthen the system's response to rising disturbances. Compared to traditional linear integration, this method can more quickly detect turning points in disturbance trends and significantly improve the system's proactive warning capabilities for sudden behaviors.

[0177] By modeling parameters such as disturbance change rate, acceleration, and fluctuation intensity, the system can not only identify whether an anomaly is currently present, but also determine whether the anomaly is increasing, stabilizing, or weakening, shifting monitoring logic from static judgment to dynamic process modeling. Nonlinear disturbance fusion logic enables the coordinated recognition of multiple minor behaviors, particularly when the behaviors are low in magnitude but highly coupled. This accurately identifies potential risks and addresses the key pain point of traditional systems, where a single action struggles to trigger an early warning. The exponential recognition mechanism and sliding window integral structure provide rapid feedback on sudden disturbances, triggering an early warning as soon as cheating behavior begins to trend, avoiding serious consequences and improving the system's ability to balance processing latency with the false positive window.

[0178] Example 5

[0179] This embodiment is explained in Example 4. Please refer to Figure 1 and Figure 2 ,Specifically: the behavior abnormality level determination module includes a ,multilevel threshold dynamic generation unit and a multilevel behavior level ,classification unit;

[0180] The multi-level threshold dynamic generation unit generates real-time changing behavior judgment thresholds based on the characteristics of the behavior mutation index Ψ and combined with dynamic adjustment factors: index threshold THr1 and mutation threshold THr2;

[0181] The index threshold THr1 is obtained by the following formula:

[0182] ;

[0183] Where μΨ represents the mean of the mutation index, η1 represents the exponential adjustment coefficient, σΨ represents the standard deviation of the mutation index, C1 represents the exponential factor, E represents the exponential attenuation coefficient, and CVΨ represents the coefficient of variation, which is obtained by the ratio of the standard deviation σΨ of the mutation index to the mean μΨ of the mutation index;

[0184] The mutation threshold THr2 is obtained by the following formula:

[0185] ;

[0186] In the formula, η2 represents the mutation adjustment coefficient, C2 represents the mutation factor, ln represents the logarithmic function, G represents, D Ψ It represents the mutation amplitude difference, which is obtained by the difference between the maximum value of the mutation index Ψ and the minimum value of the mutation index Ψ. G represents the amplitude sensitivity factor.

[0187] The multi-level behavior classification unit compares the obtained index threshold THr1 and mutation threshold THr2 with the behavior mutation index Ψ in real time, classifies the behavior risk level, and obtains the classification label L;

[0188] Risk level classification is obtained by matching in the following ways:

[0189] When the behavior mutation index Ψ ≤ the index threshold THr1, it indicates the first risk level and normal state;

[0190] When the index threshold THr1 < behavior mutation index Ψ ≤ mutation threshold THr2, it indicates the second risk level, suspicious state;

[0191] When the mutation threshold THr2 is less than the behavior mutation index Ψ, it indicates the third risk level, an abnormal state;

[0192] The classification label L is obtained by matching in the following way;

[0193] When the risk level is at the first risk level, the classification label L = 0, indicating a normal behavior state, and the system only records no response;

[0194] When the risk level is at the second risk level, the classification label L=1, indicating a suspicious state, and the system issues a soft warning.

[0195] When the risk level is at the third risk level, the classification label L=2, indicating a serious abnormality, the system triggers an alarm, video marking and manual review.

[0196] The dynamic warning and behavior feedback module uses the behavior label L as the main trigger condition, and combines the behavior mutation index Ψ and the disturbance coupling index RZH to obtain the behavior response strength RS and implement differentiated response strategies;

[0197] The behavioral response intensity RS is obtained by the following formula:

[0198] ;

[0199] Where H1 represents the classification response coefficient, H2 represents the behavior mutation weight coefficient, H3 represents the coupling index response coefficient, and L(t) represents the classification label at time t. represents the indicator function, which is 1 if the condition is met and 0 otherwise, and θRZH represents the abnormal threshold of coupling strength;

[0200] Differentiated response strategies are matched and obtained in the following ways:

[0201] When the behavioral response intensity RS=0, no strategy is executed;

[0202] When the behavioral response intensity RS = 1, the differentiated response strategy is behavioral recording;

[0203] When the behavioral response intensity RS=2, the differentiated response strategy is video marking and soft prompting;

[0204] When the behavioral response intensity RS ≥ 3, the differentiated response strategies are alarm, report generation, and video locking.

[0205] This embodiment adopts a dynamic threshold construction unit to adjust the behavior judgment threshold in real time according to different scenarios and different candidate states. Through the dynamic linkage of the mean, standard deviation, coefficient of variation and mutation amplitude difference, the exponential threshold and mutation threshold are automatically generated to achieve adaptive recognition of the behavior mutation index Ψ. This mechanism overcomes the problems of poor adaptability and high misjudgment rate caused by the use of static fixed thresholds in traditional systems. This module divides the behavioral risk level based on two dynamic thresholds, and identifies the candidate status from "normal" to "suspicious" to "abnormal" in layers, which improves the granularity and scientific classification of the behavior recognition system. Compared with the traditional two-classification system, this mechanism effectively reduces the false positives of slightly abnormal behaviors and improves the fault tolerance of the system.

[0206] The dynamic early warning module uses the classification label L as its core trigger condition and integrates two key risk parameters: the mutation index Ψ and the coupling index RZH to construct the behavioral response strength index RS. This comprehensive response model automatically triggers different strategies based on the response level, achieving an intelligent transition from passive identification to active response. A multi-level threshold mechanism effectively classifies the risk level of micro-perturbations. This not only identifies significant cheating actions but also appropriately labels and pre-records suspicious behaviors in "critical behavior states," ensuring the monitoring system's application of a "properly lenient and strict" strategy.

[0207] By combining the classification label L with the behavioral response strength RS, the system records the behavioral level, response strategy, and risk factors for each abnormal event, thereby building a full-cycle behavioral data chain, providing a complete basis for subsequent manual verification, report archiving, and system learning. This not only completes the process from data extraction to risk identification, but also realizes a closed-loop early warning and control chain of "identification-classification-response-traceability," breaking through the limitation of traditional systems that can only "identify but not coordinate and process," significantly enhancing the system's actual control value.

[0208] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. Intelligent examination monitoring system, characterized by: It includes image time series sampling and high-frame reconstruction module, multi-parameter micro-motion trajectory extraction module, disturbance coupling behavior recognition module, behavior jump window recognition module, behavior abnormality level determination module and dynamic warning and behavior feedback module; The image time series sampling and high-frame reconstruction module collects the video stream during the examination through a camera and performs high-frame rate interpolation and reconstruction to obtain a high-precision image sequence I(t); The multi-parameter micro-motion trajectory extraction module extracts features from the high-precision image sequence I(t) and fits it into a temporal feature set PF; The disturbance coupling behavior identification module performs time series fitting on the acquired time series feature set PF and constructs the disturbance coupling index RZH; The behavior jump window identification module analyzes the disturbance mutation degree of the disturbance coupling index RZH within the time sliding window T, and performs abnormal mutation detection on the disturbance coupling index RZH to obtain the behavior mutation index Ψ; The behavior abnormality level determination module compares the obtained behavior mutation index Ψ with the multi-level threshold, classifies the behavior risk level, and obtains the classification label L; The dynamic warning and behavior feedback module executes a differentiated strategy based on the classification label L, behavior mutation index Ψ and disturbance coupling index RZH.

2. The intelligent examination monitoring system according to claim 1, characterized in that: The image temporal sampling and high frame reconstruction module includes a motion vector estimation unit and a temporal image interpolation unit; The motion vector estimation unit collects the video frame sequence in the video stream through the camera and extracts the motion vector field M between adjacent frames. The motion vector field M represents the spatial displacement trend of each pixel point (x, y) between consecutive frames. The expression formula of the motion vector field M is as follows: ； Where ∂ represents the sign of the partial derivative, vi represents the brightness value of a pixel in the i-th frame image, uxi represents the horizontal motion velocity component of the pixel, and vyi represents the vertical motion velocity component of the pixel; The temporal image interpolation unit is used to perform temporal interpolation processing on the video frame sequence based on the motion vector field M, insert multiple intermediate frames between adjacent frames, and obtain a high-precision image sequence I(t); The pixel value of the intermediate frame is obtained by the following formula: ； Where vi(j)(x, y) represents the j-th interpolated frame between the i-th frame and the i+1-th frame, vi+1 represents the brightness value of a pixel in the i+1-th frame image, and λj represents the interpolation coefficient.

3. The intelligent examination monitoring system according to claim 2, characterized in that: The multi-parameter micro-motion trajectory extraction module includes a structure posture tracking unit and a dynamic fitting unit; The structural posture tracking unit performs dynamic recognition of human body structural areas based on the high-precision image sequence I(t) and extracts the examinee's posture points, including the head angle trajectory Pot, the shoulder disturbance rate Pbs, and the gaze jump amplitude Pdm. The head angle trajectory Pot is obtained by the following formula: ； Where, (xno, yno) represents the coordinates of the nose tip point, (xne, yne) represents the coordinates of the neck point, and arctan represents the inverse tangent function; The shoulder disturbance rate Pbs is obtained by the following formula: ； Where d represents the integral sign, SL represents the left shoulder vector, and SR represents the right shoulder vector; The left shoulder vector SL is obtained by the following formula: ； Where, (xL, yL) represents the coordinates of the left shoulder point; The right shoulder vector SR is obtained by the following formula: ； Where, (xR, yR) represents the coordinates of the right shoulder point; The gaze transition amplitude Pdm is obtained by the following formula: ； Where (gx(t), gy(t)) represents the mapping coordinates of the gaze projection direction at time t on the image plane, and (gx(t-1), gy(t-1)) represents the mapping coordinates of the gaze projection direction at time t-1 on the image plane.

4. The intelligent examination monitoring system according to claim 3, characterized in that: The dynamic fitting unit fits the acquired head angle trajectory Pot, shoulder disturbance rate Pbs, and gaze jump amplitude Pdm to obtain the initial feature set PC. It then uses a dynamic smoothing fitting algorithm to remove image noise fluctuations and obtain the temporal feature set PF. The time series feature set PF is obtained by the following formula: ； Where PCo(ti) represents the oth item of data in the initial feature set PC at time ti, W represents the sliding window size, and PFo(t) represents the oth item of data in the time series feature set PF at time t.

5. The intelligent examination monitoring system according to claim 4, characterized in that: The disturbance coupling behavior identification module includes a disturbance parameter derivative extraction unit and a nonlinear disturbance fusion index construction unit; The disturbance parameter derivative extraction unit extracts features from the temporal feature set PF, including the head angle change rate Vpo, shoulder disturbance acceleration Apb, and gaze jump change rate Vpd; The head angle change rate Vpo is obtained by taking the first derivative of the head angle change rate Vpo at time t; The shoulder disturbance acceleration Apb is obtained by taking the second derivative of the shoulder disturbance rate Pbs at time t; The gaze jump change rate Vpd is obtained by taking the first-order derivative of the gaze jump amplitude Pdm at time t; The nonlinear perturbation fusion index construction unit nonlinearly combines the acquired head angle change rate Vpo, shoulder perturbation acceleration Apb and gaze jump change rate Vpd to obtain the perturbation coupling index RZH; The disturbance coupling index RZH is obtained by the following formula: ； Where, represent the preset weight values ​​of the head angle change rate Vpo, shoulder disturbance acceleration Apb and gaze jump change rate Vpd, respectively, and .

6. The intelligent examination monitoring system according to claim 5, characterized in that: The behavior jump window recognition module includes a sliding window disturbance statistics unit and an exponential behavior recognition unit; The sliding window disturbance statistics unit performs statistical processing on the disturbance coupling index RZH within the sliding time window, and dynamically calculates the average value μRZH and the fluctuation amplitude ΔRZH of the disturbance coupling index RZH; The average value μRZH is obtained by the following formula: ； Where T represents the time sliding window size, RZH(ti) represents the disturbance coupling index at time ti; The fluctuation amplitude ΔRZH is obtained by the following formula: ； Where RZH(t) represents the perturbation coupling index at time t, and μRZH(t) represents the average value at time t.

7. The intelligent examination monitoring system according to claim 6, characterized in that: The exponential behavior recognition unit performs weighted integration of the disturbance coupling index RZH within the sliding window to obtain the behavior mutation index Ψ; The behavioral mutation index Ψ is obtained by the following formula: ； Where Ψ(t) represents the behavior mutation index at time t, e represents a constant, γ represents the exponential enhancement coefficient, and d represents the integral sign.

8. The intelligent examination monitoring system according to claim 7, characterized in that: The behavior abnormality level determination module includes a multi-level threshold dynamic generation unit and a multi-level behavior level classification unit; The multi-level threshold dynamic generation unit generates real-time changing behavior judgment thresholds based on the characteristics of the behavior mutation index Ψ and combined with dynamic adjustment factors: index threshold THr1 and mutation threshold THr2; The index threshold THr1 is obtained by the following formula: ； Where μΨ represents the mean of the mutation index, η1 represents the exponential adjustment coefficient, σΨ represents the standard deviation of the mutation index, C1 represents the exponential factor, E represents the exponential attenuation coefficient, and CVΨ represents the coefficient of variation, which is obtained by the ratio of the standard deviation σΨ of the mutation index to the mean μΨ of the mutation index; The mutation threshold THr2 is obtained by the following formula: ； In the formula, η2 represents the mutation adjustment coefficient, C2 represents the mutation factor, ln represents the logarithmic function, G represents, D Ψ It represents the mutation amplitude difference, which is obtained by the difference between the maximum value of the mutation index Ψ and the minimum value of the mutation index Ψ. G represents the amplitude sensitivity factor.

9. The intelligent examination monitoring system according to claim 8, characterized in that: The multi-level behavior classification unit compares the obtained index threshold THr1 and mutation threshold THr2 with the behavior mutation index Ψ in real time, classifies the behavior risk level, and obtains the classification label L; Risk level classification is obtained by matching in the following ways: When the behavior mutation index Ψ ≤ the index threshold THr1, it indicates the first risk level and normal state; When the index threshold THr1 < behavior mutation index Ψ ≤ mutation threshold THr2, it indicates the second risk level, suspicious state; When the mutation threshold THr2 is less than the behavior mutation index Ψ, it indicates the third risk level, an abnormal state; The classification label L is obtained by matching in the following way; When the risk level is at the first risk level, the classification label L = 0, indicating a normal behavior state, and the system only records no response; When the risk level is at the second risk level, the classification label L=1, indicating a suspicious state, and the system issues a soft warning. When the risk level is at the third risk level, the classification label L=2, indicating a serious abnormality, the system triggers an alarm, video marking and manual review.

10. The intelligent examination monitoring system according to claim 1, characterized in that: The dynamic warning and behavior feedback module uses the behavior label L as the main trigger condition, and combines the behavior mutation index Ψ and the disturbance coupling index RZH to obtain the behavior response strength RS and implement differentiated response strategies; The behavioral response intensity RS is obtained by the following formula: ； Where H1 represents the classification response coefficient, H2 represents the behavior mutation weight coefficient, H3 represents the coupling index response coefficient, and L(t) represents the classification label at time t. represents the indicator function, which is 1 if the condition is met and 0 otherwise, and θRZH represents the abnormal threshold of coupling strength; Differentiated response strategies are matched and obtained in the following ways: When the behavioral response intensity RS=0, no strategy is executed; When the behavioral response intensity RS = 1, the differentiated response strategy is behavioral recording; When the behavioral response intensity RS=2, the differentiated response strategy is video marking and soft prompting; When the behavioral response intensity RS ≥ 3, the differentiated response strategies are alarm, report generation, and video locking.

Images