CN-122023078-A - Anti-cheating monitoring method for online examination system

CN122023078ACN 122023078 ACN122023078 ACN 122023078ACN-122023078-A

Abstract

The invention discloses an anti-cheating monitoring method of an online examination system, which relates to the technical field of anti-cheating of examination, and comprises the steps of synchronously collecting an audio and video of an examination room through a multi-view pick-up camera array, positioning sound source coordinates through audio noise reduction, human voice screening and combining time delay difference and loudness attenuation law, mapping a cheating area to a multi-view video pixel area through space coordinate conversion, unifying perspective conversion to a main view coordinate system to realize cross-modal association of the audio and video, respectively designing sub-models for audio, facial and action characteristics, carrying out independent cheating suspicion screening, dynamically distributing three types of characteristic weights through a multi-head attention mechanism to fuse comprehensive cheating probability, combining time sequence association analysis to identify collective cheating and monitoring collaborative cheating, encrypting and uploading cheating judgment results and audio and video evidence to a alliance chain storage certificate, generating a tamper-proof evidence chain and pushing a grading early warning. The method effectively improves the accuracy and scene suitability of cheating identification, and provides technical guarantee for online examination fairness.

Inventors

HUA ZHONGHAI
WU JIAN
LIU DENGFENG
ZHANG SHUO

Assignees

山东乐闻信息科技有限公司

Dates

Publication Date: 20260512
Application Date: 20260202

Claims (10)

1. The anti-cheating monitoring method of the online examination system is characterized by comprising the following steps of: collecting real-time audio and video in an examination room through a multi-view pick-up camera array, and obtaining a multi-view audio and video set through unified time stamp calibration; Performing noise suppression and voice activity detection on the audio stream in the multi-view audio-video set, extracting voice features, screening cheating related voice, positioning a cheating related region by combining multi-view voice delay differences and loudness attenuation rules, and storing the cheating related voice features and the region coordinates in an associated manner to form an audio feature set; Mapping the cheating related region coordinates to a multi-view video pixel region, intercepting corresponding image data and unifying the image data to a main view coordinate system, and extracting facial features and action features of personnel in the region based on a target detection algorithm to form a regional personnel feature set; Constructing a cheating judgment model based on a deep learning framework, inputting an audio feature set and a regional personnel feature set into the model for independent cheating suspicion screening, dynamically distributing weights through a multi-head attention mechanism, optimizing dynamic weight vectors by adopting a particle swarm optimization algorithm, finally fusing to obtain comprehensive cheating suspicion probability, and determining the cheating type by combining time sequence association analysis to obtain a cheating judgment result; and (3) associating the cheating judgment result with the audio and video stream and the judgment log to form an evidence packet, encrypting the evidence packet, uploading the evidence packet to a alliance chain storage certificate, generating hierarchical early warning information and pushing the hierarchical early warning information to a designated terminal.
2. The anti-cheating monitoring method of an online examination system of claim 1, wherein the process of unifying the positioning of the cheating-related area and the multi-view image to the main view coordinate system comprises: Calculating time delay differences among the multi-view pick-up cameras based on the cross-correlation function to derive multi-view distance difference information sets for the screened cheating related voice; Measuring the actual measurement loudness under each view angle through an audio acquisition device, establishing a mapping relation between the loudness and the sound source distance according to the inverse square law of spherical wave propagation and the sound level-sound intensity logarithmic relation, and reversely pushing the sound source distance range by combining an environmental attenuation correction coefficient; And cross-verifying and positioning the cheating related area through the multi-view angle distance difference information set and the sound source distance range, calling the pre-calibrated pickup camera parameters to map the three-dimensional coordinates to each view angle video pixel area, intercepting corresponding image data after distortion correction, and then executing perspective transformation through a homography matrix to be unified to a main view angle coordinate system.
3. The method for anti-cheating monitoring of an online examination system according to claim 1, wherein the construction of the cheating judgment model and the calculation process of the comprehensive cheating suspicion probability comprise the following steps: Constructing a cheating judgment model containing three types of special sub-models, wherein an audio feature screening sub-model adopts a light convolutional neural network architecture, a facial feature screening sub-model adopts a CNN+transducer encoder architecture, and an action feature screening sub-model adopts a gating circulation unit architecture; And splicing the three types of features into a joint feature matrix through linear mapping, splitting the matrix according to the number of preset attention heads, calculating the feature association strength through dot product attention, splicing the output of each attention head, generating a dynamic weight vector through linear transformation, and obtaining the comprehensive cheating suspicion probability by weighting and fusing the single feature cheating suspicion probability based on the vector.
4. A method of anti-cheating monitoring for an online examination system as claimed in claim 1 or 3, wherein the determining of the type of cheating comprises: when the comprehensive cheating suspicion probability exceeds a threshold value, associating multi-person characteristic data and audio data of the cheating related area and the surrounding area under the same time stamp, and extracting a time sequence data slice covering the cheating period; Calculating facial feature cosine similarity and action feature dynamic time regular similarity among people, verifying a collective cheating candidate group, calculating audio homology through the Mel frequency spectrum feature cosine similarity, verifying a time sequence calling relationship, and confirming collective cheating; and combining the identity characteristics of the prisoner, calculating the characteristic correlation between the prisoner and the cheating personnel, and verifying the cooperative cheating of the prisoner.
5. The anti-cheating monitoring method of an online examination system of claim 1, wherein the screening process of the cheating-related voice comprises: Filtering environmental noise on the audio stream by adopting a recursive least square adaptive noise suppression algorithm, performing voice activity detection by a double-threshold judgment method combining an energy threshold and a zero crossing rate threshold, and removing a mute section to reserve effective voice; and windowing the effective voice, extracting Mel frequency spectrum characteristics through fast Fourier transform, synchronously extracting acoustic characteristics of fundamental frequency and speech speed, splicing the acoustic characteristics into characteristic vectors, inputting the characteristic vectors into a pre-trained CNN-LSTM model, and screening cheating related voice.
6. The anti-cheating monitoring method of an online examination system according to claim 1 or 2, wherein the forming process of the regional personnel feature set comprises: For the image unified to the main view angle coordinate system, adopting a self-adaptive histogram equalization algorithm to improve local contrast, and combining with a Retinex algorithm to separate illumination and reflection components so as to enhance image details; locating personnel in the image area through a target detection algorithm, extracting facial feature vectors of the personnel by adopting a face recognition model, extracting skeleton point coordinates of the head, the shoulder, the elbow and the wrist through a human body posture estimation algorithm, calculating a motion track, a joint angle and a motion frequency to generate motion feature vectors, and classifying the motion feature vectors according to personnel ID to form a regional personnel feature set.
7. The anti-cheating monitoring method of an online examination system of claim 1, wherein the encrypting and uploading process of the evidence package comprises: The cheating judgment result is associated with the corresponding audio/video fragment and the judgment process log through a unique evidence packet ID, and is packaged into an index and data format complete evidence packet; And (3) encrypting the evidence packet by adopting an SM4 national encryption algorithm, uploading the evidence packet to a alliance chain formed by nodes of an examination organization party, a supervision organization and examination points together, and ensuring the safety of evidence transmission and storage.
8. The anti-cheating monitoring method of an online examination system according to claim 1 or 7, wherein the evidence chain generation and early warning information pushing process comprises the following steps: Triggering an intelligent contract built in the alliance chain after uploading the evidence packet, automatically recording the hash value of the evidence packet, uploading time and node identity information by the intelligent contract, and generating an untampereable evidence chain; The intelligent contract is matched with the early warning level according to the confidence coefficient of the cheating judgment result, and early warning information comprising the early warning level, the evidence chain inquiry address, the cheating core information and the processing advice is generated and directionally pushed to the appointed terminal of the examination organization party, the supervision institution and the examination point.
9. The anti-cheating monitoring method of an online examination system according to claim 2, wherein the process of back-pushing of the sound source distance range and locating the cheating-related area comprises the steps of: Based on the inverse square law of spherical wave propagation, namely the square attenuation of sound intensity along with distance, and combining the logarithmic relation between sound level and sound intensity, introducing an environmental attenuation correction coefficient, and establishing a mapping relation between actual measurement loudness and sound source distance; And (3) reversely deducing the minimum and maximum value intervals of the sound source distance through the mapping relation, performing cross-validation on the minimum and maximum value intervals and the multi-view distance difference information set obtained based on the cross-correlation function, screening coordinate subsets meeting all view distance ranges, taking intersection sets, and finally positioning the coordinate subsets to the cheating related area.
10. The anti-cheating monitoring method of an online examination system of claim 1, wherein the process of acquiring the multi-view audio-video set comprises: the multi-view pick-up camera array adopts a layout combining full coverage of a main view and an auxiliary view, and each pick-up camera is matched with an integrated pick-up microphone to synchronously collect audio; Accessing a local server through a power over ethernet switch, and completing unified timestamp calibration by adopting a network time protocol; And high-efficiency video coding is adopted for video streams, advanced audio coding compression is adopted for audio streams, and the video streams are named and stored according to a pick-up camera ID-timestamp rule to form a multi-view audio-video set covering a full examination room.

Description

Anti-cheating monitoring method for online examination system Technical Field The invention relates to the technical field of anti-cheating examination, in particular to an anti-cheating monitoring method of an online examination system. Background With the wide application of online examination modes in the fields of academic education, professional qualification certification and the like, the limitation of anti-cheating monitoring technology is increasingly prominent, and mainly two core technology pain points exist, namely, on one hand, the audio and video information is analyzed independently, a space association mechanism is lacked, abnormal signals can only be captured roughly, the cheating main body can not be locked through sound source positioning and video pixel region accurate binding, the problems of audio and video disjointing and cheating region positioning ambiguity exist, on the other hand, the cheating judgment is dependent on a single mode characteristic or a fixed weight fusion mode, characteristic differences of different cheating behaviors cannot be adapted, for example, a transfer slip takes limb actions as core characteristics, secret number communication takes audio as core characteristics, the importance of combining multiple faces and action association characteristics is needed for cooperative peeping, the importance of each characteristic can not be adjusted dynamically, and the analysis capability of complex scenes such as collective cheating, collaborative cheating and the like is lacked, so that the cheating judgment rate is higher, and the online examination requirement on large scale and high fairness is difficult to meet. Disclosure of Invention The invention aims to provide an anti-cheating monitoring method of an online examination system, which achieves the purposes of improving the cheating recognition precision, strengthening evidence public confidence and maintaining online examination order by synchronously processing and positioning cheating areas through audio and video, dynamically weighting, fusing and judging cheating and alliance chain evidence storage early warning. The technical scheme for realizing the purpose of the invention is as follows: A cheating prevention monitoring method of an online examination system comprises the following steps: collecting real-time audio and video in an examination room through a multi-view pick-up camera array, and obtaining a multi-view audio and video set through unified time stamp calibration; Performing noise suppression and voice activity detection on an audio stream in an audio-video set, extracting voice characteristics, screening cheating related voice, positioning a cheating related area by combining multi-view voice delay difference and loudness attenuation law, and storing the cheating related voice characteristics and the area coordinates in an associated manner to form an audio characteristic set; Mapping the cheating related region coordinates to a multi-view video pixel region, intercepting corresponding image data and unifying the image data to a main view coordinate system, and extracting facial features and action features of personnel in the region based on a target detection algorithm to form a regional personnel feature set; The method comprises the steps of constructing a cheating judgment model based on a deep learning framework, inputting an audio feature set and a regional personnel feature set into the model to perform independent cheating suspicion screening, dynamically distributing weights through a multi-head attention mechanism to integrate cheating suspicion probability, and combining time sequence association analysis to determine the cheating type to obtain a cheating judgment result; and (3) associating the cheating judgment result with the audio and video stream and the judgment log to form an evidence packet, encrypting the evidence packet, uploading the evidence packet to a alliance chain storage certificate, generating hierarchical early warning information and pushing the hierarchical early warning information to a designated terminal. Further, aiming at the screened cheating related voice, firstly aligning the voice segments collected by the multi-view pick-up camera according to a unified timestamp, intercepting a fixed-length audio segment meeting the time delay difference calculation requirement, traversing all time offsets based on a cross-correlation function, calculating a signal similarity peak value among the multi-view pick-up cameras, precisely acquiring time delay difference data, further deriving distance difference information from a sound source to each camera by combining standard sound velocity at normal temperature, measuring equivalent continuous sound level at each view angle as actual measurement loudness through a loudness meter module arranged in an audio collection device, pre-collecting an environmental attenuation correction coefficient based on the acoustic c