An examination cheating intelligent detection method and system based on behavior recognition

Abstract

The present application relates to a kind of based on behavior recognition's examination cheating intelligent detection method and system, it is related to examination cheating technical field, including the following steps, to examination video stream extraction continuous image frame, carry out human body area detection and obtain examinee human body frame;Based on the frame positioning pose key point coordinate, constructs upper body action trajectory sequence;Accordingly, identify suspicious action fragment, if there is, then in the time period to adjacent examinee spatial relative position carries out dynamic correlation analysis, generates the multi-person collaborative behavior relationship diagram;When at least two examinees in the diagram present consistent abnormal action mode in the same time period, it is judged as suspected cheating behavior, and output contains the alarm information of time stamp and the examinee identification involved in, solve the technical problem that traditional technology often neglects the spatial interaction between examinees in examination room environment, it is difficult to effectively identify complex cheating mode.

Description

Technical Field

[0001] This invention relates to the field of exam cheating technology, and in particular to an intelligent method and system for detecting exam cheating based on behavior recognition. Background Technology

[0002] With the widespread application of online examinations and intelligent proctoring systems, traditional methods relying on manual proctoring are no longer sufficient to meet the demands for large-scale, efficient, and accurate examination supervision. Especially in scenarios such as distance education and qualification certification, cheating behaviors are becoming increasingly covert, collaborative, and diversified, rendering detection methods relying solely on video playback or simple rule-based judgments increasingly inadequate. Therefore, there is an urgent need to introduce artificial intelligence technology, particularly vision-based behavior recognition methods, to achieve automatic perception and intelligent analysis of abnormal candidate behavior.

[0003] Current mainstream cheating detection methods mostly focus on analyzing isolated actions of a single examinee, such as frequent head-down movements and shifting gaze, lacking the ability to model potential collaborative cheating behaviors among multiple individuals. These methods often overlook the spatial interactions between examinees in the examination environment, making it difficult to effectively identify complex cheating patterns requiring multiple-person cooperation, such as passing notes, using coded gestures, or simultaneously flipping through exam papers. Furthermore, existing systems have relatively coarse modeling of the temporal characteristics of actions when processing continuous video streams, easily misclassifying normal behavior as abnormal, resulting in a high false alarm rate and impacting practical application effectiveness. Summary of the Invention

[0004] The purpose of this invention is to at least partially solve one of the technical problems existing in the prior art.

[0005] To achieve the above objectives, this invention provides an intelligent detection method for exam cheating based on behavior recognition, comprising the following steps: The examination room video stream is subjected to frame sequence extraction to obtain continuous image frames, and human body region detection is performed based on the image frames to obtain the candidate's human body bounding box; Based on the candidate's body bounding box, the candidate's posture key points in each image frame are located to obtain the posture key point coordinates, and an upper body movement trajectory sequence is constructed based on the posture key point coordinates. Based on the upper body movement trajectory sequence, determine whether there are any suspicious movement segments. If so, perform dynamic correlation analysis on the spatial relative positions between adjacent candidates based on the suspicious movement segments to obtain a multi-person collaborative behavior relationship diagram. When at least two candidates exhibit consistent abnormal behavior patterns within the same time period in the multi-person collaborative behavior relationship graph, it is determined that suspected cheating behavior has occurred within the corresponding time period, and a cheating alarm message containing a timestamp and the identifier of the candidate involved is output.

[0006] Furthermore, frame sequence extraction is performed on the examination room video stream to obtain continuous image frames, and human body region detection is performed based on the continuous image frames to obtain the candidate's human body bounding box, including: The examination room video stream is processed by a preset video stream parser to extract frame sequences, resulting in continuous image frames. The continuous image frames are then preprocessed to obtain a preprocessed image. The preprocessed image is initially screened for human body regions to obtain candidate human body region maps. The candidate human body region maps are then subjected to edge detection and contour closure processing to obtain closed contour human body maps. The closed contour human body image is verified for authenticity to obtain a real human body region image. Based on the real human body region image, the minimum bounding rectangle is used to define the candidate's human body frame.

[0007] Furthermore, edge detection and contour closure processing are performed on the candidate body region map to obtain a closed contour human body map, including: Gradient calculation is performed on the candidate body region map to obtain an edge intensity map, and threshold segmentation is performed on the edge intensity map to obtain an initial edge map; Based on the initial edge map, breakpoints are connected to obtain an edge link map, and the edge link map is filled with regions to obtain a complete contour map; The complete contour map is smoothed to obtain a smooth contour map, and the smooth contour map is validated by region to obtain a closed contour human body map.

[0008] Furthermore, based on the candidate's body bounding box, the pose key points of the candidate in each image frame are located to obtain the coordinates of the pose key points, including: The candidate's body frame is divided into regions to obtain an upper body region image, and the contrast of the upper body region image is enhanced to obtain an enhanced upper body image. Feature point detection is performed on the enhanced upper body image to obtain candidate joint points, and false points are removed from the candidate joint points to obtain the filtered joint points. Based on the selected joint points, the posture key point coordinates of the examinee in each image frame are located and mapped to obtain the joint point coordinates. The joint point coordinates are then classified and organized to obtain the posture key point coordinates.

[0009] Furthermore, based on the selected key points, the pose key point coordinates of the examinee in each image frame are located and mapped to obtain the key point coordinates, including: After filtering, the pixel coordinates of the joint points are extracted to obtain a set of joint pixels. The coordinate system of the set of joint pixels is then calibrated to obtain standard coordinate points. The connection relationship of the human skeleton is established based on the standard coordinate points to generate a skeleton connection diagram, and the skeleton connection diagram is scaled and normalized to obtain a standard skeleton diagram. Based on the standard skeleton diagram, key points of the examinees in each image frame are paired to obtain a pose node group, and the pose node group is subjected to spatial projection transformation to obtain a mapping coordinate set. Based on the mapped coordinate set, joint points are classified and labeled to obtain a joint type table. Then, the joint type table is integrated with coordinates to obtain joint point coordinates.

[0010] Furthermore, based on the coordinates of the aforementioned posture key points, a sequence of upper body movement trajectories is constructed, including: The coordinates of the attitude key points are sorted according to frame time sequence to obtain a time sequence coordinate column, and the time sequence coordinate column is interpolated and completed to obtain a completed coordinate column; The completed coordinate column is split according to joint type to obtain single joint coordinate groups, and the single joint coordinate groups are connected by continuous line segments to obtain single joint trajectory lines; The single-joint trajectory line is aligned with the time axis to obtain the aligned trajectory line, and the aligned trajectory line is combined frame by frame to obtain the upper body movement trajectory sequence.

[0011] Furthermore, based on the suspicious action segments, a dynamic correlation analysis is performed on the spatial relative positions between adjacent candidates to obtain a multi-person collaborative behavior relationship diagram, including: Based on the candidates in the suspicious action segments, the positional distances between the candidates and their corresponding adjacent candidates are calculated to obtain the candidate position matrix. Then, the proximity relationship analysis of the candidate position matrix is ​​performed to obtain the candidate adjacency graph. Based on the candidate's adjacency graph, temporal action matching is performed to obtain the action correlation matrix; The action correlation matrix is ​​divided into time windows to obtain a time period correlation table, and the frequency of the time period correlation table is statistically analyzed to obtain a correlation strength map. Group behavior analysis is performed based on the association strength map to obtain a behavior cluster map, and the behavior cluster map is spatiotemporally integrated to obtain a multi-person collaborative behavior relationship map.

[0012] Furthermore, based on the candidate adjacency graph, temporal action matching is performed to obtain an action correlation matrix, including: Trajectory extraction is performed on candidates and their corresponding adjacent candidates in the candidate adjacency graph. The upper body movement trajectory segments of adjacent candidate pairs within the suspicious action segments are obtained. The coordinate change sequence of all upper body posture key points in each upper body movement trajectory segment is extracted according to the time axis to obtain the temporal characteristics of the candidate's actions. Among them, the adjacent candidate pairs are candidate combinations determined based on the spatial positional relationship of the candidate's human body frame. The time axis is aligned for the coordinate change sequence of each key point of the candidate's action sequence characteristics. The start time difference, duration ratio and motion direction overlap of different parts of the action are calculated to obtain the action coordination parameters. The motion coordination parameters are weighted and calculated to obtain the motion correlation degree of adjacent candidates in each time window. The motion correlation degree is then arranged in a matrix according to the correspondence between adjacent candidates and time windows to obtain the motion correlation matrix.

[0013] This invention also provides an intelligent examination cheating detection system based on behavior recognition, comprising: The extraction module is used to extract frame sequences from the examination room video stream to obtain continuous image frames, and to perform human body region detection based on the image frames to obtain the candidate's human body bounding box. The positioning module is used to locate the posture key points of the examinee in each image frame based on the examinee's human body frame, obtain the coordinates of the posture key points, and construct the upper body movement trajectory sequence based on the coordinates of the posture key points. The determination module is used to determine whether there are suspicious action segments based on the upper body movement trajectory sequence. If there are, the module performs dynamic correlation analysis on the spatial relative positions between adjacent candidates based on the suspicious action segments to obtain a multi-person collaborative behavior relationship diagram. The determination module is used to determine that suspected cheating has occurred in the corresponding time period when at least two candidates exhibit the same abnormal action pattern in the multi-person collaborative behavior relationship graph, and outputs cheating alarm information including timestamp and the identifier of the candidate involved.

[0014] This invention provides an intelligent detection method for exam cheating based on behavior recognition, comprising the following steps: extracting frame sequences from an exam video stream to obtain continuous image frames, and performing human body region detection based on the image frames to obtain a candidate's body bounding box; locating posture key points of the candidates in each image frame based on the candidate's body bounding box to obtain posture key point coordinates, and constructing an upper body motion trajectory sequence based on the posture key point coordinates; determining whether there are suspicious action segments based on the upper body motion trajectory sequence; if so, performing dynamic correlation analysis on the spatial relative positions between adjacent candidates based on the suspicious action segments to obtain a multi-person collaborative behavior relationship graph; when the multi-person... When at least two candidates exhibit consistent abnormal action patterns within the same time period in the collaborative behavior relationship graph, suspected cheating behavior is determined to have occurred within the corresponding time period. A cheating alarm message containing a timestamp and the identifier of the candidate involved is output. This solves the technical problem that traditional technologies often ignore the spatial interaction relationships between candidates in the examination environment, making it difficult to effectively identify complex cheating patterns. It achieves strong robustness of the action trajectory sequence constructed based on posture key points, and can work stably under actual examination room interference conditions such as changes in lighting and partial occlusion. At the same time, the temporal modeling method preserves the dynamic evolution characteristics of the actions, which helps to capture cheating behavior patterns with temporal continuity. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a schematic diagram of the steps of an intelligent detection method for exam cheating based on behavior recognition in one embodiment of the present invention; Figure 2 This is a schematic diagram of the steps of an intelligent examination cheating detection system based on behavior recognition in one embodiment of the present invention; The objectives, features, and advantages of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0017] The embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention. The step numbers in the following embodiments are set only for ease of explanation, and there is no limitation on the order between the steps. The execution order of each step in the embodiments can be adaptively adjusted according to the understanding of those skilled in the art.

[0018] The following describes in detail, with reference to the accompanying drawings, an intelligent detection method for exam cheating based on behavior recognition according to an embodiment of the present invention. First, the intelligent detection method for exam cheating based on behavior recognition according to an embodiment of the present invention will be described in detail with reference to the accompanying drawings.

[0019] Figure 1 This invention provides an intelligent detection method for exam cheating based on behavior recognition, comprising the following steps: Step S1: Extract the frame sequence from the examination room video stream to obtain continuous image frames, and perform human body region detection based on the image frames to obtain the candidate's human body bounding box.

[0020] Specifically, the process begins by decoding the raw video stream captured by surveillance cameras deployed in the examination room frame by frame at a preset sampling frequency (e.g., 10 frames per second) to obtain a series of sequentially arranged image frames. Then, each frame is input into a pre-trained object detection model (such as YOLOv5 or Faster R-CNN). This model takes the entire image as input, locates the human body regions of all visible examinees in the image space, and outputs a bounding box for each detected examinee, i.e., the examinee's body bounding box, whose coordinates are typically (x, y, x). min , y min , x max , y max This is used to precisely define the position range of each candidate in the current frame. For example, in a standardized written test, when the video stream contains 30 candidates, the system can generate 30 corresponding candidate body bounding boxes in a certain frame through human body region detection. This ensures that subsequent pose key point localization is only performed within these boxes, avoiding the background or other interfering objects from affecting the analysis accuracy, thereby providing an accurate spatial constraint basis for constructing individual motion trajectories.

[0021] Step S2: Based on the candidate's body frame, locate the key points of the candidate's posture in each image frame, obtain the coordinates of the key points of posture, and construct the upper body movement trajectory sequence based on the coordinates of the key points of posture.

[0022] Specifically, based on the candidate's body bounding box, pose key points are located in each image frame to obtain pose key point coordinates. Then, an upper body motion trajectory sequence is constructed based on these coordinates. The specific implementation process is as follows: In each image frame, the obtained candidate body bounding box is used as the input region, cropped, or directly fed into a pre-trained pose estimation model (such as HRNet or OpenPose). This model identifies and outputs the two-dimensional pixel coordinates of a preset number of upper body key points (including but not limited to the tip of the nose, eyes, shoulders, elbows, and wrists) within the candidate body bounding box; subsequently, for the same candidate in… The coordinates of key points in consecutive image frames are arranged sequentially in time to form multidimensional time series data with the frame sequence as the time axis, namely, the upper body movement trajectory sequence. Each time step contains the spatial location information of all upper body key points at that moment. For example, in the video monitoring of a standardized written test, if a candidate frequently lowers his head and moves his right hand repeatedly between frames 100 and 120, the system will generate a continuous upper body movement trajectory sequence based on the coordinate changes of key points such as the shoulders, neck, and right wrist in each frame. This sequence is used to determine subsequent suspicious action segments, thereby ensuring that the action analysis is based on accurate and temporally consistent posture data.

[0023] Step S3: Based on the upper body movement trajectory sequence, determine whether there are any suspicious movement segments. If so, perform dynamic correlation analysis on the spatial relative positions between adjacent candidates based on the suspicious movement segments to obtain a multi-person collaborative behavior relationship diagram.

[0024] Specifically, the upper body movement trajectory sequence of each candidate is first input into a preset abnormal movement discrimination model or rule engine. This model identifies subsequences that are continuous in time intervals and conform to suspicious behavior patterns, i.e., suspicious movement segments, based on the temporal change characteristics of key point coordinates (such as frequent head drooping, abnormal lateral arm movement, sudden shoulder angle changes, etc.). When at least one candidate is detected to have a suspicious movement segment, the system immediately extracts the time window corresponding to the segment and obtains the center coordinates of the candidate's body frame or the torso position in the posture key points of all candidates within this time window. This allows the Euclidean distance or relative orientation between any two adjacent candidates to be calculated. The system then constructs a graph structure with examinees as nodes and spatial proximity and action synchronization as edge weights, namely a multi-person collaborative behavior relationship graph. For example, in a standardized written test, if the system detects suspicious action segments in frames 150 to 170 where examinee A and examinee B both rapidly move their right wrists to the left and simultaneously turn their heads in the direction of the other, and the distance between the center of their human bounding boxes during this period is always less than a set threshold (such as 80 cm), then a high-weighted association edge is established between examinee A and examinee B in the multi-person collaborative behavior relationship graph to represent their potential collaborative cheating behavior, thereby providing a structured basis for subsequent judgment.

[0025] Step S4: When at least two candidates exhibit consistent abnormal action patterns in the multi-person collaborative behavior relationship graph within the same time period, it is determined that suspected cheating behavior has occurred within the corresponding time period, and a cheating alarm message containing a timestamp and the identifier of the candidate involved is output.

[0026] Specifically, when at least two candidates exhibit consistent abnormal action patterns within the same time period in the multi-person collaborative behavior relationship graph, suspected cheating behavior is determined to have occurred within the corresponding time period, and a cheating alarm message containing a timestamp and the identifier of the candidate involved is output. The specific implementation process of this step is as follows: First, all edges and their associated node pairs in the multi-person collaborative behavior relationship graph are traversed. Within the time window corresponding to the identified suspicious action segments, the dynamic features of the upper body action trajectory sequences of the connected candidates are compared, including time-series indicators such as key point movement direction, speed change, and posture angle. If the similarity of the action patterns of the two candidates exceeds a preset threshold (e.g., calculated by Dynamic Time Warping (DTW) or cosine similarity), it is determined that they exhibit consistent abnormal action patterns within that time period. Once the conditions of "at least two candidates" and "at the same time" are met, the cheating behavior is determined to have occurred within the same time period. If the system detects three conditions—"segment," "consistent abnormal action pattern"—it will trigger the judgment logic. The start and end frames of that time segment will be converted into absolute timestamps in the video stream. Combined with the unique identifier assigned to the examinee's body frame during system initialization (such as seat number or detection ID), a structured cheating alarm message will be generated. For example, in a standardized written exam, if, between frames 200 and 215 (corresponding to timestamps 08:42:15–08:42:18), the system determines that examinees with seats 12 and 13 both rapidly move their right hands downwards and simultaneously tilt their heads to the right, and their action trajectory similarity reaches 0.92, exceeding the set threshold of 0.85, a cheating alarm message will be immediately output. This message will include the timestamp "08:42:15–08:42:18" and the identifiers of the examinees involved, "Seat_12, Seat_13," for invigilators to verify in real time or collect evidence afterward.

[0027] In a specific embodiment, frame sequence extraction is performed on the examination room video stream to obtain continuous image frames, and human body region detection is performed based on the continuous image frames to obtain the candidate's human body bounding box, including: The examination room video stream is processed by a preset video stream parser to extract frame sequences, resulting in continuous image frames. The continuous image frames are then preprocessed to obtain a preprocessed image. The preprocessed image is initially screened for human body regions to obtain candidate human body region maps. The candidate human body region maps are then subjected to edge detection and contour closure processing to obtain closed contour human body maps. The closed contour human body image is verified for authenticity to obtain a real human body region image. Based on the real human body region image, the minimum bounding rectangle is used to define the candidate's human body frame.

[0028] Specifically, a preset video stream parser extracts the frame sequence of the examination room video stream. This parser decodes the H.264 encoded video captured in real-time or offline from the examination room monitoring camera frame by frame according to a set sampling rate (e.g., 15 frames per second), thus obtaining a continuous sequence of image frames arranged strictly in chronological order. Subsequently, the continuous image frames undergo normalization preprocessing operations, including size normalization (e.g., uniform scaling to 1280×720 pixels), illumination equalization (using the CLAHE algorithm to enhance local contrast), and noise reduction (using Gaussian filtering or median filtering to suppress sensor noise), ultimately outputting a clear and uniformly formatted preprocessed image. Next, the preprocessed image is input into a lightweight convolutional neural network (e.g., MobileNet-SSD) or a traditional background subtraction model for preliminary human body region screening. This model scans and marks all regions that may contain human bodies within the entire image domain, generating a candidate human body region map represented in binary mask form. Based on this, The Canny edge detection algorithm is applied to the candidate human body region image to extract contour boundaries. Morphological closing operations (such as dilation followed by erosion using a 5×5 rectangular structuring element) are then used to connect and fill broken or discontinuous edges, resulting in a closed contour human body image with complete boundaries and internal connectivity. Subsequently, a realism verification step is performed on the closed contour human body image. This step calculates the geometric features of each closed contour (e.g., whether the area is between 0.05 and 0.5 times the total image area, whether the aspect ratio is between 0.3 and 2.0, and whether the contour convex hull ratio is greater than 0.85), and combines this with contextual temporal consistency (e.g., whether a contour in the current frame persists in adjacent frames and its position changes smoothly). False detection areas caused by shadows, reflections from tables and chairs, or hanging objects are removed, retaining areas that conform to the real human body shape and movement patterns, forming a real human body region image. Finally, based on each valid human body contour in the real human body region image, its minimum bounding rectangle is calculated. This rectangle is defined by (x...)... min , y min , x max ,y maxThe system precisely encloses the corresponding candidate's body area in coordinate form and assigns a unique ID (e.g., Seat_01, Seat_02, etc., in spatial order from left to right), thus obtaining the final candidate's body bounding box. For example, in a standardized written exam, the exam room video stream contains 30 candidates sitting neatly at their desks. After processing the above process at frame 100, the system successfully filters out 30 candidate areas from the pre-processed image. After edge closure and realism verification, two false outlines caused by reflections from the podium are eliminated. Finally, precise body bounding boxes that fit the upper body and torso of the remaining 30 real candidates are generated, ensuring that subsequent pose key point localization is only performed within these boxes, avoiding background interference, and laying a reliable visual perception foundation for the entire cheating detection process.

[0029] In a specific embodiment, edge detection and contour closure processing are performed on the candidate body region map to obtain a closed contour human body map, including: Gradient calculation is performed on the candidate body region map to obtain an edge intensity map, and threshold segmentation is performed on the edge intensity map to obtain an initial edge map; Based on the initial edge map, breakpoints are connected to obtain an edge link map, and the edge link map is filled with regions to obtain a complete contour map; The complete contour map is smoothed to obtain a smooth contour map, and the smooth contour map is validated by region to obtain a closed contour human body map.

[0030] Specifically, gradient calculation is performed on the candidate body region map. This calculation uses the Sobel operator to convolve the image pixel values ​​in the horizontal and vertical directions to obtain the gradient magnitude and direction of each pixel. The gradient magnitude is then normalized to generate an edge intensity map. Next, a dual-threshold segmentation strategy (e.g., a high threshold of 0.25 and a low threshold of 0.1) is applied to the edge intensity map. Strong edge points above the high threshold are retained, while weak edge points connected to strong edge points but between the high and low thresholds are also retained. The remaining pixels are set as background, thus obtaining an initial edge map. Subsequently, breakpoint connection is performed based on the initial edge map. This operation involves traversing all unconnected edge endpoints and searching within a set neighborhood radius (e.g., 8 pixels) for another endpoint with an angle less than 30 degrees between the two endpoints. If the condition is met, the Bresenham line algorithm or spline interpolation is used to fill in the missing edge between the two points, thereby generating a more continuous edge link map. Based on this, the edge link map is filled with regions using flood filling. The Fill algorithm selects a seed point from within each closed or nearly closed edge ring and expands outwards until it encounters the edge boundary, marking all pixels inside as foreground to form a complete contour map. Then, the complete contour map is smoothed by using Gaussian filtering or moving average to filter the contour boundary coordinate sequence, or by employing the Douglas-Peucker algorithm to simplify and smooth the contour polygon vertices, removing high-frequency jitter caused by noise or jagged edges, thus obtaining a smooth contour map. Finally, the smooth contour map undergoes region verification. This verification process follows preset geometric and morphological rules (such as the contour area not being less than 2000 pixels, the contour perimeter to area ratio being within a reasonable human body proportion range, and the number of contour convexity defects not exceeding 3), eliminating non-human-like regions (such as pseudo-contours formed by table legs or stacked books) caused by false detection or overfilling, retaining only valid contours that conform to the characteristics of a real human body, thus obtaining a closed contour human body image. For example, in the 150th frame of a standardized written test, the candidate's body region map contains 32 preliminary detection regions. Two of these regions correspond to the stack of documents on the podium. After gradient calculation and threshold segmentation, although these regions produce edge responses, they cannot form closed loops due to the lack of continuous structure during the breakpoint connection stage. After region filling, they still present a fragmented form. However, in the candidate's region, due to the stable human posture and clear outline, a complete closed region is formed after breakpoint connection and filling. Then, the jagged effect of the arm edges is eliminated by boundary smoothing, and false outlines that are too small or abnormally shaped are eliminated by region verification. Finally, 30 closed outlines that conform to the human body shape are retained to form a closed outline human body map, which provides an accurate geometric basis for subsequent authenticity verification and candidate body frame generation.

[0031] In a specific embodiment, the pose key points of the examinee in each image frame are located based on the examinee's body bounding box to obtain the coordinates of the pose key points, including: The candidate's body frame is divided into regions to obtain an upper body region image, and the contrast of the upper body region image is enhanced to obtain an enhanced upper body image. Feature point detection is performed on the enhanced upper body image to obtain candidate joint points, and false points are removed from the candidate joint points to obtain the filtered joint points. Based on the selected joint points, the posture key point coordinates of the examinee in each image frame are located and mapped to obtain the joint point coordinates. The joint point coordinates are then classified and organized to obtain the posture key point coordinates.

[0032] Specifically, based on the candidate's body bounding box, the pose key points of the candidate in each image frame are located to obtain the coordinates of the pose key points. This process first requires dividing the candidate's body bounding box into regions. This step involves calculating the geometric center of the candidate's body bounding box and dividing the upper body region according to a preset ratio (e.g., 1:2). Next, in order to improve the accuracy of feature extraction, the upper body region needs to be contrast-enhanced. Here, histogram equalization or CLAHE (Contrast-Limited Adaptive Histogram Equalization) technology is used to adjust the distribution of image pixel values, thereby obtaining the enhanced upper body image. Subsequently, feature point detection is performed on the enhanced upper body image. This step typically employs techniques such as Harris corner detection, SIFT feature descriptors, or more modern deep learning models like OpenPose to identify potential candidate joints. These joints represent major joint locations in the human body, such as the elbow and shoulder, and each joint is assigned a confidence score to measure its reliability. However, considering the possibility of false points, i.e., falsely detected non-joint points, the candidate joints are then filtered out. This step filters out points that do not conform to the anatomical structure of the human body or have a confidence score below the threshold, based on the geometric constraints between joints and a confidence score threshold (e.g., set to 0.5), thus obtaining the filtered joints. After completing the above preparations, the pose key point coordinates of the examinee in each image frame can be located and mapped based on the selected joint points. This stage often uses a deep neural network model to predict the precise coordinate position of each joint point in the original image and maps these coordinates back to the original image size to ensure matching with the overall image. In this process, regression analysis or direct heatmap estimation methods may be used to improve the accuracy of coordinate positioning. Finally, in order to facilitate subsequent processing and analysis, the joint point coordinates need to be classified and organized according to different parts of the body (such as the head and limbs) to form a complete human pose representation system, thereby obtaining the pose key point coordinates. For example, in a standardized test monitoring scenario, suppose we are processing the 150th frame of an image, which contains a student writing answers. After the aforementioned processing steps, the student's body outline is first determined, then the upper body region is separated, and contrast enhancement is used to make the boundary between the arm and the pen in the image clearer. In the feature point detection stage, the algorithm identifies multiple candidate joints, including the wrist and elbow, but some points are incorrectly marked due to background interference. Therefore, false points are removed, and only the valid joints that truly belong to the student's body are retained. Based on these filtered joints, the system accurately locates the position of the student's hand that is writing and the posture of the rest of the body. Finally, this information is organized into an easy-to-understand posture key point coordinate form for subsequent assessment of whether the student has abnormal behavior such as cheating attempts.The entire process not only enhances the accuracy and robustness of the monitoring system, but also provides technical support for achieving automated examination room management.

[0033] In a specific embodiment, based on the selected joint points, the pose key point coordinates of the examinee in each image frame are located and mapped to obtain the joint point coordinates, including: After filtering, the pixel coordinates of the joint points are extracted to obtain a set of joint pixels. The coordinate system of the set of joint pixels is then calibrated to obtain standard coordinate points. The connection relationship of the human skeleton is established based on the standard coordinate points to generate a skeleton connection diagram, and the skeleton connection diagram is scaled and normalized to obtain a standard skeleton diagram. Based on the standard skeleton diagram, key points of the examinees in each image frame are paired to obtain a pose node group, and the pose node group is subjected to spatial projection transformation to obtain a mapping coordinate set. Based on the mapped coordinate set, joint points are classified and labeled to obtain a joint type table. Then, the joint type table is integrated with coordinates to obtain joint point coordinates.

[0034] Specifically, pixel coordinates are extracted from the selected joints, meaning the two-dimensional position (x, y) of each retained joint in the original image coordinate system is read from the image, forming a set of joint pixel points in pixels. Then, the joint pixel point set undergoes coordinate system calibration. This calibration process converts the original image coordinates into a standard coordinate system with the upper left corner of the candidate's body frame as the origin and a unit length consistent with the image resolution, thereby eliminating offsets caused by image scaling or cropping and obtaining standard coordinate points. Next, a human skeleton connection relationship is established based on the standard coordinate points. This relationship follows a predefined human topology (such as the 18-point skeleton model defined by OpenPose), sequentially connecting adjacent key points such as the shoulder and elbow, and the elbow and wrist, to generate a skeleton connection diagram. Then, the skeleton connection diagram is scaled and normalized. Specifically, the coordinates of all joint points are divided by the torso length (such as the distance from the midpoint between the two shoulders to the midpoint of the hip) to ensure comparability of candidate skeletons of different heights or distances from the camera, thus obtaining a standard skeleton diagram. Based on this, keypoint pairing is performed on examinees in each image frame using the standard skeleton diagram. This pairing process uses the Hungarian algorithm or nearest neighbor matching strategy to spatiotemporally associate the keypoints detected in the current frame with the skeletons of examinees with known identities in the previous frame, ensuring that the keypoints of the same examinee are consistently assigned in consecutive frames, thus obtaining a pose node group. Subsequently, a spatial projection transformation is performed on the pose node group. This transformation uses the intrinsic parameter matrix of the examination room camera and pre-calibrated ground plane parameters to backproject the two-dimensional image coordinates to approximate planar coordinates in the three-dimensional space of the examination room (such as the plane at the height of the desktop). The XY coordinates are used to obtain a mapping coordinate set, or the coordinates are affine aligned under a unified image reference system. Next, the joints are classified and labeled based on the mapping coordinate set. That is, each coordinate point is assigned a corresponding semantic label according to its connection role in the skeleton connection diagram (such as whether it is the left shoulder, right wrist, nose tip, etc.), forming a joint type table. Finally, the joint type table is integrated by binding the semantic label of each joint point to its corresponding mapping coordinate, and arranging them in a fixed order (such as according to the OpenPose standard number 0~17) to obtain structured joint coordinates.For example, in frame 200 of a standardized written exam, the system has obtained the joint points of a candidate after pseudopoint removal, including seven valid points: left shoulder, right shoulder, left elbow, right elbow, left wrist, right wrist, and nose tip. After pixel coordinate extraction, the positions of these points in the image are recorded as (320, 180), (360, 182), etc. The coordinate system is then calibrated to convert them into standard coordinates based on the candidate's body frame. A skeleton connection diagram is then constructed according to human skeletal rules, and the scale is normalized using the distance between the two shoulders as the unit. Subsequently, combined with the candidate's identity ID in frame 199, key points are paired to form posture node groups, and these points are projected onto a unified examination room plane coordinate system through camera calibration parameters. Finally, the system labels each point with types such as "right wrist" and "nose tip," integrates them into an ordered list, and outputs complete joint point coordinates for subsequent construction of upper body motion trajectory sequences to support intelligent identification of cheating behavior.

[0035] In a specific embodiment, an upper body motion trajectory sequence is constructed based on the coordinates of the posture key points, including: The coordinates of the attitude key points are sorted according to frame time sequence to obtain a time sequence coordinate column, and the time sequence coordinate column is interpolated and completed to obtain a completed coordinate column; The completed coordinate column is split according to joint type to obtain single joint coordinate groups, and the single joint coordinate groups are connected by continuous line segments to obtain single joint trajectory lines; The single-joint trajectory line is aligned with the time axis to obtain the aligned trajectory line, and the aligned trajectory line is combined frame by frame to obtain the upper body movement trajectory sequence.

[0036] Specifically, the attitude keypoint coordinates are sorted in frame time sequence. That is, all attitude keypoint coordinates extracted from consecutive image frames are arranged in ascending order according to the timestamp of their respective frames, forming a time-series coordinate column organized in chronological order. Each item contains the corresponding frame number, joint type identifier, and two-dimensional or three-dimensional coordinate value. Subsequently, the time-series coordinate column is interpolated and completed. This operation is used to address the situation where some joints are missing in individual frames due to occlusion, detection failure, or low confidence in attitude estimation. Linear interpolation, spline interpolation, or Kalman filtering based on motion models are used to estimate reasonable coordinate values ​​at the location of the missing frame, thereby generating a seamless completed coordinate column to ensure the continuity of subsequent trajectories. Next, the completed coordinate column is split according to joint type. That is, according to the semantic tags defined in the joint type table (such as "left shoulder", "right wrist", "nose tip", etc.), all coordinate points in the completed coordinate column are classified into their respective categories to form multiple independent single joint coordinate groups. Each single joint coordinate group contains only the coordinate sequence of the same type of joint in all frames. Then, the single joint coordinate groups are connected by continuous line segments. This step constructs a single joint trajectory line that reflects the movement path of the joint in the spatiotemporal dimension by connecting the coordinate points of adjacent frames in the same single joint coordinate group with straight line segments. Its essence is a broken line or a multi-segment smooth curve composed of discrete points. Based on this, the single-joint trajectory lines are aligned on the time axis. This alignment operation unifies the start and end frames of all single-joint trajectory lines, ensuring that the trajectories of different joints have the same number of frames and time step within the same time window. If the trajectory length of a certain joint is insufficient, boundary values ​​or zero values ​​are added at the beginning and end to achieve alignment, thereby obtaining aligned trajectory lines. Finally, the aligned trajectory lines are combined frame by frame, that is, the coordinate values ​​of all upper body joints (such as shoulders, elbows, wrists, neck, head, etc.) at the same time step (i.e., the same frame) are packaged into a high-dimensional vector and arranged in the order of frames, finally obtaining a structured upper body motion trajectory sequence. For example, in a standardized written test, the system processes 51 frames of video, from frame 100 to frame 150. After pose estimation, it obtains the coordinates of 12 upper body key points for each candidate. In frame 123, the right wrist was not detected because the candidate lowered their head too quickly. The system then performs linear interpolation to complete the coordinates of this point between frames 122 and 124. Subsequently, the 51 coordinates of the "right wrist" category are grouped into single-joint coordinate sets and connected into trajectory lines. These are then aligned with the trajectory lines of the other 11 joints on the time axis (ensuring that each is 51 frames). Finally, each frame forms a 24-dimensional vector (12 points × 2 coordinates). After the 51 frames are stitched together, a complete upper body motion trajectory sequence is formed for subsequent identification of suspicious motion segments.

[0037] In a specific embodiment, a dynamic correlation analysis is performed on the spatial relative positions between adjacent candidates based on the suspicious action segments to obtain a multi-person collaborative behavior relationship diagram, including: Based on the candidates in the suspicious action segments, the positional distances between the candidates and their corresponding adjacent candidates are calculated to obtain the candidate position matrix. Then, the proximity relationship analysis of the candidate position matrix is ​​performed to obtain the candidate adjacency graph. Based on the candidate's adjacency graph, temporal action matching is performed to obtain the action correlation matrix; The action correlation matrix is ​​divided into time windows to obtain a time period correlation table, and the frequency of the time period correlation table is statistically analyzed to obtain a correlation strength map. Group behavior analysis is performed based on the association strength map to obtain a behavior cluster map, and the behavior cluster map is spatiotemporally integrated to obtain a multi-person collaborative behavior relationship map.

[0038] Specifically, based on the candidates in the suspicious action segments, the positional distances between the candidates and their corresponding adjacent candidates are calculated. That is, in each frame covered by the suspicious action segments, the center coordinates of the candidate's body bounding box of all involved candidates are extracted, and the Euclidean distance between any two candidates is calculated. These distances are organized into a two-dimensional array structure according to time steps and candidate pairs to form a candidate position matrix. Subsequently, the candidate position matrix is ​​subjected to proximity analysis. This analysis sets a spatial proximity threshold (e.g., 70 cm). If the distance between a candidate pair in multiple consecutive frames is less than the threshold, it is determined that there is a stable spatial proximity relationship, which is recorded in the form of an undirected graph, with nodes representing candidate identifiers and edges representing proximity relationships, thus obtaining a candidate adjacency graph. Next, temporal action matching is performed based on the candidate adjacency graph. Specifically, for each pair of candidates connected by an edge in the graph, the upper body motion trajectory sequence within the suspected action segment is extracted. The consistency of their action patterns is calculated using Dynamic Time Warping (DTW) or keypoint motion vector similarity. The action similarity values ​​of all candidate pairs are then filled into a matrix according to row and column correspondence to obtain an action correlation matrix. Then, the action correlation matrix is ​​divided into time windows, further subdividing the entire suspected action segment into several sub-time windows (e.g., every 5 frames per window). The sub-windows in which the action correlation of each candidate pair exceeds a preset threshold are recorded, forming a time-segment association table. Based on this, frequency statistics are performed on the time-segment association table, counting the number of times each candidate pair is marked as highly correlated in multiple sub-windows, or counting the total frequency of a single candidate participating in highly correlated associations. This quantifies the intensity of their collaborative behavior, generating an association strength map expressed in the form of a numerical matrix or heatmap. Subsequently, group behavior analysis is performed based on the association strength map. This analysis uses graph clustering algorithms (such as Louvain community detection) to identify candidate subgroups with association strength significantly higher than background noise. Each subgroup represents a potential collaborative cheating group, thus obtaining a behavior cluster map. Finally, the behavior cluster map is spatiotemporally integrated, that is, the active intervals, spatial location distribution (from candidate body bounding boxes), and action synchronization features (from action correlation matrix) of each cluster in the time dimension are merged into a unified weighted directed graph structure, where nodes are candidates, edge weights comprehensively reflect spatial proximity, action consistency, and temporal overlap, and timestamp information is retained, finally outputting a multi-person collaborative behavior relationship graph.For example, in a standardized written test, the system detected multiple candidates exhibiting abnormal head-down and hand movements in frames 180–200. Positional distance calculations revealed that the pairwise distances between Seat_08, Seat_09, and Seat_10 were all less than 65 centimeters, forming a candidate adjacency graph. Temporal action matching showed that Seat_08 and Seat_09 had a similarity of 0.92 in frames 185–190, while Seat_09 and Seat_10 had a similarity of 0.89 in frames 188–193, forming highly correlated pairs. The time-segment association table recorded that their collaboration occurred in two overlapping sub-windows, with frequency statistics showing association strengths of 2 and 2, respectively. Group behavior analysis clustered them into the same behavioral cluster, and finally, the relationship between the three was represented by strong connections in the multi-person collaborative behavior relationship graph for subsequent cheating detection.

[0039] In a specific embodiment, temporal action matching is performed based on the candidate's adjacency graph to obtain an action correlation matrix, including: Trajectory extraction is performed on candidates and their corresponding adjacent candidates in the candidate adjacency graph. The upper body movement trajectory segments of adjacent candidate pairs within the suspicious action segments are obtained. The coordinate change sequence of all upper body posture key points in each upper body movement trajectory segment is extracted according to the time axis to obtain the temporal characteristics of the candidate's actions. Among them, the adjacent candidate pairs are candidate combinations determined based on the spatial positional relationship of the candidate's human body frame. The time axis is aligned for the coordinate change sequence of each key point of the candidate's action sequence characteristics. The start time difference, duration ratio and motion direction overlap of different parts of the action are calculated to obtain the action coordination parameters. The motion coordination parameters are weighted and calculated to obtain the motion correlation degree of adjacent candidates in each time window. The motion correlation degree is then arranged in a matrix according to the correspondence between adjacent candidates and time windows to obtain the motion correlation matrix.

[0040] Specifically, trajectory extraction is performed on candidates and their corresponding adjacent candidates in the candidate adjacency graph. This involves traversing each edge in the candidate adjacency graph and obtaining the upper body movement trajectory segments of the two candidates connected to it within the time window covered by the suspicious action segment. These trajectory segments are subsequences extracted by frame number from the previously constructed upper body movement trajectory sequence. Subsequently, the coordinate change sequence of all upper body posture key points (such as left shoulder, right shoulder, left elbow, right elbow, left wrist, right wrist, nose tip, etc.) in each upper body movement trajectory segment is extracted according to the time axis. The (x, y) coordinates of each key point in consecutive frames are arranged in chronological order to form a multidimensional time series, i.e., the temporal characteristics of candidate actions. Among them, adjacent candidate pairs are candidate combinations determined based on the spatial positional relationship of the candidate's body frame. For example, if the center distance between the body frames of two candidates is less than a preset threshold (such as 80 cm) in a certain frame, and this condition is continuously met within the suspicious action segment, then they are included in the candidate adjacency graph and regarded as adjacent candidate pairs. Next, the time axis of the key point coordinate change sequence of the candidate's action timing features is aligned. This alignment operation ensures that the trajectories of the two candidates being compared have the same start and end frames and frame number. If there is a length difference, it is completed by linear interpolation or end-to-end extension. Based on this, the start time difference, duration ratio, and motion direction overlap of different parts of the action are calculated: the start time difference refers to the time offset of the first time that the amplitude of the action of the same part (such as the right wrist) of the two candidates exceeds the static threshold; the duration ratio is the ratio of the number of frames of each action; the motion direction overlap is calculated by averaging the cosine of the angle between the displacement vectors of the corresponding key points of the two candidates within the time window to measure the consistency of the action direction, thereby obtaining a set of quantitative indicators, namely the action coordination parameters. Then, the motion coordination parameters are weighted and calculated using preset weighting coefficients (e.g., 0.4 for start time difference, 0.3 for duration ratio, and 0.3 for motion direction overlap). After normalizing the three parameters, they are weighted and summed to obtain a value between 0 and 1, which serves as the motion correlation degree between adjacent candidate pairs within the time window. Finally, the motion correlation degree is arranged in a matrix according to the correspondence between adjacent candidate pairs and time windows. That is, all adjacent candidate pairs are used as row and column indices of the matrix, and each element is filled with the motion correlation degree value within its corresponding time window. If a candidate pair does not appear in the window, it is set to 0, thus forming a structured motion correlation matrix.For example, in a standardized written test, the system identified a pair of adjacent candidates (Seat_12, Seat_13) within a suspicious action segment in frames 200–215. After extracting the coordinate change sequence of their right wrist key points, the system found through time axis alignment that both candidates' actions started at frame 202, lasted for 6 frames, and had a displacement direction angle of 12 degrees. The calculated start time difference was 0 frames, the duration ratio was 1.0, the motion direction overlap was 0.98, and the weighted action correlation was 0.97. This value was filled into the position of the 12th row and 13th column of the action correlation matrix for subsequent threshold segmentation to determine whether there was collaborative cheating.

[0041] The above describes the intelligent detection method for exam cheating based on behavior recognition in the embodiments of the present invention. The following describes the intelligent detection system for exam cheating based on behavior recognition in the embodiments of the present invention. Please refer to [link / reference]. Figure 2 One embodiment of the intelligent exam cheating detection system based on behavior recognition in this invention includes: The extraction module 21 is used to extract the frame sequence from the examination room video stream to obtain continuous image frames, and to perform human body region detection based on the image frames to obtain the candidate's human body bounding box. The positioning module 22 is used to locate the posture key points of the examinee in each image frame based on the examinee's body frame, obtain the posture key point coordinates, and construct an upper body movement trajectory sequence based on the posture key point coordinates. The determination module 23 is used to determine whether there are suspicious action segments based on the upper body movement trajectory sequence. If there are, the spatial relative positions between adjacent candidates are dynamically correlated based on the suspicious action segments to obtain a multi-person collaborative behavior relationship diagram. The determination module 24 is used to determine that suspected cheating behavior has occurred in the corresponding time period when at least two candidates exhibit the same abnormal action pattern in the multi-person collaborative behavior relationship graph, and output cheating alarm information including timestamp and the identifier of the candidate involved.

[0042] In this embodiment, the specific implementation of each module in the above system embodiment is described in the above method embodiment, and will not be repeated here.

Claims

1. A method for intelligent detection of exam cheating based on behavior recognition, characterized in that, Includes the following steps: The examination room video stream is subjected to frame sequence extraction to obtain continuous image frames, and human body region detection is performed based on the image frames to obtain the candidate's human body bounding box; Based on the candidate's body bounding box, the candidate's posture key points in each image frame are located to obtain the posture key point coordinates, and an upper body movement trajectory sequence is constructed based on the posture key point coordinates. Based on the upper body movement trajectory sequence, determine whether there are any suspicious movement segments. If so, perform dynamic correlation analysis on the spatial relative positions between adjacent candidates based on the suspicious movement segments to obtain a multi-person collaborative behavior relationship diagram. When at least two candidates exhibit consistent abnormal behavior patterns within the same time period in the multi-person collaborative behavior relationship graph, it is determined that suspected cheating behavior has occurred within the corresponding time period, and a cheating alarm message containing a timestamp and the identifier of the candidate involved is output.

2. The intelligent detection method for exam cheating based on behavior recognition according to claim 1, characterized in that, The examination room video stream is subjected to frame sequence extraction to obtain continuous image frames, and human body region detection is performed based on the continuous image frames to obtain the candidate's human body bounding box, including: The examination room video stream is processed by a preset video stream parser to extract frame sequences, resulting in continuous image frames. The continuous image frames are then preprocessed to obtain a preprocessed image. The preprocessed image is initially screened for human body regions to obtain candidate human body region maps. The candidate human body region maps are then subjected to edge detection and contour closure processing to obtain closed contour human body maps. The closed contour human body image is verified for authenticity to obtain a real human body region image. Based on the real human body region image, the minimum bounding rectangle is used to define the candidate's human body frame.

3. The intelligent detection method for exam cheating based on behavior recognition according to claim 2, characterized in that, Edge detection and contour closure processing are performed on the candidate body region map to obtain a closed contour human body map, including: Gradient calculation is performed on the candidate body region map to obtain an edge intensity map, and threshold segmentation is performed on the edge intensity map to obtain an initial edge map; Based on the initial edge map, breakpoints are connected to obtain an edge link map, and the edge link map is filled with regions to obtain a complete contour map; The complete contour map is smoothed to obtain a smooth contour map, and the smooth contour map is validated by region to obtain a closed contour human body map.

4. The intelligent detection method for exam cheating based on behavior recognition according to claim 1, characterized in that, Based on the candidate's body bounding box, pose key points of the candidate in each image frame are located to obtain the coordinates of the pose key points, including: The candidate's body frame is divided into regions to obtain an upper body region image, and the contrast of the upper body region image is enhanced to obtain an enhanced upper body image. Feature point detection is performed on the enhanced upper body image to obtain candidate joint points, and false points are removed from the candidate joint points to obtain the filtered joint points. Based on the selected joint points, the posture key point coordinates of the examinee in each image frame are located and mapped to obtain the joint point coordinates. The joint point coordinates are then classified and organized to obtain the posture key point coordinates.

5. The intelligent detection method for exam cheating based on behavior recognition according to claim 4, characterized in that, Based on the selected key points, the pose key point coordinates of the examinee in each image frame are located and mapped to obtain the key point coordinates, including: After filtering, the pixel coordinates of the joint points are extracted to obtain a set of joint pixels. The coordinate system of the set of joint pixels is then calibrated to obtain standard coordinate points. The connection relationship of the human skeleton is established based on the standard coordinate points to generate a skeleton connection diagram, and the skeleton connection diagram is scaled and normalized to obtain a standard skeleton diagram. Based on the standard skeleton diagram, key points of the examinees in each image frame are paired to obtain a pose node group, and the pose node group is subjected to spatial projection transformation to obtain a mapping coordinate set. Based on the mapped coordinate set, joint points are classified and labeled to obtain a joint type table. Then, the joint type table is integrated with coordinates to obtain joint point coordinates.

6. The intelligent detection method for exam cheating based on behavior recognition according to claim 1, characterized in that, Based on the coordinates of the aforementioned posture key points, a sequence of upper body movement trajectories is constructed, including: The coordinates of the attitude key points are sorted according to frame time sequence to obtain a time sequence coordinate column, and the time sequence coordinate column is interpolated and completed to obtain a completed coordinate column; The completed coordinate column is split according to joint type to obtain single joint coordinate groups, and the single joint coordinate groups are connected by continuous line segments to obtain single joint trajectory lines; The single-joint trajectory line is aligned with the time axis to obtain the aligned trajectory line, and the aligned trajectory line is combined frame by frame to obtain the upper body movement trajectory sequence.

7. The intelligent detection method for exam cheating based on behavior recognition according to claim 1, characterized in that, Based on the suspected action fragments, a dynamic correlation analysis of the spatial relative positions between adjacent candidates is performed to obtain a multi-person collaborative behavior relationship diagram, including: Based on the candidates in the suspicious action segments, the positional distances between the candidates and their corresponding adjacent candidates are calculated to obtain the candidate position matrix. Then, the proximity relationship analysis of the candidate position matrix is ​​performed to obtain the candidate adjacency graph. Based on the candidate's adjacency graph, temporal action matching is performed to obtain the action correlation matrix; The action correlation matrix is ​​divided into time windows to obtain a time period correlation table, and the frequency of the time period correlation table is statistically analyzed to obtain a correlation strength map. Group behavior analysis is performed based on the association strength map to obtain a behavior cluster map, and the behavior cluster map is spatiotemporally integrated to obtain a multi-person collaborative behavior relationship map.

8. The intelligent detection method for exam cheating based on behavior recognition according to claim 7, characterized in that, Based on the candidate's adjacency graph, temporal action matching is performed to obtain an action correlation matrix, including: Trajectory extraction is performed on candidates and their corresponding adjacent candidates in the candidate adjacency graph. The upper body movement trajectory segments of adjacent candidate pairs within the suspicious action segments are obtained. The coordinate change sequence of all upper body posture key points in each upper body movement trajectory segment is extracted according to the time axis to obtain the temporal characteristics of the candidate's actions. Among them, the adjacent candidate pairs are candidate combinations determined based on the spatial positional relationship of the candidate's human body frame. The time axis is aligned for the coordinate change sequence of each key point of the candidate's action sequence characteristics. The start time difference, duration ratio and motion direction overlap of different parts of the action are calculated to obtain the action coordination parameters. The motion coordination parameters are weighted and calculated to obtain the motion correlation degree of adjacent candidates in each time window. The motion correlation degree is then arranged in a matrix according to the correspondence between adjacent candidates and time windows to obtain the motion correlation matrix.

9. An intelligent detection system for exam cheating based on behavior recognition, characterized in that, The intelligent detection method for exam cheating based on behavior recognition, as described in any one of claims 1 to 8, comprises: The extraction module is used to extract frame sequences from the examination room video stream to obtain continuous image frames, and to perform human body region detection based on the image frames to obtain the candidate's human body bounding box. The positioning module is used to locate the posture key points of the examinee in each image frame based on the examinee's human body frame, obtain the coordinates of the posture key points, and construct the upper body movement trajectory sequence based on the coordinates of the posture key points. The determination module is used to determine whether there are suspicious action segments based on the upper body movement trajectory sequence. If there are, the module performs dynamic correlation analysis on the spatial relative positions between adjacent candidates based on the suspicious action segments to obtain a multi-person collaborative behavior relationship diagram. The determination module is used to determine that suspected cheating has occurred in the corresponding time period when at least two candidates exhibit the same abnormal action pattern in the multi-person collaborative behavior relationship graph, and outputs cheating alarm information including timestamp and the identifier of the candidate involved.

Images