Construction site face recognition safety protection method based on artificial intelligence

Abstract

The present application relates to the technical field of face recognition, in particular to a construction site face recognition safety protection method based on artificial intelligence, comprising the following steps: obtaining a face image and a skeleton coordinate, extracting an identity feature and synchronously recording, analyzing a posture change to judge an abnormal trend, matching a railing area to determine a crossing behavior, encrypting a feature field according to a permission, and generating an abnormal event record in combination with recognition information. In the present application, face features and skeleton coordinates are extracted, binding is completed in combination with recognition time, identity and behavior data consistency is improved, ankle-to-spine direction change and knee joint displacement trend are extracted, pre-boundary-crossing abnormal motion recognition accuracy is enhanced, skeleton coordinates and railing boundaries are compared, crossing behavior is judged in real time, misrecognition probability is reduced, feature fields are dynamically segmented according to permission levels, data security is ensured, terminal numbers and time information are fused to generate behavior records, and recognition response efficiency and reliability are improved.

Description

Technical Field

[0001] This invention relates to the field of facial recognition technology, and in particular to a facial recognition security protection method for construction sites based on artificial intelligence. Background Technology

[0002] The field of facial recognition technology encompasses the processes of feature extraction, analysis, and matching recognition of human facial images using computer vision and artificial intelligence methods. The core components of this technology include facial image acquisition, image preprocessing, face detection, facial feature extraction, feature comparison, and identity verification. Facial recognition technology is widely used in security monitoring, identity verification, and behavior monitoring. In recent years, it has gradually integrated convolutional neural network algorithms and edge computing technology, expanding into high-risk application areas such as construction site management and public safety. In construction site management, facial recognition-based identity verification systems can effectively assist in on-site personnel management. Combined with behavior analysis and risk warning algorithms, it can build a dynamic security protection system, becoming an important component of intelligent construction site management.

[0003] Among them, the AI-based construction site facial recognition safety protection method refers to a comprehensive construction site safety management method that combines facial detection, behavior recognition, and risk assessment. The technical aspects involved in this patent cover personnel identification, violation detection, and safety risk level determination. Specifically, it involves: acquiring real-time facial images of personnel at the construction site via cameras; inputting these images into a convolutional neural network-based facial detection algorithm for identity verification; then, using a behavior analysis algorithm to track their behavioral trajectories and identifying behaviors such as not wearing safety helmets or illegally crossing boundaries based on skeletal key point detection; finally, a risk assessment algorithm, combined with preset level rules, determines the corresponding safety risk level and transmits it to the management platform. Furthermore, this method introduces an edge computing desensitization mechanism, completing feature extraction and original image destruction within the attendance terminal, retaining only the encrypted facial feature code, and using the national cryptographic SM4 algorithm combined with job permissions to generate differentiated keys for local encryption and access control; combined with an AI behavior analysis model, it identifies and determines abnormal personnel entry and exit behaviors through spatiotemporal trajectories.

[0004] Traditional image processing-based personnel identification and behavior analysis suffer from weak information temporality and unstable dynamic recognition. The lack of synchronized processing of image and skeletal information results in a lack of rigorous correspondence between identity verification results and behavioral trajectories, hindering high-precision scene association. Single face detection often relies on static feature extraction, leading to unstable image recognition performance in complex environments with moving individuals, and is highly susceptible to interference from occlusion, lighting changes, and other factors. Existing behavior recognition paths rely on instantaneous skeletal point state judgments, ignoring the trend of pose changes between consecutive frames, failing to accurately depict the evolution of boundary-crossing behaviors, and prone to false alarms and missed detections. For boundary-crossing detection, comparisons are mostly based on the target's static position and boundary settings, lacking motion trend intervention mechanisms, making it difficult to intervene in potential violations in advance. Regarding information processing security, the encryption granularity of feature fields is not finely controlled, and the ability to dynamically adjust encryption strategies based on permission levels is lacking, posing a risk of data leakage. In generating abnormal behavior results, recognition events are not uniformly bound to basic metadata such as recognition time and terminal number, resulting in insufficient data recording granularity and limiting the completeness and accuracy of subsequent anomaly analysis and audit tracking. The aforementioned deficiencies can lead to an expansion of safety identification blind spots in dynamic operating environments, increasing the misjudgment rate and response delay of the management system. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by proposing an artificial intelligence-based facial recognition security protection method for construction sites.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a construction site facial recognition security protection method based on artificial intelligence, comprising the following steps:

[0007] S1: Acquire the image sequence collected by the face terminal at the construction site entrance and exit, extract the face region coordinates, feature vectors and recognition time information of the people in the image, combine the key skeleton point coordinate information in the standing state for synchronous recording, and bind it with the identity information to generate the recognition identity frame structure data group.

[0008] S2: Identify continuous image frames in the identity frame structure data group, extract the direction of the line connecting the ankle to the spine, calculate the direction difference between frames, extract the vertical displacement of the knee joint within the reference height range of the railing area, determine whether the direction difference continuously exceeds the threshold and the displacement continuously increases, and if the condition is met, establish an abnormal movement trend marker.

[0009] S3: Based on the skeletal coordinates recorded in the abnormal movement trend markers, match the safety boundary range of the railing set in the construction site boundary coordinate set, and compare whether there is a continuous crossing behavior area at the coordinate points. If a crossing occurs, generate a spatial boundary crossing record.

[0010] S4: Call the personnel identity information corresponding to the spatial boundary crossing record, obtain the permission level in the operation permission control table, process the face feature vector field into segments according to the level, mark the encryption level, and generate a feature vector field encryption grouping table.

[0011] As a further embodiment of the present invention, the identification frame structure data group includes face region coordinates, feature vectors, identification time information, skeleton point coordinates, and identity information; the abnormal movement trend markers include direction difference, vertical displacement, and skeleton coordinates; the spatial boundary crossing record includes the crossing behavior area, the railing safety boundary range, and the coordinate point change trajectory; and the feature vector field encryption grouping table includes face feature vector fields, encryption labels, and encryption levels.

[0012] As a further aspect of the present invention, the specific steps of S1 are as follows:

[0013] S101: Obtain the image sequence collected by the face terminal at the construction site entrance and exit, extract the face region coordinates and time information of the people in the image, calculate the image gradient change rate based on the gray distribution and edge point density, perform normalization processing, and generate image face coordinate and time dataset.

[0014] S102: Based on the image face coordinates and time dataset, call the key bone point coordinates in the corresponding image frame, match the positional relationship between key points of the face and upper limbs, and perform synchronous comparison in time order to obtain the time-synchronized bone joint data volume.

[0015] S103: Based on the amount of time-synchronized skeleton joint data, combined with the face region location and image features, extract feature vectors from each frame of data in the image sequence, bind them with identity information and time data, and generate an identity recognition frame structure data group.

[0016] As a further aspect of the present invention, the specific steps of S2 are as follows:

[0017] S201: Based on the identification frame structure data group, call continuous image frames, extract the line connecting the ankle and the spine, calculate the angle difference between the directions of the lines between adjacent frames, obtain the direction change amount corresponding to each frame, and establish a sequence value of direction change amount.

[0018] S202: Based on the direction change sequence value, filter the vertical coordinate values ​​of the knee joint in the railing area image frames that are within the reference height range, calculate the vertical displacement between adjacent frames, and establish the railing area displacement change sequence value;

[0019] S203: Based on the sequence value of the directional change and the sequence value of the displacement change in the railing area, determine whether the directional change continuously exceeds the directional change judgment threshold and whether the vertical displacement continuously increases. If the condition is met, generate an abnormal motion trend indicator.

[0020] As a further aspect of the present invention, the specific calculation formula for calculating the vertical displacement between adjacent frames is as follows:

[0021] ;

[0022] Calculate the dynamic weighted displacement value and establish a sequence of displacement changes in the railing area;

[0023] in, Represents the normalized displacement. This represents the vertical coordinate value of the knee joint in the i-th frame. This represents the vertical coordinate value of the knee joint in the (i-1)th frame. Represents the median value of the reference height range. The symbol represents the function, and the subscript i indicates the time sequence number of consecutive image frames.

[0024] As a further aspect of the present invention, the specific steps of S3 are as follows:

[0025] S301: Extract the bone coordinates for each time series based on the abnormal motion trend marker, call the railing boundary range, determine the area where the bone coordinate point is located, record the direction, and generate the directional position value of the bone coordinate boundary.

[0026] S302: Based on the directional position value of the bone coordinate boundary, merge and statistically analyze the directional identifiers of the time sequence points, screen for directional jumps, extract the time sequence intervals corresponding to the jump segments, and obtain the boundary crossing trend identification segment sequence.

[0027] S303: Based on the boundary crossing trend identification segment sequence, call the skeletal trajectory to calculate the boundary intersection position, statistically analyze the continuity index, determine whether a crossing behavior is constituted, and generate a spatial boundary crossing record.

[0028] As a further aspect of the present invention, the specific steps of S4 are as follows:

[0029] S401: Call the personnel identity information corresponding to the spatial boundary crossing record, compare it with the identity identifier field in the operation permission control table, extract the permission level field corresponding to the matching item, obtain the personnel's level classification identifier based on the correspondence between the level field and the type field, calculate the level classification identifier parameter, and generate the permission level classification value.

[0030] S402: Based on the permission level classification value, call the facial feature vector field associated with the person's identity, perform structural division of the feature encoding, divide the feature vector field into several feature segments according to the feature dimension truncation rules set by the level classification, and attach corresponding field labels to obtain the feature segment label field group.

[0031] S403: Based on the field labels in the feature segment marker field group and the mapping rules set in the encryption level comparison parameter table, select the encryption level coefficient corresponding to each group of fields, attach an encryption level identifier to the feature segment, and generate a feature vector field encryption grouping table.

[0032] As a further aspect of the present invention, the specific calculation formula for the calculation level classification identifier parameter is as follows:

[0033] ;

[0034] in, The comprehensive evaluation value representing the permission level classification. The value of the field representing the current permission level. The base value of the representative type field. Represents the measured value of the current type field. This represents a dynamic correction coefficient based on historical permission adjustment records. The equilibrium constant is a unit quantity. This represents the number of historical adjustments to a field of the associated type. Representing the The weighting factor for the next historical adjustment. This represents the absolute value of the difference between type fields. This indicates the standardization process for the dynamic correction coefficient.

[0035] As a further aspect of the present invention, the method further includes:

[0036] S5: Combine the encrypted tags in the encrypted grouping table with the feature vector field, call the identification time and terminal number information in the identification frame structure data group, construct the abnormal behavior record item, and combine the identity code, identification time point, behavior type and encryption level to generate an abnormal identification event output data block;

[0037] The anomaly detection event output data block includes the identity code, identification time, behavior type, and encryption level.

[0038] As a further aspect of the present invention, the specific steps of S5 are as follows:

[0039] S501: Based on the feature vector field, encrypt the encrypted label field in the grouping table, call the feature code index and matching number content in the feature vector field, classify and organize according to the group number, construct the feature vector set corresponding to the group, and obtain the group vector feature value set;

[0040] S502: Based on the grouped vector feature value set, call the identification time and terminal number fields in the identification frame structure data group, sort the terminal numbers according to the identification time, extract data pairs that meet the identification time interval conditions, and determine the behavior type by combining the group number and feature value difference to obtain the behavior type identification group;

[0041] S503: Based on the group number of the behavior type identification group, combine the identity code, identification time point, behavior type and encryption level to construct the corresponding data structure and generate an abnormal identification event output data block.

[0042] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0043] In this invention, facial regions, feature vectors, and skeletal point coordinates are extracted synchronously from image sequences and bound together with the recognition time to improve the spatiotemporal consistency of identity and behavior data. The changes in the direction from the ankle to the spine and the displacement trend of the knee joint are extracted, and posture differences are continuously tracked to enhance the recognition accuracy of abnormal actions before crossing the boundary. By comparing the skeletal coordinates with the boundary range of the railing, the crossing behavior is judged in real time to reduce the probability of false recognition. The feature vector fields are dynamically segmented according to the permission level, and the encryption marking method is adjusted to ensure data security. The terminal number and time information are integrated to generate a structured behavior record, which enhances the response efficiency and credibility of the recognition. Attached Figure Description

[0044] Figure 1 This is a schematic diagram of the steps of the present invention. Detailed Implementation

[0045] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0046] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0047] Please see Figure 1 An AI-based facial recognition security method for construction sites includes the following steps:

[0048] S1: Acquire the image sequence collected by the face terminal at the construction site entrance and exit, extract the face region coordinates, feature vectors and recognition time information of the people in the image, combine the key skeleton point coordinate information in the standing state for synchronous recording, and bind it with the identity information to generate the recognition identity frame structure data group.

[0049] S2: Call the continuous image frames in the identification frame structure data group, extract the direction of change of the line connecting the ankle to the spine, calculate the direction difference between frames, and extract the vertical displacement data of the knee joint within the reference height range of the railing area. Determine whether the change of the direction difference continuously exceeds the direction change judgment threshold and whether the displacement is continuously increasing. If the condition is met, establish an abnormal movement trend marker.

[0050] S3: Based on the skeletal coordinates recorded in the abnormal movement trend markers, match the safety boundary range of the railing set in the construction site boundary coordinate set, and compare whether there is a continuous crossing behavior area at the coordinate points. If a crossing occurs, generate a spatial boundary crossing record.

[0051] S4: Call the personnel identity information corresponding to the spatial boundary crossing record, obtain the permission level in the operation permission control table, segment the face feature vector field according to the level, mark the encryption level, and form a feature vector field encryption grouping table.

[0052] S5: Combine the encrypted tags in the encrypted grouping table with the feature vector field, call the identification time and terminal number information in the identification frame structure data group, construct the abnormal behavior record item, and combine the identity code, identification time point, behavior type and encryption level to generate the abnormal identification event output data block.

[0053] The identification frame structure data group includes face region coordinates, feature vectors, identification time information, skeleton point coordinates, and identity information. Abnormal movement trend indicators include direction difference, vertical displacement, and skeleton coordinates. Spatial boundary crossing records include crossing behavior area, railing safety boundary range, and coordinate point change trajectory. The feature vector field encryption group table includes face feature vector field, encryption label, and encryption level. The abnormal identification event output data block includes identity code, identification time point, behavior type, and encryption level.

[0054] The specific steps of S1 are as follows:

[0055] S101: Obtain the image sequence collected by the face terminal at the construction site entrance and exit, extract the face region coordinates and time information of the people in the image, calculate the image gradient change rate based on the gray distribution and edge point density, perform normalization processing, and generate image face coordinate and time dataset.

[0056] When the facial recognition terminal at the construction site entrance collects image sequences, it first reads the image content frame by frame according to a preset shooting frequency, such as 10 frames per second. Using common image recognition tools, it identifies and locates the visible facial regions of people in each frame, obtaining the coordinates of the top-left corner and the width and height parameters of the facial bounding box (e.g., horizontal position 320, vertical position 240, width and height 100), forming four-element coordinate data. Simultaneously, it extracts the acquisition time of each frame as timestamp information and establishes a correspondence between time and coordinates. For image quality analysis, it extracts the grayscale distribution information of the image frame by frame, calculating the average grayscale value and dispersion by traversing all pixel grayscale values. Then, it uses an edge detection algorithm to extract the edge point distribution density, judging the clarity of the image structure and contour by the ratio of the total number of edge points to the total number of pixels in the image. To determine the clarity, for example, an edge point density of 4%, and to standardize the image state, the image change feature values ​​are normalized. The lowest and highest grayscale and edge index values ​​in a certain stage of the image sequence are selected as interval references. The change rate of each frame image is standardized and mapped, converting it into a small value between 0 and 1. For example, if the change value of a frame image is 15, within the range of 5 to 30, the normalized value is approximately 0.45. Combining the coordinates of the identified face regions with time data, an image information dataset is constructed. Each record contains the face bounding box coordinates, time point, and standardized image change index. For example, the record item could be: face position 320, 240, 100, 100, time point is 10:01:12 AM, and image change index is 0.45. Finally, a combined dataset of face localization and time series under continuous images is formed.

[0057] S102: Based on the image face coordinates and time dataset, call the coordinates of key bone points in the corresponding image frame, match the positional relationship between key points of the face and upper limbs, and perform synchronous comparison in time order to obtain the temporally synchronized bone joint data volume.

[0058] Based on the established face coordinate and time dataset, the coordinates of key points of the human skeleton are further extracted from the corresponding frame images. A pose estimation algorithm framework is then used to process the image frames, identifying key coordinate points of the head, shoulders, elbows, wrists, and other upper limb parts in each frame. The horizontal and vertical positions of each key point are obtained, such as 40° and 22° for the left shoulder and 50° and 30° for the right wrist. Combined with the face bounding box center positions, such as 37° and 26°, the spatial distance between the face center and each upper limb key point is calculated. The relative positions of these key points to the face center are then analyzed. The system determines the motion state in the image and arranges the skeletal points and face data frame by frame in a time sequence. The skeletal data recorded at each time point consists of key point coordinates and time points, maintaining a uniform format. The continuous frame data is uniformly arranged at fixed time intervals, such as forming a set of skeletal data every 0.1 seconds, forming a continuous skeletal joint structure data stream. For example, if 3 seconds are recorded continuously starting from 10:01:12, 30 sets of data are formed. Each set contains the relative positional relationship between key points such as shoulders, elbows, and wrists and the face, as well as the corresponding time points, forming a multi-dimensional skeletal sequence set.

[0059] S103: Based on the amount of time-synchronized skeleton joint data, combined with the face region location and image features, extract feature vectors from each frame of data in the image sequence, bind them with identity information and time data, and generate an identity recognition frame structure data group.

[0060] Based on the obtained skeletal structure data, and combining the facial region and overall image features, each frame of the image is vectorized. Observable features are transformed into a set of numerical feature vectors. Each vector contains the average grayscale value, image edge density, and the relative distances between key points on the face and upper limb. A uniform scaling transformation converts all features to standard values ​​between 0 and 1. For example, an average grayscale value of 120, within a preset range of 50 to 200, is converted to 0.47; an edge density of 4% is converted to 0.0. 6. The distance between the face and the left shoulder is 50 pixels. If the maximum distance is set to 100 pixels, it is converted to 0.5. After combination, feature vectors such as 0.47, 0.06, 0.5, and 0.55 are formed. Combined with the corresponding frame image time point and identity recognition mark, the feature vector is bound into a structured data record. Each data includes identity number, time point and standard feature vector. For example, the image frame corresponding to number 001 at 10:01:12 contains the above feature vector data. The frames are superimposed one by one to form a complete frame-level structured dataset, which is convenient for subsequent processing and analysis.

[0061] The specific steps of S2 are as follows:

[0062] S201: Based on the identity frame structure data group, call continuous image frames, extract the line connecting the ankle and the spine, calculate the angle difference between the directions of the lines between adjacent frames, obtain the direction change amount corresponding to each frame, and establish a sequence value of direction change amount.

[0063] Based on the identity frame structure data set, the first step is to use an image processing tool with skeletal point recognition capabilities to identify the target individual in a series of consecutive image frames. Key human node positions are extracted from the images, especially the ankle point and the center point of the spine. The pixel coordinates or spatial coordinates of these points in the image are used as input parameters. The human orientation vector of the frame is formed by calculating the spatial line connecting the ankle point and the center point of the spine. Then, the orientation difference between the current frame and the previous frame is constructed frame by frame. The change in direction is obtained by comparing the degree of change of the angle between the two vectors, thus forming a complete time series data. Taking a real image sequence as an example, with a frame rate set at 30 frames per second and a total observation time of 10 seconds, a total of 300 frames are involved. For each frame, skeletal point localization and line direction calculation are required. In each frame, a line segment is formed by the ankle point and the center point of the spine, and the angle difference between this line segment and the line segment in the previous frame is recorded. The list of all angle change values ​​is the direction change sequence, which is suitable for reflecting the angle fluctuations in continuous motion. For example, large twisting or rotating movements that occur during certain actions will be manifested as continuous large angle increases or decreases in this sequence.

[0064] S202: Based on the direction change sequence value, filter the vertical coordinate values ​​of the knee joint in the image frame of the railing area that are within the reference height range, calculate the vertical displacement between adjacent frames, and establish the displacement change sequence value of the railing area;

[0065] The specific formula for calculating the vertical displacement between adjacent frames is as follows:

[0066] ;

[0067] Calculate the dynamic weighted displacement value and establish a sequence of displacement changes in the railing area;

[0068] in, Represents the normalized displacement (unit: meters). Represents the vertical coordinate value of the knee joint in the i-th frame (unit: meters). Represents the vertical coordinate value of the knee joint in the (i-1)th frame (unit: meters). Represents the median value of the reference height range (unit: meters). The sign function is represented by 1 when x>0 and -1 when x<0, and the subscript i represents the time series number of consecutive image frames;

[0069] and The vertical pixel coordinates of the knee joint points in the image coordinate system of the i-th and (i-1)-th frames were extracted using video frame image processing algorithms (such as the OpenPose pose estimation model). After calibration and transformation, the actual physical coordinate values ​​(unit: meters) were obtained. In actual measurement data, the range of motion of the knee joint in adult subjects is typically 0.2-0.6 meters (based on ergonomic data). Coordinate values ​​from two consecutive frames were taken. rice, rice;

[0070] The median knee height within the baseline range is determined based on measurements of the subject's knee joint height while standing statically. Medical research shows that the median knee joint height in adults standing naturally is 0.45 meters (range ± 0.05 meters). rice;

[0071] : Symbolic function, when Output 1 at time. The output is -1, which is automatically determined by the displacement direction and does not require manual assignment.

[0072] Calculation process:

[0073] Calculation of original displacement difference:

[0074] ;

[0075] Calculation of numerator:

[0076] ;

[0077] Calculation of denominator terms:

[0078] ;

[0079] Normalization coefficient calculation:

[0080] ;

[0081] Sign function determination:

[0082] ;

[0083] Final displacement:

[0084] ;

[0085] Parameter settings are based on:

[0086] Meters: Knee height of the 50th percentile in adult males based on anthropometry datasets;

[0087] and Values: Calibration data from the "Walking" action sequence in the publicly available dataset Human3.6M;

[0088] Denominator structure design: Squared terms eliminate the influence of positive and negative deviations. To prevent division by zero errors, the square root operation maintains the unit of measurement as meters;

[0089] Numerical Result Analysis:

[0090] Calculation results The meter represents the displacement normalized to the reference height. Its physical meaning is: after eliminating the influence of individual height differences, the actual effective knee joint displacement between adjacent frames is 4.99 cm. This value will be input as a component of the displacement change sequence into the subsequent motion pattern analysis module.

[0091] The calculation logic of this formula constructs a dynamic weighting mechanism for the displacement through multiple steps: firstly, through... Obtain the original displacement difference and the median of the reference height. Multiplication achieves dimensional standardization; the denominator adopts... The structure includes a squaring operation to eliminate the directional nature of height deviation, adding the square of the reference height to form a stable denominator, and using the square root to ensure dimensional consistency with the numerator; absolute value operations extract displacement amplitude characteristics, and a sign function is used. By preserving the original displacement direction characteristics, the displacement calculation ultimately includes both normalization under the reference height constraint and maintains the physical meaning of the motion direction.

[0092] The vertical displacement between adjacent frames characterizes the instantaneous changes in the vertical movement of the knee joint in a video sequence. Specifically, it refers to the vector displacement calculated by the algebraic difference of the vertical coordinates of the knee joint in two consecutive frames. This value contains both displacement amplitude and direction information: the absolute value reflects the spatial change distance of the knee joint in the vertical dimension between adjacent frames, while the sign function preserves the direction of motion (positive values ​​indicate upward displacement, and negative values ​​correspond to downward movement). By introducing a normalization process with a reference height constraint, this displacement effectively eliminates scale bias caused by individual height differences or changes in shooting angle, forming a motion feature parameter with spatiotemporal continuity and physical comparability.

[0093] S203: Based on the sequence value of the direction change and the sequence value of the displacement change in the railing area, determine whether the direction change continuously exceeds the direction change judgment threshold and whether the vertical displacement continuously increases. If the condition is met, generate an abnormal motion trend indicator.

[0094] Combining the two data sequences mentioned above, reasonable judgment criteria need to be set when determining whether there is an abnormal motion trend. For the directional change, a reference angle value needs to be set as a threshold, which can be set to 15 degrees. When the directional change in consecutive image frames is higher than this value and continues to exceed a set number of frames (e.g., more than 5 consecutive frames), it is considered that the directional change is too large, constituting one of the judgment bases. For the vertical displacement, it is necessary to analyze whether its sequence shows a continuous increasing or continuously expanding trend. The judgment method can be set so that the displacement value for more than 5 consecutive frames decreases frame by frame (i.e., the height value decreases, indicating that the height of the human body relative to the ground increases frame by frame). Numerically, this can be expressed as a continuous decrease in vertical coordinate values, such as from 352, 346, 341, 337, 332, etc. If both of the above conditions are met simultaneously, the image frames within that time period can be marked as having an abnormal trend, and an abnormality flag variable is output for identification. This is usually set to a binary form, for example, 1 indicates that an abnormal trend has been detected, and 0 indicates no abnormality. This flag can be attached to the relevant information of the image sequence for use in subsequent image segment event marking or motion behavior recognition processing.

[0095] The specific steps for S3 are as follows:

[0096] S301: Extract the skeletal coordinates for each time series based on the abnormal motion trend markers, call the railing boundary range, determine the area where the skeletal coordinate points are located, record the direction, and generate the directional position value of the skeletal coordinate boundary.

[0097] Extracting skeletal coordinates for each time series based on abnormal motion trend markers requires identifying the coordinates of major human joints frame by frame using video images or infrared depth sensing devices, such as the head, neck, shoulders, elbows, wrists, hips, knees, and ankles, and generating a continuous set of skeletal points in chronological order. After identification, corresponding motion trend features need to be extracted for each joint coordinate, typically including velocity changes, joint angle fluctuations, and body center of gravity shift trajectories. These are then compared with historical samples to determine if the current coordinate sequence exhibits an abnormal trend. Using the torso center point as a reference point, the posture direction is determined by combining the relative positions between joints; for example, the angle between the line connecting the shoulders and the line connecting the hips can reflect the forward / backward or left / right directional trend. The railing boundary can be pre-defined as a fixed geometric boundary, such as a planar or rectangular area in space, with a range of 2 to 5 meters in the x-direction and 1.5 to 4.5 meters in the y-direction. The coordinates of the skeletal center point are extracted in each frame to determine if it falls within this boundary area, and this is recorded in conjunction with the orientation angle. For example, if the waist point is 3.1 meters and the shoulder direction is 75 degrees, it indicates that the person is currently within the area and facing east-northeast. The recorded value is set as a combination of area status and orientation, such as "inside, 75 degrees". This value serves as a boundary directional position marker for the skeletal coordinates and is recorded frame by frame to form a complete time series.

[0098] S302: Based on the directional position values ​​of the skeletal coordinate boundary, the directional identifiers of time series points are merged and statistically analyzed to screen for directional jumps, extract the time series intervals corresponding to the jump segments, and obtain the boundary crossing trend recognition segment sequence.

[0099] First, a directional partitioning standard is established, dividing the 0-360 degree range into several main directional segments. For example, each segment is divided into eight sections: East, Northeast, North, Northwest, West, Southwest, South, and Southeast, grouped by 45 degrees. The orientation angles recorded in each frame are then categorized. Time frames showing consecutive occurrences of the same direction are grouped into a single segment, forming a stable directional interval. A significant change between consecutive directional segments, such as a jump from South to North, is defined as a directional jump. To determine the magnitude of the jump, an angle difference threshold, such as 45 degrees, can be set. If the directional difference between consecutive segments exceeds this value, it is considered a jump event. Further, boundary state judgment is incorporated. If, during a directional jump, a skeleton point moves from outside the region into the region, or moves in the opposite direction, the segment is identified as a boundary crossing trend recognition segment. For example, if the direction changes from West to East in frames 12 to 16, and the skeleton point also moves from outside the boundary to inside, this segment is recorded as a boundary crossing trend recognition interval and added to the segment sequence for subsequent processing.

[0100] S303: Identify the segment sequence based on the boundary crossing trend, call the skeletal trajectory to calculate the boundary intersection position, statistically analyze the continuity index, determine whether a crossing behavior is constituted, and generate a spatial boundary crossing record.

[0101] After obtaining the boundary crossing trend recognition segment sequence, it is necessary to further combine it with skeletal trajectory data to calculate the spatial location of the boundary intersection point. Extract the continuous coordinate sequence within the recognition segment and examine the changes of the skeletal center point inside and outside the boundary in chronological order. When a skeletal point is outside the boundary in one frame and inside the boundary in another frame in two consecutive time frames, it can be determined that a boundary crossing has occurred, and the coordinates of the crossing point are estimated on the line connecting the two points. Position interpolation can be used to roughly estimate the location of the crossing by analyzing the trend of point changes. For example, if the coordinates in frame 15 are 2.5 meters and in frame 16 are 3.1 meters, and the boundary is set at 3 meters, then it can be considered that the boundary has been crossed in this segment, and the estimated intersection point can be obtained by further combining the vertical coordinates. Statistically analyze the frequency and duration of such intersection points throughout the entire recognition segment, and set a minimum consecutive frame standard for a valid crossing, for example, 3 frames. If there are more than 3 consecutive frames showing crossing behavior, it is considered a valid crossing. Finally, the spatial coordinates, direction changes, and time intervals of the crossing points are merged to form a crossing record item. For example, in a certain segment, between frames 14 and 18, a crossing occurs from outside the region into the region, with the direction changing from west-southwest to east-northeast. The crossing point is estimated to be 3 meters horizontally and 2.3 meters vertically. This information is then fully recorded for subsequent analysis and processing.

[0102] The specific steps of S4 are as follows:

[0103] S401: Call the personnel identity information corresponding to the spatial boundary crossing record, compare it with the identity identifier field in the operation permission control table, extract the permission level field corresponding to the matching item, obtain the personnel's level classification identifier based on the correspondence between the level field and the type field, calculate the level classification identifier parameter, and generate the permission level classification value.

[0104] The specific formula for calculating the classification identifier parameters is as follows:

[0105] ;

[0106] in, The comprehensive evaluation value (dimensionless) representing the classification of authority levels. The numerical value (dimensionless) represents the current permission level field. The base value (dimensionless) of the representative type field. Represents the measurement value (dimensionless) of the current type field. Represents a dynamic correction coefficient (dimensionless) based on historical permission adjustment records. It is the unit dimensionless equilibrium constant (with a value of 1). This represents the number of times the related type field has been adjusted historically. Representing the The weighting factor (dimensionless) of the next historical adjustment. This represents the absolute value of the difference between type fields. This indicates the standardization process for dynamic correction coefficients;

[0107] Permission level field value The system obtains the user's current permission level code in real time through the permission management system. The coding rule is an integer value (levels 1-5 correspond to basic permissions to administrator permissions); for example, if user A's current permission level is level 3, then... ;

[0108] Type field baseline value According to the company's "Permission Type Classification Standard" document, the baseline value of the type field is fixed as the maximum allowed value of the type code. Referring to Section 5.2 of the standard document, the upper limit of the type code is... ;

[0109] Current type field measurement value The user behavior analysis system collects the distribution of user operation types over the past 30 days and calculates their type characteristic values. For example, user A's operation frequency in the audit module, data export module, and configuration modification module is 45 times, 30 times, and 25 times respectively, and is calculated using the standard weighted formula. ;

[0110] Dynamic correction coefficient Based on the permission audit logs, the number of times and the magnitude of historical permission adjustments were calculated using the formula. calculate;

[0111] User A has 3 privilege upgrade records in the past 12 months, then ;

[0112] Unit dimension equilibrium constant According to the standardization requirements of the International System of Units, set Used to maintain the consistency of the dimensions of the formula;

[0113] Number of historical adjustments : Directly read the total number of user permission change events recorded in the permission management system, and the number of times user A's permissions have been adjusted. ;

[0114] Weighting factors According to the "Permission Weight Calculation Specification", the weight factor for each adjustment is determined by the adjustment time decay coefficient and the adjustment magnitude.

[0115] User A adjusted the weighting factors three times as follows:

[0116] First adjustment (12 months ago): ;

[0117] Second adjustment (6 months ago): ;

[0118] Third adjustment (1 month ago): ;

[0119] Formula calculation process:

[0120] Calculate the numerator: ;

[0121] Calculate the denominator: ;

[0122] Calculate the fractional terms: Calculate the summation term: ;

[0123] Comprehensive calculation: ;

[0124] Result Analysis: This result This represents the comprehensive evaluation value for user A's permission level classification. According to Article 7.3 of the "Permission Level Implementation Specification," when... When the permission review process is triggered, this value reflects that the match between the user's permission level and operation behavior exceeds the normal threshold, and manual review is required to confirm the rationality of the permission allocation.

[0125] The formula's operation logic uses multiplication. The numerical value of the permission level is coupled and amplified with the difference in the type field to highlight the synergistic effect of the core parameters; square root calculation is used. Nonlinear compression is applied to the dynamic correction coefficients to suppress the impact of extreme value fluctuations on the calculation results; summation operation By linearly superimposing historical adjustment weighting factors, a cumulative parameter effect over time is achieved. The numerator constructs the basic authority assessment quantity, the denominator implements dynamic stability control, and the superposition term introduces historical behavior corrections. All three are normalized. To ensure the additivity of operational items, a comprehensive authority assessment system is ultimately formed;

[0126] The level classification identifier parameter (H) is a multi-dimensional fusion quantitative indicator in the access control system. Essentially, it integrates discrete access characteristics into a continuous and comparable evaluation value through structured operations. This parameter comprises three core dimensions: basic access dimension (H...). This reflects the degree of linear deviation between the current permission status and the baseline configuration, and dynamically adjusts the dimension ( ) Damping control of permission fluctuations is implemented through nonlinear transformation, historical correction dimension ( The time-series weighted cumulative historical adjustment trajectory is adopted. Parameter design is achieved through dimensional unification (introducing...). It enables the comparability and superposition of cross-dimensional data, and the final output value represents the comprehensive permission adaptability of personnel under the current spatiotemporal conditions, providing a quantifiable decision-making basis for permission classification.

[0127] S402: Based on the permission level classification value, call the facial feature vector field associated with the person's identity, perform structural division of the feature encoding, divide the feature vector field into several feature segments according to the feature dimension truncation rules set by the permission level classification, and attach corresponding field labels to obtain the feature segment label field group.

[0128] Based on the Class_O classification identifier, it is necessary to extract its associated facial feature encoding field. The feature vector matching ID 10087 is retrieved from the identity information database. For example, this vector is a 128-dimensional numerical set. According to the pre-defined segmentation rules of Class_O, it is divided into several segments, such as the first 64 dimensions and the last 64 dimensions. The first segment is labeled F_O_A, and the second segment is labeled F_O_B. The system performs segmentation on the facial feature vector according to this rule and labels each segment accordingly. In a real-world scenario, if the personnel feature encoding is a set of 128 floating-point values ​​from 0.12 to 0.87, the first to 64 dimensions are extracted according to the segmentation rules to form the first segment, labeled F_O_A, and the second segment is formed from the 65th to 128th dimensions, labeled F_O_B. The resulting structure contains two vector encodings and their labels: F_O_A corresponds to the first half of the encoding, and F_O_B corresponds to the second half. This process completes the feature field segmentation and labeling operation under the control of the class dimension.

[0129] S403: Based on the field labels in the feature segment marker field group and the mapping rules set in the encryption level reference parameter table, select the encryption level coefficient corresponding to each group of fields, add encryption level identifiers to the feature segments, and generate a feature vector field encryption grouping table.

[0130] Based on the generated feature segment labels F_O_A and F_O_B, and combined with the encryption level comparison parameter configuration table, the system assigns an encryption level coefficient to each label according to the set rules. For example, under Class_O permission, F_O_A is set to level 2 and F_O_B is set to level 3. The system reads the field labels sequentially and matches them with the preset rules, adding an encryption level value to each feature code to form a complete encryption field group. In the actual scenario, the encryption level of the first 64-dimensional vector represented by F_O_A is 2, and the encryption level of the last 64-dimensional vector represented by F_O_B is 3. The system records the label, value, and encryption level of each vector segment to obtain a structured encryption feature group, which forms an encryption level field set. Each field contains a clear feature segment and its level value. For example, F_O_A has an encryption level of 2, and F_O_B has an encryption level of 3, realizing the face encoding encryption marking operation driven by level classification.

[0131] The specific steps of S5 are as follows:

[0132] S501: Based on the feature vector field, encrypt the encrypted label field in the grouping table, call the feature code index and matching number content in the feature vector field, complete the classification and sorting according to the group number, construct the feature vector set corresponding to the group, and obtain the group vector feature value set;

[0133] The encrypted tag field in the feature vector encryption grouping table can be used as the initial operation object. When calling the feature code index and matching number content, the index value in the field should be extracted one by one, such as the index number IDX001, IDX002, etc., and combined with the matching numbers mapped in the data table, such as M001 to M100, to form a one-to-one or many-to-one relationship. Based on the mapping relationship, the connection processing between the encrypted tag and the matching number is completed. Then, the group number field defined in the table is read and classified according to the range Group01 to Group10. Each matching number is assigned to its group. After classification, feature aggregation is performed on all feature vectors in the same group, and the similarity ratio is used. Compared to other methods, the feature vectors are represented by three-dimensional values. For example, vector A is represented by 0.2, 0.4, and 0.1, and vector B is represented by 0.3, 0.1, and 0.5. The similarity between the two vectors is obtained by multiplying each of the three-dimensional values, summing the results, and dividing by the product of the square and the square root of the sum. A similarity of 0.82 indicates a high degree of similarity. If a similarity threshold of 0.75 is set, the vectors can be grouped into the same set. This group ultimately includes vector objects such as V1, V2, V5, and V6, and is labeled with the group number, such as Group03. This completes the construction of the feature vector set for all groups, forming the feature value set corresponding to each group, which is used in subsequent behavior recognition steps.

[0134] S502: Based on the grouped vector feature value set, call the identification time and terminal number fields in the identification frame structure data group, sort the terminal numbers according to the identification time, extract data pairs that meet the identification time interval conditions, and determine the behavior type by combining the group number and feature value difference to obtain the behavior type identification group;

[0135] After constructing the grouped vector feature set, the identity frame structure data group is read, and the identification time and terminal number fields are extracted. The identification time field is in standard timestamp format, such as 08:31:05. The terminal number field is such as TID007, TID010, etc. The identification time for each terminal number is arranged in chronological order to construct a time series list. For example, the identification time series for TID007 is 08:31:05, 08:31:30, and 08:31:55. The time interval between any two adjacent time points is calculated. If the time interval threshold is set to 60 seconds, then... Data pairs with an interval of 60 seconds or less are retained to form an identification sequence. The differences between each feature vector in the sequence are compared in turn. The absolute value of the difference in each feature dimension is taken and summed to obtain the total difference. If the difference range for behavior type identification is set to 0.3 to 0.6, then when the difference is 0.5, it is classified as type B behavior. This process is repeated to complete the classification and judgment of behavior type for each terminal number in each time period. For example, TID007 is identified as type B between 08:31:05 and 08:31:55, forming a behavior identification group. For example, G1 contains two records, which represent the behavior type at different time points.

[0136] S503: Based on the group number of the behavior type identification group, combine the identity code, identification time point, behavior type and encryption level to construct the corresponding data structure and generate an abnormal identification event output data block;

[0137] Based on the identified behavior type groups, the identity code, identification time, behavior type, and encryption level fields of each record are combined and constructed according to the group number. The identity code is a unique identifier, such as ID_9384. The identification time is taken from the aforementioned timestamp field. The behavior type is such as type A or B. The encryption level field is set according to the behavior difference. For example, it is set to be marked as level 3 when the total difference of the feature vector exceeds 0.6, otherwise it is level 2. If the difference of a certain record is 1.2, which exceeds the upper limit, it is marked as level 3. Then, data entries in the form of a structure are constructed. Each record contains 5 fields and is merged to form an output data block. For example, one record is G2, ID_0003, 08:32:25, type B, level 3, and another record is G2, ID_0003, 08:32:40, type A, level 2. Multiple records form an identification event output data block, which is used for data stream processing by subsequent processing modules.

[0138] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. An artificial intelligence-based construction site face recognition safety protection method, characterized in that, Comprise the following steps: S1: Obtain the image sequence collected by the site entrance face terminal, extract the face region coordinates, feature vector and identification time information of the personnel in the image, record the key skeleton point coordinate information in the standing state synchronously, and bind with the identity information to generate an identification identity frame structure data group; S2: Identify the continuous image frames in the identity frame structure data group, extract the ankle to spine connection direction, calculate the frame direction difference value, extract the vertical displacement of the knee joint in the reference height range of the railing area, judge whether the direction difference value continuously exceeds the threshold value and the displacement continuously increases, and if the conditions are met, an abnormal motion trend marker is established; S3: According to the skeleton coordinates recorded in the abnormal motion trend marker, match the railing safety boundary range set in the site boundary coordinate set, compare whether the coordinate points exist continuous crossing behavior area, and if crossing occurs, generate a space boundary crossing record; S4: Call the personnel identity information corresponding to the space boundary crossing record, obtain the permission level in the work permission control table, segment process the face feature vector field according to the level, mark the encryption level, and generate a feature vector field encryption grouping table; S401: Call the personnel identity information corresponding to the space boundary crossing record, and compare with the identity identification field in the work permission control table, extract the permission level field corresponding to the matching item, obtain the level classification identification of the personnel according to the corresponding relationship between the level field and the type field, calculate the level classification identification parameter, and generate a permission level classification value; S402: Based on the permission level classification value, call the face feature vector field associated with the personnel identity, divide the feature code structure, divide the feature vector field into several feature sections according to the feature dimension interception rule set by the level classification, and add the corresponding field label to obtain a feature section marker field group; S403: According to the field label in the feature section marker field group, select the encryption level coefficient corresponding to each field group according to the mapping rule set in the encryption level parameter table, add the encryption level identification to the feature section, and generate a feature vector field encryption grouping table; The specific calculation formula of the level classification identification parameter is: ； wherein, a comprehensive evaluation value representing a classification of the authority level, a numerical value representing the current authority level field, a reference value representing the type field, a measured value representing the current type field, a dynamic correction coefficient based on a history authority adjustment record, a unit dimension balancing constant, a number of history adjustments associated with the type field, a weight factor representing the nth history adjustment, an absolute value amount representing a difference in the type field, and a standardization process of the dynamic correction coefficient.

2. The construction site face recognition security protection method based on artificial intelligence according to claim 1, characterized in that, The identification identity frame structure data group comprises face region coordinates, feature vector, identification time information, skeleton point coordinates and identity information, the abnormal motion trend marker comprises direction difference value, vertical displacement and skeleton coordinates, the space boundary crossing record comprises crossing behavior area, railing safety boundary range and coordinate point change track, and the feature vector field encryption grouping table comprises face feature vector field, encryption label and encryption level.

3. The construction site face recognition security protection method based on artificial intelligence according to claim 1, characterized in that, The specific steps of S1 are: S101: Obtain the image sequence collected by the site entrance face terminal, extract the face region coordinates and time information of the personnel in the image, calculate the image gradient change rate according to the gray distribution and edge point density, perform normalization processing, and generate an image face coordinate and time data set; S102: Based on the image face coordinates and time data set, the key skeleton point coordinates in the corresponding image frame are called, the position relationship between the face and the upper limb key points is matched, and the time sequence synchronization skeleton joint data is obtained by synchronous comparison in time sequence; S103: According to the time sequence synchronization skeleton joint data, the feature vector of each frame of data in the image sequence is extracted in combination with the face region position and image features, and is bound with the identity information and time data to generate the recognition identity frame structure data group.

4. The construction site face recognition safety protection method based on artificial intelligence according to claim 3, characterized in that, The specific steps of S2 are: S201: Based on the recognition identity frame structure data group, the continuous image frames are called, the connecting line between the ankle and the spine is extracted, the included angle difference of the connecting line direction between adjacent frames is calculated, the direction change value corresponding to each frame is obtained, and the direction change value sequence value is established; S202: Based on the direction change value sequence value, the vertical coordinate value of the knee joint in the railing area image frame within the reference height range is screened, the vertical displacement between adjacent frames is calculated, and the railing area displacement change sequence value is established; S203: According to the direction change value sequence value and the railing area displacement change sequence value, it is judged whether the direction change continuously exceeds the direction change judgment threshold value and the vertical displacement continuously increases, and the abnormal motion trend mark value is generated when the conditions are met.

5. The construction site face recognition security protection method based on artificial intelligence according to claim 4, characterized in that, The specific calculation formula of the vertical displacement between adjacent frames is: ； The dynamic weighted displacement value is calculated, and the railing area displacement change sequence value is established; wherein, represents a normalized displacement amount, represents an i-th frame vertical coordinate value of the knee joint, represents an i-1-th frame vertical coordinate value of the knee joint, represents a reference height range median value, represents a sign function, and subscript i indicates a time series number of consecutive image frames.

6. The construction site face recognition security protection method based on artificial intelligence according to claim 4, characterized in that, The specific steps of S3 are: S301: Based on the abnormal motion trend mark value, the skeleton coordinates under each time sequence are extracted, the railing boundary range is called, the region where the skeleton coordinate point is located is judged, and the direction orientation is recorded, and the skeleton coordinate boundary directional position value is generated; S302: Based on the skeleton coordinate boundary directional position value, the direction identification of the time sequence point is merged and counted, the direction jump situation is screened, the time sequence interval corresponding to the jump section is extracted, and the boundary crossing trend identification section sequence is obtained; S303: According to the boundary crossing trend identification section sequence, the boundary intersection point position is calculated by calling the skeleton trajectory, the continuity index is counted, it is judged whether the crossing behavior is constituted, and the space boundary crossing record is generated.

7. The construction site face recognition security protection method based on artificial intelligence according to claim 1, characterized in that, The method further comprises: S5: The encryption label in the encryption grouping table is combined with the feature vector field, the identification time and terminal number information in the recognition identity frame structure data group are called, the abnormal behavior record item is constructed, and the identity code, identification time point, behavior type and encryption level are combined to generate the abnormal identification event output data block; The abnormal identification event output data block comprises identity code, identification time point, behavior type, encryption level.

8. The construction site face recognition security protection method based on artificial intelligence according to claim 7, characterized in that, The specific steps of S5 are: S501: Based on the encryption label field in the feature vector field encryption grouping table, the feature code index and matching number content in the feature vector field are called, the classification arrangement is completed according to the grouping number, the feature vector set corresponding to the grouping is constructed, and the grouping vector feature value set is obtained; S502: Based on the set of eigenvalues of the grouping vector, call the identification time and terminal number field in the identification frame structure data group, sort the terminal number according to the identification time, extract the data pair whose identification time interval meets the condition, combine the grouping number and eigenvalue difference to judge the behavior type, and obtain the behavior type identification group; S503: According to the grouping number of the behavior type identification group, combine the identity code, identification time point, behavior type and encryption level to construct the corresponding data structure, and generate abnormal identification event output data block.

Images