A visual large model-based safety warning method

Abstract

The present application relates to the technical field of intelligent security and protection, and particularly relates to a safety early warning method based on a visual large model, comprising the following steps: a visual large model identifies a visual corresponding relationship between a current work object and visible safety measures in a picture, and generates a work-measure anchoring unit; it is judged whether the visible safety measures are truly applied to the current work object, and when the visible safety measures only form a corresponding relationship with adjacent objects, historical work objects or non-current work positions, a borrowing compliance risk unit is generated; a safety early warning result is generated according to the borrowing compliance risk unit. The present application solves the borrowing compliance problem which is widespread in industrial sites but difficult to be found by existing visual systems, that is, the safety measures exist visually, the site seems to be compliant on the surface, but the actual protection relationship has been mismatched.

Description

Technical Field

[0001] This invention relates to the field of intelligent security technology, and in particular to a security early warning method based on a large visual model. Background Technology

[0002] With the advancement of intelligent transformation of industrial sites, video surveillance-based safety early warning technology has been widely applied in scenarios such as working at heights, hoisting operations, electrical maintenance, confined space operations, and hazardous area management. Currently, most visual safety early warning systems mainly use target detection, behavior recognition, or rule matching methods to identify obvious violations such as not wearing safety helmets, personnel crossing boundaries, the presence of smoke or fire, and abnormal approach of equipment, thereby achieving on-site safety risk warnings.

[0003] Existing technologies typically assume that the presence of safety measures in the visual field indicates a protected state for the current operation. For example, when a system detects safety ropes, warning tapes, tags, grounding wires, or protective nets in the visual field, it often directly determines that the current site meets safety requirements. However, in real industrial settings, a truly effective protective relationship does not necessarily exist between safety measures and the current work object. Especially in situations involving multiple personnel working simultaneously, multiple pieces of equipment operating concurrently, and dynamic shifts in the work area, a hidden risk state that is difficult for existing technologies to identify can easily arise: the phenomenon of "borrowed compliance." Borrowed compliance refers to a situation where safety measures are visually present, but these measures do not actually apply to the current work object; instead, they still apply to adjacent objects, historical work objects, or historical work areas. For example, during high-altitude operations, the safety rope may appear to be attached, but it is actually attached to a worker on the side; during hoisting operations, the warning tape may still surround the old hoisting area, but the hoisting path may have changed; during equipment maintenance, the tag lock may still correspond to the equipment that was previously maintained, but the current equipment may have been switched; in confined space operations, the monitoring personnel may still be visible, but they may actually be the previous group of workers. Summary of the Invention

[0004] This invention provides a safety early warning method based on a visual large model. It uses a visual large model to establish the real interaction relationship between safety measures and work objects, and further identifies whether safety measures have mismatched positions, historical residues, or failed due to position migration, thereby realizing the identification and proactive correction of compliance risks related to position mismatch.

[0005] A security early warning method based on a large visual model includes the following steps: S1. Input the continuous operation screen of the safety production site into the visual big model. The visual big model identifies the visual correspondence between the current operation object and the visible safety measures in the screen and generates operation-measure anchoring units. S2. Based on the operation-measure anchoring unit, determine whether the visible safety measures are actually applied to the current operation object. When the visible safety measures only correspond to adjacent objects, historical operation objects, or non-current operation locations, generate a borrowing compliance risk unit. S3. Generate a security warning result based on the borrowing compliance risk unit, and output the mismatch security measures, actual protected objects, currently unprotected objects, and correction instructions corresponding to the borrowing compliance risk unit.

[0006] Optionally, S1 further includes extracting frames and timestamping them on the continuous operation screen to form a video frame sequence with time sequence identifiers, and inputting the video frame sequence into the visual big model.

[0007] Optionally, the large visual model performs parallel segmentation of the work object instance and detection of safety measures for each frame, extracting the first spatial region features of the current work object and the second spatial region features of the visible safety measures; for the same frame, it calculates the visual association strength between the first spatial region features and the second spatial region features, wherein the visual association strength includes at least one of spatial intersection-union ratio, distance between key component connection points, and relative pose consistency.

[0008] Optionally, it also includes a preset association threshold, forming candidate anchor pairs with the visual association strength of the work objects exceeding the preset association threshold and the visible safety measures in the same frame, and performing cross-frame target tracking and association relationship confirmation on the candidate anchor pairs in multiple consecutive frames, and determining the candidate anchor pairs that simultaneously satisfy spatial continuity and temporal stability as the work-measure anchoring units.

[0009] Optionally, S2 includes obtaining the current job-measure anchoring unit and parsing out the first instance identifier of the current job object and the second instance identifier of the visible safety measure from it; Based on the first instance identifier and the second instance identifier, a search is performed in the historical anchoring association records formed by cross-frame target tracking to determine the historical operation object identifier and adjacent object identifier in which the visible safety measures were continuously in effect in the historical frame.

[0010] Optionally, for the current frame, calculate the first current association strength between the second spatial region feature of the visible safety measures and the first spatial region feature of the current work object, and calculate the second current association strength between the second spatial region feature and the adjacent object region, the historical work object region, and the non-current work location region.

[0011] Optionally, if the first current association strength is lower than a preset effective action threshold, and at least one second current association strength is higher than the first current association strength, it is determined that the visible safety measures are not actually applied to the current work object, but are applied to adjacent objects, historical work objects, or non-current work locations. The borrowing compliance risk unit is generated based on the judgment result. The borrowing compliance risk unit includes at least the mismatch security measure identifier corresponding to the visible security measure, the actual protected object identifier, the currently unprotected object identifier, and the borrowing type.

[0012] Optionally, the mismatch security measure identifier, the actual protected object identifier, the currently unprotected object identifier, and the borrowing type can be extracted from the borrowing compliance risk unit; Based on the borrowing type, the current security risk level is determined according to the preset mapping relationship between borrowing type and security risk level; The safety warning result is generated, which includes the current safety risk level, the mismatched safety measure identifier, the actual protected object identifier, and the currently unprotected object identifier. The result is then displayed visually on the continuous operation screen and / or alarmed via an on-site alarm device.

[0013] Optionally, based on the borrowing type, the identifier of the currently unprotected object, and the current work scenario information, rule matching is performed in a preset safety correction rule base to generate a corresponding correction handling instruction.

[0014] Optionally, the corrective action instruction includes: a measure reset instruction to adjust the mismatched safety measures from the actual protected object to the currently unprotected object, and / or an emergency shutdown instruction to control the currently unprotected object to suspend operation; the corrective action instruction is sent to the corresponding safety measure control terminal or work equipment control system through a communication interface to execute the corrective action.

[0015] The beneficial effects of this invention are: This invention, through the construction of a progressive safety early warning mechanism comprising operation and measure anchoring units, misalignment compliance risk units, and corrective action instructions, achieves the identification of hidden risks—that safety measures exist but fail to actually protect the current work object. This differs from existing technologies that only detect whether targets such as safety helmets, warning tapes, and safety ropes appear in the image. This invention addresses the misalignment compliance problem commonly found in industrial sites but difficult to detect by existing vision systems; that is, situations where safety measures are visually present and the site surface appears compliant, but the actual protection relationship has been mismatched.

[0016] This invention, by introducing cross-frame target tracking, historical anchoring and association recording, and a current association strength comparison mechanism, enables the large-scale visual model to not only identify target relationships within a single frame but also understand the continuous effectiveness of safety measures over time. This solves the problem that traditional visual algorithms cannot determine whether safety measures have failed or migrated. By comparing the first and second current association strengths, this invention can identify whether safety measures are still effective on historical work objects, adjacent objects, or historical work locations, avoiding misjudgments caused by factors such as short-term occlusion, target proximity, or historical area residue. Compared to the static detection method in existing technologies that considers safety effective simply by observing safety measures, this invention enables dynamic and continuous verification of safety protection relationships, improving the authenticity and accuracy of safety warning results.

[0017] This invention achieves closed-loop safety control from risk identification to proactive correction by establishing a mapping mechanism between borrowing types and risk levels, as well as a safety correction rule base. It can not only output mismatch safety measures, actual protected objects, and currently unprotected objects, but also automatically generate corresponding measure reset instructions or emergency shutdown instructions based on different borrowing types, and perform proactive intervention through communication interface linkage with the operating equipment control system. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in this invention 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 for this invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a schematic diagram of the method flow according to an embodiment of the present invention; Figure 2 This is a schematic diagram illustrating the generation of correction instructions according to an embodiment of the present invention. Detailed Implementation

[0020] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. For some well-known technologies, those skilled in the art may also use other alternative methods to implement the invention. Moreover, the accompanying drawings are only for more specific description of the embodiments and are not intended to specifically limit the present invention.

[0021] like Figures 1-2 As shown, a security early warning method based on a large visual model includes the following steps: S1. Input the continuous operation screen of the safety production site into the visual big model. The visual big model identifies the visual correspondence between the current operation object and the visible safety measures in the screen and generates operation-measure anchoring units.

[0022] The visual large model of the present invention includes the following parts: Timing video input module: used to receive video frame sequences with timing identifiers and establish time correlation between consecutive frames.

[0023] Visual semantic encoding module: used to extract visual semantic features such as people, equipment, hoisting objects, and safety measures from video frames.

[0024] The task object instance segmentation module is used to identify and segment task objects such as construction personnel, equipment, vehicles, and hoisted objects, and generate a set of task object regions.

[0025] Safety measure target detection module: used to detect safety measures such as safety ropes, warning tapes, signs, and barriers, and generate a set of safety measure areas.

[0026] Spatial Feature Extraction Module: Used to extract spatial features such as the area boundaries, key connection points, movement direction, and extension direction of the work object and safety measures.

[0027] Visual association analysis module: used to calculate the visual association strength between the work object and the safety measures, and to determine whether the two form a real protective relationship.

[0028] Cross-frame correlation tracking module: used to continuously track candidate anchor pairs based on consecutive video frames, and analyze spatial continuity and temporal stability.

[0029] In this embodiment, the specific components include: S11. Time-sequential preprocessing of continuous operation screens.

[0030] First, continuous operational footage from the safety production site is acquired. This footage can originate from fixed surveillance cameras, PTZ cameras, mobile inspection terminals, or industrial edge vision acquisition devices. To ensure that the subsequent large-scale visual model can identify the dynamic relationships during the operation, frame extraction and timestamp annotation are performed on the continuous operational footage.

[0031] Specifically, the video stream is sampled at a preset sampling frequency to form a video frame sequence: ;in, Represents a video frame sequence. Indicates the first One video frame, This indicates the total number of video frames after frame extraction.

[0032] Subsequently, a corresponding acquisition time is appended to each video frame to generate a video frame sequence with time sequence identifiers: ;in, This represents a sequence of video frames with timing identifiers. Indicates the first The timestamp corresponding to each video frame.

[0033] After timestamp annotation, the video frame sequence with time-series identifiers is input into the visual big data model. Instead of treating each video frame as an independent static image for separate identification, adjacent video frames are grouped into a continuous temporal input window according to their timestamp order. This allows the visual big data model to simultaneously reference the object position, motion state, and safety measure status from previous video frames when processing the current video frame. In practice, the visual big data model retains the object identification, motion trajectory, posture change trend, and safety measure correspondence from previous video frames, and continuously matches the same object in subsequent video frames to identify whether the work status has continuously changed. For example, it can identify whether a person changes from "safety rope attached" to "unattached," whether the warning tape moves synchronously with the hoisting area, and whether the tag locking loses its corresponding relationship after equipment switching.

[0034] To reduce visual interference caused by dust, strong light, low illumination at night, and equipment reflections in industrial settings, brightness normalization and local contrast enhancement can be performed on video frames before inputting them into the large visual model, thereby improving the stability of subsequent object and safety measure recognition.

[0035] When brightness normalization is implemented, the overall brightness distribution of the current video frame is statistically analyzed, and the deviation between the average brightness of the current image and the preset reference brightness is calculated. When low light at night, local overexposure, backlight, or strong welding light are detected, the brightness of each pixel in the video frame is uniformly stretched or compressed to ensure that video frames at different times maintain a similar overall brightness range, preventing the visual model from misidentifying the same object as different targets due to changes in ambient light. Local contrast enhancement, after brightness normalization, divides the video frame into multiple local regions and enhances the grayscale difference between the target edge and the background in each local region to strengthen the visibility of small targets such as safety ropes, warning tapes, signs, and equipment edges. For areas obscured by dust, smoke, or equipment reflections, the local texture changes in the target outline area are enhanced first, enabling the visual model to more stably identify the true correspondence between safety measures and work objects.

[0036] S12. Parallel identification of work objects and safety measures.

[0037] After receiving a sequence of video frames with time-stamped identifiers, the visual big data model first reads the video frames one by one according to their timestamps. Within each video frame, a unified visual analysis task is established, consisting of two parallel branches: the first branch is for task object instance segmentation, and the second branch is for safety measure target detection. Task object instance segmentation and safety measure target detection are performed in parallel for each video frame. Task objects include hoisted objects, construction workers, mobile equipment, maintenance equipment, transport vehicles, high-altitude work platforms, edge structures, or hazardous areas; safety measures include warning tapes, safety ropes, signs, barriers, protective covers, grounding wires, monitoring personnel, or protective nets.

[0038] The first branch: After receiving the video frame sequence with time-series identifiers, the visual big model reads the current video frame in time stamp order and inputs it into the task object instance segmentation branch. The task object instance segmentation branch performs visual encoding on the current video frame, extracting visual semantic regions such as personnel outlines, equipment edges, hoisted object shapes, vehicle outlines, and work platform boundaries. Based on the preset task object category prompts, the branch classifies the above visual semantic regions, identifying personnel, equipment, hoisted objects, vehicles, maintenance parts, work platforms, and hazardous area boundaries related to the current safe production operation as task objects. Different individuals within the same category are distinguished as instances, each task object is assigned an independent instance number, and corresponding instance masks, bounding boxes, center point positions, category labels, and recognition confidence scores are generated. The recognition results of all task object instances in the same video frame are summarized to form the task object region set corresponding to that video frame.

[0039] The second branch: The visual big model synchronously launches the safety measure target detection branch in the same current video frame. First, based on the preset safety measure category prompts, it locks the types of safety measures that need to be detected, including warning tape, safety rope, hanging signs, isolation fences, protective covers, grounding wires, monitoring personnel, protective nets, warning signs, and temporary barriers. It searches for target candidate regions in the current video frame, focusing on locating candidate regions with color markings, strip shapes, hanging shapes, connection shapes, barrier shapes, or sign shapes. It determines the safety measure category of each candidate region, excluding ordinary cables, background signs, equipment nameplates, or irrelevant barriers that are similar in appearance to safety measures but do not have a safety constraint function. It generates detection results for confirmed safety measure targets, recording their bounding boxes, visible range, connection points, extension direction, attached objects, category labels, and detection confidence. It summarizes the detection results of all safety measure targets in the same video frame to form the set of safety measure regions corresponding to that video frame.

[0040] Regarding the first video frames The set of task object regions output by the large visual model is represented as follows: ;in, Indicates the first The set of task object regions in a video frame. Indicates the first The first video frame Each task object area, This indicates the number of job objects in the current video frame.

[0041] The set of safety measure regions output simultaneously is represented as follows: ;in, Indicates the first A set of security measure regions in a video frame. Indicates the first The first video frame Each security measures area This indicates the number of security measures in the current video frame.

[0042] Subsequently, the first spatial region features of the work object area and the second spatial region features of the safety measures area were extracted respectively.

[0043] First, the spatial region features are extracted as follows: Instance mask parsing is performed on the work object region. First, the instance mask of the work object output by the large visual model is read, and the pixel regions belonging to the same work object in the instance mask are taken as the target region. Minimum bounding rectangle fitting is performed on the target region to obtain the bounding box of the work object, and the center position is calculated based on the horizontal and vertical coordinates of the bounding box. Contour extraction is performed on the edges of the instance mask to obtain the outer contour curve of the work object, and the contour shape is formed based on the aspect ratio, area, perimeter, principal axis direction, and contour curvature of the outer contour curve. Existing keypoint detection models are called for different types of work objects, specifically including: OpenPose human keypoint detection model, Keypoint R-CNN keypoint detection model, or YOLO-Pose keypoint detection model. For construction workers, OpenPose or YOLO-Pose is preferred; for industrial objects such as hoisted objects, vehicles, and robotic arms, Keypoint R-CNN can be used. For personnel, key points are extracted from the head, torso, hands, waist, and feet; for hoisted objects, key points are extracted from the lifting points, corner points, and center of gravity areas; for vehicles, key points are extracted from the front, rear, wheels, or blind spot boundaries; and for equipment under maintenance, key points are extracted from switches, valves, ports, or maintenance surfaces. The current pose direction is determined based on the direction of key point arrangement, the main axis direction of the target area, or the center displacement direction in consecutive frames, forming the first spatial area feature of the work object.

[0044] Extraction of Second Spatial Region Features: Target boundary analysis is performed on the safety measure region. The safety measure detection box or safety measure mask output by the large visual model is read, and the corresponding region is cropped from the original video frame as a candidate safety measure region. The corresponding geometric extraction method is selected based on the type of safety measure: For linear safety measures such as warning tapes, grounding wires, and safety ropes, skeletonization or line segment fitting is performed to obtain their extension path and orientation; for planar safety measures such as guardrails, protective nets, and protective covers, boundary contour extraction and closed area estimation are performed to obtain their coverage area; for signage safety measures such as signs and warning signs, rectangular boundary fitting and text / icon area localization are performed to obtain their attachment area and orientation. Then, the connection positions of the safety measures are extracted through endpoint detection, intersection detection, or local contact area identification, such as safety rope hanging points, grounding wire clamping points, sign hanging points, or guardrail closing points. Subsequently, the attachment area is determined by combining the contact area between the safety measure and the surface of surrounding work objects or equipment. The coverage area, connection position, safety measure orientation, attachment area, and boundary morphology are summarized to form the second spatial region features of the safety measures.

[0045] The first spatial region features include at least the region bounding box, center position, outline shape, key component positions, and current pose direction.

[0046] The characteristics of the second spatial area include at least the coverage of security measures, the connection location of security measures, the orientation of security measures, the attachment area of ​​security measures, and the boundary shape of security measures.

[0047] Through the above, the visual big model can not only identify what exists in the image, but also identify what objects the security measures are applied to.

[0048] S13, Calculation of visual association strength.

[0049] To determine whether safety measures are actually effective on the current work object, the visual correlation strength between the work object area and the safety measure area in the same video frame is calculated.

[0050] For any work object area Area with safety measures Its visual association strength is defined as: ;in, Indicates the area of ​​the work object Area with safety measures The strength of visual association between them Indicates the intersection-union ratio of spaces. Indicates the similarity of distances between connection points of key components. This indicates relative pose consistency. This represents the corresponding weighting coefficient.

[0051] When safety measures emphasize actual contact relationships, increase the β corresponding to the distance between critical component connection points; When security measures emphasize regional coverage relationships, increase the spatial intersection ratio corresponding to α. When safety measures emphasize directional consistency, the corresponding γ is increased to improve relative pose consistency.

[0052] For example: For high-altitude safety rope splicing scenarios: α=0.2, β=0.5, γ=0.3 can be used, with greater emphasis on the accuracy of the splicing distance; For scenarios involving warning zones and fencing: values ​​of α=0.5, β=0.2, and γ=0.3 can be used, emphasizing the coverage of different areas. For hoisting and obstacle avoidance scenarios: α=0.3, β=0.2, γ=0.5 can be used, with greater emphasis on consistency of movement direction.

[0053] 1. The spatial intersection-union ratio is calculated as follows: ; This indicates the overlapping area between the work area and the safety measure area. This indicates the combined area of ​​the work area and the safety measure area.

[0054] 2. The similarity of the distance between connection points of key components is calculated as follows: ;in, This represents the Euclidean distance between the critical connection point of the work object and the connection point of the safety measures. The preset maximum reference distance refers to the maximum permissible spatial deviation range when determining whether safety measures are actually effective on the work object. ; This represents the average historical effective connection distance. Critical connection points on a work object refer to the areas on the work object where, theoretically, they should form a physical connection, constraint relationship, or protective effect with safety measures. They indicate the locations where safety measures should actually function. Different work objects correspond to different critical connection points. For construction workers, critical connection points may be the safety belt buckle location, waist attachment area, hand operating area, or head protection area. For hoisted objects, critical connection points may be the lifting lugs, hook connections, wire rope stress points, or the center of gravity hoisting area. For maintenance equipment, critical connection points may be the tag locking location, valve locking location, electrical grounding terminal, or protective cover installation location. For vehicles, critical connection points may be the blind spot edge area, the door entry / exit area, or the brake isolation area. Essentially, it refers to the locations where safety measures should actually be attached, connected, or covered. Key points of the target structure are extracted using a key point detection model. For personnel, key points of the shoulders, waist, and hands are extracted using a human body key point model, and the waist area is identified as the key connection point for the safety rope. For hoisted objects, the location of the lifting lugs is extracted using contour corner point detection and structural recognition algorithms. For equipment, the locations of valve handles, tag holes, or grounding points are extracted using edge detection and local structural recognition. The connection location of safety measures refers to the area where safety measures actually connect, attach, restrict, or protect the work object. This includes the contact point between the safety rope and the personnel's safety belt hook, the clamping position of the grounding wire and the equipment's grounding terminal, the position where the tag is hung on the equipment's switch handle, the position where the warning tape is fixed to the isolation column, and the position where the protective net covers the edge of the adjacent structure. Essentially, it is the location where the safety measures truly exert their restraining effect. Different extraction methods are used depending on the type of safety measure: for linear safety measures, skeleton line extraction and endpoint detection are used to obtain the end connection area; for hanging safety measures, rectangular edge detection and suspension point positioning are used; for area-based safety measures such as protective nets and guardrails, contour boundary and contact area detection are used; for grounding wires, steel wire ropes, etc., connection endpoints can be determined through color segmentation and line segment tracking. Subsequently, the actual attachment position is analyzed in conjunction with the adjacent target area to form the connection position of the safety measure.

[0055] 3. Relative pose consistency is calculated as follows: ;in, This represents the angle between the direction of movement of the work object and the direction of extension of the safety measures. Wherein: The direction of movement of the work object refers to its spatial movement trend within consecutive video frames. It is used to determine whether safety measures move synchronously with the work object or maintain effective protection. This includes personnel moving to higher ground, hoisted objects moving along the hoisting path, maintenance vehicles approaching the work area, and robotic arms extending into hazardous areas. Specifically, this can be calculated by observing changes in the target's center point within consecutive frames. Indicates the first If the target's center position is within the frame, then the motion direction vector is: ;in, Indicates the direction of movement of the work object. Indicates the center position of the current frame. This indicates the center position of the next frame. Then, the direction vector is normalized to obtain the standard motion direction.

[0056] The direction of safety measure extension refers to the actual direction of action or constraint of the safety measure in space. This includes the extension of safety ropes from the attachment point towards personnel, the extension of warning tape along the isolation boundary, the extension of grounding wires along the equipment grounding connection, and the extension of protective netting along the perimeter. Essentially, it is the spatial deployment direction of the force or protection range of the safety measure.

[0057] Let the starting point of the safety measure framework line be: ; The destination is: ; The direction of the safety measures extension is as follows: ;in, This represents the vector indicating the direction of the safety measures.

[0058] By doing the above, we can avoid the vision system from misjudging that the current operation is in a protected state simply because the safety measures "appear in the picture". Instead, we can further determine whether there is a real visual correspondence between the safety measures and the current operation object.

[0059] S14, Generation of Operation-Measure Anchoring Units.

[0060] S141. For work objects and safety measures whose visual correlation strength exceeds a preset correlation threshold within the same video frame, form candidate anchor pairs.

[0061] The criteria for determination are: ;in, This indicates a preset association threshold. ; This represents the average visual association strength of historically valid anchored samples. This represents the standard deviation of the visual association strength of historically valid anchored samples. This represents the safety margin factor; when strict prevention of missed inspections is required on-site, it should be reduced. This lowers the threshold; when strict false alarm prevention is required on-site, it increases... This increases the threshold. .

[0062] Candidate anchor pairs are generated after the conditions are met: ;in, Indicates the first A set of candidate anchor pairs in a video frame.

[0063] S142. Subsequently, cross-frame target tracking and correlation confirmation are performed on the candidate anchor pairs. Specifically, the visual big data model continuously tracks the candidate anchor pairs based on changes in target position, regional morphology, and motion trajectory continuity in consecutive video frames. In practice: the visual big data model uses the candidate anchor pairs in the current video frame as the initial tracking objects and records the position of the work object, the position of the safety measure, the spatial distance between them, the regional overlap, and the relative direction relationship within each candidate anchor pair; it continues to search for the target region closest to the candidate anchor pair in subsequent video frames and determines whether the work object and the safety measure still maintain their original correspondence. Wherein: When the target location changes, the visual big data model compares the movement trends of the center position of the work object and the center position of the safety measures in the preceding and following video frames. For example, when the construction worker moves forward, if the safety rope attachment point, the boundary of the warning tape, or the protected area also moves synchronously or maintains the corresponding position, it is considered that the two are still related; if the worker has moved to a new work area, but the safety measures remain in the original area, it is considered that the candidate anchoring pair may have failed.

[0064] For changes in region morphology, the large visual model continuously compares changes in region outline, area ratio, edge shape, and target orientation in preceding and following video frames. For example, when a safety rope is actually attached, its connection area will typically experience continuous stretching as the person moves; if the direction of the safety rope suddenly changes, the area breaks, or it separates from the person's area in subsequent video frames, it indicates that the original anchoring relationship is no longer stable.

[0065] Regarding the continuity of motion trajectories, the large visual model records the motion trajectories of the work object and safety measures in consecutive video frames and determines whether their trajectories maintain a consistent temporal progression. For example, in a real protected state, the movement trajectory of personnel and the extension direction of the safety rope usually have a continuous correspondence, while the temporary safety measures often cannot form a continuous trajectory synchronously with the work object.

[0066] S143. After completing continuous tracking, the visual big model begins to determine whether the candidate anchor pairs meet the spatial continuity condition and the temporal stability condition.

[0067] Spatial continuity condition: The spatial continuity condition is mainly used to determine whether safety measures continuously act on the same work object in space. Specifically, the large visual model continuously detects changes in target spacing, region overlap, and directional consistency in consecutive video frames. If the spatial distance between the work object and the safety measures is consistently less than the allowable offset range, the region overlap relationship does not show a significant break, and the extension direction of the safety measures continuously corresponds to the movement direction of the work object, then the spatial continuity condition is considered satisfied. Conversely, if there is significant target separation, abrupt change in connection direction, detachment of the covered area, or the safety measures being transferred to other objects, then spatial continuity is considered interrupted.

[0068] Temporal stability condition: The large visual model counts the number of frames in which candidate anchor pairs maintain effective association within a continuous tracking window, and calculates the association stability. ;in, Indicates the stability of the association. This indicates the number of video frames in which candidate anchor pairs maintain a correlation. This represents the total number of video frames in the continuous tracking window. When the candidate anchor pair maintains effective association in most continuous video frames without long-term breaks, frequent switching of associated objects, or drastic fluctuations, it is considered to meet the time stability condition.

[0069] When a candidate anchoring pair simultaneously satisfies both spatial continuity and temporal stability conditions, it is determined as an operation-measure anchoring unit. Spatial continuity ensures that safety measures always maintain a stable spatial correspondence with the same work object; temporal stability prevents incorrect anchoring due to brief personnel obstruction, camera shake, sudden changes in lighting, or momentary false detections.

[0070] S2. Based on the operation-measure anchoring unit, determine whether the visible safety measures are actually applied to the current operation object. When the visible safety measures only correspond to adjacent objects, historical operation objects, or non-current operation positions, generate a borrowing compliance risk unit.

[0071] In this embodiment, the specific components include: S21, Analysis of Operation-Measure Anchoring Unit.

[0072] Obtain the task-measure anchoring unit corresponding to the current moment, and parse the correspondence between the current task object and the visible safety measures from the task-measure anchoring unit.

[0073] S211, For the current video frame The corresponding task-measure anchoring unit is denoted as: ;in, This represents the current task-measure anchoring unit. Indicates the current task object. This indicates the currently visible security measures.

[0074] S212. Extract the first instance identifier from the current job object. The first instance identifier is a unique identifier assigned to the current work object, used to continuously distinguish different work objects in consecutive video frames and maintain the continuity of the same work object's identity throughout the temporal process. Even if the work object moves, the camera's viewing angle changes, the work object is briefly occluded, or the work object's posture changes, it can still be identified that the current target is still the same work object. For example: the same construction worker always corresponds to the same first instance identifier in different video frames; the same hoisted object maintains the same first instance identifier during hoisting; the same maintenance equipment still maintains the same first instance identifier under different angles. Essentially, it is a unique identification number for the work object in the temporal video.

[0075] S213. Extract the second instance identifier from the currently visible security measures. The second instance identifier is a unique identifier assigned to a visible safety measure. It is used to continuously distinguish different safety measures across consecutive video frames and maintain the continuity of the same safety measure's identity throughout the timeline. In other words, it not only needs to know that a safety measure exists in the frame, but also that the current safety measure is the same as the previous one. For example: the same safety rope maintains the same second instance identifier across consecutive video frames; the same warning tape maintains the same second instance identifier during hoisting; the same tag continuously maintains the same second instance identifier during maintenance.

[0076] Essentially, it is a unique identifier for security measures within a time-series video.

[0077] The first instance identifier is used to uniquely identify the continuity of the identity of the current work object in consecutive video frames; the second instance identifier is used to uniquely identify the continuous existence of the current security measures in consecutive video frames.

[0078] The above methods can establish a temporal correlation between safety measures and work objects, providing target identity constraints for subsequent borrowing relationship identification.

[0079] S22. Historical anchored association record retrieval.

[0080] Based on the first instance identifier and the second instance identifier, a historical association retrieval is performed in the historical anchoring association records formed by cross-frame target tracking to determine the scope of objects in which the currently visible security measures are actually continuously effective in historical video frames.

[0081] S221. Construct historical anchored association records: ;in, This represents the set of historical anchored records. Indicates an instance identifier for a security measure. Indicates the identifier of the historical associated object instance. Indicates the time of historical video frames. This indicates the correlation strength in the corresponding historical video frame.

[0082] Specifically, the construction of historical anchoring association records begins with the operation-measure anchoring units in consecutive video frames. After completing the visual association analysis between the operation object and the safety measure in each video frame, the operation objects and safety measures that meet the association conditions are combined to form the operation-measure anchoring unit of the current frame. Simultaneously, the first instance identifier corresponding to the operation object, the second instance identifier corresponding to the safety measure, the current video frame number, and the current association strength are recorded. The operation-measure anchoring units in the current frame are then input into the cross-frame association management module, which determines whether the current anchoring relationship belongs to the same continuous association relationship as the existing anchoring relationships in historical video frames. During the determination, the consistency of instance identifiers, spatial location continuity, motion trajectory continuity, and safety measure instance continuity between the current operation object and historical objects are compared. When the above relationships meet the continuous tracking conditions, the current anchoring relationship is considered a temporal continuation of the historical anchoring relationship, and the association result in the current video frame is appended to the corresponding historical record.

[0083] Using the second instance identifier as the primary association index, an independent historical anchoring association chain is established for each safety measure. This means that the same safety measure is continuously recorded chronologically, detailing which work objects it has acted upon at different times, for how long it has lasted, and how the association strength has changed. For example, if the same safety rope is continuously associated with worker A in early video frames, but gradually moves closer to worker B in later video frames, the historical association interval between the safety rope and A, and the new association interval between it and B, are recorded respectively. Similarly, if the same warning tape surrounds a hoisting area for a long time, its historical protected area is continuously recorded. If the hoisting area has moved subsequently but the warning tape has not moved synchronously, the historical anchoring association record will still retain the continuous association relationship between the warning tape and the old area. Ultimately, the historical associated objects, association duration, association strength changes, and association area changes formed by all safety measures at different time periods constitute the historical anchoring association record. This record is used to determine whether the current safety measure is still within the historical protection relationship, thereby identifying compliance risks associated with borrowing.

[0084] S222, Identify examples of current security measures Using the primary key as the retrieval point, a time-backtracking retrieval is performed on the historical anchored association records to obtain the historical anchored association chain: ;in, This represents a set of object identifiers that have a historical connection with current security measures. Indicates the identifier of historically associated objects.

[0085] S223. Classify historically associated objects based on the duration and stability of historical associations: When an object maintains a high correlation (0.8) with the current safety measures for a long period of time in consecutive historical video frames, it is identified as a historical operation object, and the correlation lasts for at least 1.2 seconds. An object is identified as an adjacent object when it has a locally stable relationship with the current security measures only within the spatial neighborhood.

[0086] For example: A safety rope had been in use for a long time for worker A, but in the current footage, worker A has left and worker B has entered the same area; A warning tape had previously surrounded the hoisting area, but the object being hoisted has now been moved to a new area; Previously, a certain equipment was designated for maintenance, but the current work has been switched to equipment B.

[0087] By retrieving historical anchoring records, it can be identified whether safety measures are still based on historical protection relationships and have not been truly migrated to the current work objects.

[0088] S23, Current association strength calculation.

[0089] Calculate the first current association strength between the current safety measure and the current work object, and the second current association strength between the current safety measure and other potential associated targets.

[0090] S231. Calculate the first current association strength between currently visible safety measures and the current work object: ;in, Indicates the current association strength. This indicates the intersection-union ratio of the current region. This indicates the similarity of the current key connection points. This indicates that the current pose is consistent.

[0091] S232. Calculate the second current association strength between the current safety measures and adjacent object areas, historical operation object areas, and non-current operation location areas respectively: ;in, Indicates the second current association strength. This represents the area intersection ratio between the safety measures and the target area. This indicates the similarity in distance between key connection points between security measures and the target area. This indicates the pose consistency between safety measures and the target area.

[0092] in: When the target area is an adjacent object area, it is used to determine whether the safety measures actually work on the neighboring objects; When the target area is a historical work area, it is used to determine whether the safety measures still maintain the historical protection relationship; When the target area is not the current work location area, it is used to determine whether the safety measures remain in the old work area.

[0093] The above approach goes beyond simply determining whether safety measures exist; it further compares their actual effectiveness on different objects and locations.

[0094] S24. Determination of borrowing relationship.

[0095] Based on the comparison between the first current association strength and the second current association strength, it is determined whether the current safety measures are actually effective on the current work object.

[0096] S241. Let the preset effective threshold be... ; This represents the minimum correlation strength threshold for determining the actual effectiveness of safety measures. ; This represents the average current association strength of the first valid anchored samples. This represents the first current association strength standard deviation of the valid anchored samples; When the following conditions are met: and If this occurs, it is determined that the current safety measures are not actually effective on the current work object. This indicates that at least one target region satisfies the condition.

[0097] In other words, the correlation between the current safety measures and the current work object is insufficient, but the correlation between the current safety measures and other objects or other areas is stronger. In this case, further determination of the borrow type is needed: 1. Adjacent Object Borrowing: When the second current correlation strength between a safety measure and an adjacent object is the highest, it is determined to be adjacent object borrowing. For example, a safety rope is actually attached to a person next to it or a warning tape is actually wrapped around nearby equipment.

[0098] 2. Historical Object Borrowing: When a safety measure maintains the highest correlation with a historical work object area, it is determined to be historical object borrowing. For example, the tag still corresponds to the historical maintenance equipment or the grounding wire is still connected to the original equipment.

[0099] 3. Location Borrowing: When safety measures remain in the historical area while the current work has been moved, it is determined to be location borrowing from a non-current work location. For example, warning tape has not moved with the hoisting area or the protective isolation area remains in the old construction location.

[0100] The above methods can be used to identify compliance risks associated with borrowing where visual safety measures exist but the actual work object is not protected.

[0101] S25, Generation of compliance risk units.

[0102] After determining the borrowing relationship, a corresponding borrowing compliance risk unit is generated. Specifically, the borrowing compliance risk unit is defined as follows: ;in, This indicates a borrowed compliance risk unit. Indicates a mismatch safety measure label. Indicates the actual protected object identifier. Indicates an object that is currently not protected. This indicates a borrow type, which must include at least the following: Borrowing from adjacent objects; Borrowing from historical objects; Borrow from a position other than the current job position.

[0103] After generating the borrowing compliance risk unit, it is output to the subsequent security warning generation step to generate the corresponding security warning result and corrective action instructions.

[0104] The core of this part is: after discovering that safety measures exist in the footage but are not actually protecting the current work object, this anomaly is compiled into a standardized risk record, also known as a misalignment compliance risk unit. This unit records: the mismatched safety measures, the actual protected object, the object currently lacking protection, and the specific type of misalignment. For example, if a safety rope is found to be still attached to a nearby worker, but the worker currently at height is not actually attached, a misalignment compliance risk unit is generated, recording that "the actual protected object corresponding to the current safety rope is the worker next to them, while the current worker is an unprotected object," and marked as "adjacent object misalignment." Another example is if the warning tape is still around the old hoisting area, but the hoisting operation has moved to a new area, generating a "misalignment outside the current work location" risk unit. The system then uses this risk unit to generate specific safety warnings and corrective instructions.

[0105] S3. Generate a security warning result based on the borrowing compliance risk unit, and output the mismatch security measures, actual protected objects, currently unprotected objects, and correction instructions corresponding to the borrowing compliance risk unit.

[0106] In this embodiment, the specific components include: S31, Analysis of compliance risk unit for borrowing.

[0107] First, obtain the borrowing compliance risk unit generated in step S2, and then parse the core risk object information corresponding to the current borrowing compliance risk from the borrowing compliance risk unit. The borrowing compliance risk unit is: ;in: Mismatch safety measure markings are used to identify safety measures in case of borrow mismatch; The actual protected object identifier is used to identify the object on which the current safety measures are actually effective; The currently unprotected object identifier is used to identify the work object that is currently lacking protection; The borrowing type is used to identify whether the current borrowing compliance risk belongs to borrowing from adjacent objects, borrowing from historical objects, or borrowing from non-current operating positions.

[0108] The mismatch safety measure identifier, the actual protected object identifier, the currently unprotected object identifier, and the borrowing type are then input into the risk level assessment for subsequent safety risk level determination.

[0109] The above methods not only output the existence of risks, but also clearly indicate: "which security measures are mismatched", "who is actually being protected", and "who is not actually being protected".

[0110] S32. Determination of safety risk level.

[0111] Based on the pre-defined mapping relationship between borrowing types and security risk levels, the security risk level corresponding to the current borrowing compliance risk is determined. Specifically, a mapping table between borrowing types and risk levels is established: ;in, This represents the set of mapping relationships between borrowing types and risk levels. Indicates borrow type, This indicates the corresponding safety risk level. The safety risk levels include at least Level 1, Level 2, and Level 3. For example, misalignment of safety ropes during high-altitude operations corresponds to Level 1 risk, misalignment of hoisting warning tape corresponds to Level 2 risk, and misalignment of temporary fencing corresponds to Level 3 risk.

[0112] Based on the current borrowing type Retrieve the set of mapping relationships The current security risk level is obtained. .

[0113] When constructing the mapping table between borrowing types and risk levels, the first step is to classify and statistically analyze historical safety accident cases, on-site hazard records, and confirmed borrowing compliance risk samples. In practice, a hazard analysis is first conducted on different borrowing types based on the degree of impact of the borrowing behavior on safety consequences. For example, if the safety rope is not actually applied to the current construction worker during high-altitude operations, a fall would directly result in personal injury or death; therefore, this type of borrowing is classified as high-risk. While there is a risk of accidental entry when temporary fencing does not fully move with the construction area, the speed of hazard transmission is slower, thus the hazard level is relatively low. Subsequently, the risk level is determined by comprehensively considering the following factors: first, the degree of personal exposure of the currently unprotected person; second, whether the hazard source is operational; third, the severity of the consequences after an accident; and fourth, whether there is sufficient buffer time for manual intervention on site. After completing the analysis, a fixed correspondence is established between each borrowing type and its corresponding risk level, and this is recorded in the borrowing type and risk level mapping table.

[0114] After the mapping table is constructed, during actual operation, when S2 generates a borrowing compliance risk unit, it first parses the borrowing type within it, and then uses that borrowing type as the search keyword to query the corresponding risk level in the borrowing type and risk level mapping table. For example, when the system identifies the current situation as "borrowing from an adjacent object of a safety rope at height," it will be directly mapped to a Level 1 risk; when it identifies it as "borrowing from a warning tape location in a hoisting area," it will be mapped to a Level 2 risk; and if it identifies it as "borrowing from a historical location of a temporary enclosure," it will be mapped to a Level 3 risk. Subsequently, the system will dynamically correct the initial risk level obtained from the mapping by combining the current location of the unprotected object, the operating status of the on-site equipment, and the current work scenario. For example, even with the same warning tape borrowing, if the current unprotected object is directly below the hoisted object, the risk level can be automatically upgraded from Level 2 to Level 1.

[0115] Table 1. Examples of the mapping between borrowing types and risk levels. Safety rope at height, adjacent object borrowing Working at height Level 1 risk Historical object borrowing of high-altitude safety rope Working at height Level 1 risk Position of hoisting warning tape hoisting operation Level 2 risk Grounding wire historical object borrowing Electrical maintenance Level 1 risk Historical objects of the listing and locking system are borrowed. Inspection and maintenance sign Level 1 risk Temporary fence location borrowing General construction area Level 3 risk Partial displacement of protective netting Construction near the edge Level 2 risk Borrowing of objects under confined space monitoring Confined space operations Level 1 risk S33. Generate a safety warning result based on the current safety risk level, mismatched safety measures identifier, actual protected object identifier, and currently unprotected object identifier.

[0116] Specifically, the safety warning result is defined as: ;in, Indicates the results of the safety warning. Indicates the current security risk level. Indicates a mismatch safety measure label. Indicates the actual protected object identifier. This indicates an object that is currently not protected.

[0117] Visual annotations are performed on the aforementioned risk objects in the continuous operation screen.

[0118] Specifically, it includes: Highlight and mark areas that are currently unprotected; The areas with mismatched safety measures are marked with a flashing box; Draw borrowing correlation indicator lines between the actual protected object and the currently unprotected object; The image displays the borrowing type and risk level information on the side of the screen.

[0119] For example: Use a red border to mark objects that are currently unprotected; Use yellow borders to indicate mismatch safety measures; Use arrows to indicate the actual direction of the safety measures.

[0120] At the same time, the on-site alarm devices will output alarms based on the current safety risk level. Specifically, Level 1 risk triggers an audible and visual alarm and equipment shutdown; Level 2 risk triggers an on-site voice alarm; and Level 3 risk triggers an alarm notification on the management terminal.

[0121] The above methods enable on-site personnel to quickly identify: "which safety measure is mismatched" and "who is currently in real danger".

[0122] S34, Corrective action instruction generated.

[0123] Based on the borrowing type, the identifier of the currently unprotected object, and the current work scenario information, rule matching is performed in the preset safety correction rule base to generate corresponding correction and handling instructions.

[0124] Build a security correction rule base: ;in, This represents the security correction rule base. Indicates borrow type, Indicates the work scenario information. This indicates the corresponding corrective action instruction.

[0125] The work scenario information includes at least high-altitude work scenarios, hoisting work scenarios, hot work scenarios, maintenance and labeling scenarios, and confined space work scenarios.

[0126] Based on the current borrowing type Current work scenario information and currently unprotected object identifiers Retrieve the security correction rule base and generate corresponding correction action instructions. Corrective action instructions should include at least the following: 1. Measures Reset Command: Used to control the redeployment of mismatched safety measures to currently unprotected objects. Examples include prompting the reattaching of safety ropes, moving warning tapes, reattaching tags and locking mechanisms, and redeploying isolation zones.

[0127] 2. Emergency Stop Command: Used to control the suspension of work on currently unprotected objects under high-risk conditions. Examples include stopping hoisting equipment, stopping the lifting of elevated work platforms, stopping power supply to maintenance equipment, and stopping robotic arm movements. Through this, control is shifted from risk identification to proactive corrective action.

[0128] When constructing the safety correction rule base, historical accident cases, hazard rectification records, and on-site standard safety operating procedures for different safety scenarios are categorized and organized. Specifically, firstly, the system is classified according to the work scenario, such as high-altitude work scenarios, hoisting operation scenarios, electrical maintenance scenarios, confined space work scenarios, and hot work scenarios. Then, within each type of work scenario, it is further classified according to the type of borrowing, such as borrowing from adjacent objects, borrowing from historical objects, and borrowing from non-current work positions. Next, for each combination of work scenario and borrowing type, the most effective correction method for the corresponding risk is analyzed, and standardized correction actions are established. For example, in high-altitude work scenarios, when a safety rope is borrowed from an adjacent object, the system records re-attaching the safety rope and suspending the lifting platform as standard correction measures. In hoisting operation scenarios, when a warning tape is borrowed, the system records redeploying the warning area and suspending the hoisting equipment as standard correction measures. After the data is compiled, the borrowing type, work scenario, risk level, applicable equipment, and corresponding corrective actions are written into the safety correction rule base to form a searchable set of rules.

[0129] When performing a safety correction rule base search, the system reads the current borrowing type from the borrowing compliance risk unit, and then reads the work scenario information corresponding to the current unprotected object identifier, such as whether the current object belongs to a worker at height, a construction worker in a hoisting area, or an electrical maintenance worker. Subsequently, using the borrowing type combined with the work scenario as joint search conditions, the system searches for the corresponding rule item in the safety correction rule base. When multiple matching rules exist, the rule with the highest risk level and the best match to the current equipment status is selected first. If the current unprotected object has entered a high-risk area, the system automatically adds an emergency stop instruction. After rule matching is completed, a corresponding correction instruction is generated and sent to the safety measures control terminal or the work equipment control system for execution. For example, when the system detects a historical object borrowing of the safety rope at height, it generates a correction instruction to immediately reattach the safety rope and suspend the high-altitude lifting platform; when it detects a historical object borrowing of the tag-locked system, it generates a correction instruction to prohibit power supply and re-tag confirmation.

[0130] Table 2 Example Table of Safety Correction Rule Base Safety rope adjacent object borrowing Working at height Level 1 risk Reattach the safety rope and suspend the lifting platform. Safety rope historical object borrowing Working at height Level 1 risk Remove old connections and redeploy connection points Warning tape position borrowing hoisting operation Level 2 risk Redeploy the warning zone and suspend hoisting operations. Grounding wire historical object borrowing Electrical maintenance Level 1 risk Do not energize or reconnect the grounding wire. Historical objects of the listing and locking system are borrowed. Inspection and maintenance sign Level 1 risk Re-label and confirm equipment Partial displacement of protective netting Construction near the edge Level 2 risk Adjust the coverage area of ​​the protective net Guardianship object borrowing Confined space operations Level 1 risk Forced evacuation and restoration of guardianship binding Fence location borrowing General construction area Level 3 risk Adjust the position of the construction site fence and provide on-site warnings and alarms. S35. Implementation of corrective measures.

[0131] The generated corrective action instructions are sent to the corresponding safety measure control terminal or the operating equipment control system via the communication interface to execute the corresponding corrective action.

[0132] Specifically, establish a communication control link: ;in. Indicates the corrective execution chain. This indicates a corrective action instruction. This refers to the target control terminal, which includes at least: Seatbelt intelligent locking terminal; Audible and visual alarm control terminal; Lifting control system; High-altitude work platform controller; Equipment shutdown control system; Industrial edge control terminal.

[0133] Subsequently, an execution command is sent to the corresponding control terminal based on the type of corrective action instruction.

[0134] For example: When the safety rope at height is detected to be misaligned, a stop lifting command is sent to the high-altitude work platform. When the hoisting warning tape is detected to be misaligned, a deceleration or suspension command is sent to the hoisting control system. When a borrowing operation is detected, a power-off command is sent to the equipment control system.

[0135] Through the above methods, automatic correction of compliance risks and proactive on-site safety intervention can be achieved.

[0136] This invention encompasses any substitutions, modifications, equivalent methods, and solutions made within the spirit and scope of this invention. To provide the public with a thorough understanding of this invention, specific details are described in detail in the following preferred embodiments; however, those skilled in the art will fully understand the invention even without these details. Furthermore, to avoid unnecessary misunderstanding of the essence of this invention, well-known methods, processes, procedures, components, and circuits are not described in detail.

[0137] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A security early warning method based on a large visual model, characterized in that, Includes the following steps: S1. Input the continuous operation screen of the safety production site into the visual big model. The visual big model identifies the visual correspondence between the current operation object and the visible safety measures in the screen and generates operation-measure anchoring units. S2. Based on the operation-measure anchoring unit, determine whether the visible safety measures are actually applied to the current operation object. When the visible safety measures only correspond to adjacent objects, historical operation objects, or non-current operation locations, generate a borrowing compliance risk unit. S3. Generate a security warning result based on the borrowing compliance risk unit, and output the mismatch security measures, actual protected objects, currently unprotected objects, and corrective action instructions corresponding to the borrowing compliance risk unit; S2 includes obtaining the current operation-measure anchoring unit and parsing out the first instance identifier of the current operation object and the second instance identifier of the visible safety measure from it; Based on the first instance identifier and the second instance identifier, a search is performed in the historical anchoring association records formed by cross-frame target tracking to determine the historical operation object identifier and adjacent object identifier in which the visible safety measures have been continuously in effect in the historical frame. For the current frame, calculate the first current association strength between the second spatial region feature of the visible safety measures and the first spatial region feature of the current work object, and calculate the second current association strength between the second spatial region feature and the adjacent object region, the historical work object region, and the non-current work location region; If the first current association strength is lower than the preset effective action threshold, and at least one second current association strength is higher than the first current association strength, it is determined that the visible safety measures are not actually applied to the current work object, but are applied to adjacent objects, historical work objects, or non-current work locations. The borrowing compliance risk unit is generated based on the judgment result. The borrowing compliance risk unit includes at least the mismatch security measure identifier corresponding to the visible security measure, the actual protected object identifier, the currently unprotected object identifier, and the borrowing type.

2. The security early warning method based on a large visual model according to claim 1, characterized in that, S1 further includes extracting frames and timestamping them on the continuous operation screen to form a video frame sequence with time sequence identifiers, and inputting the video frame sequence into the visual big model.

3. The security early warning method based on a large visual model according to claim 2, characterized in that, The visual big model performs parallel segmentation of the work object instance and detection of the safety measure target for each frame, and extracts the first spatial region features of the current work object and the second spatial region features of the visible safety measures. For the same frame, the visual association strength between the first spatial region features and the second spatial region features is calculated. The visual association strength includes at least one of the following: spatial intersection-over-union ratio, distance between key component connection points, and relative pose consistency.

4. The security early warning method based on a large visual model according to claim 3, characterized in that, It also includes a preset association threshold, which forms candidate anchoring pairs with the visual association strength of the work objects that exceed the preset association threshold in the same frame and the visible safety measures. In multiple consecutive frames, the candidate anchoring pairs are subjected to cross-frame target tracking and association relationship confirmation. The candidate anchoring pairs that simultaneously satisfy spatial continuity and temporal stability are determined as the work-measure anchoring units.

5. A security early warning method based on a large visual model according to claim 1, characterized in that, Extract the mismatch security measure identifier, the actual protected object identifier, the currently unprotected object identifier, and the borrowing type from the borrowing compliance risk unit; Based on the borrowing type, the current security risk level is determined according to the preset mapping relationship between borrowing type and security risk level; The safety warning result is generated, which includes the current safety risk level, the mismatched safety measure identifier, the actual protected object identifier, and the currently unprotected object identifier. The result is then displayed visually on the continuous operation screen and / or alarmed via an on-site alarm device.

6. A security early warning method based on a large visual model according to claim 5, characterized in that, Based on the borrowing type, the identifier of the currently unprotected object, and the current work scenario information, rule matching is performed in the preset safety correction rule base to generate corresponding correction instructions.

7. A security early warning method based on a large visual model according to claim 6, characterized in that, The corrective action instructions include: a measure reset instruction to adjust the mismatched safety measures from the actual protected object to the currently unprotected object, and / or an emergency shutdown instruction to control the currently unprotected object to suspend operations; the corrective action instructions are sent to the corresponding safety measure control terminal or work equipment control system through a communication interface to execute the corrective action.

Images