Computer vision-based construction safety monitoring system
The computer vision-based construction safety monitoring system solves the problems of image quality degradation and unstable recognition under complex construction site conditions, improves the accuracy and real-time performance of construction safety monitoring, and enhances the availability and alarm accuracy of the monitoring system by adopting a dynamic adjustment strategy.
Patent Information
- Application Number
- CN202611249370.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-18
- Publication Date
- 2026-11-17
AI Technical Summary
When faced with complex conditions such as dust, strong light, obstruction, and vibration, the safety monitoring system at the construction site suffers from decreased image quality, unstable recognition results, false alarms or missed alarms, and the monitoring strategy cannot be updated in a timely manner, resulting in a reduction in the accuracy and real-time performance of construction safety monitoring.
The computer vision-based construction safety monitoring system enables dynamic adjustment of risk management and monitoring strategies for construction areas through modules such as construction area division, video operation status detection, safety monitoring configuration table generation, on-site risk assessment, multi-camera verification, and result feedback. These modules include adaptive adjustment of camera priority, resolution, detection category, algorithm level, and frame rate.
It improves the accuracy and real-time performance of construction safety monitoring, manages risk objects by unifying spatial benchmarks, adaptively adjusts monitoring strategies, enhances continuous tracking capabilities across cameras, reduces false alarms and missed alarms, and forms a closed-loop linkage mechanism.
Smart Images

Figure CN123123541A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of computer vision and construction safety monitoring, specifically to a construction safety monitoring system based on computer vision. Background Technology
[0002] Currently, safety monitoring at construction sites typically relies on fixed cameras combined with manual inspections. When faced with complex conditions such as dust, strong light, obstructions, vibrations, and fluctuations in the computing power of edge nodes, problems such as decreased image quality, unstable recognition results, and false alarms or missed alarms are prone to occur. Because existing monitoring methods lack a linkage adjustment mechanism for site area risks, video operation status, and multi-camera verification results, monitoring strategies often cannot be updated in a timely manner with changes on site, reducing the accuracy and real-time performance of construction safety monitoring. Summary of the Invention
[0003] The purpose of this invention is to provide a computer vision-based construction safety monitoring system that avoids problems such as image quality degradation, unstable recognition results, and false alarms or missed judgments when facing complex working conditions such as dust, strong light, obstruction, vibration, and fluctuations in computing power at edge nodes. Furthermore, it is easier to update the monitoring strategy in a timely manner through a linkage adjustment mechanism based on on-site area risks, video operation status, and multi-camera verification results, thereby improving the accuracy and real-time performance of construction safety monitoring.
[0004] The objective of this invention can be achieved through the following technical solutions:
[0005] A computer vision-based construction safety monitoring system includes:
[0006] The construction area division module is used to divide the construction area according to the calibration and construction information, and record the area boundary, basic risk weight, associated cameras and risk rules;
[0007] The video operation status detection module communicates with each camera and edge node to obtain video and node resource information, and calculates the amount of lighting degradation, dust interference coefficient, motion blur coefficient, packet loss rate, effective frame rate, occlusion ratio and trajectory continuity.
[0008] The safety monitoring configuration table generation module communicates with the construction area division module and the video operation status detection module to generate a safety monitoring configuration table with a version number. The table includes camera priority, resolution, detection category, algorithm level, frame rate, confidence threshold, risk conversion parameters, review rules, review cameras, camera intrinsic and extrinsic parameter matrix containing camera focal length, optical center coordinates and rotation and translation vectors, risk retention duration and maximum processing latency.
[0009] The on-site risk assessment module is used to perform image recognition according to the effective configuration table, detect and track construction targets and dynamic equipment, calculate the spatial relationship between the target and the dangerous area, edge openings and dynamic equipment, and output the target hazard level.
[0010] The multi-camera verification module is used to perform frame-time alignment, coordinate conversion, and association with the target trajectories of different cameras, and accordingly confirm, maintain, reduce, or eliminate the target's danger level.
[0011] The result feedback module is connected to the safety monitoring configuration table generation module, the on-site risk assessment module, and the multi-camera verification module. It is used to output risk results and handling information, and to revise the next version of the configuration table based on the verification and handling results and subsequent trajectories.
[0012] Preferably, the video running status detection module is used for:
[0013] Extract the current image frame and video transmission information from the video of the associated camera, extract the reference clear frame and reference lighting data from the video data of the historical normal period, and obtain the target occlusion information and trajectory record from the on-site risk assessment module;
[0014] The system detects the amount of illumination degradation, dust interference coefficient, ghosting coefficient, packet loss rate, effective frame rate, occlusion ratio, and trajectory continuity, and obtains the CPU utilization, memory utilization, and bandwidth utilization of the edge nodes to determine the operating conditions.
[0015] When the dust interference coefficient exceeds the limit, the lighting degradation exceeds the limit, the occlusion ratio exceeds the limit, the effective frame rate is zero, or any of the central processing unit utilization rate, memory utilization rate, or bandwidth utilization rate exceeds the limit, and each exceeds the limit for a preset confirmation period, the corresponding conditions of dust enhancement, lighting abnormality, severe occlusion, video interruption, or computing power shortage are determined.
[0016] As a preferred option, after determining that an operating condition has occurred, the safety monitoring configuration table generation module performs the following steps:
[0017] Identify the affected cameras, construction areas, operational status, and duration;
[0018] When the dust interference coefficient exceeds the limit, the resource adjustment strategy corresponding to the dust interference is to reduce the resolution, reduce the processing frame rate, or switch to a lightweight target detection network. When the CPU utilization, memory utilization, or bandwidth utilization exceeds the limit, the resolution reduction strategy is preferred.
[0019] Based on the affected construction area, retrieve the corresponding basic risk weights, risk rules, associated cameras, and the safety monitoring configuration table of the previous valid version;
[0020] Processing resources are allocated in order of priority from high to low based on basic risk weights. The maximum processing latency is limited according to a preset safety latency threshold. The system generates camera priority, input resolution, processing frame rate, detection category, recognition algorithm level, recognition confidence threshold, verification camera, risk retention duration, and maximum processing latency.
[0021] The generated safety monitoring configuration table is sent to the corresponding edge node, and the corresponding processing mode is exited after the working condition is lifted and the preset lifting time period is continuously met.
[0022] As a preferred option, the process of building and running the security monitoring configuration table includes:
[0023] Set the version number, validity period, and rollback version for each security monitoring configuration table, and enable the edge nodes to verify the integrity of the data packets and return version confirmation information after receiving them;
[0024] The edge nodes that return version confirmation information load the new version and execute the on-site risk assessment module and the multi-camera verification module, while the edge nodes that do not return version confirmation information continue to execute the previous valid version.
[0025] Receive processing time, resource utilization, task failure information, and risk identification results returned by edge nodes;
[0026] The system analyzes operational metrics within a preset time window. If the new version causes the false negative rate in the corresponding region to be higher than the baseline false negative rate of the previous effective version, the maximum processing latency to exceed the preset latency limit, the CPU utilization rate, memory utilization rate, or bandwidth utilization rate of the edge nodes to exceed the corresponding preset load limit, or the verification consistency rate to be lower than the preset consistency rate lower limit, the new version will be stopped and the system will revert to the previous effective version.
[0027] As a preferred option, the on-site risk assessment module obtains the target recognition confidence level from the image recognition results and obtains the regional basic risk weight. Peripheral risk level Equipment risk level Occlusion ratio and target determination error value ;
[0028] in, This indicates the basic risk weight of the construction area where the target is located;
[0029] The non-negative normalized value is obtained by dividing the difference between the preset edge distance threshold and the distance from the target to the edge opening by the edge distance threshold. When the distance from the target to the edge opening is greater than the edge distance threshold, Values ;
[0030] The non-negative normalized value is obtained by dividing the difference between the preset device distance threshold and the distance from the target to the dynamic device by the device distance threshold. When the distance from the target to the dynamic device is greater than the device distance threshold, Values ;
[0031] Indicates the proportion of the target that is obscured; ;
[0032] According to the formula Calculate and determine the risk level score ,in This is a preset occlusion amplification factor used to quantify occlusion interference weights. Let be the preset error amplification factor used for quantization error interference weights, and satisfy . .
[0033] As a preferred option, a first risk limit is preset. Second risk boundary ,and ;
[0034] When the risk level is rated When the target identification confidence level is not lower than the current identification confidence level threshold and the first preset frame number threshold is met continuously, the target danger level is set to or restored to the safe level.
[0035] When the risk level score meets Furthermore, when the second preset frame count threshold is met consecutively, the target danger level is set to the proximity level.
[0036] When the risk level score meets Or, when the target enters a dangerous area and the spatial relationship is reliable, the target's danger level is set to danger level;
[0037] Targets at the danger level only meet the danger level score within the third consecutive preset frame threshold. And when there is no continuous obstruction and the verification results confirm safety, it will be restored to the safe level;
[0038] When a target is missed for a short period of time or the occlusion ratio exceeds the preset occlusion limit, and the number of continuous tracking frames before occlusion reaches the fourth preset frame threshold, the target danger level before occlusion is maintained, and the associated camera is requested to be reviewed after the occlusion time exceeds the risk maintenance duration.
[0039] As a preferred option, the multi-camera verification module acquires the time difference of the acquisition of the target to be verified in different cameras, the coordinate conversion error, the direction of movement and the degree of similarity in appearance;
[0040] Targets in different cameras are identified as the same target only when the acquisition time difference is less than or equal to the preset time difference, the coordinate conversion error is less than or equal to the preset distance difference, the angle of movement direction is less than or equal to the preset angle threshold, and the appearance similarity is greater than or equal to the preset matching threshold.
[0041] The danger level is maintained as long as the associated camera confirms that the target is still in the danger zone; the danger level is raised to a higher level if the risk is confirmed to be escalating.
[0042] When the associated camera confirms that the target has left the danger zone, it is upgraded to a safe level according to the continuous safety assessment criteria.
[0043] When the appearance similarity is lower than the preset matching threshold, the original high-risk status is not lifted, and a review is initiated again within the risk retention period. If a reliable association cannot be achieved after the risk retention period, a manual review request is generated.
[0044] As a preferred embodiment, the result feedback module is connected to the alarm terminal and the manual review server to output the target number, trajectory number, area number, risk type, risk level, start time, duration, evidence frame, review source, security monitoring configuration table version and suggested handling information. High-risk results are sent to the alarm terminal, and results with insufficient identification confidence and unable to complete multi-camera review are sent to the queue of the manual review server.
[0045] The results feedback module also updates the security monitoring configuration table parameters based on the feedback results, and takes effect after verification; if verification fails, it reverts to the previous valid version.
[0046] The beneficial effects of this invention are:
[0047] 1. This invention incorporates edge openings, hoisting paths, and vehicle access routes into a unified spatial benchmark through a construction area division module, and records area boundaries, basic risk weights, associated cameras, and risk rules to achieve zoning management and priority allocation of risk objects at the construction site; through a video operation status detection module, it jointly detects lighting degradation, dust interference, image ghosting, packet loss rate, effective frame rate, occlusion ratio, trajectory continuity, and edge node resource occupation, and adopts a continuous confirmation mechanism to determine the working condition, which can avoid false switching caused by short-term dust, instantaneous occlusion, and transmission fluctuations;
[0048] 2. This invention maintains monitoring availability and real-time performance in complex environments by generating and dynamically distributing versioned security monitoring configuration tables, and adaptively adjusting input resolution, processing frame rate, detection category, recognition algorithm level, verification rules, and maximum processing latency in conjunction with conditions such as dust enhancement and computing power constraints. Furthermore, the on-site risk assessment module integrates regional basic risk weights, edge risk quantities, equipment risk quantities, occlusion ratios, and recognition confidence levels, and performs risk maintenance and level assessment for short-term missed detections and occluded targets, thereby improving alarm accuracy and reducing false alarms and missed detections.
[0049] 3. This invention completes frame time alignment, coordinate conversion, and target association through a multi-camera verification module, and confirms, maintains, reduces, or eliminates the danger level accordingly. This enhances the continuous tracking capability across cameras and suppresses single-view recognition bias. At the same time, the result feedback module writes the verification results, manual handling results, and subsequent trajectory corrections back to the configuration table of the next version, and automatically rolls back when the performance of the new version deteriorates, thus forming a closed-loop linkage mechanism covering detection, judgment, verification, feedback, and update. Attached Figure Description
[0050] To more clearly illustrate the technical solutions in the embodiments of the present 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 some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort:
[0051] Figure 1 This is a module architecture diagram of a computer vision-based building construction safety monitoring system in an embodiment of this application. Detailed Implementation
[0052] To make the above and other objects, features and advantages of the present invention more apparent and understandable, preferred embodiments will be described in detail below with reference to the accompanying drawings;
[0053] Please see Figure 1 A computer vision-based construction safety monitoring system includes:
[0054] The construction area division module is used to divide the construction area according to the calibration and construction information, and record the area boundary, basic risk weight, associated cameras and risk rules;
[0055] The video operation status detection module communicates with each camera and edge node to obtain video and node resource information, and calculates the amount of lighting degradation, dust interference coefficient, motion blur coefficient, packet loss rate, effective frame rate, occlusion ratio and trajectory continuity.
[0056] The safety monitoring configuration table generation module communicates with the construction area division module and the video operation status detection module to generate a safety monitoring configuration table with a version number. The table includes camera priority, resolution, detection category, algorithm level, frame rate, confidence threshold, risk conversion parameters, review rules, review cameras, camera intrinsic and extrinsic parameter matrix containing camera focal length, optical center coordinates and rotation and translation vectors, risk retention duration and maximum processing latency.
[0057] The on-site risk assessment module is used to perform image recognition according to the effective configuration table, detect and track construction targets and dynamic equipment, calculate the spatial relationship between the target and the dangerous area, edge openings and dynamic equipment, and output the target hazard level.
[0058] The multi-camera verification module is used to perform frame-time alignment, coordinate conversion, and association with the target trajectories of different cameras, and accordingly confirm, maintain, reduce, or eliminate the target's danger level.
[0059] The result feedback module is connected to the safety monitoring configuration table generation module, the on-site risk assessment module, and the multi-camera verification module. It is used to output risk results and handling information, and to revise the next version of the configuration table based on the verification and handling results and subsequent trajectories.
[0060] In a preferred embodiment, the video operation status detection module is used to: extract the current image frame and video transmission information from the video of the associated camera, extract the reference clear frame and reference lighting data from the video data of historical normal time periods, and obtain target occlusion information and trajectory records from the on-site risk assessment module;
[0061] The system detects the amount of illumination degradation, dust interference coefficient, ghosting coefficient, packet loss rate, effective frame rate, occlusion ratio, and trajectory continuity, and obtains the CPU utilization, memory utilization, and bandwidth utilization of the edge nodes to determine the operating conditions.
[0062] When the dust interference coefficient exceeds the limit, the lighting degradation exceeds the limit, the occlusion ratio exceeds the limit, the effective frame rate is zero, or any of the central processing unit utilization rate, memory utilization rate, or bandwidth utilization rate exceeds the limit, and each exceeds the limit for a preset confirmation period, the corresponding conditions of dust enhancement, lighting abnormality, severe occlusion, video interruption, or computing power shortage are determined.
[0063] In this embodiment, the construction area division module is executed by the area configuration server deployed on site; the server first receives camera calibration information, construction layout, location of edge openings, hoisting path and vehicle access information, and then divides the site into several risk areas;
[0064] Each risk zone includes the zone boundary, basic risk weight, associated cameras, and applicable risk rules. The basic risk weight is used to characterize the basic level of danger when personnel approach the edge, enter the vicinity of the opening, or approach dynamic equipment in that zone.
[0065] After the construction area is divided, the area number and spatial boundary are written into the area configuration table and sent to the subsequent risk assessment module and review module as a unified spatial benchmark.
[0066] The video operation status detection module is executed by the edge node; the edge node receives continuous video streams from each camera and extracts the current image frame, video transmission information and node resource information in each system cycle; the current image frame is used to calculate the illumination degradation, dust interference coefficient, image motion coefficient, occlusion ratio and trajectory continuity;
[0067] The video transmission information includes the data packet sequence number and arrival time. The packet loss rate is calculated as the percentage of the total number of missing sequence numbers in the current system cycle relative to the total number of expected received sequence numbers. The effective frame rate is calculated by dividing the number of fully decoded image frames in the current system cycle by the system cycle duration.
[0068] Node resource information is used to read the CPU utilization, memory utilization, and bandwidth utilization. The method for calculating the amount of illumination degradation is to compare the global brightness average of the current image frame with the reference clear frame selected in the historical normal period, and extract the difference in grayscale distribution between the two to obtain the degree of deviation relative to the reference illumination data. Specifically, the absolute value of the difference between the global brightness average of the current frame and the brightness of the reference illumination data is divided by the brightness of the reference illumination data, and the percentage obtained is used as the amount of illumination degradation.
[0069] The dust interference coefficient is calculated by applying the Laplacian operator to the image to calculate the edge gradient variance. The ratio of the decrease in the variance of the current frame relative to the variance of the reference normal frame is used as the quantification result of the high-frequency detail attenuation. That is, the variance of the reference normal frame is subtracted from the variance of the current frame and then divided by the variance of the reference normal frame.
[0070] The image motion blur coefficient identifies the motion trailing at the edge of the target by inter-frame difference. Specifically, the current frame is differentially processed with the previous adjacent frame. A continuous pixel connected region with monotonically decreasing gray values is extracted in the opposite direction of motion along the outer edge of the target as the trailing region. The maximum pixel span of this connected region in the opposite direction of motion is taken as the residual pixel width, and then divided by the pixel width of the target itself to obtain the result.
[0071] The occlusion ratio is calculated based on the proportion of the visible area within the detection box. Specifically, the target detection network is used to extract key feature points of the target, including the coordinates of the target's head, shoulders, hips, and feet. After extraction, the maximum and minimum horizontal coordinates and the maximum and minimum vertical coordinates of all visible key feature points are taken, and the visible bounding box is generated based on the extreme value range of these coordinates. The visible bounding box area is then subtracted from the area of the prior bounding box with the standard aspect ratio of the target category preset by the system, and then divided by the area of the prior bounding box.
[0072] Trajectory continuity is determined by whether the rate of change of target coordinates between adjacent frames exceeds the extreme value of normal movement, and by whether the target identifier is lost, thus determining an interruption.
[0073] Normal motion extrema are the maximum pixel displacement thresholds that are converted from the theoretical maximum motion velocity of the target's physical type to the image plane. The specific conversion function is as follows:
[0074]
[0075] in, The unit is pixel; The theoretical maximum physical velocity of the target, expressed in m / s; This refers to the camera's focal length, in mm. The time interval between adjacent frames, in seconds; This is an estimated depth distance between the target and the camera, in meters. The physical size of a single pixel in an image sensor, expressed in m / pixel;
[0076] Reference clear frames and reference lighting data are extracted only from historical videos that are unobstructed, alarm-free, and have satisfactory image quality, in order to avoid writing abnormal images into the reference baseline;
[0077] When the dust interference coefficient exceeds the limit, the lighting degradation exceeds the limit, the occlusion ratio exceeds the limit, the effective frame rate is zero, or any of the central processing unit utilization rate, memory utilization rate, or bandwidth utilization rate exceeds the limit, the threshold for each indicator is calibrated according to the on-site environment. For example, if the dust interference coefficient drops by 40% or the lighting brightness deviates from the normal benchmark by more than 30%, the video running status detection module will not immediately switch the processing mode, but will continue to accumulate the results in the continuous confirmation window.
[0078] A flip only occurs when the abnormal state meets the cumulative threshold based on the state machine counter. Specifically, an abnormal counter is set independently for each of the above abnormal indicators. If the abnormal indicator exceeds the limit in each system cycle, the counter is incremented by 1; otherwise, it is decremented by 1 with a lower limit of 0.
[0079] When the count value of any abnormal counter continuously accumulates to the upper limit of the number of cycles corresponding to the preset confirmation period, usually set to 3 to 5 system cycles to filter out occasional interference such as brief dust, the system state machine triggers a flip, and the corresponding camera or node is determined to be in a state of increased dust, abnormal lighting, severe obstruction, video interruption, or computing power shortage.
[0080] The judgment result is sent to the safety monitoring configuration table generation module to adjust the subsequent detection category, input resolution, processing frame rate and review rules; in addition, when the calculation result of the dust interference coefficient is negative, it is treated as 0 to ensure that the coefficient is always non-negative and is not used as the triggering basis for the dust enhancement condition.
[0081] In a preferred embodiment, after determining that an operating condition has occurred, the safety monitoring configuration table generation module specifically performs the following steps:
[0082] Identify the affected cameras, construction areas, operational status, and duration;
[0083] When the dust interference coefficient exceeds the limit, the resource adjustment strategy corresponding to the dust interference is to reduce the resolution, reduce the processing frame rate, or switch to a lightweight target detection network. When the CPU utilization, memory utilization, or bandwidth utilization exceeds the limit, the resolution reduction strategy is preferred.
[0084] Based on the affected construction area, retrieve the corresponding basic risk weights, risk rules, associated cameras, and the safety monitoring configuration table of the previous valid version;
[0085] Processing resources are allocated in order of priority from high to low based on basic risk weights. The maximum processing latency is limited according to a preset safety latency threshold. The system generates camera priority, input resolution, processing frame rate, detection category, recognition algorithm level, recognition confidence threshold, verification camera, risk retention duration, and maximum processing latency.
[0086] The generated safety monitoring configuration table is sent to the corresponding edge node, and the corresponding processing method is exited after the working condition is lifted and the preset lifting time period is continuously met.
[0087] The process of building and running the security monitoring configuration table includes: setting the version number, validity period and rollback version for each security monitoring configuration table, and enabling the edge node to verify the integrity of the data packet and return version confirmation information after receiving it;
[0088] The edge nodes that return version confirmation information load the new version and execute the on-site risk assessment module and the multi-camera verification module, while the edge nodes that do not return version confirmation information continue to execute the previous valid version.
[0089] Receive processing time, resource utilization, task failure information, and risk identification results returned by edge nodes;
[0090] The system will collect operational metrics within a preset time window. If the new version causes the false negative rate in the corresponding region to be higher than the baseline false negative rate of the previous effective version, the maximum processing latency to exceed the preset latency limit, the CPU utilization rate, memory utilization rate, or bandwidth utilization rate of the edge node to exceed the corresponding preset load limit, or the verification consistency rate to be lower than the preset consistency rate lower limit, the new version will be stopped and the system will be restored to the previous effective version.
[0091] The on-site risk assessment module obtains the target recognition confidence level from the image recognition results and obtains the regional basic risk weight. Peripheral risk level Equipment risk level Occlusion ratio and target determination error value ;
[0092] in, This indicates the basic risk weight of the construction area where the target is located;
[0093] The non-negative normalized value is obtained by dividing the difference between the preset edge distance threshold and the distance from the target to the edge opening by the edge distance threshold. When the distance from the target to the edge opening is greater than the edge distance threshold, Values ;
[0094] The non-negative normalized value is obtained by dividing the difference between the preset device distance threshold and the distance from the target to the dynamic device by the device distance threshold. When the distance from the target to the dynamic device is greater than the device distance threshold, Values ;
[0095] Indicates the proportion of the target that is obscured; ;
[0096] According to the formula Calculate and determine the risk level score ,in This is a preset occlusion amplification factor used to quantify occlusion interference weights. Let be the preset error amplification factor used for quantization error interference weights, and satisfy . ;
[0097] Preset first risk limit Second risk boundary ,and When the risk level is scored When the target identification confidence level is not lower than the current identification confidence level threshold and the first preset frame number threshold is met continuously, the target danger level is set to or restored to the safe level.
[0098] When the risk level score meets Furthermore, when the second preset frame count threshold is met consecutively, the target danger level is set to the proximity level.
[0099] When the risk level score meets Or, when the target enters a dangerous area and the spatial relationship is reliable, the target's danger level is set to danger level;
[0100] Targets at the danger level only meet the danger level score within the third consecutive preset frame threshold. And when there is no continuous obstruction and the verification results confirm safety, it will be restored to the safe level;
[0101] When a target is missed for a short period of time or the occlusion ratio exceeds the preset occlusion limit, and the number of continuous tracking frames before occlusion reaches the fourth preset frame threshold, the target danger level before occlusion is maintained, and the associated camera is requested to be reviewed after the occlusion time exceeds the risk maintenance time.
[0102] The multi-camera verification module acquires the time difference of the target to be verified in different cameras, coordinate conversion error, direction of motion and similarity of appearance;
[0103] Targets in different cameras are identified as the same target only when the acquisition time difference is less than or equal to the preset time difference, the coordinate conversion error is less than or equal to the preset distance difference, the angle of movement direction is less than or equal to the preset angle threshold, and the appearance similarity is greater than or equal to the preset matching threshold.
[0104] The danger level is maintained as long as the associated camera confirms that the target is still in the danger zone; the danger level is raised to a higher level if the risk is confirmed to be escalating.
[0105] When the associated camera confirms that the target has left the danger zone, it is upgraded to a safe level according to the continuous safety assessment criteria.
[0106] When the appearance similarity is lower than the preset matching threshold, the original high-risk status is not lifted, and a review is initiated again within the risk retention period. If a reliable association cannot be achieved after the risk retention period, a manual review request is generated.
[0107] The result feedback module communicates with the alarm terminal and the manual review server to output target number, trajectory number, area number, risk type, risk level, start time, duration, evidence frame, review source, security monitoring configuration table version and suggested handling information. It sends high-risk results to the alarm terminal and sends results with insufficient identification confidence and unable to complete multi-camera review to the queue of the manual review server.
[0108] The results feedback module also updates the security monitoring configuration table parameters based on the feedback results, and takes effect after verification; if verification fails, it reverts to the previous valid version.
[0109] In this embodiment, the safety monitoring configuration table generation module is executed by the regional configuration server. This module first reads the affected cameras, affected construction areas, working condition levels and durations, and then combines them with the previous valid version of the safety monitoring configuration table to generate a new configuration table.
[0110] If the current operating condition is dust-enhanced, prioritize reducing the detection categories that are highly dependent on details, and adjust the input resolution, processing frame rate, and recognition algorithm level so that edge nodes can still maintain usable processing capabilities under dust interference.
[0111] If the current operating conditions are characterized by limited computing power, a resolution reduction strategy will be prioritized to reduce the processing burden per frame, and camera priorities will be adjusted simultaneously to ensure that cameras in high-risk areas receive processing resources first; the specific configuration generation rules are as follows:
[0112] Establish a preset tiered adjustment mapping table. When the system determines that it is under computing power pressure and the working condition level is level one, for example, when the CPU utilization rate is between 85% and 95%, the input resolution is reduced from 1080p to 720p and the processing frame rate is reduced from 25 frames / second to 15 frames / second.
[0113] If the operating condition level is more severe, such as when the CPU utilization reaches or exceeds 95%, the recognition algorithm is further switched from a high-precision network that extracts multi-dimensional appearance features of the target. This high-precision network takes the three-channel red-green-blue image matrix of the current resolution as input features and the output vector contains not only the target bounding box coordinates but also a multi-dimensional appearance feature vector with an extraction dimension of 128. The algorithm is then switched to a lightweight target detection network that only outputs the target bounding box coordinates. This lightweight network uses depthwise separable convolution. The input is the same, but the output vector only contains the bounding box coordinates and a one-dimensional confidence score. It no longer outputs high-dimensional features, thereby compressing the computation time to within the original safe latency threshold.
[0114] The high-precision network mentioned above specifically uses a fifty-layer residual network as the backbone extraction network, and adds a region of interest alignment layer and a fully connected layer to output 128-dimensional features; the lightweight object detection network specifically adopts a second-generation lightweight mobile neural network architecture.
[0115] Both networks are trained using image datasets collected from actual construction sites and with manually annotated bounding boxes. The loss function during the training phase... Generalized intersection-union loss from bounding box regression Cross-entropy loss with classification Weighted composition, the specific formula is as follows
[0116]
[0117] Among them, the weighting coefficient , ;
[0118] To prevent the target trajectory from breaking due to the cross-model feature dimension discontinuity during the downgrade switching, the edge nodes are forced to retain the target bounding box and motion prediction vector output by the high-precision network in the previous frame during the transition frame of the algorithm level switching. They are also associated with the bounding box generated by the lightweight network in the current frame by performing cross-over ratio and Kalman filter bidirectional matching to inherit the original target trajectory number.
[0119] The state vector of the Kalman filter is set as follows: , respectively represent the horizontal coordinates of the center point of the target bounding box. Vertical coordinates of the center point Aspect Ratio ,high and the rate of change of each variable The observation vector is set to The state transition model uses uniform linear kinematics equations;
[0120] When the configuration table is generated, the basic risk weight, risk rules, associated cameras and configuration content of the previous effective version of the affected construction area are retrieved first. Then, the basic risk weights are sorted from high to low to ensure that cameras, detection categories and review rules in high-risk areas are allocated first.
[0121] To meet the maximum processing latency constraint, the specific allocation calculation method is as follows: the edge nodes pre-store the baseline time corresponding to each detection category and algorithm level under different input resolutions, and the module allocates high-precision recognition algorithms and high frame rates to the associated cameras in order of high to low basic risk weights;
[0122] After each allocation, the baseline latency of each camera is accumulated to estimate the total latency under this configuration. If the estimated total latency is greater than the safe latency threshold, the resolution reduction and frame rate reduction strategies are executed sequentially for the cameras with lower weights until the estimated total latency meets the constraints.
[0123] The camera priority here is used to determine which video stream to process first when resources are scarce; the recognition algorithm level is used to determine whether to use a lighter or higher precision recognition method; the risk retention time is used to limit the state retention time when the target disappears for a short time.
[0124] Maximum processing latency is used to limit the longest allowed path from image input to risk output; after the configuration table is generated, the server sends it to the corresponding edge node. The edge node first checks the integrity of the data packet and then returns version confirmation information; the node that returns confirmation information loads the new version and performs risk judgment and review, while the node that does not return confirmation information continues to use the previous valid version;
[0125] During operation, edge nodes periodically transmit processing time, resource utilization, task failure information, and risk identification results; the configuration table generation module collects these indicators within a preset time window and compares them with the corresponding operating indicators of the previous valid version.
[0126] The corresponding area false alarm rate is calculated by the system based on the ratio of the number of times the current camera target trajectory is abnormally interrupted, that is, disappears before reaching the normal exit, and is successfully verified by the associated camera or subsequently recaptured, to the total number of tracked targets, thus providing a quantitative self-evaluation basis.
[0127] If the new version causes an increase in the false negative rate in the corresponding area, such as an increase of more than 5% higher than the baseline false negative rate of the previous effective version, or the maximum processing latency or node load exceeds the limit range, where the preset latency upper limit is set to 300 to 500 milliseconds based on the real-time requirements of security monitoring, the preset load upper limit is set to 85% to 90%, or the verification consistency rate decreases, such as falling below the preset consistency rate lower limit of 80%, then the system's forced rollback state machine mechanism will be triggered to stop the execution of the new version and revert to the previous effective version in order to prevent the processing strategy from continuing to spread to more cameras;
[0128] In this rollback control process, the server sends a rollback instruction with the highest priority interruption flag to the abnormal edge node. After the edge node intercepts the instruction, it asynchronously reloads the static configuration parameters of the previous valid version in the background. During this period, the current pipeline is maintained until the transition frame is completed. Only the feature register of the new version is cleared, but the target state buffer of the previous valid frame is forcibly retained to maintain the temporal continuity of the Kalman filter during the physical hysteresis time, so as to achieve a deterministic closed-loop reset of the version state.
[0129] In this embodiment, the on-site risk assessment module determines the level of danger score. At the same time, it is achieved through the linear fusion and penalty mechanism of multi-dimensional risk factors to expose its underlying computational relationship and avoid the uninterpretability of the algorithm;
[0130] When calculating, the amount of near-edge risk is included. and equipment risk Take the maximum value as the spatial risk quantity; then combine this spatial risk quantity with the regional basic risk weight. Add them together to get the base score;
[0131] Use occlusion ratio and target determination error value As the basis for punishment, preset occlusion amplification factor and error amplification factor are introduced respectively; the specific calculation formula is as follows:
[0132] Risk level rating It equals the product of the base score and the penalty multiplier, where the penalty multiplier is a value of 1 plus the occlusion percentage. The product of the magnification factor and the occlusion factor, plus the target determination error value. The product of the error amplification factor and the error amplification factor; the amplification factor is used to quantify the weight of different penalty factors on the final risk assessment, and its specific value is determined based on historical sample regression analysis.
[0133] When the target is For the region with a value of 0.3, calculate It is 0.6 and If the value is 0, the spatial risk level is taken as 0.6, and the base score is 0.9; if the system identifies... It is 0.2 and The value is set to 0.1, because occlusion has a higher feature interference weight in visual evaluation, so it is preset to be assigned to... The occlusion magnification factor, such as 0.5, is greater than that assigned to... The error amplification factor, such as 0.2, results in a penalty multiplier of 1 plus... That is, 1.12, resulting in the final score. It is 0.9 multiplied by 1.12, which is 1.008;
[0134] At the same time, regarding the first risk limit Second risk boundary The tolerance levels can be preset to 0.8 and 1.2 respectively; in the above scoring... In instances where the value is 1.008, the value is between and If a target continuously meets the second preset frame rate threshold, it can be directly classified as a proximity level.
[0135] In the case of target occlusion, when the occlusion ratio exceeds the preset occlusion limit, such as when the visible area of the detection box is less than 50%, a short-term hold mechanism is triggered. Through this structured logic decomposition and boundary threshold setting, the multivariate fusion operation is ensured to have a quantitative implementation basis and the feature fusion processing efficiency is improved.
[0136] For the appearance similarity in the multi-camera verification module, its calculation mechanism needs to adapt to the heterogeneous algorithm state of the edge nodes under different resource conditions; when the edge nodes corresponding to the two associated cameras are running high-precision networks, the 128-dimensional multi-dimensional appearance feature vector output by the network is directly reused to calculate the cosine similarity.
[0137] When any node is downgraded to a lightweight network that does not output high-dimensional features due to computing power constraints, the hue, saturation, brightness, and color histogram feature vectors of the target to be verified in the images of the two cameras are extracted.
[0138] The hue, saturation, and brightness of the color space are quantized according to preset intervals, such as dividing the H component into 32 segments. and The components are divided into 48 segments. After the color histogram is statistically analyzed and normalized, the frequency of each component segment is concatenated into a 128-dimensional one-dimensional feature vector. When there are nodes that run a lightweight network in a downgraded manner, the hue, saturation, and brightness color histogram feature vectors of the target to be verified in both video frames are forcibly extracted. The cosine similarity between these two homogeneous 128-dimensional color feature vectors is calculated as the degree of appearance similarity, thereby compatibility with the output dimension differences of different networks.
[0139] Meanwhile, to address the inconsistent processing frame rates that may result from different edge nodes not synchronously confirming the new version, such as a mixed running of 15 frames / second and 25 frames / second, the verification module uses the nearest neighbor timestamp matching method when performing frame time alignment, selecting the frame with the smallest absolute difference in acquisition timestamps between the two video streams as the association benchmark.
[0140] To address the spatial mapping problem, the multi-camera verification module retrieves the camera intrinsic and extrinsic parameter matrices from the configuration table and uses perspective transformation to project the midpoint of the bottom edge of the detection frame of the target to be verified in the images of two different cameras onto a unified on-site two-dimensional plane reference. It then calculates the Euclidean distance between the two projection points in this unified coordinate system as the coordinate conversion error.
[0141] After completing the target association mapping, if the target danger level recalculated by the associated camera from its own perspective is higher than the level output by the plaintiff police camera, then the risk escalation is confirmed.
[0142] If it is determined that the target has left, a continuous safety countdown will begin. Only when the cosine similarity is greater than the preset matching threshold, such as 0.85, and the acquisition time difference is less than the preset time difference, and the coordinate conversion error is less than the preset distance difference and the angle of motion direction is less than the preset angle threshold, will the mapping with the target be completed, thus providing a rigorous and quantifiable underlying benchmark calculation basis for heterogeneous time series and data dimensions across cameras.
[0143] In addition, the amount of risk at the border and equipment risk When the calculation result is negative, it is recorded as 0 to ensure the non-negativity of the normalization result. , , , and All are dimensionless scalars, therefore the risk level score It is also a dimensionless score.
[0144] For any part not mentioned in this invention, existing technologies can be used or referenced. Of course, the above description is not a limitation of this invention, and this invention is not limited to the examples given above. Any changes, modifications, additions or substitutions made by those skilled in the art within the scope of this invention should also be within the protection scope of this invention.
Claims
1. A computer vision-based construction safety monitoring system, characterized in that, include: The construction area division module is used to divide the construction area according to the calibration and construction information, and record the area boundary, basic risk weight, associated cameras and risk rules; The video operation status detection module communicates with each camera and edge node to obtain video and node resource information, and calculates the amount of lighting degradation, dust interference coefficient, motion blur coefficient, packet loss rate, effective frame rate, occlusion ratio and trajectory continuity. The safety monitoring configuration table generation module communicates with the construction area division module and the video operation status detection module to generate a safety monitoring configuration table with a version number. The table includes camera priority, resolution, detection category, algorithm level, frame rate, confidence threshold, risk conversion parameters, review rules, review cameras, camera intrinsic and extrinsic parameter matrix including camera focal length, optical center coordinates and rotation and translation vectors, risk retention duration and maximum processing latency. The on-site risk assessment module is used to perform image recognition according to the effective configuration table, detect and track construction targets and dynamic equipment, calculate the spatial relationship between the target and the dangerous area, edge openings and dynamic equipment, and output the target hazard level. The multi-camera verification module is used to perform frame-time alignment, coordinate conversion, and association with the target trajectories of different cameras, and accordingly confirm, maintain, reduce, or eliminate the target's danger level. The result feedback module is connected to the safety monitoring configuration table generation module, the on-site risk assessment module, and the multi-camera verification module. It is used to output risk results and handling information, and to revise the next version of the configuration table based on the verification and handling results and subsequent trajectories.
2. The computer vision-based construction safety monitoring system according to claim 1, characterized in that, The video running status detection module is used for: Extract the current image frame and video transmission information from the video of the associated camera, extract the reference clear frame and reference lighting data from the video data of historical normal time periods, and obtain the target occlusion information and trajectory record from the on-site risk assessment module; The system detects the amount of illumination degradation, dust interference coefficient, ghosting coefficient, packet loss rate, effective frame rate, occlusion ratio, and trajectory continuity, and obtains the CPU utilization, memory utilization, and bandwidth utilization of the edge nodes to determine the operating conditions. When the dust interference coefficient exceeds the limit, the lighting degradation exceeds the limit, the occlusion ratio exceeds the limit, the effective frame rate is zero, or any of the central processing unit occupancy rate, memory occupancy rate, and bandwidth occupancy rate exceeds the limit, and each exceeds the limit for a preset confirmation period, the corresponding conditions of dust enhancement, lighting abnormality, severe occlusion, video interruption, or computing power shortage are determined to have occurred.
3. The computer vision-based construction safety monitoring system according to claim 2, characterized in that, After determining that the aforementioned operating condition has occurred, the safety monitoring configuration table generation module performs the following steps: Identify the affected cameras, construction areas, operational status, and duration; When the dust interference coefficient exceeds the limit, the resource adjustment strategy corresponding to the dust interference is to reduce the resolution, reduce the processing frame rate, or switch to a lightweight target detection network. When the CPU utilization, memory utilization, or bandwidth utilization exceeds the limit, the resolution reduction strategy is preferred. Based on the affected construction area, retrieve the corresponding basic risk weights, risk rules, associated cameras, and the safety monitoring configuration table of the previous valid version; Processing resources are allocated in order of priority from high to low based on basic risk weights. The maximum processing latency is limited according to a preset safety latency threshold. The system generates camera priority, input resolution, processing frame rate, detection category, recognition algorithm level, recognition confidence threshold, verification camera, risk retention duration, and maximum processing latency. The generated safety monitoring configuration table is sent to the corresponding edge node, and the corresponding processing mode is exited after the working condition is lifted and the preset lifting time period is continuously met.
4. The computer vision-based construction safety monitoring system according to claim 3, characterized in that, The process of building and running the security monitoring configuration table includes: Set the version number, validity period, and rollback version for each security monitoring configuration table, and enable the edge nodes to verify the integrity of the data packets and return version confirmation information after receiving them; The edge nodes that return version confirmation information load the new version and execute the on-site risk assessment module and the multi-camera verification module, while the edge nodes that do not return version confirmation information continue to execute the previous valid version. Receive processing time, resource utilization, task failure information, and risk identification results returned by edge nodes; The system analyzes operational metrics within a preset time window. If the new version causes the false negative rate in the corresponding region to be higher than the baseline false negative rate of the previous effective version, the maximum processing latency to exceed the preset latency limit, the CPU utilization rate, memory utilization rate, or bandwidth utilization rate of the edge nodes to exceed the corresponding preset load limit, or the verification consistency rate to be lower than the preset consistency rate lower limit, the new version will be stopped and the system will revert to the previous effective version.
5. The computer vision-based construction safety monitoring system according to claim 4, characterized in that, The on-site risk assessment module obtains the target recognition confidence level from the image recognition results and obtains the regional basic risk weight. Peripheral risk level Equipment risk level Occlusion ratio and target determination error value ; in, This indicates the basic risk weight of the construction area where the target is located; The non-negative normalized value is obtained by dividing the difference between the preset edge distance threshold and the distance from the target to the edge opening by the edge distance threshold. When the distance from the target to the edge opening is greater than the edge distance threshold, Values ; The non-negative normalized value is obtained by dividing the difference between the preset device distance threshold and the distance from the target to the dynamic device by the device distance threshold. When the distance from the target to the dynamic device is greater than the device distance threshold, Values ; Indicates the proportion of the target that is obscured; ; According to the formula Calculate and determine the risk level score ,in This is a preset occlusion amplification factor used to quantify occlusion interference weights. Let be the preset error amplification factor used for quantization error interference weights, and satisfy . .
6. The computer vision-based construction safety monitoring system according to claim 5, characterized in that, Preset first risk limit Second risk boundary ,and ; When the risk level is rated When the target identification confidence level is not lower than the current identification confidence level threshold and the first preset frame number threshold is met continuously, the target danger level is set to or restored to the safe level. When the risk level score meets Furthermore, when the second preset frame count threshold is met consecutively, the target danger level is set to the proximity level. When the risk level score meets Or, when the target enters a dangerous area and the spatial relationship is reliable, the target's danger level is set to danger level; Targets at the danger level only meet the danger level score within the third consecutive preset frame threshold. And when there is no continuous obstruction and the verification results confirm safety, it will be restored to the safe level; When a target is missed for a short period of time or the occlusion ratio exceeds the preset occlusion limit, and the number of continuous tracking frames before occlusion reaches the fourth preset frame threshold, the target danger level before occlusion is maintained, and the associated camera is requested to be reviewed after the occlusion time exceeds the risk maintenance duration.
7. The computer vision-based construction safety monitoring system according to claim 6, characterized in that, The multi-camera verification module acquires the time difference of the target to be verified in different cameras, coordinate conversion error, direction of movement and similarity of appearance; Targets in different cameras are identified as the same target only when the acquisition time difference is less than or equal to a preset time difference, the coordinate conversion error is less than or equal to a preset distance difference, the angle of movement direction is less than or equal to a preset angle threshold, and the appearance similarity is greater than or equal to a preset matching threshold. The danger level is maintained when the associated camera confirms that the target is still in a dangerous area, and is raised to a higher level when the risk is confirmed to be escalating. When the associated camera confirms that the target has left the danger zone, it is converted to a safe level according to the continuous safety determination criteria; When the appearance similarity is lower than the preset matching threshold, the original high-risk status is not lifted, and a review is initiated again within the risk retention period. If a reliable association cannot be achieved after the risk retention period, a manual review request is generated.
8. The computer vision-based construction safety monitoring system according to claim 7, characterized in that, The result feedback module is connected to the alarm terminal and the manual review server. It is used to output the target number, trajectory number, area number, risk type, risk level, start time, duration, evidence frame, review source, security monitoring configuration table version and suggested handling information. It sends high-risk results to the alarm terminal and sends results with insufficient confidence and unable to complete multi-camera review to the queue of the manual review server. The result feedback module also updates the security monitoring configuration table parameters based on the feedback results, and takes effect after verification; if verification fails, it reverts to the previous valid version.
