A construction area safety monitoring method and system based on target identification

Abstract

The application discloses a kind of construction area safety monitoring method and system based on target identification, it is related to image recognition technical field, including: to the continuous monitoring video stream of construction area is frame by frame analysis, identify and continuously track multiple monitoring targets, form space-time state sequence;Cross target correlation analysis is carried out, and interactive behavior sub-sequence is extracted;Risk behavior mode library including multiple risk primitive patterns is constructed;Interactive behavior sub-sequence is matched with risk primitive pattern in multidimension, and risk event instance is judged according to adaptive threshold value, risk event in preset period is counted, risk exposure density distribution and behavior safety deviation degree index are calculated, and construction area dynamic safety situation atlas is generated.The application solves the technical problem that potential risk interactive behavior in construction area is difficult to automatically identify and quantitatively evaluate in the prior art, achieves the technical effect of discovering potential safety risk in construction area in advance and realizing risk situation quantization presentation.

Description

Technical Field

[0001] This invention relates to the field of image recognition technology, and specifically to a method and system for safety monitoring of construction areas based on target recognition. Background Technology

[0002] Within construction areas, personnel, machinery, materials, and environmental elements such as openings and edges are densely intertwined. Safety risks are often not static, single targets, but rather dynamic interactions such as personnel approaching equipment, personnel entering hazardous work areas, and the intersection of hoisted objects and workers' trajectories. Traditional construction safety monitoring relies heavily on manual inspections, post-event video review, or single-target identification alarms. These methods typically only detect violations or obviously abnormal targets, making it difficult to continuously track the position, posture, movement trends, and interactions between multiple targets. Furthermore, there is a lack of means to uniformly model, match, judge, and quantify potential risk interactions, resulting in the difficulty in timely detection and assessment of risk events within the construction area that have not yet escalated into accidents. Summary of the Invention

[0003] This application provides a construction area safety monitoring method and system based on target recognition, which is used to address the technical problem that it is difficult to automatically identify and quantify the interaction behaviors of potential risks in construction areas in the prior art.

[0004] In view of the above problems, this application provides a construction area safety monitoring method and system based on target recognition.

[0005] The first aspect of this application provides a construction area safety monitoring method based on target identification, the method comprising:

[0006] The system acquires and analyzes the monitoring video stream of the construction area frame by frame over a continuous time period, identifies and continuously tracks multiple monitoring targets, and generates a spatiotemporal state sequence for each target, including its category, location, attitude, motion vector, and spatial occupancy. Cross-target correlation analysis is performed on the spatiotemporal state sequences of the multiple monitoring targets to extract interactive behavior subsequences of each target within a preset time window. Each interactive behavior subsequence includes at least the interaction relationship between two monitoring targets. A risk behavior pattern library containing multiple risk primitive patterns is constructed, where each risk primitive pattern can define a combination of risk features of one or more monitoring targets in the spatiotemporal state sequence and / or interactive behavior subsequences. The interactive behavior subsequences are matched with the risk primitive patterns in multiple dimensions. When the matching degree exceeds an adaptive threshold, a risk event instance is determined. Based on all risk event instances determined within a preset period, a risk exposure density distribution and behavioral safety deviation index are calculated. Based on the risk exposure density distribution and behavioral safety deviation index, a dynamic safety situation map of the construction area is generated.

[0007] A second aspect of this application provides a construction area safety monitoring system based on target recognition, the system comprising: The system includes a target identification module for acquiring and analyzing frame-by-frame surveillance video streams of the construction area over a continuous time period, identifying multiple monitoring targets for continuous tracking, and generating a spatiotemporal state sequence for each target, including its category, location, attitude, motion vector, and spatial occupancy. A correlation analysis module performs cross-target correlation analysis on the spatiotemporal state sequences of the multiple monitoring targets, extracting interactive behavior sub-sequences of each target within a preset time window. These interactive behavior sub-sequences include at least the interaction relationship between two monitoring targets. A pattern library construction module constructs a risk behavior pattern library containing multiple risk primitive patterns, where each risk primitive pattern can define a combination of risk features of one or more monitoring targets in the spatiotemporal state sequence and / or interactive behavior sub-sequences. A matching module performs multi-dimensional matching between the interactive behavior sub-sequences and the risk primitive patterns; when the matching degree exceeds an adaptive threshold, a risk event instance is determined to have occurred. A calculation module calculates the risk exposure density distribution and behavioral safety deviation index based on all risk event instances determined within a preset period, and generates a dynamic safety situation map of the construction area based on these indicators.

[0008] One or more technical solutions provided in this application have at least the following technical effects or advantages: This application acquires and analyzes frame-by-frame surveillance video streams of the construction area over a continuous time period, identifies and continuously tracks multiple monitoring targets, and generates a spatiotemporal state sequence for each target, including its category, location, attitude, motion vector, and spatial occupancy. It then performs cross-target correlation analysis on the spatiotemporal state sequences of the multiple monitoring targets, extracting interactive behavior subsequences of each target within a preset time window. These interactive behavior subsequences include at least the interaction relationship between two monitoring targets. A risk behavior pattern library containing multiple risk primitive patterns is constructed, where each risk primitive pattern can define a combination of risk characteristics of one or more monitoring targets in the spatiotemporal state sequence and / or interactive behavior subsequences. The interactive behavior subsequences are matched with the risk primitive patterns in multiple dimensions; when the matching degree exceeds an adaptive threshold, a risk event instance is determined. Based on all risk event instances determined within a preset period, a risk exposure density distribution and a behavioral safety deviation index are calculated, and a dynamic safety situation map of the construction area is generated based on these indicators. This invention addresses the technical problem of the difficulty in automatically identifying and quantifying potential risk interactions within construction areas in existing technologies. By continuously tracking multiple monitoring targets in a surveillance video stream and determining risk event instances based on the matching of interaction behaviors between targets with risk primitive patterns, it achieves the technical effect of early detection of potential safety risks in construction areas and quantitative presentation of risk status. Attached Figure Description

[0009] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying 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.

[0010] Figure 1 A schematic diagram of a construction area safety monitoring method based on target recognition provided in this application embodiment; Figure 2 This is a schematic diagram of a construction area safety monitoring system based on target recognition, provided as an embodiment of this application.

[0011] Figure labeling: Target recognition module 11, Association analysis module 12, Pattern library construction module 13, Matching module 14, Calculation module 15. Detailed Implementation

[0012] This application provides a construction area safety monitoring method and system based on target recognition. It addresses the technical problem of the difficulty in automatically identifying and quantifying the interactive behaviors of potential risks in construction areas in the prior art. By continuously tracking multiple monitoring targets in the monitoring video stream and determining risk event instances based on the matching of interactive behaviors between targets and risk primitive patterns, it achieves the technical effect of discovering potential safety risks in construction areas in advance and realizing the quantitative presentation of risk status.

[0013] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0014] It should be noted that any variation of the terms "comprising" and "having" is intended to cover non-exclusive inclusion, for example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or modules that are not explicitly listed or that are inherent to such processes, methods, products, or devices.

[0015] Example 1, as Figure 1 As shown, this application provides a construction area safety monitoring method based on target recognition, the method comprising: Step S100: Obtain the monitoring video stream of the construction area in a continuous time period, perform frame-by-frame analysis, identify multiple monitoring targets for continuous tracking, and generate a spatiotemporal state sequence for each monitoring target, including its category, location, attitude, motion vector, and spatial occupancy range.

[0016] In this embodiment, fixed cameras, tower crane cameras, vehicle cameras, or mobile inspection cameras installed within the construction area continuously capture images of the construction area to obtain a monitoring video stream of the construction area over a continuous time period. After acquiring the monitoring video stream, it is decoded, continuous video frames are extracted according to a preset frame rate, and the acquisition time of each video frame is recorded, so that each video frame corresponds to a specific time point.

[0017] Before frame-by-frame analysis, a target recognition model is established. Sample images of the construction area, including construction workers, machinery and equipment, transport vehicles, hoisted objects, materials, protective facilities, and hazardous area markers, are collected. Objects in the sample images are labeled with categories, bounding boxes, and poses to form a training sample set. The training sample set is then input into the target recognition model for training, enabling the model to learn the appearance, contour, shape, and pose features of different monitored targets. When the category, location, and pose output by the target recognition model meet the preset accuracy requirements, the trained target recognition model is obtained.

[0018] Next, consecutive video frames are input sequentially into the trained target recognition model. The target recognition model extracts features from each video frame and identifies multiple monitoring targets in the video frame based on the extracted features, outputting the category, location, and attitude of each monitoring target. Here, category indicates the object type to which the monitoring target belongs, location indicates the coordinate position of the monitoring target in the video frame or construction area, and attitude indicates the body orientation and working posture of construction personnel, or the orientation and operating status of machinery, transport vehicles, and hoisted objects.

[0019] After identifying multiple monitoring targets, the monitoring targets in adjacent video frames are matched. The monitoring targets in the previous video frame are compared with those in the next video frame. If they are of the same category, their positional distance is less than a preset distance threshold, and their positional changes conform to a continuous direction of movement, they are determined to be the same monitoring target and are identified by the same target identifier. By repeating the above matching process on consecutive video frames, the motion trajectory of each monitoring target within a continuous time period is obtained, enabling continuous tracking of multiple monitoring targets.

[0020] Then, based on the position coordinates of the same monitored target in adjacent video frames and the acquisition time, the motion vector of the monitored target is calculated. The motion vector includes the direction of movement and the speed of movement. Based on the position bounding box, target outline, and spatial coordinates of the construction area, the spatial occupancy range of the monitored target is determined. The spatial occupancy range represents the spatial area occupied by the monitored target within the construction area. Finally, the category, position, attitude, motion vector, and spatial occupancy range of each monitored target in each video frame are arranged in chronological order of acquisition time to generate a spatiotemporal state sequence for each monitored target.

[0021] Step S200: Perform cross-target correlation analysis on the spatiotemporal state sequences of the multiple monitoring targets, and extract the interaction behavior subsequences of each monitoring target within a preset time window. The interaction behavior subsequences include at least the interaction relationship between two monitoring targets.

[0022] In this embodiment of the application, when performing cross-target correlation analysis on the spatiotemporal state sequences of multiple monitoring targets, the spatial relationship changes between any two monitoring targets are calculated in real time based on the category, position, attitude, motion vector and spatial occupancy range of each monitoring target in a continuous time period. The changes in spatial relationship are used to determine whether the two monitoring targets have entered the predefined interactive interest area. When the two monitoring targets enter the predefined interactive interest area, a preset time window is triggered. Within the preset time window, the spatiotemporal state sequences corresponding to the two monitoring targets are extracted synchronously. The extracted spatiotemporal state sequences are then feature-encoded to obtain interactive behavior subsequences that characterize the interaction relationships between the two monitoring targets, such as mutual approach, relative motion, spatial intersection, approach or distance.

[0023] Furthermore, the method provided in the application embodiment, which performs cross-target correlation analysis on the spatiotemporal state sequences of the multiple monitoring targets and extracts the interaction behavior sub-sequences of each monitoring target within a preset time window, further includes: In the spatiotemporal state sequence, the spatial relationship changes between any two monitored targets are detected in real time. The spatial relationship changes include at least the evolution of distance, relative orientation, and the angle of motion trend. When the spatial relationship changes meet preset conditions, indicating that the two monitored targets enter a predefined interactive interest area, a preset time window is triggered. Within the preset time window, the spatiotemporal state sequences of the two monitored targets are extracted synchronously. Feature extraction and encoding are performed on the synchronously extracted spatiotemporal state sequences of the two monitored targets to generate an interactive behavior subsequence representing the interaction between the two.

[0024] In this embodiment, any two monitoring targets are processed in pairs according to the acquisition time in the spatiotemporal state sequence. Specifically, the position coordinates, motion vector, and spatial occupancy of the first monitoring target at the current acquisition time are first read from the spatiotemporal state sequence, and then the position coordinates, motion vector, and spatial occupancy of the second monitoring target at the same acquisition time are read. The position coordinates are represented by the spatial coordinates of the construction area, including horizontal and vertical coordinates; the motion vector includes horizontal and vertical motion components; and the spatial occupancy is represented by the target boundary area. After reading, the horizontal coordinates of the second monitoring target are subtracted from the horizontal coordinates of the first monitoring target to obtain the horizontal coordinate difference, and the vertical coordinates of the second monitoring target are subtracted from the vertical coordinates of the first monitoring target to obtain the vertical coordinate difference. The horizontal and vertical coordinate differences are squared respectively, and the two squared results are added together and then squared to obtain the distance between the two monitoring targets at the current acquisition time. Subsequently, the above distance calculation is repeated for multiple consecutive acquisition times according to the acquisition time sequence, and the distance of the later acquisition time is compared with the distance of the previous acquisition time to form the evolution result of the distance increasing, decreasing, or remaining constant over time.

[0025] When calculating the relative orientation, the position coordinates of the first monitoring target are used as the reference point, and the position coordinates of the second monitoring target are used as the comparison point. The relative orientation is determined based on the difference in the lateral and longitudinal coordinates of the second monitoring target relative to the first monitoring target. Specifically, when the absolute value of the lateral coordinate difference is less than a preset lateral deviation threshold and the longitudinal coordinate difference is greater than 0, the second monitoring target is determined to be in front of the first monitoring target; when the absolute value of the lateral coordinate difference is less than a preset lateral deviation threshold and the longitudinal coordinate difference is less than 0, the second monitoring target is determined to be behind the first monitoring target; when the absolute value of the longitudinal coordinate difference is less than a preset longitudinal deviation threshold and the lateral coordinate difference is greater than 0, the second monitoring target is determined to be to the right of the first monitoring target; when the absolute value of the longitudinal coordinate difference is less than a preset longitudinal deviation threshold and the lateral coordinate difference is less than 0, the second monitoring target is determined to be to the left of the first monitoring target; when the lateral coordinate difference is greater than 0 and the longitudinal coordinate difference is greater than 0, the second monitoring target is determined to be to the right front of the first monitoring target; when the lateral coordinate difference is less than 0 and the longitudinal coordinate difference is greater than 0, the second monitoring target is determined to be to the left front of the first monitoring target; when the lateral coordinate difference is greater than 0 and the longitudinal coordinate difference is less than 0, the second monitoring target is determined to be to the right rear of the first monitoring target; when the lateral coordinate difference is less than 0 and the longitudinal coordinate difference is less than 0, the second monitoring target is determined to be to the left rear of the first monitoring target. The above judgment is repeated for multiple consecutive collection times in chronological order, and the changes in relative orientation at each collection time are recorded to obtain the evolution of relative orientation.

[0026] When calculating the motion trend angle, the motion vectors of the first and second monitoring targets are read at the current acquisition time. For the motion vector of the first monitoring target, its lateral and longitudinal motion components are read, and the first motion direction is determined based on these components. For the motion vector of the second monitoring target, its lateral and longitudinal motion components are read, and the second motion direction is determined based on these components. Specifically, the direction angles of the two motion vectors relative to the lateral coordinate axis of the construction area are first determined, and then the absolute value of the difference between the two direction angles is taken. When the absolute value of the difference is greater than 180°, the absolute value of the difference is subtracted from 360° to obtain the final angle value. When the absolute value of the difference is less than or equal to 180°, the absolute value of the difference is used as the final angle value. The final angle value is the motion trend angle between the two monitoring targets at the current acquisition time. The above calculation is repeated for multiple consecutive acquisition times according to the acquisition time sequence, and the changes in the motion trend angle at each acquisition time are recorded to obtain the evolution of the motion trend angle.

[0027] After obtaining the evolution of distance, relative orientation, and the angle of motion trend, these are compared with preset conditions. In this process, it is first determined whether the distance between the two monitored targets is less than a preset distance threshold; if the distance is less than the preset distance threshold, it is further determined whether the relative orientation falls within a preset orientation range; if the relative orientation falls within the preset orientation range, it is further determined whether the angle of motion trend falls within a preset angle range. The preset angle ranges include the facing approach angle range, the intersecting approach angle range, and the parallel approach angle range. The facing approach angle range is used to determine whether the movement directions of the two monitored targets are relative; the intersecting approach angle range is used to determine whether the movement directions of the two monitored targets intersect; and the parallel approach angle range is used to determine whether the two monitored targets move in similar directions. When the distance, relative orientation, and angle of motion trend all meet the corresponding preset conditions, it is determined that the two monitored targets have entered a predefined interactive interest area.

[0028] When two monitored targets are determined to have entered a predefined interactive interest area, the acquisition time that first meets the preset conditions is determined as the trigger time, and this trigger time is used as the center point of the time window to trigger a preset time window. During this process, the start time of the preset time window is determined according to a preset backtracking duration, and the end time is determined according to a preset duration; for example, if the trigger time is 10 seconds, the backtracking duration is 3 seconds, and the duration is 5 seconds, then the preset time window is from 7 seconds to 15 seconds. Within the preset time window, using the acquisition time as the alignment benchmark, the category, position, attitude, motion vector, and spatial occupancy range under the same acquisition time are extracted from the spatiotemporal state sequences of the first and second monitored targets, respectively. When the acquisition times of the two spatiotemporal state sequences are not completely consistent, they are aligned using the acquisition time corresponding to the same video frame number, or the state record closest to the target acquisition time is selected for alignment. After alignment, the spatiotemporal state sequences of the two monitored targets synchronously extracted within the preset time window are obtained.

[0029] Finally, feature extraction and encoding are performed on the spatiotemporal state sequences of the two monitored targets. In this process, firstly, based on the positional changes of the two monitored targets within a preset time window, the shortest distance, average distance, and rate of change of distance between the two monitored targets are calculated to generate the spatiotemporal proximity features of the interaction behavior subsequence. Then, based on the motion vectors and continuous positional changes of the two monitored targets, the relative velocity, relative acceleration, and the degree of convergence or divergence of their trajectories are calculated to generate the relative motion features of the interaction behavior subsequence. Subsequently, based on the directional and angular relationships between the motion vector of one monitored target and the spatial occupancy of the other monitored target, the interaction geometric features in the interaction behavior subsequence are generated. Finally, the spatiotemporal proximity features, relative motion features, and interaction geometric features are combined and encoded according to the temporal order within the preset time window to form an interaction behavior subsequence that can characterize the interaction relationship between the two monitored targets.

[0030] Furthermore, in the method provided in the application embodiments, feature extraction and encoding are performed on the spatiotemporal state sequences of two monitored targets simultaneously to generate the interaction behavior subsequence representing the interaction between the two targets, which further includes: Based on the spatiotemporal state sequences of the two monitored targets, the shortest distance, average distance, and rate of change of distance between the two monitored targets within a preset time window are calculated to generate the spatiotemporal proximity characteristics of the interaction behavior subsequence. Based on the spatiotemporal state sequences of the two monitored targets, the relative velocity, relative acceleration, and degree of convergence or divergence of motion trajectories are calculated to generate the relative motion characteristics of the interaction behavior subsequence. Based on the spatiotemporal state sequences of the two monitored targets, the direction and angle of the motion vector of any monitored target relative to the spatial occupancy of the other monitored target are calculated to generate the interaction geometric characteristics in the interaction behavior subsequence. Based on the spatiotemporal proximity characteristics, the relative motion characteristics, and the interaction geometric characteristics, the interaction behavior subsequence is constructed.

[0031] In this embodiment, based on the spatiotemporal state sequence of two monitored targets extracted synchronously, the position coordinates of the first and second monitored targets at each acquisition time are read according to the acquisition time sequence within a preset time window. For any acquisition time, the horizontal coordinate of the second monitored target is first subtracted from the horizontal coordinate of the first monitored target to obtain the horizontal coordinate difference, and then the vertical coordinate of the second monitored target is subtracted from the vertical coordinate of the first monitored target to obtain the vertical coordinate difference. Subsequently, the horizontal coordinate difference and the vertical coordinate difference are squared respectively, and the two squared results are added together and then squared to obtain the distance between the two monitored targets at that acquisition time. The distances corresponding to all acquisition times within the preset time window are calculated in the above manner to form a distance sequence. The distance with the smallest value in the distance sequence is selected as the shortest distance between the two monitored targets. All distance values ​​in the distance sequence are added together and divided by the distance quantity to obtain the average distance between the two monitored targets. The distance of the later acquisition time is subtracted from the distance of the previous acquisition time, and then divided by the time interval between the two acquisition times to obtain the distance change rate corresponding to adjacent acquisition times. The shortest distance, average distance, and distance change rate corresponding to each collection time are organized to generate the spatiotemporal proximity characteristics of the interaction behavior subsequence. The shortest distance represents the closest interval between two monitoring targets within a preset time window, the average distance represents the overall interval level between two monitoring targets within a preset time window, and the distance change rate represents the speed and direction of the distance between two monitoring targets changing over time.

[0032] Based on the spatiotemporal state sequences of two monitored targets extracted synchronously, the motion vectors of the first and second monitored targets at each acquisition time are read. These motion vectors include lateral velocity components and longitudinal velocity components. For any acquisition time, the lateral velocity component of the second monitored target is subtracted from the lateral velocity component of the first monitored target to obtain the lateral relative velocity component; the longitudinal velocity component of the second monitored target is subtracted from the longitudinal velocity component of the first monitored target to obtain the longitudinal relative velocity component. The lateral and longitudinal relative velocity components are squared respectively, and the sum of the two squares is then taken as the square root to obtain the relative velocity of the two monitored targets at that acquisition time. Subsequently, the lateral relative velocity component of the later acquisition time is subtracted from the lateral relative velocity component of the earlier acquisition time, and divided by the time interval between the two acquisition times to obtain the lateral relative acceleration component; the longitudinal relative velocity component of the later acquisition time is subtracted from the longitudinal relative velocity component of the earlier acquisition time, and divided by the time interval between the two acquisition times to obtain the longitudinal relative acceleration component; the lateral and longitudinal relative acceleration components are then squared respectively, and the sum of the two squares is then taken as the square root to obtain the relative acceleration. To determine the degree of convergence or divergence in motion trajectories, the position coordinates of the first monitored target are connected in chronological order to form the first motion trajectory, and the position coordinates of the second monitored target are connected to form the second motion trajectory. The distance changes between the two monitored targets at adjacent acquisition times are compared. If the distance at a later acquisition time is less than the distance at a earlier acquisition time, it is recorded as convergence; if the distance at a later acquisition time is greater than the distance at a earlier acquisition time, it is recorded as divergence. The ratio of the number of convergences within a preset time window to the total number of comparisons is used as the degree of convergence, and the ratio of the number of divergences to the total number of comparisons is used as the degree of divergence. The relative velocity, relative acceleration, and the degree of convergence or divergence of the motion trajectories are then processed to generate the relative motion characteristics of the interactive behavior subsequence.

[0033] Based on the spatiotemporal state sequences of two monitored targets extracted simultaneously, the position and motion vector of any one monitored target at each acquisition time are read, and the spatial occupancy range of the other monitored target at the same acquisition time is also read. Taking the spatial occupancy range from the first monitored target to the second monitored target as an example, the representative point of the spatial occupancy range of the second monitored target is first determined; when the spatial occupancy range is a rectangular area, the coordinates of the left, right, upper, and lower boundaries of the rectangular area are read. The left and right boundary coordinates are added together and divided by 2 to obtain the lateral coordinates of the representative point, and the upper and lower boundary coordinates are added together and divided by 2 to obtain the longitudinal coordinates of the representative point; when the spatial occupancy range is a polygonal area, the lateral and longitudinal coordinates of each vertex of the polygonal area are read, and the lateral coordinates of all vertices are... The horizontal coordinates of the representative point are obtained by summing the vertical coordinates of all vertices and dividing by the number of vertices. The vertical coordinates of the representative point are then obtained by summing the vertical coordinates of all vertices and dividing by the number of vertices. When the spatial area is circular, the center coordinates are directly used as the representative point coordinates. When the spatial area is an irregular area represented by a pixel mask, the spatial coordinates of the construction area corresponding to all pixels within the mask area are read. The horizontal coordinates of all pixels are summed and divided by the number of pixels to obtain the horizontal coordinates of the representative point. The vertical coordinates of all pixels are summed and divided by the number of pixels to obtain the vertical coordinates of the representative point. After determining the representative point, the horizontal coordinates of the first monitoring target are subtracted from the horizontal coordinates of the representative point to obtain the horizontal component of the pointing direction. The vertical coordinates of the first monitoring target are subtracted from the vertical coordinates of the representative point to obtain the vertical component of the pointing direction. This gives the direction from the first monitoring target to the spatial area occupied by the second monitoring target. Next, read the lateral and longitudinal velocity components of the motion vector of the first monitoring target. Calculate the angle between the motion vector direction and the aforementioned pointing direction. First, multiply the lateral velocity component of the motion vector by the lateral component of the pointing direction to obtain the first product. Then, multiply the longitudinal velocity component of the motion vector by the longitudinal component of the pointing direction to obtain the second product. Add the first and second products to obtain the direction product. Then, square the lateral and longitudinal velocity components of the motion vector respectively, add the two squares and take the square root to obtain the length of the motion vector. Square the lateral and longitudinal components of the pointing direction respectively, add the two squares and take the square root to obtain the length of the pointing direction. Multiply the motion vector length by the pointing direction length to obtain the length product. Divide the direction product by the length product to obtain the cosine of the angle. Perform an inverse cosine operation on the cosine of the angle to obtain the angle between the motion vector of the first monitoring target and the spatial area occupied by the second monitoring target. The motion vector of the second monitoring target is calculated relative to the spatial occupancy of the first monitoring target using the same process, along with the direction and angle between them. These are then organized according to the acquisition time sequence to generate the interactive geometric features in the interactive behavior subsequence.

[0034] Finally, the spatiotemporal proximity features, relative motion features, and interactive geometric features are arranged according to the acquisition time sequence within a preset time window. For each acquisition time, the distance change rate, relative velocity, relative acceleration, convergence or divergence markers, and the direction and angle of the motion vector relative to the spatial occupancy of another monitored target are recorded as a set of features. Simultaneously, the shortest distance, average distance, convergence degree, and divergence degree corresponding to the preset time window are recorded as the overall feature record for that preset time window. The feature records of each acquisition time are combined with the overall feature record to construct an interactive behavior subsequence.

[0035] Step S300: Construct a risk behavior pattern library containing multiple risk primitive patterns, wherein each risk primitive pattern can define a combination of risk characteristics of one or more monitoring targets in spatiotemporal state sequences and / or interaction behavior subsequences.

[0036] In this embodiment, when constructing a risk behavior pattern library containing multiple risk primitive patterns, the causes of accidents and hazards are first deconstructed based on historical safety accident cases, hazard reports, and work safety procedures. The monitoring target types, spatial relationships, motion relationships, and temporal evolution relationships corresponding to different risk scenarios are extracted to form multiple basic risk scenarios. Then, for each basic risk scenario, at least one combination of target types, a set of spatiotemporal constraints, and a temporal evolution rule constituting the core risk are defined. The target type combination is used to limit the categories of monitoring targets participating in risk behavior; the spatiotemporal constraints are used to limit the position, distance, attitude, motion vector, and spatial occupancy relationship of the monitoring targets in the spatiotemporal state sequence and / or interactive behavior subsequence; and the temporal evolution rule is used to limit the order of change of risk behavior within a continuous time period. Finally, the target type combination, spatiotemporal constraints, and temporal evolution rule are bound together to form a risk primitive pattern. Multiple risk primitive patterns are stored in the risk behavior pattern library, enabling each risk primitive pattern to define a combination of risk characteristics of one or more monitoring targets in the spatiotemporal state sequence and / or interactive behavior subsequence.

[0037] Furthermore, the method provided in the application embodiments, which constructs a risk behavior pattern library containing multiple risk primitive patterns, also includes: Based on historical safety accident cases, hazard reports, and operational safety procedures, the causes of accidents and hazards are deconstructed to form multiple basic risk scenarios. For each basic risk scenario, at least one combination of target types, a set of spatiotemporal constraints, and a temporal evolution rule constituting its core risk are defined. The target type combination, the spatiotemporal constraints, and the temporal evolution rule are bound together to form a risk primitive pattern, which is then stored in the risk behavior pattern library.

[0038] Furthermore, the method provided in the application embodiments also includes: The risk primitive pattern includes at least one of the following constraints: constraint one, it must include at least two monitoring targets of a specific category; constraint two, the relative position, distance or motion envelope of the at least two monitoring targets in the spatiotemporal state sequence must satisfy a preset spatial relationship inequality; constraint three, the interaction state exhibited by the at least two monitoring targets in the interaction behavior subsequence must follow a preset order or prohibit the occurrence of a preset combination.

[0039] In this embodiment, based on historical safety accident cases, hazard reports, and operational safety procedures, the types of accidents, hazard types, operational processes, and violations related to the construction area are first organized. The location of the accident, the involved parties, the movement process of the parties, changes in distance between parties, changes in operational status, and the accident result recorded in each historical safety accident case and hazard report are extracted and correlated with the safety distances, prohibited areas, prohibited cross-operation relationships, equipment operation restrictions, and personnel operation requirements stipulated in the operational safety procedures. Through this correlation, the causes of accidents and hazards are broken down into risk factors that can be described by the monitoring targets, such as construction workers approaching operating machinery, hoisted objects passing over personnel, vehicles reversing and approaching personnel, and personnel entering areas with inadequate protection at edge openings, thus forming multiple basic risk scenarios. These basic risk scenarios, derived from the deconstruction of the causes of accidents and hazards, represent a specific source of construction safety risk.

[0040] For each basic risk scenario, define at least one combination of target types constituting its core risk, a set of spatiotemporal constraints, and a temporal evolution rule. The target type combination limits the categories of monitoring targets that must participate in risk assessment within that basic risk scenario, such as construction workers and transport vehicles, construction workers and hoisted objects, construction workers and adjacent areas, engineering machinery and protective facilities, etc. The spatiotemporal constraints limit the relative positions, distances, motion envelopes, postures, and spatial occupancy relationships of each monitoring target in the spatiotemporal state sequence within the target type combination. For example, the distance between two monitoring targets is less than a safe distance threshold; the position of construction workers falls within the motion envelope of machinery; the spatial occupancy of hoisted objects overlaps with the spatial occupancy of personnel in the vertical projection. The temporal evolution rule limits the temporal sequence of the risk formation process. For example, it may first occur when personnel enter the equipment's operating range, then when the equipment moves towards the personnel, and subsequently, the distance between them continuously decreases; or it may prohibit the interaction state combination of a vehicle reversing and personnel being within a safe distance behind the vehicle.

[0041] When defining the constraints for risk characteristic combinations, at least one constraint must be set according to the risk primitive model judgment. Constraint one requires the inclusion of at least two monitoring targets of a specific category. That is, the risk primitive model will only proceed to subsequent judgments if monitoring targets of the specified category simultaneously exist in the spatiotemporal state sequence and / or interactive behavior sub-sequence. Constraint two requires that the relative positions, distances, or motion envelopes of at least two monitoring targets in the spatiotemporal state sequence satisfy a preset spatial inequality. For example, the distance between the two monitoring targets must be less than a preset safety distance; the position coordinates of one monitoring target must be within the motion envelope boundary of the other monitoring target; or the distance between the spatial occupancy ranges of the two monitoring targets must be less than a preset distance threshold. Constraint three requires that the interactive states exhibited by at least two monitoring targets in the interactive behavior sub-sequence must follow a preset order or prohibit preset combinations. For example, the interactive states must occur in the order of approaching, entering a danger zone, and continuous approach; or the combination of personnel remaining in front of the equipment's direction of travel while the equipment continues to move is prohibited.

[0042] After defining the target type combination, spatiotemporal constraints, and temporal evolution rules, these three elements are bound together to form a risk primitive pattern. During binding, the target type combination is determined as the object matching condition of the risk primitive pattern, the spatiotemporal constraints as the spatial matching condition, and the temporal evolution rules as the temporal matching condition. These three types of conditions are then stored as risk feature combinations within the same risk primitive pattern. Subsequently, various risk primitive patterns generated from multiple basic risk scenarios are stored in a risk behavior pattern library, ensuring that each risk primitive pattern in the library corresponds to a specific cause of an accident or potential hazard.

[0043] Step S400: Perform multi-dimensional matching between the interaction behavior subsequence and the risk primitive pattern. When the matching degree exceeds the adaptive threshold, determine that a risk event instance has occurred.

[0044] In this embodiment, when performing multi-dimensional matching between the interaction behavior subsequence and the risk primitive pattern, firstly, feature vectors that characterize the interaction relationship between two monitoring targets are extracted from the interaction behavior subsequence. Then, the feature vectors are compared one by one with each feature dimension defined by the risk primitive pattern to be matched, and the conformity score of each feature dimension is calculated. Next, the dimensional weights of each feature dimension are dynamically adjusted in combination with the current construction stage, operation type, and environmental complexity, so that key risk features under different construction scenarios have corresponding weights in the matching process. Subsequently, the conformity scores of each feature dimension are weighted and summed according to the adjusted dimensional weights to obtain the total matching degree between the interaction behavior subsequence and the risk primitive pattern. Finally, the total matching degree is compared with the adaptive threshold calculated according to the current scenario. When the total matching degree exceeds the adaptive threshold, the matching is determined to be successful, and the corresponding risk event instance is generated.

[0045] Furthermore, in the method provided in the application embodiment, the interaction behavior subsequence is matched with the risk primitive pattern in multiple dimensions, and when the matching degree exceeds an adaptive threshold, a risk event instance is determined to have occurred. This further includes: The feature vectors of the interaction behavior subsequences are extracted and compared one by one with each feature dimension defined by the risk primitive pattern to be matched, and the conformity score of each feature dimension is calculated. The dimensional weights of each feature dimension are dynamically adjusted according to the current construction stage, operation type and environmental complexity. The conformity scores of each feature dimension are weighted and summed according to the adjusted dimensional weights to obtain the total matching degree between the interaction behavior subsequence and the risk primitive pattern. The total matching degree is compared with an adaptive threshold calculated according to the current scenario. When the total matching degree exceeds the adaptive threshold, the match is determined to be successful and a corresponding risk event instance is generated.

[0046] In this embodiment, feature vectors of interactive behavior subsequences are extracted and compared one by one with each feature dimension defined by the risk primitive pattern to be matched. When calculating the conformity score for each feature dimension, the shortest distance, average distance, distance change rate, relative speed, relative acceleration, degree of convergence or divergence of motion trajectory, and direction and angle of motion vector relative to the spatial area occupied by another monitored target are first read from the interactive behavior subsequence and arranged in a fixed order to form feature vectors. Then, the target type combination, distance threshold, speed threshold, acceleration threshold, convergence threshold, divergence threshold, pointing angle threshold, and time sequence evolution rules corresponding to the risk primitive pattern are read, and the conformity score is calculated item by item. When the target type combination is consistent, the conformity score for this dimension is 1; when inconsistent, it is 0. For the distance dimension, the score is calculated by dividing the distance threshold by the actual distance; a result greater than 1 is counted as 1. For the distance change rate dimension, the score is calculated by dividing the number of distance change rates less than 0 by the total number of distance change rates. For the relative velocity dimension, the score is calculated by dividing the actual relative velocity by the velocity threshold; a result greater than 1 is counted as 1. For the relative acceleration dimension, the score is calculated by dividing the actual relative acceleration by the acceleration threshold; a result greater than 1 is counted as 1. For the convergence or divergence of motion trajectories dimension, the score is calculated by dividing the actual convergence or divergence by the corresponding threshold; a result greater than 1 is counted as 1. For the direction and angle dimension, the score is calculated by dividing the pointing angle threshold by the actual angle; a result greater than 1 is counted as 1. For the temporal evolution dimension, the score is calculated by dividing the number of states conforming to the risk primitive pattern's prescribed order by the total number of prescribed states.

[0047] When dynamically adjusting the dimensional weights of each feature dimension based on the current construction stage, operation type, and environmental complexity, initial weights are first set for the target type combination dimension, distance dimension, distance change rate dimension, relative velocity dimension, relative acceleration dimension, convergence or divergence of motion trajectory dimension, direction and angle dimension, and time series evolution dimension. Then, the values ​​corresponding to the current construction stage, operation type, and environmental complexity are read, and the initial weights are adjusted accordingly. When the current construction stage is a high-risk stage, the initial weights of the distance dimension, distance change rate dimension, and time series evolution dimension are multiplied by 1.2 respectively; when the current construction stage is a normal stage, the initial weights of each feature dimension remain unchanged. When the operation type is vehicle transportation, the weights of the relative velocity dimension and direction and angle dimension are multiplied by 1.2 respectively; when the operation type is hoisting, the weights of the direction and angle dimension and time series evolution dimension are multiplied by 1.2 respectively; when the operation type is edge work or high-altitude work, the weights of the target type combination dimension, distance dimension, and time series evolution dimension are multiplied by 1.2 respectively. Environmental complexity is calculated using the number of monitored targets, the number of overlapping operations, and the number of obstructed areas. Specifically, within the same monitoring cycle of the current construction area, the total number of continuously tracked monitored targets is counted, denoted as the target number; the number of different types of operations existing simultaneously within the same spatial area is counted, denoted as the overlapping operation number; and the number of areas in the video feed where the monitored target's spatial area is partially covered due to obstruction by equipment, components, fences, or scaffolding is counted, denoted as the obstructed area number. The target number is divided by a preset target number threshold to obtain the target number ratio; the overlapping operation number is divided by a preset overlapping operation number threshold to obtain the overlapping operation ratio; and the obstructed area number is divided by a preset obstructed area number threshold to obtain the obstructed area ratio. The target number ratio, overlapping operation ratio, and obstructed area ratio are added together and divided by 3 to obtain the environmental complexity value; when the environmental complexity value is greater than 1, it is counted as 1. When the environmental complexity value is greater than or equal to 0.7, the environmental complexity is considered high; when the environmental complexity value is less than 0.7, the environmental complexity is considered low. When the environment is complex, the weights of the target type combination dimension, distance dimension, direction and angle dimension, and time sequence evolution dimension are multiplied by 1.1 respectively. After adjustment, the weights of all dimensions involved in matching are added together to obtain the total weights. Then, each adjusted dimension weight is divided by the total weights to obtain the normalized dimension weights.

[0048] Subsequently, the consistency scores of each feature dimension are weighted and summed according to the adjusted dimension weights to obtain the total matching degree between the interaction behavior subsequence and the risk primitive pattern. In this process, the consistency score of each feature dimension is multiplied by its corresponding normalized dimension weight to obtain the weighted score for that feature dimension. Then, the weighted scores of the target type combination dimension, distance dimension, distance change rate dimension, relative velocity dimension, relative acceleration dimension, convergence or divergence of motion trajectory dimension, direction and angle dimension, and time series evolution dimension are added together to obtain the total matching degree. If the risk primitive pattern does not define a certain feature dimension, that feature dimension does not participate in the consistency score calculation and weighted summation; only the feature dimensions participating in the matching are weighted and normalized, so that the total matching degree is calculated from the feature dimensions actually defined in the risk primitive pattern.

[0049] Finally, the total matching degree is compared with the adaptive threshold calculated based on the current scenario. When the total matching degree exceeds the adaptive threshold, the match is considered successful, and a corresponding risk event instance is generated. Specifically, the basic threshold corresponding to the risk primitive pattern is first read, and then the basic threshold is adjusted according to the current construction stage, operation type, and environmental complexity. During adjustment, the threshold deduction for high-risk construction stages is 0.05, the threshold deduction for hazardous operation types is 0.05, and the threshold deduction for high environmental complexity is 0.03; for ordinary construction stages, ordinary operation types, and low environmental complexity, the threshold increase is 0.03. The adaptive threshold is obtained by subtracting the corresponding threshold deduction or adding the corresponding threshold increase to the basic threshold. For example, if the basic threshold is 0.80, and the current construction stage is a high-risk construction stage, the operation type is a hazardous operation type, and the environmental complexity is high, the adaptive threshold is 0.80 minus 0.05 minus 0.05 minus 0.03, resulting in 0.67. The total matching degree is then compared with the adaptive threshold. When the total matching degree is greater than the adaptive threshold, it is determined that the interaction behavior subsequence is successfully matched with the risk primitive pattern, and a corresponding risk event instance is generated. The risk event instance records the risk primitive pattern identifier, the monitoring target identifier participating in the matching, the occurrence time, the occurrence location, the total matching degree, the conformity score of each feature dimension, the dimension weight of each feature dimension, and the corresponding risk type.

[0050] Step S500: Based on all risk event instances identified within a preset period, calculate the risk exposure density distribution and behavioral safety deviation index, and generate a dynamic safety situation map of the construction area based on the risk exposure density distribution and behavioral safety deviation index.

[0051] In this embodiment, when calculating the risk exposure density distribution based on all risk event instances identified within a preset period, the construction area is first divided into multiple spatial grids in space, and the risk event instances within the preset period are assigned to the corresponding spatial grids according to their occurrence locations. Then, the frequency and duration of occurrence of risk event instances in each spatial grid are counted, and the risk exposure density of each spatial grid is calculated by combining the preset risk level weights corresponding to each risk event instance. Finally, the risk exposure densities of all spatial grids are arranged according to their spatial locations in the construction area to generate a risk exposure density distribution that reflects the degree of risk concentration at different spatial locations in the construction area.

[0052] Based on all risk event instances identified within a preset period, when calculating the behavioral safety deviation index, the target assessment object is first determined, and the frequency of risk event instances corresponding to various risk primitive patterns triggered by the target assessment object within the preset period is statistically analyzed to obtain the actual trigger frequency distribution. Then, the baseline trigger frequency distribution corresponding to various risk primitive patterns of the target assessment object during historical safe operations or under the safety specifications of similar operations is obtained. Subsequently, the difference measure between the actual trigger frequency distribution and the baseline trigger frequency distribution is calculated, and this difference measure is used as the behavioral safety deviation index of the target assessment object within the preset period to characterize the degree of deviation of the current risk behavior triggering situation of the target assessment object from the baseline safety state.

[0053] Finally, based on the risk exposure density distribution and behavioral safety deviation index, a dynamic safety situation map of the construction area is generated. In this process, the risk exposure density distribution is first overlaid on a 3D digital map of the construction area in the form of a heat map, according to the positional relationship of each spatial grid within the construction area, forming a spatial risk layer reflecting the degree of risk aggregation at different spatial locations. Then, the behavioral safety deviation index is associated with corresponding work teams, job types, or individuals in the form of labels, forming a behavioral safety layer reflecting the behavioral safety status of different assessment objects. Subsequently, the spatial risk layer and the behavioral safety layer are merged and linked to a timeline, unifying the spatial risk information and behavioral safety information at different times to generate a dynamic safety situation map of the construction area.

[0054] Furthermore, in the method provided in the application embodiments, calculating the risk exposure density distribution based on all risk event instances determined within a preset period further includes: The construction area is divided into multiple spatial grids. Risk event instances occurring in each spatial grid within a preset period are counted. Based on the frequency of occurrence, duration, and preset risk level weight of the risk event instances, the risk exposure density of each grid is calculated. A risk exposure density distribution is generated based on the risk exposure density of all spatial grids.

[0055] In this embodiment, when spatially dividing the construction area into grids, the spatial boundary coordinates of the construction area are first read to determine the minimum and maximum coordinates in the horizontal direction, as well as the minimum and maximum coordinates in the vertical direction. Then, the area is divided at equal intervals along the horizontal and vertical directions according to a preset grid size, generating multiple spatial grids. Each spatial grid records the grid number, left boundary coordinates, right boundary coordinates, lower boundary coordinates, upper boundary coordinates, and grid area. The grid area is obtained by multiplying the horizontal length and vertical width of the spatial grid. Through this grid division, the location of each risk event instance within the construction area can be assigned to the corresponding spatial grid based on its coordinate range.

[0056] Next, when counting risk event instances occurring in each spatial grid within a preset period, all risk event instances identified within the preset period are first read, and the occurrence location, occurrence time, end time, risk primitive mode type, and preset risk level weight of each risk event instance are extracted. For any risk event instance, the horizontal coordinate of its occurrence location is compared with the left and right boundary coordinates of each spatial grid, and the vertical coordinate of its occurrence location is compared with the lower and upper boundary coordinates of each spatial grid. When the horizontal coordinate of the occurrence location is greater than or equal to the left boundary coordinate and less than the right boundary coordinate of a spatial grid, and the vertical coordinate of the occurrence location is greater than or equal to the lower boundary coordinate and less than the upper boundary coordinate of a spatial grid, the risk event instance is included in that spatial grid. After all risk event instances are assigned, the number of risk event instances included in each spatial grid is counted to obtain the occurrence frequency of risk event instances in each spatial grid.

[0057] The risk exposure density of each grid is then calculated. First, the duration of each risk event instance within the spatial grid is calculated, obtained by subtracting the occurrence time from the end time of the risk event instance. When a risk event instance has only a single occurrence time, a data collection interval is used as the duration. Then, the preset risk level weight corresponding to the risk event instance is read. The preset risk level weight is set corresponding to the risk primitive mode type; for example, different preset risk level weights are assigned for situations such as construction workers entering the equipment operating range, vehicles reversing and approaching construction workers, and hoisted objects passing over construction workers. For each risk event instance within the same spatial grid, its duration is multiplied by the preset risk level weight to obtain the risk exposure value of a single risk event instance. The risk exposure values ​​of all risk event instances within the spatial grid are summed to obtain the total risk exposure value of the spatial grid. Finally, the total risk exposure value is divided by the grid area of ​​the spatial grid to obtain the risk exposure density of the spatial grid.

[0058] Finally, when generating the risk exposure density distribution based on the risk exposure densities of all spatial grids, the risk exposure density of each spatial grid is stored in correspondence with its grid number and grid boundary coordinates. The risk exposure density of each spatial grid is then filled into the corresponding spatial location according to the horizontal and vertical arrangement order of the spatial grids within the construction area. If no risk event instance is recorded for a spatial grid within a preset period, the risk exposure density of that spatial grid is set to 0. If multiple risk event instances are recorded within a spatial grid, the risk exposure value of each risk event instance is calculated separately, summed, and then divided by the grid area. After assigning values ​​to all spatial grids, a risk exposure density distribution covering the construction area is generated.

[0059] Furthermore, the method provided in the application embodiments, which calculates the behavioral safety deviation index based on all risk event instances determined within a preset period, further includes: For the target assessment object, the frequency of risk event instances corresponding to various risk primitive patterns triggered within the preset period is statistically analyzed to obtain the actual trigger frequency distribution; the baseline trigger frequency distribution of various risk primitive patterns of the target assessment object during historical safe operation or under the safety specifications of similar operations is obtained; the difference metric between the actual trigger frequency distribution and the baseline trigger frequency distribution is calculated, and the difference metric is the behavioral safety deviation index of the assessment object within the preset period.

[0060] In this embodiment, during the calculation of the behavioral safety deviation index for the target assessment object, the work group, job type, or individual corresponding to the target assessment object is first determined, and records associated with the target assessment object are read from all risk event instances identified within a preset period. The association relationship is determined based on the monitoring target identifier, work group identifier, job type identifier, or individual identifier recorded in the risk event instance; for example, if the monitoring target identifier matched in a certain risk event instance corresponds to a certain construction worker, then the risk event instance is counted as the individual, job type, or work group to which the construction worker belongs. After the association is completed, the risk event instances are classified according to the risk element pattern type, the number of risk event instances corresponding to the same risk element pattern is accumulated, and the accumulated number of each type of risk element pattern is divided by the statistical duration of the preset period to obtain the trigger frequency of each type of risk element pattern within the preset period; then, the trigger frequencies are arranged in a fixed order according to the risk element patterns to obtain the actual trigger frequency distribution.

[0061] After obtaining the actual trigger frequency distribution, the baseline trigger frequency distribution corresponding to the target assessment object is then obtained. For target assessment objects with historical safety operation records, risk event instance records during their historical safety operation periods are read and classified and statistically analyzed according to the same risk primitive pattern type as the actual trigger frequency distribution. The trigger count of each risk primitive pattern during the historical safety operation period is divided by the statistical duration of the historical safety operation period to obtain the corresponding baseline trigger frequency. For target assessment objects lacking historical safety operation records, the allowed trigger count or allowed trigger frequency corresponding to each risk primitive pattern in the safety specifications for similar operations is read. When the safety specifications provide an allowed trigger count, the allowed trigger count is divided by the corresponding statistical duration of the specification to obtain the baseline trigger frequency. When the safety specifications provide an allowed trigger frequency, this allowed trigger frequency is directly used as the baseline trigger frequency. The baseline trigger frequencies of each risk primitive pattern are arranged in the same order as the actual trigger frequency distribution to obtain the baseline trigger frequency distribution.

[0062] After determining the actual trigger frequency distribution and the benchmark trigger frequency distribution, they are aligned item by item according to the same risk element pattern type, so that each risk element pattern corresponds to an actual trigger frequency and a benchmark trigger frequency. Then, for each risk element pattern, the actual trigger frequency is subtracted from the benchmark trigger frequency to obtain the frequency difference; the absolute value of the frequency difference is taken to obtain the deviation value of that risk element pattern; the preset risk level weight corresponding to that risk element pattern is then read, and the deviation value is multiplied by the preset risk level weight to obtain the weighted deviation value of that risk element pattern; the weighted deviation values ​​of all risk element patterns are summed to obtain the difference measure between the actual trigger frequency distribution and the benchmark trigger frequency distribution. This difference measure is determined as the behavioral safety deviation index of the target assessment object within a preset period.

[0063] Furthermore, in the method provided in the application embodiments, generating a dynamic safety situation map of the construction area based on the risk exposure density distribution and behavioral safety deviation index further includes: The risk exposure density distribution is overlaid as a heat map on a three-dimensional digital map of the construction area to form a spatial risk layer; the behavioral safety deviation index is labeled and associated with the corresponding work team, type of work or individual to form a behavioral safety layer; the spatial risk layer and the behavioral safety layer are merged and associated with a time axis to generate a dynamic safety situation map.

[0064] In this embodiment, when overlaying the risk exposure density distribution as a heatmap onto a 3D digital map of the construction area, the 3D digital map of the construction area is first read, and the spatial coordinate system, floor structure, work area boundary, passage area boundary, and equipment layout are extracted from the 3D digital map. Then, the grid number, grid boundary coordinates, and risk exposure density value corresponding to each spatial grid in the risk exposure density distribution are read, and the boundary coordinates of each spatial grid are converted to the spatial coordinate system of the 3D digital map, so that the spatial grid has a corresponding display position in the 3D digital map. Subsequently, a preset heatmap color mapping table is read. The preset heatmap color mapping table includes risk exposure density ranges and corresponding display colors; for example, a risk exposure density of 0 corresponds to green, a risk exposure density greater than 0 and less than a preset first density threshold corresponds to light yellow, a risk exposure density greater than or equal to the first density threshold and less than a preset second density threshold corresponds to orange, and a risk exposure density greater than or equal to the second density threshold corresponds to red. The corresponding display color is found based on the risk exposure density value of each spatial grid, and this display color is filled into the corresponding spatial grid area in the 3D digital map, forming a spatial risk layer overlaid on the 3D digital map of the construction area in the form of a heatmap.

[0065] Next, when associating the behavioral safety deviation index with corresponding work teams, job types, or individuals in the form of tags, the calculation results of the behavioral safety deviation index are first read, along with the identification information and spatial association information of the corresponding target assessment object. The target assessment object includes work teams, job types, or individuals; the identification information includes the work team identifier, job type identifier, or individual identifier; and the spatial association information includes the work area coordinates of the work team, the work area coordinates of the job type, or the current location coordinates of the individual. Then, tag content is generated according to preset tag fields. The tag content includes the target assessment object identifier, the behavioral safety deviation index value, and the corresponding time. For work teams, the tag is bound to the center coordinates of the work area of ​​that work team; for job types, the tag is bound to the center coordinates of the work area of ​​that job type; and for individuals, the tag is bound to the current location coordinates of that individual. After binding is complete, each tag is loaded into its corresponding position on the 3D digital map, forming a behavioral safety layer.

[0066] When merging the spatial risk layer and the behavioral safety layer, both layers are first loaded into the same 3D digital map coordinate system and overlaid according to their display order. The spatial risk layer is set as the bottom layer, and the behavioral safety layer as the top layer, ensuring that the heatmap area corresponding to the risk exposure density and the label corresponding to the behavioral safety deviation index are simultaneously displayed on the 3D digital map of the construction area. Then, according to a preset statistical time interval, a time stamp is set for each group of spatial risk layers and behavioral safety layers, and layers with the same time stamp are bound together as a single layer data set; for example, a layer data set is generated per hour, per shift, or per day. The layer data sets are then written into the timeline in chronological order, ensuring that each time node on the timeline corresponds to a single spatial risk layer and a single behavioral safety layer.

[0067] When generating the dynamic safety situation map, the layer data corresponding to each time node is read sequentially along the timeline, and the corresponding spatial risk layer and behavioral safety layer are loaded into the 3D digital map. When a time node changes, the heatmap colors, spatial risk layers, and behavioral safety layers in the 3D digital map are updated synchronously. If the risk exposure density of a spatial grid corresponding to a certain time node changes, the display color of that spatial grid is re-determined according to the preset heatmap color mapping table. If the behavioral safety deviation index of a certain target assessment object changes, the index value in its label is updated. Through the above layer loading, color mapping, label binding, and timeline association, a dynamic safety situation map of the construction area is generated.

[0068] In summary, the embodiments of this application have at least the following technical effects: This application acquires and analyzes frame-by-frame surveillance video streams of the construction area over a continuous time period, identifies and continuously tracks multiple monitoring targets, and generates a spatiotemporal state sequence for each target, including its category, location, attitude, motion vector, and spatial occupancy. It then performs cross-target correlation analysis on the spatiotemporal state sequences of the multiple monitoring targets, extracting interactive behavior subsequences of each target within a preset time window. These interactive behavior subsequences include at least the interaction relationship between two monitoring targets. A risk behavior pattern library containing multiple risk primitive patterns is constructed, where each risk primitive pattern can define a combination of risk characteristics of one or more monitoring targets in the spatiotemporal state sequence and / or interactive behavior subsequences. The interactive behavior subsequences are matched with the risk primitive patterns in multiple dimensions; when the matching degree exceeds an adaptive threshold, a risk event instance is determined. Based on all risk event instances determined within a preset period, a risk exposure density distribution and a behavioral safety deviation index are calculated, and a dynamic safety situation map of the construction area is generated based on these indicators. This invention addresses the technical problem of the difficulty in automatically identifying and quantifying potential risk interactions within construction areas in existing technologies. By continuously tracking multiple monitoring targets in a surveillance video stream and determining risk event instances based on the matching of interaction behaviors between targets with risk primitive patterns, it achieves the technical effect of early detection of potential safety risks in construction areas and quantitative presentation of risk status.

[0069] Example 2 is based on the same inventive concept as the target recognition-based construction area safety monitoring method in the foregoing examples, such as... Figure 2 As shown, this application provides a construction area safety monitoring system based on target recognition. The system and method embodiments in this application are based on the same inventive concept. The system includes: The target identification module 11 is used to acquire and analyze the monitoring video stream of the construction area frame by frame over a continuous time period, identify multiple monitoring targets for continuous tracking, and generate a spatiotemporal state sequence for each monitoring target, including its category, location, attitude, motion vector, and spatial occupancy range. The association analysis module 12 is used to perform cross-target association analysis on the spatiotemporal state sequences of the multiple monitoring targets, extract the interaction behavior subsequences of each monitoring target within a preset time window, and the interaction behavior subsequences include at least the interaction relationship between two monitoring targets. The pattern library construction module 13 is used to construct a risk behavior pattern library containing multiple risk primitive patterns, wherein each risk primitive pattern can define a combination of risk features of one or more monitoring targets in the spatiotemporal state sequence and / or interaction behavior subsequences. The matching module 14 is used to perform multi-dimensional matching between the interaction behavior subsequences and the risk primitive patterns, and when the matching degree exceeds an adaptive threshold, a risk event instance is determined to have occurred. The calculation module 15 is used to calculate the risk exposure density distribution and behavioral safety deviation index based on all risk event instances determined within a preset period, and generate a dynamic safety situation map of the construction area based on the risk exposure density distribution and behavioral safety deviation index.

[0070] Furthermore, the system is also used to implement the following functions: In the spatiotemporal state sequence, the spatial relationship changes between any two monitored targets are detected in real time. The spatial relationship changes include at least the evolution of distance, relative orientation, and the angle of motion trend. When the spatial relationship changes meet preset conditions, indicating that the two monitored targets enter a predefined interactive interest area, a preset time window is triggered. Within the preset time window, the spatiotemporal state sequences of the two monitored targets are extracted synchronously. Feature extraction and encoding are performed on the synchronously extracted spatiotemporal state sequences of the two monitored targets to generate an interactive behavior subsequence representing the interaction between the two.

[0071] Furthermore, the system is also used to implement the following functions: Based on the spatiotemporal state sequences of the two monitored targets, the shortest distance, average distance, and rate of change of distance between the two monitored targets within a preset time window are calculated to generate the spatiotemporal proximity characteristics of the interaction behavior subsequence. Based on the spatiotemporal state sequences of the two monitored targets, the relative velocity, relative acceleration, and degree of convergence or divergence of motion trajectories are calculated to generate the relative motion characteristics of the interaction behavior subsequence. Based on the spatiotemporal state sequences of the two monitored targets, the direction and angle of the motion vector of any monitored target relative to the spatial occupancy of the other monitored target are calculated to generate the interaction geometric characteristics in the interaction behavior subsequence. Based on the spatiotemporal proximity characteristics, the relative motion characteristics, and the interaction geometric characteristics, the interaction behavior subsequence is constructed.

[0072] Furthermore, the system is also used to implement the following functions: Based on historical safety accident cases, hazard reports, and operational safety procedures, the causes of accidents and hazards are deconstructed to form multiple basic risk scenarios. For each basic risk scenario, at least one combination of target types, a set of spatiotemporal constraints, and a temporal evolution rule constituting its core risk are defined. The target type combination, the spatiotemporal constraints, and the temporal evolution rule are bound together to form a risk primitive pattern, which is then stored in the risk behavior pattern library.

[0073] Furthermore, the system is also used to implement the following functions: The risk primitive pattern includes at least one of the following constraints: constraint one, it must include at least two monitoring targets of a specific category; constraint two, the relative position, distance or motion envelope of the at least two monitoring targets in the spatiotemporal state sequence must satisfy a preset spatial relationship inequality; constraint three, the interaction state exhibited by the at least two monitoring targets in the interaction behavior subsequence must follow a preset order or prohibit the occurrence of a preset combination.

[0074] Furthermore, the system is also used to implement the following functions: The feature vectors of the interaction behavior subsequences are extracted and compared one by one with each feature dimension defined by the risk primitive pattern to be matched, and the conformity score of each feature dimension is calculated. The dimensional weights of each feature dimension are dynamically adjusted according to the current construction stage, operation type and environmental complexity. The conformity scores of each feature dimension are weighted and summed according to the adjusted dimensional weights to obtain the total matching degree between the interaction behavior subsequence and the risk primitive pattern. The total matching degree is compared with an adaptive threshold calculated according to the current scenario. When the total matching degree exceeds the adaptive threshold, the match is determined to be successful and a corresponding risk event instance is generated.

[0075] Furthermore, the system is also used to implement the following functions: The construction area is divided into multiple spatial grids. Risk event instances occurring in each spatial grid within a preset period are counted. Based on the frequency of occurrence, duration, and preset risk level weight of the risk event instances, the risk exposure density of each grid is calculated. A risk exposure density distribution is generated based on the risk exposure density of all spatial grids.

[0076] Furthermore, the system is also used to implement the following functions: For the target assessment object, the frequency of risk event instances corresponding to various risk primitive patterns triggered within the preset period is statistically analyzed to obtain the actual trigger frequency distribution; the baseline trigger frequency distribution of various risk primitive patterns of the target assessment object during historical safe operation or under the safety specifications of similar operations is obtained; the difference metric between the actual trigger frequency distribution and the baseline trigger frequency distribution is calculated, and the difference metric is the behavioral safety deviation index of the assessment object within the preset period.

[0077] Furthermore, the system is also used to implement the following functions: The risk exposure density distribution is overlaid as a heat map on a three-dimensional digital map of the construction area to form a spatial risk layer; the behavioral safety deviation index is labeled and associated with the corresponding work team, type of work or individual to form a behavioral safety layer; the spatial risk layer and the behavioral safety layer are merged and associated with a time axis to generate a dynamic safety situation map.

[0078] It should be noted that the order of the embodiments described above is for descriptive purposes only and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this specification. The processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some implementations, multitasking and parallel processing are possible or may be advantageous.

[0079] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for safety monitoring of construction areas based on target recognition, characterized in that, The method includes: The system acquires and analyzes the monitoring video stream of the construction area frame by frame over a continuous time period, identifies multiple monitoring targets for continuous tracking, and generates a spatiotemporal state sequence for each monitoring target, including its category, location, attitude, motion vector, and spatial occupancy range. Cross-target correlation analysis is performed on the spatiotemporal state sequences of the multiple monitoring targets to extract the interaction behavior subsequences of each monitoring target within a preset time window. The interaction behavior subsequences include at least the interaction relationship between two monitoring targets. Construct a risk behavior pattern library containing multiple risk primitive patterns, wherein each risk primitive pattern can define a combination of risk characteristics of one or more monitoring targets in spatiotemporal state sequences and / or interaction behavior subsequences; The interaction behavior subsequence is matched with the risk primitive pattern in multiple dimensions. When the matching degree exceeds the adaptive threshold, a risk event instance is determined to have occurred. Based on all risk event instances identified within a preset period, the risk exposure density distribution and behavioral safety deviation index are calculated, and a dynamic safety situation map of the construction area is generated based on the risk exposure density distribution and behavioral safety deviation index.

2. The construction area safety monitoring method based on target recognition as described in claim 1, characterized in that, Cross-target correlation analysis is performed on the spatiotemporal state sequences of the multiple monitoring targets to extract the interaction behavior subsequences of each monitoring target within a preset time window, including: In the spatiotemporal state sequence, the changes in the spatial relationship between any two monitored targets are detected in real time. The changes in the spatial relationship include at least the evolution of distance, relative orientation, and the angle between the motion trends. When the change in spatial relationship meets the preset conditions, indicating that the two monitoring targets enter the predefined interactive attention area, a preset time window is triggered, and within the preset time window, the spatiotemporal state sequence of the two monitoring targets is extracted synchronously. Feature extraction and encoding are performed on the spatiotemporal state sequences of two monitored targets simultaneously to generate interactive behavior subsequences that characterize their interaction.

3. The construction area safety monitoring method based on target recognition as described in claim 1, characterized in that, Feature extraction and encoding are performed on the spatiotemporal state sequences of two monitored targets simultaneously to generate the interaction behavior subsequence representing their interaction, including: Based on the spatiotemporal state sequences of the two monitored targets, the shortest distance, average distance, and distance change rate between the two monitored targets within a preset time window are calculated to generate the spatiotemporal proximity characteristics of the interaction behavior subsequence. Based on the spatiotemporal state sequences of the two monitored targets, the relative velocity, relative acceleration, and degree of convergence or divergence of motion trajectories are calculated to generate the relative motion characteristics of the interaction behavior subsequence. Based on the spatiotemporal state sequences of the two monitored targets, the direction and angle of the motion vector of any monitored target relative to the spatial occupancy of the other monitored target are calculated to generate the interactive geometric features in the interactive behavior subsequence; Based on the spatiotemporal proximity features, the relative motion features, and the interactive geometric features, the interactive behavior subsequence is constructed.

4. The construction area safety monitoring method based on target recognition as described in claim 1, characterized in that, Construct a risk behavior pattern library containing multiple risk primitive patterns, including: Based on historical safety accident cases, hazard reports, and operational safety procedures, the causes of accidents and hazards are deconstructed to form multiple basic risk scenarios; For each basic risk scenario, define at least one combination of target types, a set of spatiotemporal constraints, and a temporal evolution rule that constitute its core risk. The target type combination, the spatiotemporal constraints, and the temporal evolution rules are bound together to form a risk primitive pattern, which is then stored in the risk behavior pattern library.

5. The construction area safety monitoring method based on target recognition as described in claim 4, characterized in that, The constraints on the risk feature combinations included in the risk primitive pattern include at least one of the following: Constraint 1: It must include at least two monitoring targets of a specific category; Constraint 2: The relative positions, distances, or motion envelopes of the at least two monitored targets in the spatiotemporal state sequence must satisfy a preset spatial inequality. Constraint 3: The interaction states exhibited by the at least two monitored targets in the interaction behavior subsequence must follow a preset order or a preset combination is prohibited.

6. The construction area safety monitoring method based on target recognition as described in claim 1, characterized in that, The interaction behavior subsequence is matched with the risk primitive pattern in multiple dimensions. When the matching degree exceeds an adaptive threshold, a risk event instance is determined to have occurred, including: Extract the feature vector of the interaction behavior subsequence and compare it one by one with each feature dimension defined by the risk primitive pattern to be matched, and calculate the conformity score of each feature dimension. The dimensional weights of each feature dimension are dynamically adjusted based on the current construction stage, operation type, and environmental complexity. The consistency scores of each feature dimension are weighted and summed according to the adjusted dimensional weights to obtain the total matching degree between the interaction behavior subsequence and the risk primitive pattern. The total matching degree is compared with an adaptive threshold calculated based on the current scenario. When the total matching degree exceeds the adaptive threshold, the matching is determined to be successful, and a corresponding risk event instance is generated.

7. The construction area safety monitoring method based on target recognition as described in claim 1, characterized in that, Based on all risk event instances identified within a preset period, the risk exposure density distribution is calculated, including: The construction area is spatially divided into grids to generate multiple spatial grids; Within a preset period, risk event instances occurring in each spatial grid are counted. Based on the frequency of occurrence, duration, and preset risk level weight of the risk event instances, the risk exposure density of each grid is calculated, and a risk exposure density distribution is generated based on the risk exposure density of all spatial grids.

8. The construction area safety monitoring method based on target recognition as described in claim 7, characterized in that, Based on all risk event instances identified within a preset period, a behavioral safety deviation index is calculated, including: For the target assessment object, the frequency of risk event instances corresponding to various risk primitive patterns triggered within the preset period is statistically analyzed to obtain the actual trigger frequency distribution; Obtain the baseline trigger frequency distribution of various risk primitive patterns for the target assessment object during historical safe operations or under the safety specifications of similar operations; Calculate the difference measure between the actual trigger frequency distribution and the benchmark trigger frequency distribution. The difference measure is the behavioral safety deviation index of the evaluation object within the preset period.

9. The construction area safety monitoring method based on target recognition as described in claim 1, characterized in that, Based on the risk exposure density distribution and behavioral safety deviation index, a dynamic safety situation map of the construction area is generated, including: The risk exposure density distribution is overlaid in the form of a heat map onto a three-dimensional digital map of the construction area to form a spatial risk layer; The behavioral safety deviation index is labeled and associated with the corresponding work group, job type or individual to form a behavioral safety layer; By merging the spatial risk layer and the behavioral safety layer, and associating them with a timeline, a dynamic security situation map is generated.

10. A construction area safety monitoring system based on target recognition, characterized in that, The system is used to execute a construction area safety monitoring method based on target recognition as described in any one of claims 1-9, and the system includes: The target recognition module is used to acquire the monitoring video stream of the construction area in a continuous time period, perform frame-by-frame analysis, identify multiple monitoring targets for continuous tracking, and generate a spatiotemporal state sequence for each monitoring target, including its category, location, attitude, motion vector and spatial occupancy range. The correlation analysis module is used to perform cross-target correlation analysis on the spatiotemporal state sequences of the multiple monitoring targets, and extract the interaction behavior subsequences of each monitoring target within a preset time window. The interaction behavior subsequences include at least the interaction relationship between two monitoring targets. The pattern library construction module is used to build a risk behavior pattern library containing multiple risk primitive patterns. Each risk primitive pattern can define a combination of risk characteristics of one or more monitoring targets in spatiotemporal state sequences and / or interaction behavior subsequences. The matching module is used to perform multi-dimensional matching between the interaction behavior subsequence and the risk primitive pattern. When the matching degree exceeds the adaptive threshold, a risk event instance is determined to have occurred. The calculation module is used to calculate the risk exposure density distribution and behavioral safety deviation index based on all risk event instances identified within a preset period, and to generate a dynamic safety situation map of the construction area based on the risk exposure density distribution and behavioral safety deviation index.

Images