Multi-view cooperative job behavior risk assessment method and system thereof

Abstract

The application discloses a multi-viewpoint cooperative work behavior risk assessment method and system, relates to the technical field of behavior risk assessment, and comprises the following steps: synchronously acquiring an initial image sequence of multiple viewpoints around a work station, extracting human body key point 2D skeleton feature data of each viewpoint, and calling environment semantic map data containing device entity 3D geometric boundary and device operation point coordinates; each 2D skeleton feature data is projected into a world coordinate system consistent with the environment semantic map through spatial coordinate mapping transformation, the spatial mapping coordinates of human body key points under each viewpoint are obtained, and the spatial mapping coordinates of different viewpoints belonging to the same human body target are spatially aggregated to generate an initial 3D fusion skeleton flow; the beneficial effects are that non-real action data generated due to shielding or perspective deviation can be accurately identified and removed, and the false alarm problem caused by the absence of logical verification in a complex space in a traditional algorithm is solved.

Description

Technical Field

[0001] This invention relates to the field of behavioral risk assessment technology, and in particular to a multi-perspective collaborative method and system for assessing operational behavioral risks. Background Technology

[0002] In modern manufacturing workshops, to ensure the personal safety of workers and standardize operating procedures, visual monitoring systems deployed around workstations are typically used to monitor personnel behavior in real time. Existing technical solutions mostly employ a simple integration of single or multiple cameras, using deep learning algorithms to extract the human posture skeleton and comparing it with predefined templates of prohibited actions to achieve preliminary identification and risk assessment of hazardous work behaviors.

[0003] However, existing assessment methods and systems mainly focus on feature extraction and pattern matching of the human body's two-dimensional or three-dimensional skeleton. When dealing with the dynamic interaction between personnel and surrounding equipment, they often ignore the semantic constraints of the relationship between human movement trajectory and physical work space. Due to the lack of deep perception of the occupancy status of physical obstacles in the work space, the system can only guess the posture through visual features when facing common industrial conditions such as complex angles, drastic changes in lighting, or partial occlusion. This results in the generated posture data being highly likely to have serious logical contradictions with the actual layout of the physical space. This behavior recognition bias caused by the lack of logical verification in the spatial dimension makes it difficult for the system to accurately remove algorithm noise caused by visual occlusion or perspective shift. As a result, the robustness of the risk assessment results is insufficient, the false alarm rate is difficult to reduce, and it cannot meet the needs of industrial-grade high-reliability safety monitoring. Summary of the Invention

[0004] In view of the above-mentioned prior art, this application is hereby proposed. Embodiments of this application provide a multi-view collaborative method and system for assessing operational behavior risks, which can accurately identify and eliminate non-realistic action data caused by occlusion or perspective shift, solving the problem of false alarms caused by the lack of logical verification in traditional algorithms within complex spaces.

[0005] According to one aspect of this application, a multi-perspective collaborative method for assessing operational behavior risk is provided, including:

[0006] Simultaneously acquire initial image sequences from multiple perspectives around the workstation, extract 2D skeleton feature data of human key points from each perspective, and retrieve environmental semantic map data containing 3D geometric boundaries of equipment entities and coordinates of equipment operation points.

[0007] The 2D skeleton feature data are projected onto a coordinate system that corresponds to the semantics of the environment through spatial coordinate mapping transformation. Figure 1Under the unified world coordinate system, the spatial mapping coordinates of key points of the human body from various perspectives are obtained, and the spatial mapping coordinates of different perspectives belonging to the same human target are spatially aggregated to generate an initial 3D fused skeleton flow.

[0008] The initial 3D fused skeleton flow is spatially topologically compared with the 3D geometric boundary of the device entity to identify physical conflict events in which the human body key points penetrate the 3D geometric boundary of the device entity in the world coordinate system, and the physical spatial displacement deviation between the key points of the human body operation limb end and the device operation point is calculated.

[0009] Based on the frequency of occurrence of the physical conflict events and the physical spatial displacement deviation, a real-time spatial reliability factor is calculated for the feature data corresponding to each viewpoint, and the feature data of each viewpoint are weighted and fused using the real-time spatial reliability factor to output the risk assessment result of the operation behavior.

[0010] Based on the risk assessment results, the viewpoints that meet the preset stability threshold of the real-time spatial credibility factor within the preset time window are selected as the benchmark viewpoints. With the benchmark viewpoints as a reference, the mapping transformation parameters of the other viewpoints with the physical conflict events are corrected online to obtain correction parameters. Based on the correction parameters, in the next sampling period, the sampling frequency of the corresponding data acquisition source and the fusion weight of the feature data of the corresponding viewpoints in the weighted fusion process are adjusted synchronously.

[0011] According to another aspect of this application, a multi-perspective collaborative operational behavior risk assessment system is provided, including:

[0012] Data acquisition module: used to synchronously acquire initial image sequences from multiple perspectives around the workstation, extract 2D skeleton feature data of human key points from each perspective, and retrieve environmental semantic map data containing 3D geometric boundaries of equipment entities and coordinates of equipment operation points.

[0013] Mapping and aggregation module: used to project the 2D skeleton feature data onto a coordinate mapping model that corresponds to the semantics of the environment. Figure 1 Under the unified world coordinate system, the spatial mapping coordinates of key points of the human body from various perspectives are obtained, and the spatial mapping coordinates of different perspectives belonging to the same human target are spatially aggregated to generate an initial 3D fused skeleton flow.

[0014] The identification module is used to perform spatial topological comparison between the initial 3D fused skeleton flow and the 3D geometric boundary of the device entity, identify physical conflict events in which the human body key points penetrate the 3D geometric boundary of the device entity in the world coordinate system, and calculate the physical spatial displacement deviation between the key points at the end of the human operating limb and the operating point of the device.

[0015] Fusion assessment module: Based on the frequency of occurrence of the physical conflict events and the physical spatial displacement deviation, it calculates the real-time spatial credibility factor for the feature data corresponding to each perspective, and uses the real-time spatial credibility factor to perform weighted fusion of the feature data of each perspective, and outputs the risk assessment result of the operation behavior.

[0016] The calibration feedback module is used to select, based on the risk assessment results, the viewpoints whose real-time spatial credibility factors meet the preset stability threshold within a preset time window as the benchmark viewpoints, and, with the benchmark viewpoints as a reference, perform online correction of the mapping transformation parameters for the other viewpoints with the physical conflict events to obtain correction parameters. Based on the correction parameters, in the next sampling period, the sampling frequency of the corresponding data acquisition source and the fusion weight of the feature data of the corresponding viewpoints in the weighted fusion process are synchronously adjusted.

[0017] According to another aspect of this application, an electronic device is provided, including a memory and a processor, the memory being used to store computer-executable instructions, and the processor being used to execute the computer-executable instructions, which, when executed by the processor, implement the steps of the method described above.

[0018] According to another aspect of this application, a computer storage medium is provided that stores computer-executable instructions thereon, which, when executed by a processor, implement the steps of the method described above.

[0019] Compared with existing technologies, the multi-view collaborative operation behavior risk assessment method and system according to the embodiments of this application can accurately identify and eliminate non-realistic action data caused by occlusion or perspective shift by identifying physical conflict events where key points penetrate the equipment boundary. This solves the problem of false alarms caused by the lack of logical verification in traditional algorithms in complex spaces. It ensures that when the system faces local camera failure or severe occlusion at a specific angle, it can automatically tilt to a high-confidence viewpoint, ensuring that the output risk assessment results maintain high stability and accuracy under changing working conditions.

[0020] By synchronously adjusting the sampling frequency and fusion weight of the next cycle, computing resources are dynamically concentrated towards high-value perspectives, which not only improves the accuracy of capturing key behaviors but also optimizes the overall operating efficiency of the system, providing highly deterministic early warning basis for industrial-grade safety management. Attached Figure Description

[0021] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0022] Figure 1 This is a schematic diagram of the overall process of the multi-perspective collaborative operational behavior risk assessment method of the present invention.

[0023] Figure 2 This is a schematic diagram illustrating the correction logic of the judgment result of the multi-perspective collaborative operational behavior risk assessment method of the present invention.

[0024] Figure 3 This is a schematic diagram illustrating the real-time spatial credibility factor generation logic of the multi-perspective collaborative operational behavior risk assessment method of the present invention. Detailed Implementation

[0025] Hereinafter, exemplary embodiments according to this application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments of this application. It should be understood that this application is not limited to the exemplary embodiments described herein.

[0026] Example 1:

[0027] Reference Figures 1-3 As an embodiment of the present invention, a multi-perspective collaborative method for assessing operational behavior risk is provided:

[0028] Figure 1 The illustration shows a multi-perspective collaborative operational risk assessment method according to an embodiment of this application, including:

[0029] Simultaneously acquire initial image sequences from multiple perspectives around the workstation, extract 2D skeleton feature data of human key points from each perspective, and retrieve environmental semantic map data containing 3D geometric boundaries of equipment entities and coordinates of equipment operation points.

[0030] Specifically, the system synchronously captures images of workers using multiple data acquisition sources (such as RGB industrial cameras or surveillance cameras) pre-deployed around the workstation. To ensure the accuracy of subsequent spatial aggregation, each data acquisition source is aligned using a timestamp synchronization protocol to obtain an initial image sequence containing different observation perspectives. Each perspective image sequence records the multi-dimensional motion projection of the worker in the same time plane. Then, computer vision processing technology is used to perform instance segmentation and human key point detection on the worker in each frame of the image. Specifically, the system uses a preset human pose estimation framework (such as HRNet, OpenPose, or a deep convolutional neural network with the same function) to scan the image, identify and locate the pixel positions of the main joints of the human body in the current perspective image coordinate system, and outputs 2D skeleton feature data corresponding to each perspective image by normalizing and feature encoding these pixel positions. These 2D skeleton feature data represent the instantaneous posture of the human body in the planar dimension in the form of coordinate vectors.

[0031] While extracting human features, the system retrieves environmental semantic map data matching the current workstation from local storage or cloud database. The environmental semantic map data is not a simple image, but contains the 3D geometric boundaries of the hardware devices (such as machine tools, robotic arms, guardrails, etc.) in the workstation in the world coordinate system, obtained in advance through CAD modeling or real-scene scanning, and the coordinates of the device operation points marked with key locations with interactive attributes on the devices (such as switch buttons, feed ports, emergency stop positions, etc.). This environmental semantic map serves as a global benchmark for subsequent physical constraint determination, providing a unified physical reference framework for multi-view 2D skeleton feature data, enabling human key points to be mapped from isolated pixel coordinates to a physically meaningful workspace.

[0032] The 2D skeleton feature data are projected onto a coordinate mapping transformation that corresponds to the semantics of the environment. Figure 1 Under the unified world coordinate system, the spatial mapping coordinates of key points of the human body from various perspectives are obtained, and the spatial mapping coordinates of different perspectives belonging to the same human target are spatially aggregated to generate an initial 3D fused skeleton flow.

[0033] It should be noted that the world coordinate system, as a globally unified physical measurement benchmark, is used to eliminate spatial pose differences between various data acquisition sources. This world coordinate system is a globally unique static reference system preset in the workstation space, typically a right-handed Cartesian coordinate system. Its spatial origin is physically anchored to a preset fixed reference point within the workstation, such as a specific corner point of the machine tool base, the geometric center of the work area ground, or a fixed support position of the protective fence; among which, shaft and The axis is preset in a plane parallel to the horizontal plane of the work station, and is used to characterize the distribution position of the operators and equipment in the horizontal space; The axis is perpendicular to the horizontal plane of the workstation and points directly upwards, used to measure the spatial position of key human body points and equipment components in the height dimension; this coordinate system serves as the base for global logic verification, supporting the 3D geometric boundaries of equipment entities in the environmental semantic map. By performing geometric calculations on the same world coordinate system, the static occupancy information of the system can eliminate data distortion caused by perspective scaling and installation angle differences from different camera perspectives, thereby providing a physically realistic discrimination environment for subsequent identification of physical conflict events where key points penetrate the device boundary.

[0034] Specifically, using calibrated camera parameters, the pixel coordinates in the 2D skeleton feature data from each viewpoint are converted into spatially mapped coordinates in the world coordinate system. To ensure the physical and logical rigor of the conversion process, a homogeneous coordinate transformation relationship is used to achieve dimensionality enhancement. The specific formula is as follows:

[0035] ;

[0036] in, The calculated spatial mapping coordinates of the j-th keypoint in the world coordinate system from the i-th viewpoint are as follows: The viewpoint index corresponds to the data acquisition source number; j represents the human body keypoint index, such as shoulder or elbow; and W represents the world coordinate system. Let be the homogeneous pixel coordinates of the j-th keypoint from the i-th viewpoint. Let be the depth scale factor of the j-th keypoint from the i-th viewpoint. Let be the camera intrinsic parameter matrix for the i-th viewpoint, and let its inverse be . Indicates the conversion ratio between pixels and length units. Let be the rotation matrix of the i-th viewpoint relative to the world coordinate system, and let its inverse be . , Let i be the translation vector of the i-th viewpoint relative to the origin of the world coordinate system;

[0037] The correlation is established by calculating the geometric proximity between spatially mapped coordinates, and a spatial distance threshold is set for the coordinates of the same key points mapped from viewpoints A and B. and Calculate their Euclidean distance. If the Euclidean distance is less than or equal to the spatial distance threshold, then the two sets of coordinates are determined to belong to the same physical joint of the same human target.

[0038] For key point clusters identified as belonging to the same human target, the system obtains the unique optimal spatial position of the key point in the world coordinate system by calculating the geometric center. The specific formula is as follows:

[0039] ;

[0040] in, Here, represents the aggregated, high-precision spatial mapping coordinates, and n represents the total number of viewpoints that effectively observed the j-th keypoint. Indicates the view index, and ;

[0041] Obtain aggregated coordinates of each key point Then, based on the preset definition of human biomechanical structure, the discrete aggregate coordinates are... Spatial topological association is performed. Specifically, the system establishes a spatial topological configuration representing the instantaneous posture of the human body according to the physical connection order of the human skeleton (e.g., vector connection of wrist keypoint coordinates with elbow keypoint coordinates, vector connection of shoulder keypoint coordinates with neck keypoint coordinates, etc.). Subsequently, the system encapsulates the spatial topological configuration corresponding to each sampling time into a temporal set according to the order of sampling timestamps, thereby forming an initial 3D fused skeleton flow S that records the trajectory of the human body's dynamic work behavior. The specific formula is as follows:

[0042] ;

[0043] in, Here, T represents the sampling time sequence number of the current sampling period, and T represents the total number of samples within the preset sampling time window. For the t-th sampling time, the aggregated coordinates are... The spatial topology set consisting of the physical connection vectors of the vectors.

[0044] The initial 3D fused skeleton flow is compared with the 3D geometric boundary of the device entity using spatial topology. Specifically, the system defines the 3D geometric boundary of the device entity as a closed spatial point set region in the world coordinate system, and aggregates the coordinates of each human key point in the initial 3D fused skeleton flow at sampling time t. Topological alignment is performed using a spatial point set inclusion determination formula, as follows:

[0045] ;

[0046] The topology comparison result represents the spatial occupancy state of the j-th keypoint relative to the device entity at time t. This is an indicator function that outputs 1 if the condition is true and 0 otherwise. Let J be the aggregate coordinates of the j-th keypoint at time t in the world coordinate system. The device entity is a closed three-dimensional space set occupied by the device entity in the world coordinate system, and its boundary is determined by the 3D geometric parameters of the device entity. Through this topological comparison process, the system can determine in real time whether the human skeleton overlaps with the space occupied by the device entity in the three-dimensional physical space, thereby providing a basis for subsequent identification of physical conflict events.

[0047] Identify physical conflict events where key human body points penetrate the 3D geometric boundaries of the device entity in the world coordinate system;

[0048] Further identification of physical conflict events includes the following steps:

[0049] Based on the displacement vectors of human body key points between adjacent sampling times, the motion trajectory flow of each human body key point is constructed. Specifically, the system obtains the aggregated coordinates of each human body key point at the current sampling time t. and retrieve the previous adjacent sampling time. Corresponding coordinates The displacement vector between two time points is calculated using the following formula:

[0050] ;

[0051] in, Given the displacement vectors between two time points, the motion trajectory flow of each key point on the human body is constructed by temporally accumulating the displacement vectors between multiple consecutive sampling time points. This trajectory flow represents the motion path line segment of human joints in three-dimensional space;

[0052] In this embodiment of the application, the system is in its initial state (i.e., the first sampling time). (and there is no previous adjacent sampling time) In the extreme case, the system executes the initialization judgment logic: the system detects the current sampling time number t, if... Then, skipping the steps of constructing displacement vectors and motion trajectory flows, the aggregated coordinates of each key point of the human body at the current moment are directly obtained. A static spatial topology comparison is performed with the 3D geometric boundaries of the device entity; at this time, the system will use the coordinates from the previous moment. The default value is the coordinate of the current time. , or displacement vector The default value is a zero vector to complete the initial encapsulation of the trajectory flow; when the system runs to... When the system starts up, it automatically switches to the normal processing mode that constructs motion trajectory flow based on displacement vectors between adjacent sampling times. This initialization process ensures that the system can still perform basic spatial penetration recognition at startup and guarantees the continuity and stability of time series data calculation.

[0053] The motion trajectory flow The intersection operation is performed with the 3D geometric boundaries of the device entities in the environmental semantic map. If the trajectory flow segment interferes with the device surface, the coordinates of the intersection point are extracted. Subsequently, the system identifies the aggregated coordinates of key points on the current human body. The formula for calculating the depth of entry into the space within the equipment entity is as follows:

[0054] ;

[0055] in, Space entry depth represents the straight-line distance from which key points of the human body penetrate into the interior of the device.

[0056] Calculate the duration of the motion trajectory flow's continuous residence within the spatial entry depth. Specifically, accumulate the number of consecutive sampling periods for each key point within the device entity in real time. Let the total number of consecutive sampling periods currently within the entity be... The duration of stay is then The calculation is as follows:

[0057] ;

[0058] in, The preset sampling period duration; when the depth of entry into the space exceeds the preset depth threshold and the duration of continuous residence exceeds the preset time threshold, a physical conflict event is determined to be triggered;

[0059] In this embodiment, the preset depth threshold is set according to the physical surface material properties of the device to be monitored and the steady-state error range of human key point recognition, and its value range is usually from 5mm to 30mm. If there is a soft protective layer on the surface of the device or the visual detection noise is large, the threshold is increased accordingly to avoid false triggering of conflict events due to slight jitter of the skeleton coordinates. The preset time threshold is determined based on the sampling period of the data acquisition source and the biomechanical characteristics of human movement, and is usually set to 2 to 5 consecutive sampling periods. By setting this time threshold, transient noise points that instantly cross the boundary of the device due to perspective shift during human key points during movement can be effectively filtered out.

[0060] While performing conflict identification, the physical spatial displacement deviation between key points at the human end of the operating limb and the operating point of the device is calculated. The specific formula is as follows:

[0061] ;

[0062] in, Let t be the spatial aggregated coordinates of key points at the extremities of the human body being manipulated. The three-dimensional coordinates of the corresponding device operation points preset in the environmental semantic map.

[0063] Based on the frequency of physical conflict events and the deviation of physical spatial displacement, a real-time spatial credibility factor is calculated for the feature data corresponding to each perspective. The feature data of each perspective are then weighted and fused using the real-time spatial credibility factor to output the risk assessment results of the operation behavior.

[0064] Specifically, the system continuously counts the total number of physical conflict events triggered by the i-th viewpoint within the current sampling time window, denoted as the occurrence frequency. The sampling time window contains T consecutive sampling periods. The system uses the latest sampling time number k to perform a backtracking analysis on that time. The conflict determination results within each cycle are summed using the following formula:

[0065] ;

[0066] Where T is the total number of samples within the preset sampling time window, representing the spatial span of the sliding window, k is the real-time sequence number of the current sampling period (i.e., the latest sampling time point), and k≥T, and t' is the summation loop variable, which starts from the window start time sequence number. Iterate up to the current time index k, This represents the conflict determination result for the i-th viewpoint at the t-th sampling time; a value of 1 indicates a trigger, otherwise 0. The cumulative frequency of physical conflicts occurring from the i-th viewpoint within the current sampling time window;

[0067] The system is based on the frequency of occurrence. The real-time spatial reliability factor of the feature data from each perspective is calculated based on the deviation from physical space displacement. The specific formula is as follows:

[0068] ;

[0069] in, Let be the real-time spatial reliability factor for the i-th viewpoint, representing the probability of accuracy of the 3D mapping result for that viewpoint. Let be the physical spatial displacement deviation of the operating end point identified from the i-th viewpoint. The spatial sensitivity coefficient is a preset value, and its dimensions are... Consistent;

[0070] Real-time spatial credibility factor obtained by calculation Feature data from each perspective (such as spatially mapped coordinates) The weighted fusion is performed to obtain the final fused feature data. The specific formula is as follows:

[0071] ;

[0072] in, The coordinates are the final spatial feature coordinates after fusion, and n is the total number of viewpoints currently participating in the evaluation. The spatial mapping coordinates of the j-th key point extracted from the i-th viewpoint;

[0073] The system logically compares the final spatial feature coordinates after fusion with the hazardous areas and standard operating procedures in the environmental semantic map. By calculating the distance between the fused skeleton and the hazardous boundary of the equipment and the compliance of the action sequence, it outputs a risk assessment result that reflects the current safety level of the operation. Since the fusion process eliminates the interference of low-confidence perspectives, the output risk assessment result has extremely high spatial accuracy and anti-interference ability.

[0074] Based on the risk assessment results, the perspectives that meet the preset stability threshold in real-time spatial credibility factors within a preset time window are selected as benchmark perspectives. Specifically, the real-time spatial credibility factors corresponding to each perspective are monitored in real time. In order to select the optimal reference source, a source that meets the real-time spatial reliability factor within a preset time window is selected. Greater than the preset stability threshold Using the perspective of the subject as the baseline, the judgment logic is as follows: If Then the i-th viewpoint is determined as the reference viewpoint. ;

[0075] in, To preset the stability threshold, Real-time spatial credibility factor selected from the benchmark candidate set The perspective with the largest mean (i.e., the baseline perspective) This indicates that within T sampling periods, the frequency of conflict and displacement deviation between the spatial mapping result and the physical boundary at this perspective are both at a low level.

[0076] Using the reference viewpoint as a guide, online correction parameters are obtained by performing mapping transformation on other viewpoints with physical conflict events. Specifically, for viewpoints other than the reference viewpoint... Other perspectives involving physical conflict events (referred to as perspectives to be corrected) (from a baseline perspective) Aggregate coordinates generated in the world coordinate system As a spatial truth reference, the original mapping transformation parameters of the viewpoint to be corrected are corrected online. Let the viewpoint to be corrected be... The original mapping transformation parameter matrix is The corrected parameters are The correction parameters are obtained by minimizing the reprojection residual. The details are as follows:

[0077] ;

[0078] in, The initial mapping transformation parameter matrix for the viewpoint to be corrected contains rotation and translation components. The correction parameters are then solved using the above calculations. Then, it is applied to the mapping transformation parameters of the viewpoint to be corrected. This process realizes the automatic online correction of camera extrinsic parameter deviations caused by external environmental interference through the spatial constraints of high-confidence viewpoints without human intervention.

[0079] Based on the correction parameters, in the next sampling period, the sampling frequency of the corresponding data acquisition source and the fusion weight of the feature data of the corresponding viewpoint in the weighted fusion process are adjusted synchronously.

[0080] The adjustment of sampling frequency and fusion weights includes the following steps:

[0081] Calculate the information complementarity between the feature data corresponding to each viewpoint and the spatial mapping coordinates under the reference viewpoint;

[0082] It should be noted that the calculation of information complementarity includes:

[0083] Determine the angles between the optical axes of each viewpoint and the reference viewpoint in the world coordinate system, and calculate the field-of-view overlap rate of each viewpoint for key points on the human body based on the angles between the optical axes. The specific formula is as follows:

[0084] ;

[0085] in, The overlap rate of the field of view coverage of key points of the human body from various perspectives. The angle between the optical axes;

[0086] Calculate the position of key human body points relative to the optical center of the corresponding data acquisition source from various viewpoints. Imaging distance The unit spatial resolution for each viewpoint is determined based on the imaging distance, using the following formula:

[0087] ;

[0088] in, This represents the unit spatial resolution for each viewpoint. Focal length Pixel size;

[0089] By combining the field of view coverage overlap rate and unit spatial resolution, the ability of each viewpoint to provide incremental features beyond the reference viewpoint is quantified to obtain the information complementarity. The specific formula is as follows:

[0090] ;

[0091] in, For information complementarity;

[0092] Based on information complementarity and correction parameters, the potential contribution value of each perspective to the risk assessment results is identified, and the specific formula is as follows:

[0093] ;

[0094] in, The contribution potential value for the i-th viewpoint. For the perspective to be corrected The translation correction deviation modulus, This is the standard deviation reference value, used for normalizing spatial deviation.

[0095] Based on their contribution potential, each perspective is prioritized, and in the next sampling period, the sampling frequency of each corresponding data acquisition source is positively correlated with its priority. The specific formula is as follows:

[0096] ;

[0097] in, Assign a sampling frequency to the i-th viewpoint in the next sampling period. The reference sampling frequency of the data acquisition source. The arithmetic mean of the contribution potential values ​​of all participating modulating perspectives;

[0098] Simultaneously, the fusion weights of the feature data from each corresponding viewpoint are adjusted proportionally during the weighted fusion process. The specific formula is as follows:

[0099] ;

[0100] in, is the weighting coefficient of the feature data of the i-th viewpoint in the next sampling period, and n is the total number of viewpoints currently participating in the evaluation.

[0101] In this embodiment, a feedback control mechanism based on the geometric quality of visual information is established to address the problem of uneven contribution of heterogeneous data to the evaluation results in a multi-view environment. First, the field of view coverage overlap rate is calculated by determining the optical axis angle, aiming to assess the spatial differences of each viewpoint relative to the reference viewpoint and ensure the identification of observation dimensions with high information increment. Second, the unit spatial resolution is determined by calculating the imaging distance, aiming to quantify the sampling accuracy of the sensor for key human features. Finally, a correction parameter is introduced to identify the contribution potential value, aiming to couple visual geometric attributes with physical deviation feedback, thereby completing the quantitative evaluation of the data value of each viewpoint in both spatiotemporal dimensions. This operation realizes a technological leap from fixed acquisition to on-demand acquisition, ensuring that core computing resources always serve high-confidence feature data.

[0102] By dynamically and positively allocating sampling frequency based on contribution potential values, the sampling density of high-value perspectives per unit time is significantly improved, thereby enhancing the real-time performance and accuracy of risk assessment from the source. By synchronously adjusting the weight coefficients according to the proportion of potential values, the weights of each perspective in the final decision can be automatically adjusted based on the real-time reliability fluctuations of the data, enhancing robustness to interference from operating conditions such as local occlusion and sudden changes in illumination. Through the coordinated adjustment of frequency and weight, the redundant computational overhead caused by low-quality perspectives is reduced, optimizing the overall operating efficiency of the multi-perspective collaborative system.

[0103] Alternative technical solutions include: First, weight allocation based on image sharpness operators, but this method ignores the three-dimensional geometric complementarity between cameras and cannot provide accurate weight constraints in complex spatial operation environments; Second, a fixed fusion method based on equal-frequency sampling of multiple video streams, which is susceptible to the negative impact of low-confidence data when dealing with situations where there are large differences in quality among different viewpoints, leading to an increase in the false alarm rate of the evaluation results, and cannot dynamically optimize the distribution of computing resources according to actual needs.

[0104] For example, consider the adjustment process of viewpoints 1, 2, and 3 within a continuous sampling period:

[0105] Let viewpoint 1 be the reference viewpoint. At the current moment, the angle between viewpoint 2 and the optical axis of viewpoint 1 Field of view coverage overlap rate focal length of viewpoint 2 Pixel size Imaging distance Its unit spatial resolution was calculated. ;

[0106] Information complementarity from perspective 2 Since the reference angle 1 makes an angle of 0 with its own optical axis, its Similarly, we can obtain perspective 3. ;

[0107] Set the standard deviation reference value If the translation correction deviation modulus of viewpoint 2 is Then the contribution potential value of perspective 2 If perspective 3 Then its Let the potential value of reference viewpoint 1 be... Set the high-confidence baseline constant to 0.00001;

[0108] Calculated arithmetic mean potential value If the reference sampling frequency Then, in the next sampling period, the allocated sampling frequency for viewpoint 2... Due to excessive spatial deviation, the sampling frequency for viewpoint 3 was reduced to approximately 16.6Hz, and the weighting coefficients were adjusted proportionally accordingly. The weighting for viewpoint 2 was also adjusted. The weight of perspective 3 From the baseline perspective 1, due to its highest potential contribution value, it holds the dominant weight. .

[0109] Figure 2 This is a schematic diagram illustrating the correction logic of the judgment result of the multi-perspective collaborative work behavior risk assessment method of the present invention;

[0110] This application further proposes a method that includes:

[0111] To obtain real-time operating status parameters of the equipment under monitoring during operation, specifically, this is achieved by communicating with the control unit (such as a PLC or robot controller) of the equipment under monitoring to obtain these parameters. Real-time running status parameters This includes, but is not limited to, encoder feedback values ​​for various moving components of the equipment (such as servo axes and telescopic arms);

[0112] The real-time operating status parameters are matched with the preset device status-operation point mapping relationship, and the real-time spatial coordinates of the corresponding device's operation point in the world coordinate system in the environmental semantic map are calculated.

[0113] It should be noted that the calculation of real-time spatial coordinates includes the following steps:

[0114] The real-time offset of each moving component in the monitored equipment relative to the equipment reference point is obtained based on real-time operating status parameters. and rotation angle ; Real-time offset and rotation angle Substituting the preset device state-operation point mapping relationship, which is a chain-like spatial transformation matrix determined by the physical connection sequence between the moving components of the device to be monitored, the relative coordinates of the operation point in the local coordinate system with the device reference point as the origin are calculated. The specific formula is as follows:

[0115] ;

[0116] in, Here, m represents the relative coordinates of the operating point in a local coordinate system with the equipment reference point as the origin, and m is the hierarchical index of the moving component. The preset fixed coordinates of the operation point in the coordinate system of the end of the moving component. It is a chain-like spatial transformation matrix, formed by cascading homogeneous transformation matrices between various components. These are the relative coordinates of the operation point in the local coordinate system.

[0117] By utilizing the pose parameters of the device reference point in the environmental semantic map in the world coordinate system, the relative coordinates are converted into the real-time spatial coordinates of the operation point in the world coordinate system. The specific formula is as follows:

[0118] ;

[0119] in, This refers to the real-time spatial coordinates of the corresponding device's operation point in the environmental semantic map at time t in the world coordinate system. The pose parameters of the device reference point in the world coordinate system (including the position vector and attitude matrix of the device reference point).

[0120] Based on real-time spatial coordinates Update the operation point positions in the environmental semantic map, and recalculate the physical spatial displacement deviation between the key points of the human operating limbs and the device operation points based on the updated environmental semantic map. The specific formula is as follows:

[0121] ;

[0122] in, The physical spatial displacement deviation between the key points of the human operating limbs and the operating points of the equipment;

[0123] The updated physical space displacement deviation is used to verify and correct the judgment results of physical conflict events. Specifically, if the original key point was identified as penetrating the geometric boundary of the device, but the calculated result is now correct, the error can be corrected. If the value is less than the preset operation trigger threshold, it indicates that the penetration behavior is a normal interaction between the person and the moving component, rather than an identification error. In this case, the system will automatically cancel or correct the physical conflict event judgment at that moment, thereby improving the accuracy of the risk assessment results under complex dynamic working conditions.

[0124] In this embodiment, the operation trigger threshold is set based on the effective trigger radius of the device operation point in the environmental semantic map and the safe operating distance required by the work specifications. Specifically, the system retrieves the physical interaction characteristics of the device under monitoring at the operation point, defines the minimum functional contact distance between the key point of the operating limb end (such as the center of the finger or palm) and the center of the device operation point as the benchmark, and adds 10% to 20% spatial redundancy as the final operation trigger threshold. When the calculated physical spatial displacement deviation... When the value is below the threshold, the system logically determines that the current personnel's limbs are within the normal range of work interaction, and the overlap between them and the equipment boundary is a compliant operation rather than an abnormal physical penetration. Based on this, the physical conflict event is canceled or corrected, thus achieving accurate removal of false alarm signals under complex interactive conditions.

[0125] This solution addresses the challenge of spatial matching between equipment elements and key human body points in the environmental semantic map under motion conditions by introducing real-time motion feedback from the equipment. At the work site, the positions of components of the monitored equipment change dynamically with the work logic. If physical conflict is determined solely based on static environmental data, normal interactions between the human body and moving components will be identified as risks. Therefore, this solution obtains the underlying encoder data of the equipment control unit and uses a chain-like spatial transformation matrix to map the real-time offset and rotation angle of the moving components to the world coordinate system, thereby achieving synchronous correction of the operation point positions in the environmental semantic map. The essence of this operation is to use equipment state parameters as constraints to dynamically reconstruct the spatial logic of visual recognition.

[0126] By calculating the coordinates of the operation point in real time and introducing an operation trigger threshold, the system can accurately identify normal interactive behaviors of operators and dangerous physical conflict events, logically eliminating misjudgments caused by component movement. The application of the chain-like spatial transformation matrix ensures the geometric rigor of the operation point in the process of transforming between the local coordinate system and the world coordinate system, providing an accurate physical reference for the calculation of displacement deviation. By setting the operation trigger threshold, it can tolerate the small measurement fluctuations of visual recognition in complex backgrounds, significantly improving the reliability of risk assessment results under equipment movement conditions.

[0127] Alternative technical solutions include: First, directly detecting human-machine contact actions based on visual recognition algorithms, but this method is prone to missed detections in complex occlusion environments and cannot provide quantitative spatial displacement values; Second, deploying physical protection sensors on the surface of all moving components, but this method increases system hardware costs and maintenance difficulty, and cannot be deeply integrated with logical coordinates in the environmental semantic map; In contrast, this solution improves the discrimination accuracy without increasing hardware costs by fusing software algorithms with existing device parameters.

[0128] Figure 3 This is a schematic diagram of the real-time spatial credibility factor generation logic of the multi-perspective collaborative operational behavior risk assessment method of the present invention.

[0129] This application further proposes a method that includes:

[0130] A standard knowledge base trained from historical safety operation data is acquired. This standard knowledge base contains the standard spatiotemporal distribution probabilities of human 3D skeleton poses under different operation stages within the environmental semantic map. Specifically, the standard knowledge base is stored in the system's storage unit and contains the standard spatiotemporal distribution probabilities of human 3D skeleton poses for different operation stages (such as loading, processing, and inspection) within the environmental semantic map. And the standard spatiotemporal distribution probability It characterizes the coordinate distribution pattern of key points of the human body in three-dimensional space and the high-frequency trajectory intervals that evolve over time under the premise of safe operation;

[0131] It should be noted that the standard knowledge base is obtained through offline feature modeling and probabilistic statistical training on massive historical safety operation data. First, the system collects standard operation video streams and corresponding synchronous sensor data completed by multiple skilled operators at different operation stages in the same work position. Human pose estimation technology is used to extract the 3D skeleton features of the operators from the above data and project them into the world coordinate system corresponding to the environmental semantic map, forming a standard pose sample set. Then, the sample set is labeled according to the preset operation stages, and statistical learning methods (such as Gaussian mixture models or probability density estimation) are used to model the three-dimensional spatial coordinates of each key point of the human body in each operation stage, thereby calculating the standard spatiotemporal distribution probability of the human 3D skeleton pose. Standard spatiotemporal distribution probability It can accurately quantify the frequency of occurrence of various joints of the human body in a specific spatial area and the probability of trajectory evolution over time under the premise of safe operation. Finally, the distribution probabilities corresponding to each operation stage are associated and stored to build a standard knowledge base for subsequent consistency verification.

[0132] Based on the motion trend of the initial 3D fusion skeleton flow and combined with the current equipment operating status at the workstation, identify the specific operation stage at present. Subsequently, it calls from the standard knowledge base and... Corresponding standard spatiotemporal distribution probability ;

[0133] Calculate the 3D skeleton pose data (i.e., spatial mapping coordinates) provided by each current viewpoint. ) and standard spatiotemporal distribution probability The consistency value between behaviors is calculated using the following formula:

[0134] ;

[0135] in, The behavioral consistency value represents the degree to which the currently visually observed action matches the standard safety action template. The feature extraction function for the current pose;

[0136] The spatial consistency factor is calculated based on the frequency of physical conflict events and the deviation of physical spatial displacement. The specific formula is as follows:

[0137] ;

[0138] in, Spatial consistency factor;

[0139] The behavioral consistency value and the spatial consistency factor are weighted and fused to generate a comprehensive real-time spatial credibility factor, as shown in the following formula:

[0140] ;

[0141] in, Comprehensive real-time spatial credibility factor.

[0142] In complex industrial operation scenarios, relying solely on a single physical conflict determination (such as displacement deviation or penetration frequency) is insufficient to distinguish between abnormal identification noise and atypical safe operations. Therefore, this invention constructs a standard knowledge base, using the probability distribution of skeleton postures in historical safe operations as a safety template in a spatiotemporal dimension. First, the corresponding standard spatiotemporal distribution probability is retrieved using the identified operation stage, and a behavioral consistency value is obtained through integration to evaluate the compliance of the current data from the perspective of action logic. Second, combined with a spatial consistency factor reflecting geometric conflict, a comprehensive real-time spatial credibility factor is generated through weighted fusion. The essence of this operation is to couple experience-based behavioral patterns with physical spatial constraints, constructing an evaluation system that includes both instantaneous geometric verification and long-term behavioral verification.

[0143] By introducing behavioral consistency values, abnormal feature data that, although not physically conflicted, deviate significantly from safe operating procedures in spatiotemporal distribution can be identified and suppressed, reducing the false negative rate of risk assessment from a behavioral logic perspective. The introduction of spatial consistency factors strengthens the real-time evaluation of viewpoint image quality, ensuring that misidentified views with high frequency of physical conflicts and large displacement deviations are automatically downweighted during multi-view fusion. Through behavioral weighting coefficients… The adjustment enables a dynamic balance between the two evaluation dimensions of action compliance and spatial authenticity, allowing the real-time spatial credibility factor to more comprehensively and objectively reflect the potential contribution of data from various perspectives to the final risk assessment results, and greatly enhancing the system's decision robustness in complex and variable operating environments.

[0144] Example 2:

[0145] This is one embodiment of the present invention, which differs from the previous embodiment in that:

[0146] A multi-perspective, collaborative operational risk assessment system, including:

[0147] Data acquisition module: used to synchronously acquire initial image sequences from multiple perspectives around the workstation, extract 2D skeleton feature data of human key points from each perspective, and retrieve environmental semantic map data containing 3D geometric boundaries of equipment entities and coordinates of equipment operation points.

[0148] Mapping and aggregation module: Used to project 2D skeleton feature data onto a coordinate mapping model that corresponds to the semantic environment. Figure 1 Under the unified world coordinate system, the spatial mapping coordinates of key points of the human body from various perspectives are obtained, and the spatial mapping coordinates of different perspectives belonging to the same human target are spatially aggregated to generate an initial 3D fused skeleton flow.

[0149] The recognition module is used to perform spatial topological comparison between the initial 3D fused skeleton flow and the 3D geometric boundary of the device entity, identify physical conflict events in which key points of the human body penetrate the 3D geometric boundary of the device entity in the world coordinate system, and calculate the physical spatial displacement deviation between the key points of the human operating limb end and the device operating point.

[0150] Fusion Assessment Module: Based on the frequency of physical conflict events and physical spatial displacement deviation, it calculates the real-time spatial reliability factor for the feature data corresponding to each perspective, and uses the real-time spatial reliability factor to weight and fuse the feature data of each perspective to output the risk assessment result of the operation behavior.

[0151] The calibration feedback module is used to select the viewpoint that meets the preset stability threshold in real time within a preset time window as the benchmark viewpoint based on the risk assessment results. With the benchmark viewpoint as a reference, the module performs online correction of the mapping transformation parameters of other viewpoints with physical conflict events to obtain correction parameters. Based on the correction parameters, the module synchronously adjusts the sampling frequency of the corresponding data acquisition source and the fusion weight of the feature data of the corresponding viewpoint in the weighted fusion process in the next sampling period.

[0152] Example 3:

[0153] In one embodiment of the present invention, which differs from the previous embodiment, the electronic device includes one or more processors and a memory.

[0154] A processor can be a central processing unit (CPU) or other form of processing unit with data processing and / or instruction execution capabilities, and can control other components in an electronic device to perform desired functions.

[0155] The memory may include one or more computer program products, which may include various forms of computer-readable storage media, such as volatile memory and / or non-volatile memory. Volatile memory may include, for example, random access memory (RAM) and / or cache memory. Non-volatile memory may include, for example, read-only memory (ROM), hard disk, flash memory, etc.

[0156] In one example, the electronic device may also include input devices and output devices, which are interconnected via a bus system and / or other forms of connection mechanisms (not shown). In addition, depending on the specific application, the electronic device may include any other suitable components.

[0157] Example 4:

[0158] Embodiments of this application may also be computer-readable storage media storing computer program instructions thereon, which, when executed by a processor, cause the processor to perform the steps described in the "Exemplary Methods" section above according to the various embodiments of this application.

[0159] Computer-readable storage media may take the form of any combination of one or more readable media. A readable medium may be a readable signal medium or a readable storage medium. A readable storage medium may, for example, include, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatuses, or devices, or any combination thereof. More specific examples of readable storage media (a non-exhaustive list) include: electrical connections having one or more wires, portable disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0160] The basic principles of this application have been described above with reference to specific embodiments. However, it should be noted that the advantages, benefits, and effects mentioned in this application are merely examples and not limitations, and should not be considered as essential features of each embodiment of this application. Furthermore, the specific details disclosed above are for illustrative and facilitative purposes only, and are not limitations. These details do not restrict the application from being implemented using the specific details described above.

[0161] The block diagrams of devices, apparatuses, devices, and systems involved in this application are merely illustrative examples and are not intended to require or imply that they must be connected, arranged, or configured in the manner shown in the block diagrams. As those skilled in the art will recognize, these devices, apparatuses, devices, and systems can be connected, arranged, and configured in any manner. Words such as “comprising,” “including,” “having,” etc., are open-ended terms meaning “including but not limited to,” and are used interchangeably with them. The terms “or” and “and” as used herein refer to the terms “and / or,” and are used interchangeably with them unless the context clearly indicates otherwise. The term “such as” as used herein refers to the phrase “such as but not limited to,” and is used interchangeably with it.

[0162] It should also be noted that in the apparatus, equipment, and methods of this application, the components or steps can be disassembled and / or recombined. These disassemblies and / or recombinations should be considered as equivalent solutions of this application.

[0163] The above description of the disclosed aspects is provided to enable any person skilled in the art to make or use this application. Various modifications to these aspects will be readily apparent to those skilled in the art, and the general principles defined herein can be applied to other aspects without departing from the scope of this application. Therefore, this application is not intended to be limited to the aspects shown herein, but rather to be accorded the widest scope consistent with the principles and novel features disclosed herein.

[0164] The above description has been given for purposes of illustration and description. Furthermore, this description is not intended to limit the embodiments of this application to the forms disclosed herein. Although numerous exemplary aspects and embodiments have been discussed above, those skilled in the art will recognize certain variations, modifications, alterations, additions, and sub-combinations thereof.

Claims

1. A multi-viewpoint cooperative work behavior risk assessment method, characterized by, include: Simultaneously acquire initial image sequences from multiple perspectives around the workstation, extract 2D skeleton feature data of human key points from each perspective, and retrieve environmental semantic map data containing 3D geometric boundaries of equipment entities and coordinates of equipment operation points. By transforming spatial coordinates, the 2D skeleton feature data are projected onto a world coordinate system consistent with the environmental semantic map to obtain the spatial mapping coordinates of human key points from each viewpoint. The spatial mapping coordinates of different viewpoints belonging to the same human target are spatially aggregated to generate an initial 3D fused skeleton flow. The initial 3D fused skeleton flow is spatially topologically compared with the 3D geometric boundary of the device entity to identify physical conflict events in which the human body key points penetrate the 3D geometric boundary of the device entity in the world coordinate system, and the physical spatial displacement deviation between the key points of the human body operation limb end and the device operation point is calculated. Based on the frequency of occurrence of the physical conflict events and the physical spatial displacement deviation, a real-time spatial reliability factor is calculated for the feature data corresponding to each viewpoint, and the feature data of each viewpoint are weighted and fused using the real-time spatial reliability factor to output the risk assessment result of the operation behavior. Based on the risk assessment results, the viewpoints that meet the preset stability threshold of the real-time spatial credibility factor within the preset time window are selected as the benchmark viewpoints. With the benchmark viewpoints as a reference, the mapping transformation parameters of the other viewpoints with the physical conflict events are corrected online to obtain correction parameters. Based on the correction parameters, in the next sampling period, the sampling frequency of the corresponding data acquisition source and the fusion weight of the feature data of the corresponding viewpoints in the weighted fusion process are adjusted synchronously.

2. The multi-perspective collaborative job behavior risk assessment method of claim 1, wherein, The method further includes: Acquire real-time operating status parameters of the equipment to be monitored during operation; The real-time operating status parameters are matched with the preset device status-operation point mapping relationship, and the real-time spatial coordinates of the corresponding device's operation point in the environmental semantic map in the world coordinate system are calculated. The calculation of the real-time spatial coordinates includes the following steps: Based on the real-time operating status parameters, obtain the real-time offset and rotation angle of each moving component in the monitored device relative to the device reference point; Substituting the real-time offset and rotation angle into the preset device state-operation point mapping relationship, the relative coordinates of the operation point in the local coordinate system with the device reference point as the origin are calculated; wherein, the device state-operation point mapping relationship is a chain-like spatial transformation matrix determined by the physical connection sequence between the moving components of the device to be monitored. Using the pose parameters of the device reference point in the environmental semantic map in the world coordinate system, the relative coordinates are converted into the real-time spatial coordinates of the operation point in the world coordinate system; The environmental semantic map is updated based on the real-time spatial coordinates, and the physical spatial displacement deviation between the key points of the human operating limb end and the device operating point is recalculated based on the updated environmental semantic map. The updated physical space displacement deviation is used to verify and correct the determination result of the physical conflict event.

3. The multi-perspective collaborative job behavior risk assessment method of claim 1, wherein, The method further includes: A standard knowledge base trained from historical safety operation data is obtained, wherein the standard knowledge base contains the standard spatiotemporal distribution probability of human 3D skeleton posture under different operation stages in the environmental semantic map; Based on the initial 3D fusion skeleton flow, the current operation stage is identified, and the corresponding standard spatiotemporal distribution probability is retrieved from the standard knowledge base; Calculate the behavioral consistency value between the 3D skeleton pose data provided by each current viewpoint and the standard spatiotemporal distribution probability; The spatial consistency factor is calculated based on the frequency of occurrence of the physical conflict events and the physical spatial displacement deviation. By weighting and fusing behavioral consistency values ​​and spatial consistency factors, a comprehensive real-time spatial credibility factor is generated.

4. The multi-perspective synergistic job behavior risk assessment method of claim 1, wherein, The adjustment of the sampling frequency and the fusion weight includes the following steps: Calculate the information complementarity between the feature data corresponding to each of the aforementioned viewpoints and the spatial mapping coordinates under the reference viewpoint; Based on the information complementarity and the correction parameters, identify the potential contribution value of each perspective to the risk assessment results; Based on the contribution potential value, each viewpoint is prioritized and sorted. In the next sampling period, the sampling frequency of each corresponding data acquisition source is positively correlated with the priority and sorted, and the fusion weight of the feature data of each corresponding viewpoint is adjusted proportionally when performing weighted fusion.

5. The multi-perspective collaborative job behavior risk assessment method of claim 4, wherein, The calculation of the information complementarity includes: Determine the angle between each viewpoint and the reference viewpoint in the world coordinate system, and calculate the field-of-view coverage overlap rate of each viewpoint to the key points of the human body based on the angle between the optical axes. Calculate the imaging distance of the human body key points relative to the corresponding data acquisition source under each of the aforementioned viewpoints, and determine the unit spatial resolution of each of the aforementioned viewpoints based on the imaging distance; By combining the field of view coverage overlap rate and the unit spatial resolution, the ability of each viewpoint to provide incremental features beyond the reference viewpoint is quantified, and the information complementarity is obtained.

6. The multi-perspective collaborative operational behavior risk assessment method according to claim 1, characterized in that, The identification of the physical conflict event includes the following steps: Based on the displacement vectors of the human body key points between adjacent sampling times, the motion trajectory flow of each human body key point is constructed; Extract the coordinates of the intersection points where the motion trajectory flow passes through the 3D geometric boundary of the device entity, and identify the spatial entry depth of the intersection point coordinates within the device entity; The duration of the motion trajectory flow within the spatial entry depth is calculated, and when the spatial entry depth exceeds a preset depth threshold and the duration of the continuous stay exceeds a preset time threshold, the physical conflict event is determined to be triggered.

7. A multi-view collaborative work behavior risk assessment system, characterized by, include: Data acquisition module: used to synchronously acquire initial image sequences from multiple perspectives around the workstation, extract 2D skeleton feature data of human key points from each perspective, and retrieve environmental semantic map data containing 3D geometric boundaries of equipment entities and coordinates of equipment operation points. Mapping and aggregation module: used to project the 2D skeleton feature data of each person into a world coordinate system consistent with the environmental semantic map through spatial coordinate mapping transformation, obtain the spatial mapping coordinates of human body key points from each viewpoint, and spatially aggregate the spatial mapping coordinates of different viewpoints belonging to the same human body target to generate an initial 3D fused skeleton stream. The identification module is used to perform spatial topological comparison between the initial 3D fused skeleton flow and the 3D geometric boundary of the device entity, identify physical conflict events in which the human body key points penetrate the 3D geometric boundary of the device entity in the world coordinate system, and calculate the physical spatial displacement deviation between the key points at the end of the human operating limb and the operating point of the device. Fusion assessment module: Based on the frequency of occurrence of the physical conflict events and the physical spatial displacement deviation, it calculates the real-time spatial credibility factor for the feature data corresponding to each perspective, and uses the real-time spatial credibility factor to perform weighted fusion of the feature data of each perspective, and outputs the risk assessment result of the operation behavior. The calibration feedback module is used to select, based on the risk assessment results, the viewpoints whose real-time spatial credibility factors meet the preset stability threshold within a preset time window as the benchmark viewpoints, and, with the benchmark viewpoints as a reference, perform online correction of the mapping transformation parameters for the other viewpoints with the physical conflict events to obtain correction parameters. Based on the correction parameters, in the next sampling period, the sampling frequency of the corresponding data acquisition source and the fusion weight of the feature data of the corresponding viewpoints in the weighted fusion process are synchronously adjusted. 8.An electronic device comprising a memory and a processor, the electronic device characterized by: The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions, which, when executed by the processor, implement the steps of the method as described in any one of claims 1 to 6.

9. A computer storage medium having stored thereon computer- executable instructions which, when executed by a computer, cause the computer to carry out the steps of claim 1. When the computer-executable instructions are executed by a processor, they implement the steps of the method as described in any one of claims 1 to 6.

Images