A computer vision-based intelligent management and control method for safety production personnel behavior

Abstract

The application discloses a kind of based on computer vision's safe production personnel behavior intelligent management and control method, belong to production safety management and control technical field.This method first extracts reference skeleton key point sequence and spatial motion trajectory according to each station standard operation video, constructs initial operation reference including core motion node sequence and tolerance boundary;In adaptive cycle, collect target personnel multi-cycle transition operation data, calculate trajectory offset and eliminate incidental anomaly, extract reasonable deviation data when offset is within tolerance, correct and generate personalized operation reference;When offset exceeds tolerance, combined with upstream and downstream station timing offset collaborative verification, determine as legal process change, then reconstruct reference, otherwise determine violation;After cycle, based on personalized reference real-time comparison determines operation compliance.The application can adapt to individual operation difference and process change, effectively reduce the misjudgment rate, improve management and control accuracy and adaptive ability.

Description

Technical Field

[0001] This invention relates to the field of production safety management and control technology, and specifically to a method for intelligent management and control of personnel behavior in production safety based on computer vision. Background Technology

[0002] In the field of industrial safety production management, relying on computer vision and artificial intelligence technologies to achieve intelligent control of worker behavior is a core development direction for enterprise safety standardization. Patent application CN112966593A discloses an enterprise safety standardization operation method based on artificial intelligence and big data. This method extracts the 3D coordinates of key skeletal points of workers by collecting scene and depth images, identifies baseline action categories, and constructs an action change map and an initial action change matrix. Simultaneously, it divides the background sub-region, constructs a bipartite graph relating the environment and actions, filters environmental influencing factors, and optimizes the action judgment model. This reduces the interference of environmental factors on action judgment and improves the accuracy of safety standardization judgments.

[0003] However, this technical solution still has obvious limitations: it adopts a fixed and uniform action judgment standard, which is difficult to adapt to the normal action differences caused by different operators due to their body shape and operating habits. It is easy to misjudge compliant operations as abnormal, resulting in a high misjudgment rate. At the same time, it only analyzes and judges the actions of a single workstation independently, which cannot adapt to the scenario of dynamic adjustment of production process. When the work process changes reasonably, the system cannot automatically adapt to the judgment rules, which easily leads to a large number of false alarms and makes it difficult to meet the precise control needs of dynamic assembly line production scenarios. Summary of the Invention

[0004] This invention provides a computer vision-based intelligent management and control method for personnel behavior in safe production, which solves the problems of existing technologies that use a unified operating benchmark to judge personnel behavior, resulting in a high misjudgment rate due to the lack of adaptation to individual differences in body shape and movement habits; and the inability to automatically identify legal process changes and dynamically update the benchmark, leading to insufficient adaptive control and accuracy.

[0005] To achieve the above objectives, this invention provides a computer vision-based intelligent management method for safety production personnel behavior, comprising the following steps: S1: For each production workstation, acquire the standard operation video of the benchmark operator at that workstation, extract the benchmark skeletal keypoint sequence and standard spatial motion trajectory, and construct an initial operation benchmark corresponding to each production workstation, wherein the initial operation benchmark includes a core motion node sequence and a tolerance boundary; S2: Within a preset adaptive period when the target personnel access the production workstation, acquire the transition operation video of the target personnel in multiple complete operation periods, and extract the transition skeletal keypoint sequence and its corresponding transition spatial motion trajectory; S3: Calculate the transition spatial motion trajectory in multiple complete operation periods relative to the initial... The trajectory offset sequence of the operation benchmark is processed and occasional anomalies are removed to obtain a corrected sequence. If the transition skeletal key point sequence triggers the corresponding core action node sequence and the corrected sequence is within the tolerance boundary, reasonable deviation data is extracted based on the corrected sequence. If the corrected sequence exceeds the tolerance boundary, a process change judgment is made according to the collaborative verification strategy, and benchmark reconstruction is triggered according to the judgment result. S4: Based on the reasonable deviation data, the tolerance boundary is corrected through the feature parameter update strategy to generate a personalized operation benchmark for the target personnel. S5: After the preset adaptive cycle ends, the real-time operation video of the target personnel is acquired, the real-time skeletal key point sequence and real-time spatial action trajectory are extracted, and they are compared with the personalized operation benchmark to determine whether there is a violation.

[0006] Optionally, the construction of the initial operation benchmark includes: performing temporal segmentation on the skeletal key point sequence in the standard operation video, identifying key action postures during the operation process, and determining the core action node sequence; extracting the distribution features of the standard spatial action trajectory within a preset time window, and calculating the motion envelope of the benchmark skeletal key points in space using a spatial clustering algorithm, and setting the motion envelope as the tolerance boundary.

[0007] Optionally, the calculation process of the trajectory offset includes: temporally aligning the transition space motion trajectory with the standard space motion trajectory; in the temporally aligned spatial coordinate system, calculating the Euclidean distance between the transition space motion trajectory and the standard space motion trajectory at each sampling time point, and taking the average value of the Euclidean distance at each sampling time point as the trajectory offset.

[0008] Optionally, the process of removing occasional anomalies in step S3 includes: aligning the trajectory offset sequences corresponding to multiple complete operation cycles according to a unified time series reference to construct a multi-cycle offset time series matrix; performing outlier detection on all offset values ​​at the same sampling time in the multi-cycle offset time series matrix to identify and mark abnormal offset points caused by occasional action disturbances; replacing the marked abnormal offset points with the median value of the remaining normal offsets at the same sampling time to remove occasional abnormal data; and integrating the offset sequences of each cycle after replacement according to the time dimension to obtain the corrected sequence.

[0009] Optionally, the process of extracting reasonable deviation data includes: within a preset adaptive period, obtaining a correction sequence generated after removing occasional anomalies; performing sliding window filtering on the correction offset quantum sequence corresponding to each complete operation cycle within the correction sequence to further filter out small random fluctuation deviations; calculating the statistical mean of the filtered multi-cycle correction offset at the corresponding sampling points on the time axis to obtain a systematic deviation vector characterizing individual body shape differences and movement habits, and establishing the systematic deviation vector as the reasonable deviation data.

[0010] Optionally, the execution process of the collaborative verification strategy is as follows: when the correction sequence exceeds the tolerance boundary, the actual occurrence time of the core action node sequence of each upstream adjacent workstation and downstream adjacent workstation is obtained respectively; the timing offset between the actual occurrence time of each core action node of the upstream adjacent workstation and the corresponding reference time is calculated, as well as the timing offset between the actual occurrence time of each core action node of the downstream adjacent workstation and the corresponding reference time; if the timing offsets of all core action nodes of the upstream and downstream adjacent workstations are in the same direction, and the difference in timing offsets between the upstream and downstream workstations is less than the preset synchronization threshold, and the process connection time interval between this workstation and the upstream and downstream adjacent workstations is still within the reference connection tolerance range, then it is determined to be a legal process change, and the operation reference reconstruction of this workstation is triggered; if the core action node time of the upstream and downstream adjacent workstations is consistent with the corresponding reference time, then it is determined to be a violation of the target personnel's operation, and the reference reconstruction is not triggered.

[0011] Optionally, the benchmark reconstruction process described in step S3 is as follows: retrieve the transition skeletal key point sequence and transitional spatial motion trajectory corresponding to all legal process changes within the preset adaptive period to construct a reconstruction sample set; determine the effective time interval using the core motion node times of upstream and downstream workstations as time anchors, re-segment the skeletal key point sequence of the sample set within this interval, identify key motion postures, and generate an updated core motion node sequence; based on the trajectory data of the effective time interval corresponding to the sample set, calculate the motion envelope of the skeletal key points using a spatial clustering algorithm to generate an updated tolerance boundary; integrate the updated core motion node sequence and tolerance boundary into a reconstruction operation benchmark, replacing the original initial operation benchmark.

[0012] Optionally, the feature parameter update strategy includes: obtaining the spatial geometric coordinates of the tolerance boundary in the initial operation benchmark, constructing a spatial translation transformation matrix based on the systematic deviation vector corresponding to the reasonable deviation data and the preset correction weight coefficient; using the spatial translation transformation matrix to perform translation transformation on the spatial geometric coordinates to generate a personalized tolerance boundary; associating and encapsulating the personalized tolerance boundary with the core action node sequence to generate the personalized operation benchmark.

[0013] Optionally, the feature comparison process includes: dynamically aligning the real-time skeletal keypoint sequence with the core action node sequence in the personalized operation benchmark, and extracting the trigger timestamps of the key nodes of the real-time operation action; performing a temporal logic comparison based on the trigger timestamps to determine the completeness of the nodes and the triggering order; mapping the real-time spatial action trajectory to a spatial coordinate system containing personalized tolerance boundaries, and performing spatial envelope collision detection; if the core action nodes are complete and triggered in sequence, and the real-time spatial action trajectory is entirely contained within the personalized tolerance boundaries, then the operation is deemed compliant; otherwise, the operation is deemed non-compliant.

[0014] Optionally, in steps S3 and S5, after determining that the operation is in violation, the method further includes: capturing real-time operation video clips before and after the violation time and synchronizing them to the background management system, and triggering an audio-visual alarm.

[0015] The present invention provides a computer vision-based intelligent management and control method for safety production personnel behavior. By performing first-order spatial fusion and first-order temporal fusion of in-furnace visual acoustic data and process flue gas data respectively, the incomplete combustion state of the solid material layer and gas phase is simultaneously acquired. Combined with a solid-gas coupling reverse verification mechanism to correct the fusion boundary conditions, the accuracy and comprehensiveness of combustion state perception are effectively improved. Based on the generated three-dimensional comprehensive combustion state space matrix of the furnace, feedforward control commands are output, realizing early prediction and precise adjustment of combustion anomalies, significantly smoothing furnace temperature fluctuations, suppressing excessive flue gas pollutants, and improving the stability and combustion efficiency of incinerator operation. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In the drawings: Figure 1 This is an overall flowchart of the intelligent management and control method for safe production personnel behavior provided in the embodiments of the present invention; Figure 2This is a flowchart of the logical decision-making process for trajectory offset processing, deviation extraction, and collaborative verification judgment in step S3 of this embodiment of the invention. Figure 3 This is a geometrical schematic diagram of the feature parameter update strategy according to an embodiment of the present invention; Figure 4 This is a schematic diagram of multi-station collaborative verification provided in an embodiment of the present invention. Detailed Implementation

[0017] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the scope of the present invention.

[0018] It should be noted that the acquisition, transmission, storage, use, and processing of data in the technical solution of this application all comply with the relevant provisions of national laws and regulations. In the embodiments of this application, certain existing industry solutions such as software, components, and models may be mentioned. These should be considered exemplary, intended only to illustrate the feasibility of implementing the technical solution of this application, and do not imply that the applicant has already used or necessarily used such solutions.

[0019] The following is combined with Figures 1-4 This invention is described in detail.

[0020] like Figure 1As shown, this embodiment of the invention provides a computer vision-based intelligent management and control method for safe production personnel behavior, including the following steps: S1: For each production workstation, acquire the standard operation video of the benchmark operator at that workstation, extract the benchmark skeletal key point sequence and standard spatial motion trajectory, and construct the initial operation benchmark corresponding to each production workstation. The initial operation benchmark includes the core motion node sequence and tolerance boundary; S2: Within a preset adaptive period when the target personnel access the production workstation, acquire the transition operation video of the target personnel in multiple complete operation cycles, and extract the transition skeletal key point sequence and its corresponding transition spatial motion trajectory; S3: Calculate the transition spatial motion trajectory in multiple complete operation cycles relative to the initial operation benchmark. The trajectory offset sequence is accurately obtained and occasional anomalies are removed to obtain a corrected sequence. If the transition skeletal key point sequence triggers the corresponding core action node sequence and the corrected sequence is within the tolerance boundary, reasonable deviation data is extracted based on the corrected sequence. If the corrected sequence exceeds the tolerance boundary, a process change judgment is made according to the collaborative verification strategy, and a benchmark reconstruction is triggered based on the judgment result. S4: Based on the reasonable deviation data, the tolerance boundary is corrected through a feature parameter update strategy to generate a personalized operation benchmark for the target personnel. S5: After the preset adaptive cycle ends, the real-time operation video of the target personnel is acquired, the real-time skeletal key point sequence and real-time spatial action trajectory are extracted, and they are compared with the personalized operation benchmark to determine whether there is a violation.

[0021] Among them, a production workstation refers to a fixed work area in the production process used to complete a corresponding process. Each production workstation corresponds to a unique technological task and standard operating procedure. A baseline operator refers to a person who can complete all the operational procedures of the corresponding workstation according to the standard process specifications. A standard operation video is video data formed by the baseline operator completing the complete operation process according to the standard process. A skeletal keypoint sequence is a data sequence formed by extracting the spatial positions of various joints of the human body using human posture recognition technology and arranging them in chronological order. A spatial motion trajectory is a human motion path formed based on the continuous changes of skeletal keypoints, used to describe the continuous change process of human movement in space. An initial operation baseline is a standard motion reference model established based on the standard operation video, including a core motion node sequence and tolerance boundaries. The core motion node sequence uses... The standard operation benchmark represents the key actions and their sequence of occurrence during the completion of the standard process; the tolerance boundary is used to limit the allowable spatial deviation range of the standard actions to accommodate the normal differences in actions between different operators; the complete operation cycle refers to the entire process of a complete production task from start to finish; the preset adaptive cycle is a learning stage used by the target personnel after entering the corresponding production station to collect data from multiple complete operation cycles and establish a personalized action model. The preset adaptive cycle is usually set to 10 complete operation cycles, and the correction weight coefficient is set to 0.9 to ensure that the benchmark fits the individual operating habits while avoiding excessive deviation from the standard process requirements; the personalized operation benchmark is an action reference model obtained by modifying the stable action characteristics formed by the target personnel while maintaining the consistency of the standard process flow.

[0022] Specifically, in step S1, by collecting standard operation videos and extracting the corresponding skeletal key point sequences and spatial motion trajectories, the standard operation process is digitally modeled, and continuous actions are converted into motion feature data that can be calculated and analyzed, providing a unified reference standard for subsequent target personnel action analysis, so that the actions of different personnel can be mapped to the same action description system for comparison.

[0023] Furthermore, in step S2, video data of multiple complete operation cycles of the target person are continuously collected within a preset adaptive period. By continuously acquiring the skeletal key point sequence and spatial motion trajectory of the target person, the system can gradually learn the stable motion characteristics formed by the target person due to different physical conditions, operating habits and motion amplitudes. This provides a continuous data foundation for establishing personalized operation benchmarks in the future, while avoiding the impact of random errors generated by a single operation on subsequent analysis results.

[0024] Furthermore, in step S3, the spatial movement trajectory of the target personnel is compared with the initial operation baseline under a unified temporal feature analysis. The spatial offset generated during the action execution is analyzed, and the triggering status of core action nodes is combined to determine whether the target personnel has completed the corresponding process according to the standard process flow. When the action flow remains consistent and the spatial offset is within a reasonable range, the stable motion offset is retained as an individual action feature. When the spatial offset exceeds the allowable range, the process relationship between upstream and downstream workstations is further combined to determine whether it belongs to the production process adjustment, thereby avoiding misjudgment due to changes in the production process and improving the accuracy of subsequent behavior recognition.

[0025] Furthermore, in step S4, based on the extracted reasonable deviation data, the allowable range of actions in the initial operation benchmark is adaptively corrected. While maintaining the consistency of the core action sequence and process flow, the action reference model is gradually adapted to the stable operating habits formed by the target personnel over a long period of time, and finally a personalized operation benchmark that conforms to the actual operating characteristics of the target personnel is formed.

[0026] Furthermore, in step S5, video data corresponding to the current operation process of the target personnel is collected in real time, and the corresponding skeletal key point sequence and spatial motion trajectory are extracted. These are then matched with a unified feature of the personalized operation benchmark. The execution sequence, integrity of the action, and range of spatial movement are comprehensively analyzed to determine whether the current operation behavior of the target personnel complies with the safety production standards, thereby achieving real-time intelligent control of the operation behavior on the production site.

[0027] Furthermore, after posture recognition, the standard operation video generates a skeletal key point sequence and spatial motion trajectory. The core motion node sequence is extracted based on the skeletal key point sequence, and the tolerance boundary is generated based on the spatial motion trajectory. The two together form the initial operation benchmark. The target personnel operation video is processed in the same way to generate a transitional skeletal key point sequence and transitional spatial motion trajectory. The systematic deviation vector is obtained through trajectory offset analysis and reasonable deviation extraction, and the tolerance boundary is updated accordingly to form a personalized operation benchmark. The real-time operation video then completes motion matching, violation judgment, and alarm output based on the personalized operation benchmark.

[0028] Preferably, the construction of the initial operation benchmark includes: performing temporal segmentation on the skeletal key point sequence in the standard operation video, identifying key action postures during the operation process, and determining the core action node sequence; extracting the distribution features of the standard spatial action trajectory within a preset time window, and calculating the motion envelope of the benchmark skeletal key points in space using a spatial clustering algorithm, and setting the motion envelope as the tolerance boundary.

[0029] Specifically, seventeen core nodes were selected for the human skeleton key points, covering the head, neck, left and right shoulders, left and right elbows, left and right wrists, left and right hips, left and right knees, left and right ankles, and the center point of the torso. Among them, the core nodes for upper limb control are the right wrist, right elbow, and right shoulder, which serve as the main basis for trajectory calculation and tolerance boundary construction. When performing temporal segmentation on the skeletal key point sequence in the standard operation video stream, the coordinates of the human skeleton key points in consecutive frames are first obtained according to the video sampling frequency; then, the displacement change and joint angle change of each skeletal key point between adjacent frames are calculated; when the displacement change or joint angle change exceeds the preset action change threshold, the corresponding moment is determined as the action stage switching point, thereby dividing the continuous operation process into multiple temporal segments with stable action characteristics. The preset action change threshold can be set according to the accuracy requirements of the workstation operation. For conventional assembly workstations, the displacement change threshold can be set to 3cm and the joint angle change threshold to 15 degrees.

[0030] Furthermore, when identifying key action postures during the operation process, the relative positional relationships and joint angles between key joints such as the human torso, shoulders, elbows, wrists, and hips are calculated for each time segment, and features such as action duration, direction of movement, and range of motion are extracted. When the corresponding features meet the preset key action judgment conditions, the posture corresponding to the time segment is determined as the key action posture. Then, the core action node sequence is generated according to the order of appearance of the key action postures on the time axis to describe the key action execution process in the standard process.

[0031] Furthermore, when extracting the distribution characteristics of standard spatial motion trajectories within a preset time window, the spatial motion trajectories are divided into time windows of fixed length. Within each time window, the spatial coordinate distribution, range of motion, and trajectory density of skeletal key points are statistically analyzed to form the local spatial characteristics of the corresponding time window, thereby characterizing the spatial variation law of human movement during standard operation.

[0032] Furthermore, when calculating the motion envelope of the baseline skeletal key points using a spatial clustering algorithm, cluster analysis is performed on the spatial coordinates of the same skeletal key point across multiple time windows. Coordinate points with a spatial distance less than a preset clustering radius are grouped into the same motion region. The preferred spatial clustering algorithm is density clustering. Multi-cycle trajectory points of the same key point are divided into density partitions, and the convex envelopes of each partition are extracted and merged to form a continuous motion envelope. The preset clustering radius can be set according to the amplitude of the operation; for conventional manual operation positions, the clustering radius is set to 6 cm. Then, the boundary range of each motion region is calculated, and multiple motion regions are merged to obtain the normal activity space of the corresponding skeletal key point during standard operation.

[0033] Furthermore, the normal activity spaces corresponding to each key skeletal point are combined according to the topological relationship of the human skeleton to form a human motion envelope covering the entire standard operation process, and this human motion envelope is defined as the tolerance boundary; wherein, the tolerance boundary is used to limit the range of normal motion deviations allowed under the premise of keeping the standard process unchanged, so as to make consistent judgments on the actions of different operators in the future.

[0034] In a preferred embodiment of the present invention, key action postures in the standard operation process are identified by temporal segmentation of the skeletal key point sequence, and tolerance boundaries are constructed by combining spatial clustering algorithms. This enables the establishment of an initial operation benchmark with a clear structure that can accurately represent the characteristics of the standard operation, providing a unified and reliable reference for subsequent comparison of target personnel action trajectories, extraction of reasonable deviations, and determination of violations, thereby improving the accuracy and stability of personnel behavior recognition.

[0035] Preferably, the calculation process of the trajectory offset includes: temporally aligning the transition space motion trajectory with the standard space motion trajectory; in the temporally aligned spatial coordinate system, calculating the Euclidean distance between the transition space motion trajectory and the standard space motion trajectory at each sampling time point, and taking the average value of the Euclidean distance at each sampling time point as the trajectory offset.

[0036] Specifically, when aligning the transition space motion trajectory with the standard space motion trajectory in terms of time sequence, the time series corresponding to the two spatial motion trajectories are first obtained, and the start and end times of the action are used as a unified time reference. Then, a correspondence is established based on the time series, and time interpolation is performed on the positions where there is a deviation in the sampling time, so that the two spatial motion trajectories have a one-to-one corresponding sampling point on a unified time axis, providing a consistent time reference for subsequent trajectory difference calculation. The time sequence alignment adopts a dynamic time warping algorithm, using the core action node as an anchor point to perform nonlinear matching of the two trajectory sequences, eliminating the time sequence deviation caused by the difference in operation rhythm, and achieving a one-to-one correspondence for each sampling point.

[0037] Furthermore, after completing the temporal alignment, the transition space motion trajectory and the standard space motion trajectory are mapped to a unified spatial coordinate system, and the spatial coordinates of the skeletal key points corresponding to the same sampling time are extracted. Then, the spatial distance between each corresponding skeletal key point is calculated, and point-by-point matching is performed according to the correspondence of each skeletal key point of the human body to obtain the overall spatial offset of the human body at each sampling time.

[0038] In this embodiment, when calculating the trajectory offset, the overall spatial offset values ​​of the human body corresponding to each sampling time are statistically analyzed in chronological order, and the spatial offset values ​​corresponding to each sampling time within the entire operation cycle are averaged to obtain the trajectory offset that reflects the degree to which the target person's overall movements deviate from the standard operation. In particular, by comprehensively analyzing the spatial offsets within the entire operation cycle, the impact of local fluctuations at a single sampling point on the calculation results can be reduced, making the obtained trajectory offset more stable and providing a reliable data foundation for subsequent reasonable deviation analysis and movement consistency judgment.

[0039] In a preferred embodiment of the present invention, by uniformly aligning the transition space motion trajectory with the standard space motion trajectory and calculating the spatial offset within the entire operation cycle under a unified time reference, the overall deviation of the target person's actions from the standard actions can be accurately reflected, reducing calculation errors caused by different sampling times or differences in action rhythms, improving the accuracy and stability of trajectory offset calculation, and providing a reliable basis for subsequent behavior determination.

[0040] like Figure 2 As shown, preferably, the process of removing occasional anomalies in step S3 includes: aligning the trajectory offset sequences corresponding to multiple complete operation cycles according to a unified time series reference to construct a multi-cycle offset time series matrix; performing outlier detection on all offset values ​​at the same sampling time in the multi-cycle offset time series matrix to identify and mark abnormal offset points caused by occasional action disturbances; replacing the marked abnormal offset points with the median value of the remaining normal offsets at the same sampling time to remove occasional abnormal data; and integrating the offset sequences of each cycle after replacement according to the time dimension to obtain the corrected sequence.

[0041] Further preferably, the process of extracting reasonable deviation data includes: within a preset adaptive period, obtaining a correction sequence generated after removing occasional anomalies; performing sliding window filtering on the correction offset quantum sequence corresponding to each complete operation cycle within the correction sequence to further filter out small random fluctuation deviations; calculating the statistical mean of the filtered multi-cycle correction offset at the corresponding sampling points on the time axis to obtain a systematic deviation vector characterizing individual body shape differences and movement habits, and establishing the systematic deviation vector as the reasonable deviation data.

[0042] Specifically, the trajectory offset sequences corresponding to multiple complete operation cycles are arranged according to a unified sampling time to construct a multi-cycle offset time series matrix. Each column in the matrix corresponds to the trajectory offset generated by multiple complete operation cycles at the same sampling time, and each row corresponds to a complete operation cycle. This establishes a unified data correspondence between different operation cycles, providing a data foundation for subsequent abnormal offset identification.

[0043] Furthermore, when performing outlier detection on the multi-cycle offset time series matrix, the distribution of all trajectory offsets corresponding to the same sampling time is statistically analyzed, and the degree of deviation of each trajectory offset from the overall distribution center is calculated. Outlier detection adopts the three-standard-deviation criterion, and the mean and standard deviation of all offsets at the same sampling time are calculated. Values ​​exceeding the range of mean plus or minus three standard deviations are judged as abnormal offset points, so as to identify abnormal data caused by occasional pauses, temporary avoidance, human body occlusion, or skeletal key point recognition errors of the target personnel.

[0044] In this embodiment, when correcting the marked abnormal offset points, the unmarked trajectory offsets at the same sampling time are used as valid data, the median value at the corresponding sampling time is calculated, and the corresponding abnormal offset points are replaced with the median value. Since the median value can better reflect the concentrated distribution of trajectory offsets at the same sampling time, it can reduce the impact of abnormal data on the overall analysis results, while maintaining the continuity of the trajectory offset sequence.

[0045] Specifically, after replacing all abnormal offset points, the trajectory offset sequences corresponding to each complete operation cycle are recombined in chronological order to form a correction sequence, and the correction sequence is filtered using a sliding window method. The sliding window moves continuously along the time axis according to the window length of 7 sampling points, and the correction offset is smoothed within each window to further reduce the local random fluctuations generated during continuous sampling and retain the stable action offset characteristics formed by multiple complete operation cycles.

[0046] Furthermore, the filtered multi-cycle correction offsets are statistically analyzed according to a unified sampling time. The statistical mean of the correction offsets corresponding to each sampling time is calculated and then combined in chronological order to form a systematic deviation vector. The systematic deviation vector is used to describe the continuous movement offset pattern of the target personnel throughout the entire operation process. Compared with the trajectory offset of a single complete operation cycle, it can more stably reflect the long-term operating habits formed by the target personnel.

[0047] Specifically, the systematic deviation vector is determined as reasonable deviation data and saved into the action feature data of the corresponding target personnel, providing a basis for subsequent personalized operation benchmark updates. Since the reasonable deviation data comes from the common statistical results of multiple complete operation cycles, it can effectively avoid the impact of occasional abnormal actions or random fluctuations on the construction of personalized operation benchmarks, and improve the accuracy and stability of subsequent behavior recognition.

[0048] Taking a correction sequence containing 300 sampling moments as an example, a sliding window with a length of 15 sampling moments is used, with one sampling moment as the window movement step size, to smooth the correction offsets within the window. After completing the calculation of all windows, the average value of the correction offsets corresponding to the same sampling moment for multiple complete operation cycles is calculated, and then combined sequentially to form a systematic deviation vector. Since this systematic deviation vector comprehensively reflects the stable action offset formed by multiple complete operation cycles, it can serve as reasonable deviation data that distinguishes the target personnel from standard operation characteristics, and can be used for the construction of subsequent personalized operation benchmarks, without affecting the final result due to a single abnormal action.

[0049] In a preferred embodiment of the present invention, outlier detection and anomaly correction are performed on multi-cycle trajectory offsets. Combined with sliding window filtering and statistical analysis to extract systematic deviation vectors, the influence of occasional abnormal actions and random fluctuations can be effectively filtered out. The long-term stable action offset features of the target personnel can be accurately extracted, providing reliable and reasonable deviation data for the subsequent construction of personalized operation benchmarks, thereby improving the accuracy and stability of behavior recognition.

[0050] like Figure 4 As shown, preferably, the execution process of the collaborative verification strategy is as follows: when the correction sequence exceeds the tolerance boundary, the actual occurrence time of the core action node sequence of each upstream adjacent workstation and downstream adjacent workstation is obtained respectively; the time offset between the actual occurrence time of each core action node of the upstream adjacent workstation and the corresponding reference time is calculated, as well as the time offset between the actual occurrence time of each core action node of the downstream adjacent workstation and the corresponding reference time; if the time offset of all core action nodes of the upstream and downstream adjacent workstations are in the same direction, and the difference in time offset between the upstream and downstream workstations is less than the preset synchronization threshold, and the process connection time interval between this workstation and the upstream and downstream adjacent workstations is still within the reference connection tolerance range, then it is determined to be a legal process change, and the operation reference reconstruction of this workstation is triggered; if the core action node time of the upstream and downstream adjacent workstations is consistent with the corresponding reference time, then it is determined to be a violation of the target personnel's operation, and the reference reconstruction is not triggered.

[0051] Further preferably, the execution process of the benchmark reconstruction in step S3 is as follows: retrieve the transition skeletal key point sequence and transitional spatial motion trajectory corresponding to all legal process changes within the preset adaptive period to construct a reconstruction sample set; determine the effective time interval using the core motion node time of the upstream and downstream workstations as the time anchor point, re-segment the skeletal key point sequence of the sample set in time within this interval, identify key motion postures, and generate an updated core motion node sequence; based on the trajectory data of the effective time interval corresponding to the sample set, calculate the motion envelope of the skeletal key points through a spatial clustering algorithm to generate an updated tolerance boundary; integrate the updated core motion node sequence and tolerance boundary into a reconstruction operation benchmark, replacing the original initial operation benchmark.

[0052] Specifically, when the correction sequence exceeds the tolerance boundary, the operation data of the upstream and downstream adjacent workstations corresponding to the current workstation are obtained respectively, and the actual occurrence time of the core action node corresponding to each workstation is extracted. A unified time sequence correspondence is established based on the actual occurrence time of the core action node and the corresponding reference time, providing a unified time sequence reference for process change judgment.

[0053] Specifically, a joint analysis is performed on the timing offset of the core action nodes corresponding to the upstream adjacent workstations, the current workstation, and the downstream adjacent workstations. The timing offset direction of each core action node is determined, and the change in the process connection time between adjacent workstations is calculated. When the core action nodes of each workstation show a consistent timing offset trend, and the process connection relationship between adjacent workstations remains stable, it can be determined that the current action change originates from the adjustment of the production process, rather than the abnormal operation of the target personnel, thereby triggering the reconstruction of the operation benchmark.

[0054] Specifically, when establishing the reconstruction sample set, all operation data determined to be legitimate process changes are selected from the preset adaptive cycle, and the corresponding transition skeleton key point sequence and transition space motion trajectory are extracted. Each group of data is organized according to the complete operation cycle to form a reconstruction sample set for updating the operation benchmark, so as to ensure that the reconstruction process is based only on operation data that conforms to the current production process.

[0055] In this embodiment, the actual occurrence time of the core action nodes of the upstream and downstream workstations is used as the time sequence anchor point to perform unified time sequence positioning on the reconstructed sample set, and the effective time sequence interval is divided according to the time range between adjacent core action nodes; within each effective time sequence interval, the skeletal key point sequence is re-segmented in time sequence, and the corresponding key action posture is extracted, and the core action node sequence adapted to the current process flow is regenerated according to the occurrence order of the key actions.

[0056] Specifically, after updating the core action node sequence, the spatial action trajectory within the effective time interval corresponding to the reconstructed sample set is statistically analyzed. The spatial coordinates corresponding to the same skeletal key point are clustered, and the normal range of motion of each skeletal key point under the current process conditions is calculated. Furthermore, combined with the topological relationship of the human skeleton, the normal range of motion corresponding to each skeletal key point is combined to form an updated motion envelope, and the updated motion envelope is determined as the new tolerance boundary.

[0057] Furthermore, the updated core action node sequence is associated and integrated with the updated tolerance boundary to generate a reconstruction operation benchmark. The original initial operation benchmark is then replaced by the reconstruction operation benchmark, ensuring that the subsequent action recognition process remains consistent with the current production process. This enhances the system's adaptability to production process adjustments while guaranteeing standard process requirements.

[0058] Taking the continuous completion of three complete production cycles on an assembly line as an example, when the core action nodes of the current workstation and its upstream and downstream adjacent workstations all experience a consistent temporal shift in the same direction within three consecutive complete production cycles, and the change in the process connection time between adjacent workstations remains within 0.2 seconds, it can be determined that the overall production cycle has been adjusted. Subsequently, the operation data corresponding to the legal process changes are used to form a reconstruction sample set, and the effective temporal interval is re-determined based on the core action nodes of the upstream and downstream workstations. The key action postures are re-extracted from each sample, and the tolerance boundary is recalculated in combination with the updated spatial action trajectory. Finally, a reconstruction operation benchmark adapted to the current production process is generated to replace the original operation benchmark, thereby realizing the system's adaptive update to the production process adjustment.

[0059] In a preferred embodiment of the present invention, by combining the action sequence of adjacent upstream and downstream workstations for collaborative verification, it is possible to accurately distinguish between the target personnel's violation of regulations and the action changes caused by production process adjustments, thus avoiding misjudging process changes as violations. Furthermore, by reconstructing the core action node sequence and tolerance boundaries based on the operation data corresponding to the legal process changes, the operation benchmark can adapt to changes in the production process in a timely manner, ensuring that subsequent behavior recognition remains consistent with the current process flow, thereby improving the adaptive capability of the operation benchmark and the accuracy and reliability of safety production behavior control.

[0060] like Figure 3 As shown, preferably, the feature parameter update strategy includes: obtaining the spatial geometric coordinates of the tolerance boundary in the initial operation benchmark, constructing a spatial translation transformation matrix based on the systematic deviation vector corresponding to the reasonable deviation data and the preset correction weight coefficient; using the spatial translation transformation matrix to perform translation transformation on the spatial geometric coordinates to generate a personalized tolerance boundary; associating and encapsulating the personalized tolerance boundary with the core action node sequence to generate the personalized operation benchmark.

[0061] Specifically, when obtaining the spatial geometric coordinates corresponding to the tolerance boundary, the coordinates of the boundary vertices corresponding to each skeletal key point constituting the tolerance boundary are extracted, and the spatial relationship between each boundary vertex is established according to the topological relationship of the human skeleton, forming a set of spatial coordinates for geometric transformation, providing a unified data foundation for subsequent tolerance boundary correction.

[0062] Specifically, when constructing the spatial translation transformation matrix, the systematic deviation vector is used as the spatial translation direction corresponding to each skeletal key point, and the translation amount in each direction is calculated in combination with the preset correction weight coefficient. The spatial translation transformation matrix is ​​generated based on the translation amount corresponding to each skeletal key point, so that different skeletal key points can be adjusted in spatial position according to the long-term movement deviation pattern formed by the target person.

[0063] Specifically, when using the spatial translation transformation matrix to correct the tolerance boundary, the spatial translation transformation matrix is ​​applied to the coordinates of the boundary vertices corresponding to each skeletal key point. The boundary vertices are then spatially translated as a whole, so that the corrected tolerance boundary maintains the original spatial structure relationship and only adjusts the overall spatial position. This allows the tolerance boundary to adapt to the stable motion offset characteristics formed by the target personnel, while maintaining the range of motion changes corresponding to the standard operation.

[0064] Specifically, after completing the tolerance boundary correction, the core action node sequence remains unchanged, and the updated tolerance boundary is re-established to correspond with the core action node sequence. Since systematic deviations mainly reflect the stable shift of the target personnel's spatial position and do not change the execution order of actions corresponding to the standard process, there is no need to regenerate the core action node sequence, thus ensuring that the standard process remains consistent.

[0065] Specifically, the updated tolerance boundaries and core action node sequences are uniformly organized, and a correlation mapping relationship is established between the two to generate personalized operation benchmarks for the corresponding target personnel. Among them, the core action node sequences are used to constrain the action execution process, and the personalized tolerance boundaries are used to describe the reasonable action offset range allowed for the target personnel. Both serve as a reference for subsequent real-time identification.

[0066] Taking the long-term movement deviation of the target personnel as an example, which is mainly manifested by the overall displacement of the right hand's range of motion to the outside of the body, after statistical analysis of multiple complete operation cycles, the key bone points corresponding to the right wrist are stably shifted by about 20 mm, and the key bone points corresponding to the right elbow are stably shifted by about 15 mm. Based on the corresponding systematic deviation vector, the spatial translation amount corresponding to each key bone point is calculated, and the spatial geometric coordinates corresponding to the tolerance boundary are synchronously adjusted using the spatial translation transformation matrix. This causes the updated tolerance boundary to be translated in the corresponding direction, while the core movement node sequence remains unchanged, thereby forming a personalized operation benchmark that adapts to the target personnel's movement habits.

[0067] In a preferred embodiment of the present invention, by using a systematic deviation vector to spatially shift and correct the tolerance boundary while keeping the core action node sequence unchanged, the operating benchmark can adapt to the long-term stable action habits formed by the target personnel without changing the standard process flow. This reduces misjudgments caused by individual action differences and improves the adaptability of personalized operating benchmarks and the accuracy of subsequent behavior recognition.

[0068] Preferably, the feature comparison process includes: dynamically aligning the real-time skeletal keypoint sequence with the core action node sequence in the personalized operation benchmark, and extracting the trigger timestamps of the key nodes of the real-time operation action; performing a temporal logic comparison based on the trigger timestamps to determine the completeness of the nodes and the triggering order; mapping the real-time spatial action trajectory to a spatial coordinate system containing personalized tolerance boundaries, and performing spatial envelope collision detection; if the core action nodes are complete and triggered in sequence, and the real-time spatial action trajectory is entirely contained within the personalized tolerance boundaries, then the operation is deemed compliant; otherwise, the operation is deemed non-compliant.

[0069] More preferably, in steps S3 and S5, after determining that the operation is in violation, the method further includes: capturing real-time operation video clips before and after the violation time and synchronizing them to the background management system, and triggering an audio-visual alarm.

[0070] Specifically, when dynamically aligning the real-time skeletal keypoint sequence with the core action node sequence in the personalized operation benchmark, the real-time acquired skeletal keypoint sequence is used as the sequence to be detected, and the core action node sequence is used as the reference sequence. The corresponding relationship is gradually established according to the order of action occurrence, and the corresponding results are continuously updated during real-time operation, so that actions at different operation speeds can be mapped to the corresponding core action nodes, thereby improving the adaptability of the action matching process.

[0071] Specifically, after completing the dynamic timing alignment, a real-time action execution timing sequence is established based on the trigger time corresponding to each core action node, and the timing logic is compared sequentially according to the predetermined execution order of the core action nodes. When each core action node is fully triggered and the triggering order meets the standard process flow, the real-time operation is determined to meet the action flow requirements. When there are missing core action nodes, repeated triggering, or abnormal triggering order, the real-time operation is determined to have an abnormal action flow.

[0072] Specifically, when performing spatial envelope collision detection on real-time spatial motion trajectories, the real-time spatial motion trajectory is mapped to a unified spatial coordinate system where the personalized tolerance boundary is located, and each sampling moment is used to determine whether each skeletal key point is located inside the corresponding tolerance boundary. When the corresponding skeletal key points are all within the tolerance boundary coverage at each sampling moment, the real-time spatial motion trajectory is determined to meet the spatial constraint requirements. When any skeletal key point exceeds the corresponding tolerance boundary for multiple consecutive sampling moments, the real-time motion is determined to have spatial deviation.

[0073] Specifically, after completing the action flow analysis and spatial envelope detection, the results of the temporal logic comparison and the spatial envelope collision detection are comprehensively analyzed. When the action execution sequence meets the standard process requirements and the real-time spatial action trajectory is always within the coverage range of the personalized tolerance boundary, the current operation is determined to comply with the safety production specifications. When any detection result does not meet the corresponding judgment condition, the current operation is determined to have a violation, and the corresponding violation judgment result is generated.

[0074] Specifically, when a violation is determined to occur in the current operation, the time of the violation determination is used as the base time point, and video data of a preset duration is obtained before and after the violation. The video segments of the corresponding time period are then extracted and saved, so that the complete operation process before and after the violation can be continuously recorded, which is convenient for subsequent analysis of the cause of the violation and tracing of responsibility.

[0075] Specifically, after saving the violation video clips, the violation judgment result, the time of violation, the corresponding production station information, and the violation video clips are simultaneously sent to the back-end management system. After receiving the violation information, the back-end management system generates a corresponding violation event record and simultaneously triggers on-site audio and visual alarms, enabling on-site management personnel to be aware of the violation in a timely manner and take action, thereby achieving real-time early warning and full-process traceability management of violations.

[0076] Taking the example of a target person missing a core action node during a certain production process, if the corresponding core action node is not triggered during the dynamic timing alignment process, and the real-time spatial motion trajectory shows that 20 consecutive sampling times exceed the personalized tolerance boundary, then the current operation is determined to be a violation. Subsequently, using the time of the violation as a benchmark, video clips of 10 seconds before and 10 seconds after the violation are automatically extracted, and the violation time, production station information, and corresponding video are simultaneously sent to the back-end management system. At the same time, an on-site audio-visual alarm is triggered, enabling managers to view the violation process and take timely action.

[0077] In summary, the computer vision-based intelligent management and control method for safe production personnel provided by this invention constructs a standard operating benchmark, combines the target personnel's action learning process within a preset adaptive period, extracts stable and reasonable deviation data, and generates personalized operating benchmarks, achieving adaptive fusion of standard process requirements and individual operating habits. Furthermore, by combining upstream and downstream workstation action sequence for collaborative verification, it can accurately distinguish between production process adjustments and target personnel violations, and dynamically reconstruct the operating benchmark based on legitimate process changes, ensuring that behavior recognition always remains consistent with the current production process. During real-time management, by comprehensively analyzing the action execution sequence and spatial motion trajectory, accurate judgment of safe production personnel's operating behavior is achieved. Combined with violation video retention, background synchronization, and on-site alarm mechanisms, real-time early warning, full-process traceability, and intelligent management of violations are realized, effectively reducing misjudgments caused by individual action differences or process adjustments, and improving the accuracy, adaptability, and reliability of safe production behavior recognition.

[0078] The above description is merely a preferred embodiment of the technical solution of the present invention and is not intended to limit the scope of protection of the present invention. The parameters such as the number of complete operation cycles, video sampling frequency, number of sampling moments, sliding window length, time length, spatial offset, and process connection time involved in the above embodiments are merely exemplary parameters given to illustrate the technical solution of the present invention. Those skilled in the art can adjust them according to different production processes, acquisition equipment performance, and application scenarios. Without departing from the concept of the present invention, any changes in the parameters should fall within the scope of protection of the present invention.

Claims

1. A method for intelligent control of personnel behavior in safe production based on computer vision, characterized in that, Includes the following steps: S1: For each production station, acquire the standard operation video of the benchmark operator at that production station, extract the benchmark skeletal key point sequence and standard spatial motion trajectory, and construct the initial operation benchmark corresponding to each production station. The initial operation benchmark includes the core motion node sequence and tolerance boundary. S2: Within the preset adaptive cycle of the target personnel accessing the production station, acquire the transition operation video of the target personnel in multiple complete operation cycles, and extract the transition skeleton key point sequence and its corresponding transition space motion trajectory. S3: Calculate the trajectory offset sequence of the transition space motion trajectory relative to the initial operation reference under multiple complete operation cycles and remove occasional anomalies to obtain the correction sequence; If the transition skeleton key point sequence triggers the corresponding core action node sequence, and the correction sequence is within the tolerance boundary, then reasonable deviation data is extracted based on the correction sequence; if the correction sequence exceeds the tolerance boundary, then process change judgment is performed according to the collaborative verification strategy, and benchmark reconstruction is triggered based on the judgment result. S4: Based on the reasonable deviation data, the tolerance boundary is corrected through the feature parameter update strategy to generate a personalized operating benchmark for the target personnel. S5: After the preset adaptive cycle ends, acquire the real-time operation video of the target person, extract the real-time skeletal key point sequence and real-time spatial motion trajectory, and compare them with the personalized operation benchmark to determine whether there is a violation.

2. The intelligent control method for the behavior of safety production personnel according to claim 1, characterized in that, The construction of the initial operating baseline includes: The skeletal key point sequence in the standard operation video is segmented temporally to identify key action postures during the operation process and determine the core action node sequence. Extract the distribution features of standard spatial motion trajectories within a preset time window, and calculate the motion envelope of the baseline skeletal key points in space using a spatial clustering algorithm, and set the motion envelope as the tolerance boundary.

3. The intelligent control method for the behavior of safety production personnel according to claim 1, characterized in that, The calculation process for the trajectory offset includes: The transition space motion trajectory is time-aligned with the standard space motion trajectory; In the time-aligned spatial coordinate system, the Euclidean distance between the transition spatial motion trajectory and the standard spatial motion trajectory at each sampling time point is calculated, and the average value of the Euclidean distance at each sampling time point is used as the trajectory offset.

4. The intelligent control method for the behavior of safety production personnel according to claim 1, characterized in that, The process of removing occasional anomalies in step S3 includes: The trajectory offset sequences corresponding to multiple complete operation cycles are aligned according to a unified time series reference to construct a multi-cycle offset time series matrix; Outlier detection is performed on all offset values ​​at the same sampling time in the multi-cycle offset time series matrix to identify and mark abnormal offset points caused by occasional action disturbances; For the marked abnormal offset points, the median value of the other normal offsets at the same sampling time is used to replace them, thus eliminating occasional abnormal data; The replacement cycle offset sequences are integrated along the time dimension to obtain the corrected sequence.

5. The intelligent control method for the behavior of safety production personnel according to claim 1, characterized in that, The process of extracting reasonable deviation data includes: Within a preset adaptive period, obtain the corrected sequence generated after removing occasional anomalies; The corrected offset quantum sequence corresponding to each complete operation cycle within the corrected sequence is subjected to sliding window filtering to further filter out small random fluctuation deviations; The statistical mean of the filtered multi-cycle correction offset is calculated at the corresponding sampling points on the time axis to obtain a systematic deviation vector that characterizes individual body shape differences and movement habits, and the systematic deviation vector is established as the reasonable deviation data.

6. The intelligent control method for the behavior of safety production personnel according to claim 1, characterized in that, The execution process of the collaborative verification strategy is as follows: When the correction sequence exceeds the tolerance boundary, the actual occurrence time of the core action node sequence of each upstream and downstream adjacent workstation is obtained respectively. Calculate the time offset between the actual occurrence time of each core action node of the upstream adjacent workstation and the corresponding reference time, and the time offset between the actual occurrence time of each core action node of the downstream adjacent workstation and the corresponding reference time. If the timing offsets of all core action nodes of the upstream and downstream adjacent workstations are in the same direction, and the timing offset difference between the upstream and downstream workstations is less than the preset synchronization threshold, and the process connection time interval between this workstation and the upstream and downstream adjacent workstations is still within the baseline connection tolerance range, then it is determined to be a legal process change, triggering the operation baseline reconstruction of this workstation. If the core action node times of adjacent upstream and downstream workstations are consistent with the corresponding baseline times, it is determined that the target personnel have violated the operation rules and the baseline reconstruction will not be triggered.

7. The intelligent control method for the behavior of safety production personnel according to claim 6, characterized in that, The execution process of the benchmark reconstruction described in step S3 is as follows: Retrieve the transition skeleton key point sequence and transition space motion trajectory corresponding to all legal process changes within the preset adaptive cycle, and construct a reconstruction sample set; The effective time interval is determined by using the core action node times of upstream and downstream workstations as time anchors. Within this interval, the skeletal key point sequence of the sample set is re-segmented in time to identify key action postures and generate an updated core action node sequence. Based on the trajectory data corresponding to the effective time interval of the sample set, the motion envelope of the skeletal key points is calculated by spatial clustering algorithm to generate updated tolerance boundaries. The updated core action node sequence and tolerance boundary are integrated into the reconstruction operation benchmark, replacing the original initial operation benchmark.

8. The intelligent control method for the behavior of safety production personnel according to claim 1, characterized in that, The feature parameter update strategy includes: Obtain the spatial geometric coordinates of the tolerance boundary in the initial operating reference; Based on the systematic deviation vector corresponding to the reasonable deviation data and the preset correction weight coefficient, a spatial translation transformation matrix is ​​constructed. The spatial geometric coordinates are translated using the spatial translation transformation matrix to generate personalized tolerance boundaries. The personalized tolerance boundary is associated and encapsulated with the core action node sequence to generate the personalized operation benchmark.

9. The intelligent control method for the behavior of safety production personnel according to claim 6, characterized in that, The process of feature comparison includes: Dynamically align the real-time skeletal keypoint sequence with the core action node sequence in the personalized operation benchmark, and extract the trigger timestamps of key nodes of the real-time operation action. Based on the trigger timestamp, a time sequence logic comparison is performed to determine the node integrity and triggering order; Real-time spatial motion trajectories are mapped to a spatial coordinate system with personalized tolerance boundaries to perform spatial envelope collision detection. If the core action nodes are complete and triggered in sequence, and the real-time spatial action trajectory is entirely contained within the personalized tolerance boundary, the operation is deemed compliant; otherwise, the operation is deemed non-compliant.

10. The intelligent control method for the behavior of safety production personnel according to claim 9, characterized in that, In steps S3 and S5, after determining that the operation is in violation, the following steps are also included: capturing real-time operation video clips before and after the violation time and synchronizing them to the background management system, and triggering an audio-visual alarm.

Images