Intelligent identification and hierarchical early warning method for unsafe behaviors in cigarette storage environment

Abstract

This invention discloses an intelligent identification and graded early warning method for unsafe behaviors in cigarette storage environments, relating to the fields of artificial intelligence and safety monitoring technology. The method includes the following steps: deploying multiple video acquisition devices within the cigarette storage operation area, fixedly installed at different angles and heights, to collect real-time image data of personnel and equipment within the operation area; performing multi-target detection processing on the collected image data, using a trained deep learning model to identify the positions and movement trajectories of personnel, forklifts, and mobile equipment respectively, and outputting the identification results in a structured format. This invention achieves accurate identification of unsafe behaviors in cigarette storage through deep learning and behavior modeling, and combines risk scoring and spatial gridding technology for dynamic graded early warning, significantly improving the real-time identification and predictive early warning capabilities for high-risk behaviors, enhancing the intelligent safety management level of storage operations, and possessing advantages such as accurate identification, efficient response, and strong deployment adaptability.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and security monitoring technology, specifically to a method for intelligent identification and graded early warning of unsafe behaviors in cigarette storage environments. Background Technology

[0002] "Intelligent identification and graded early warning of unsafe behaviors in cigarette storage environments" refers to the use of artificial intelligence technologies (such as video surveillance, image recognition, and behavior analysis algorithms) in the storage operation area of ​​a cigarette factory to automatically identify violations or dangerous actions (such as not wearing a safety helmet, entering restricted areas, and driving a forklift without authorization) by employees during handling, stacking, and forklift operations. At the same time, it classifies different levels (such as low risk, medium risk, and high risk) according to the degree of danger of different behaviors and issues corresponding early warning notices to managers or operators in a timely manner. This realizes intelligent supervision and risk intervention for operational safety, thereby reducing the occurrence of safety accidents and improving the inherent safety level of storage operations.

[0003] Existing technologies have the following shortcomings: Most existing unsafe behavior recognition and early warning systems for cigarette storage environments rely on static rules or simple target detection algorithms, resulting in weak capabilities for recognizing complex interactive behaviors. For example, when an employee temporarily turns back to retrieve an item and briefly overlaps with a forklift approaching at high speed in a blind spot, the system, lacking a deep understanding of the employee's intent and dynamic trajectory, often fails to promptly identify this as a potentially high-risk event. This leads to missed opportunities for early warning, easily resulting in collisions with potentially serious injuries or even death.

[0004] The information disclosed in the background section is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent identification and graded early warning method for unsafe behaviors in cigarette storage environments. By using deep learning and behavioral modeling, it achieves accurate identification of unsafe behaviors in cigarette storage. Combined with risk scoring and spatial gridding technology, it provides dynamic graded early warning, significantly improving the real-time identification and predictive early warning capabilities for high-risk behaviors. This enhances the level of intelligent safety management in storage operations and has advantages such as accurate identification, efficient response, and strong deployment adaptability, thereby solving the problems mentioned in the background technology.

[0006] To achieve the above objectives, the present invention provides the following technical solution: an intelligent identification and graded early warning method for unsafe behaviors in cigarette storage environments, comprising the following steps: Multiple video acquisition devices are deployed in the cigarette storage area, fixedly installed at different angles and heights, to collect real-time image data of personnel and equipment in the area. The collected image data is processed for multi-target detection. A trained deep learning model is used to identify the position and motion trajectory of personnel, forklifts and mobile devices, and the recognition results are output in a structured manner. Establish a behavior recognition module based on temporal feature analysis to perform correlation analysis on target trajectories in different time periods and identify high-risk behavior patterns including personnel turning back, sudden stops, and reverse crossing. A spatial risk mapping model is constructed to divide the warehouse area into regional grids, label the risk weight of each area, and compare the relative positional relationship and behavioral status of personnel and equipment in the grid in real time. Based on the preset hazard level determination rules, the identified unsafe behaviors are classified into levels according to behavior category, risk level and location of occurrence, and corresponding warning level labels are generated; For different levels of unsafe behavior, warning information is sent to on-site safety personnel and the management system through audible and visual alarms, mobile push notifications, and background pop-ups to achieve tiered intervention.

[0007] Preferably, the video acquisition device is an industrial-grade camera with infrared night vision capability and high frame rate function, with a frame rate of not less than 60 frames / second and a resolution of not less than 1080p, to ensure clear and stable image acquisition of personnel and equipment under low light and high-speed motion conditions. When deploying the data acquisition device, the warehouse movement line is used as a reference, and the installation angle is controlled between 30 and 60 degrees to cover the high-frequency areas of forklift operations and high-risk activity areas of personnel. Overlapping coverage is also applied to the blind spots of the camera's field of view to achieve seamless continuity and regional complementarity of image data.

[0008] Preferably, the deep learning model adopts a multi-label classification structure with ResNet50 as the backbone network, and introduces an attention mechanism module to enhance local features of human body posture, specifically including enhanced extraction of head orientation, hand movements, and walking posture details. Furthermore, the recognition model is retrained on a real-world cigarette storage environment dataset through transfer learning to adapt to application scenarios with complex backgrounds including stacked shelves, light reflection, and smoke interference, thereby improving the model's recognition accuracy and robustness in real-world storage environments.

[0009] Preferably, the behavior recognition module adopts a sequence behavior modeling structure based on LSTM, and combines a time window sliding strategy to extract a set of personnel and equipment trajectory points within 5 consecutive seconds, with a sampling frequency of 10 times per second, to construct dynamic trajectory sequence data; The system is trained and classified using multidimensional features such as trajectory change amplitude, velocity vector direction change rate, and path intersection density. It is used to identify whether there is an intention to turn back, cross, or stop suddenly, and to make logical inferences to determine whether it belongs to the defined "high-risk behavior template library".

[0010] Preferably, the spatial risk mapping model adopts a regular grid structure, dividing the storage area into 1-meter × 1-meter square units. Each square is assigned a basic risk factor. The value of the basic factor is calculated based on historical accident density, equipment activity frequency and personnel density statistics. The risk level in each square is dynamically adjusted in real time by combining the current behavior status and interaction scenario. The regional heat map is updated every 60 seconds and the distribution of high-risk areas is displayed in real time with color gradients to guide people to avoid danger.

[0011] Preferably, the spatial risk level update mechanism adopts a prediction method based on a Markov chain state transition model combined with a dynamic risk gain function, specifically including the following steps: Construct a set of states, let the set of spatial risk states be . Each status corresponds to the risk level of one grid in the storage area, starting from the lowest level. To the highest level , This is the current state; The statistical state transition probability is calculated by extracting the number of state transitions from historical trajectory data and then calculating the state transition probability matrix, where each element represents the transition from state... Transition to state The probability of is calculated using the following expression: In the formula, From historical data, from the state Transition to state Number of times, From historical data, from the state Transition to state Number of times, These are elements in the static state transition probability matrix, representing the transition from state... Transition to state The probability of; A risk triggering factor is introduced, and its reinforcement coefficient for state transition is calculated based on the average current behavior risk score within each grid. The calculation expression is as follows: In the formula, It is the first All spatial grids corresponding to each state at time t Average behavioral risk score It is the maximum behavioral risk score threshold. It is the current time. The risk amplification coefficient; The dynamic state transition matrix is ​​calculated to construct a dynamically corrected state transition matrix, which is used to reflect the risk-inducing effect. The calculation expression is as follows: In the formula, These are elements in the dynamically corrected state transition matrix. It is the Kronecker function; The spatial risk state distribution for the next time step is updated using the state vector, calculated as follows: In the formula, It is the state of a certain spatial grid region in the next time step. The probability, It is the entire dynamically corrected state transition matrix. ; If the risk level corresponding to the updated state of any grid exceeds the set threshold ,Right now If the risk is high, a high-level early warning response will be triggered for that grid area, and the area will be marked as a high-risk zone in the risk heat map.

[0012] Preferably, the warning level label is divided into three levels: high risk, medium risk and low risk. The risk level is jointly assessed based on the location of the unsafe behavior, the number of people, the speed of the equipment, and historical risk records. High-risk behaviors include people entering restricted areas and coming into close proximity to forklifts; medium-risk behaviors include not wearing safety equipment or speeding; and low-risk behaviors include slightly crossing the line or making a brief stop. By constructing a multi-factor decision tree model, node conditions are determined for each type of behavior, thereby determining the corresponding warning level. The priority of sending warning information is sorted in order of level.

[0013] Preferably, the behavioral risk assessment adopts a weighted dynamic scoring mechanism, and the specific steps are as follows: Collect behavioral characteristic indicators and calculate the raw risk score. The calculation expression is as follows: In the formula, It is the rate of change of the velocity vector. It is the duration of the behavior. It is a relatively close distance. It is the frequency of crossover. It is a behavior category code. It is the original risk score. It is the weighting coefficient of the rate of change of the velocity vector. It is a weighting coefficient for the duration of the behavior. It is a weighting coefficient based on relative proximity. It is the weighting coefficient for crossover frequency. These are the weighting coefficients for behavior category encoding. satisfy ; Introducing a scenario adjustment factor, and weighting it by combining the current area's historical risk density and equipment density, the calculation expression is as follows: In the formula, It is a scene adjustment factor. It is the density of historical risks. It is equipment density. and These are all adjustment coefficients, used to control historical risk density. With equipment density Its influence in overall risk compensation, and meets the following requirements. ; The final risk score is output, calculated using the following formula: In the formula, It is the final risk score, used for subsequent warning level determination.

[0014] The technical effects and advantages provided by the present invention in the above technical solution are as follows: This invention effectively overcomes the problem of traditional systems struggling to identify dynamic interactive behaviors in complex warehousing scenarios by introducing a deep learning multi-object detection model and a temporal behavior modeling algorithm. The system can not only identify the trajectory information of personnel and equipment such as forklifts in real time, but also capture high-risk behavioral characteristics such as temporary personnel turning back or reversing direction, and achieve early intervention by combining behavioral intent analysis. Compared with existing early warning systems based on static rules, this invention significantly improves the accuracy and timeliness of behavior recognition, especially demonstrating stronger robustness in situations involving multiple object intersections and visual blind spots.

[0015] This invention constructs a spatial grid risk mapping model and introduces a behavioral risk scoring mechanism. This allows for real-time quantification of the risk level of each unsafe behavior, categorized into high, medium, and low levels. Simultaneously, a prediction mechanism based on Markov state transitions and time weighting is employed to model the risk evolution trend, enabling the early warning mechanism to go beyond reactive responses and predict potential hazards in advance. This tiered early warning mechanism ensures differentiated responses to different risky behaviors, improving the efficiency of the management system in handling safety hazards and the rationality of resource allocation.

[0016] This invention achieves comprehensive intelligent monitoring of the safety status of personnel and equipment in cigarette storage environments by constructing a closed-loop system encompassing image acquisition, behavior analysis, and intelligent early warning. The system is highly automated and adaptable, automatically adjusting its early warning strategy based on changes in the on-site environment and simultaneously triggering early warning information through multiple terminals (such as audio-visual equipment, a back-end platform, and mobile devices), ensuring "immediate alert upon detection." This intelligent linkage mechanism significantly reduces problems such as human error and missed reports, and delayed responses, fundamentally improving the inherent safety level and intelligent management capabilities of warehousing operations. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0018] Figure 1 This is a flowchart of the intelligent identification and graded early warning method for unsafe behaviors in cigarette storage environments according to the present invention. Detailed Implementation

[0019] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided so that the description of this disclosure will be more complete and fully convey the concept of the exemplary embodiments to those skilled in the art.

[0020] This invention provides, for example Figure 1 The intelligent identification and graded early warning method for unsafe behaviors in cigarette storage environments, as shown, includes the following steps: Multiple video acquisition devices are deployed in the cigarette storage area, fixedly installed at different angles and heights, to collect real-time image data of personnel and equipment in the area. The video acquisition device is an industrial-grade camera with infrared night vision capability and high frame rate function, with a frame rate of no less than 60 frames / second and a resolution of no less than 1080p, used to ensure clear and stable image acquisition of personnel and equipment in low light and high-speed motion conditions. When deploying the data acquisition device, the warehouse movement line is used as a reference, and the installation angle is controlled between 30 and 60 degrees to cover the high-frequency areas of forklift operations and high-risk activity areas of personnel. Overlapping coverage is also applied to the blind spots of the camera's field of view to achieve seamless continuity and regional complementarity of image data.

[0021] The collected image data is processed for multi-target detection. A trained deep learning model is used to identify the position and motion trajectory of personnel, forklifts and mobile devices, and the recognition results are output in a structured manner. The deep learning model adopts a multi-label classification structure with ResNet50 as the backbone network and introduces an attention mechanism module to enhance local features of human body posture, including the enhanced extraction of head orientation, hand movements, and walking posture details. Furthermore, the recognition model is retrained on a real-world cigarette storage environment dataset through transfer learning to adapt to application scenarios with complex backgrounds including stacked shelves, light reflection, and smoke interference, thereby improving the model's recognition accuracy and robustness in real-world storage environments.

[0022] Establish a behavior recognition module based on temporal feature analysis to perform correlation analysis on target trajectories in different time periods and identify high-risk behavior patterns including personnel turning back, sudden stops, and reverse crossing. The behavior recognition module adopts a sequence behavior modeling structure based on LSTM (Long Short-Term Memory Network) and combines a time window sliding strategy to extract a set of trajectory points of people and equipment within 5 consecutive seconds, with a sampling frequency of 10 times per second to construct dynamic trajectory sequence data; The system is trained and classified using multidimensional features such as trajectory change amplitude, velocity vector direction change rate, and path intersection density. It is used to identify whether there is an intention to turn back, cross, or stop suddenly, and to make logical inferences to determine whether it belongs to the defined "high-risk behavior template library".

[0023] A spatial risk mapping model is constructed to divide the warehouse area into regional grids, label the risk weight of each area, and compare the relative positional relationship and behavioral status of personnel and equipment in the grid in real time. The spatial risk mapping model adopts a regular grid structure, dividing the storage area into 1-meter × 1-meter square units. Each square is assigned a basic risk factor. The value of the basic factor is calculated based on historical accident density, equipment activity frequency and personnel density statistics. It also dynamically combines the current behavior status and interaction scenario to adjust the risk level in each square in real time. The regional heat map is updated every 60 seconds and the distribution of high-risk areas is displayed in real time with color gradients to guide people to avoid danger.

[0024] The space risk level update mechanism adopts a prediction method based on a Markov chain state transition model combined with a dynamic risk gain function, specifically including the following steps: Construct a set of states, let the set of spatial risk states be . Each status corresponds to the risk level of one grid in the storage area, starting from the lowest level. To the highest level , This is the current state; The statistical state transition probability is calculated by extracting the number of state transitions from historical trajectory data and then calculating the state transition probability matrix, where each element represents the transition from state... Transition to state The probability of is calculated using the following expression: In the formula, From historical data, from the state Transition to state The number of times is used as the statistical basis for constructing the state transition matrix. From historical data, from the state Transition to state Number of times, These are elements in the static state transition probability matrix, representing the transition from state... Transition to state The probability of; A risk triggering factor is introduced, and its reinforcement coefficient for state transition is calculated based on the average current behavior risk score within each grid. The calculation expression is as follows: In the formula, It is the first All spatial grids corresponding to each state at time t The average behavioral risk score, calculated by the downstream behavior recognition model, reflects the current risk level of the state. It is the maximum behavioral risk score threshold. It is the current time. The risk enhancement coefficient is used to increase the probability of transitioning to a high-risk state; The dynamic state transition matrix is ​​calculated to construct a dynamically corrected state transition matrix, which is used to reflect the risk-inducing effect. The calculation expression is as follows: In the formula, These are elements in the dynamically corrected state transition matrix, representing the transition from state [state name missing] after the introduction of the current risk enhancement factor. Transition to state The corrected probability, It is the Kronecker function, used to maintain the self-stability of non-risk states, when The value is 1 if the condition is met, otherwise it is 0. The spatial risk state distribution for the next time step is updated using the state vector, calculated as follows: In the formula, It is the state of a certain spatial grid region in the next time step. The probability, It is the entire dynamically corrected state transition matrix. ; If the risk level corresponding to the updated state of any grid exceeds the set threshold ,Right now If the risk is high, a high-level early warning response will be triggered for that grid area, and the area will be marked as a high-risk zone in the risk heat map.

[0025] Based on the preset hazard level determination rules, the identified unsafe behaviors are classified into levels according to behavior category, risk level and location of occurrence, and corresponding warning level labels are generated; The warning level label is divided into three levels: high risk, medium risk, and low risk. The risk level is jointly assessed based on the location of the unsafe behavior, the number of people, the speed of the equipment, and historical risk records. High-risk behaviors include people entering restricted areas and coming into close proximity to forklifts; medium-risk behaviors include not wearing safety equipment or speeding; and low-risk behaviors include slightly crossing the line or making a brief stop. By constructing a multi-factor decision tree model, node conditions are determined for each type of behavior, thereby determining the corresponding warning level. The priority of sending warning information is sorted in order of level.

[0026] Behavioral risk assessment employs a weighted dynamic scoring mechanism, with the following specific steps: Collect behavioral characteristic indicators and calculate the raw risk score. The calculation expression is as follows: In the formula, It is the velocity vector rate of change, which represents the degree of change in the direction of velocity of the monitored object (such as a person or forklift) per unit time. It measures whether there are abnormal changes in its movement trajectory, such as sharp turns or sudden reversals. This refers to the duration of the abnormal behavior, measured in seconds. A shorter duration indicates a more urgent need for response time. This refers to the relative proximity distance, which is the minimum spatial distance between an actor (such as a person) and a potential hazard (such as a forklift), measured in meters. The smaller the value, the closer the two are, and the higher the risk. This refers to the frequency of intersections, which indicates the number of times personnel and equipment trajectories cross each other within a unit time window (e.g., 5 seconds). The higher the frequency, the more chaotic the working environment and the higher the risk of collisions. It is a behavior category code, which means that each unsafe behavior category is assigned a fixed risk weight code. It is the original risk score. It is the weighting coefficient of the rate of change of the velocity vector. It is a weighting coefficient for the duration of the behavior. It is a weighting coefficient based on relative proximity. It is the weighting coefficient for crossover frequency. These are the weighting coefficients for behavior category encoding. satisfy ; Introducing a scenario adjustment factor, and weighting it by combining the current area's historical risk density and equipment density, the calculation expression is as follows: In the formula, It is a scenario adjustment factor used to introduce dynamic risk compensation for scenarios. Historical risk density indicates the frequency of unsafe events occurring in the grid area where the current action is located over a period of time. The higher the value, the more frequent the historical accidents and the higher the basic risk. Equipment density refers to the number of active forklifts and other mobile equipment per unit area in a given area. Higher density indicates higher traffic risk. and These are all adjustment coefficients, used to control historical risk density. With equipment density Its influence in overall risk compensation, and meets the following requirements. ; The final risk score is output, calculated using the following formula: In the formula, It is the final risk score, used for subsequent warning level determination.

[0027] For different levels of unsafe behavior, warning information is sent to on-site safety personnel and the management system through audible and visual alarms, mobile push notifications, and background pop-ups to achieve tiered intervention.

[0028] Implementation Method 1: This implementation method addresses the common problems in cigarette storage environments, such as confined spaces, numerous blind spots, and mixed operations involving equipment and personnel. It achieves accurate detection of personnel, forklifts, and other moving targets by deploying multi-angle, multi-type video acquisition devices and combining them with a deep learning image recognition model. The system deployment follows the principles of comprehensive coverage, complementary perspectives, and no blind spots to ensure the continuity and integrity of image data across the entire work area.

[0029] When deploying a video capture system in a warehouse, the first step is to plan the camera placements in detail based on the warehouse layout map. This typically includes key areas such as main aisles, intersections, areas near high-bay racking, and forklift turning points, where high-angle cameras should be installed to obtain a clear overhead view and avoid ground obstruction. Additionally, wide-angle cameras installed in corners should be used to supplement coverage in confined spaces or areas prone to obstruction. To ensure image quality at night and in low-light conditions, the cameras should have infrared illumination and low-light shooting capabilities.

[0030] Each camera is connected to an edge computing unit, responsible for initial data compression and real-time object detection processing. The system employs lightweight object detection algorithms such as YOLOv5. The pre-trained model is trained using transfer learning based on public datasets and combined with real-world warehouse samples, significantly improving object recognition accuracy. The object detection module can output structured information such as the position, size, movement trend, and category of each object in the image in real time. For example, it can detect a forklift with coordinates (x1, y1) moving northwest at a speed of 0.7 m / s.

[0031] To ensure system robustness, video data from multiple cameras will be fused using frame synchronization technology to construct a panoramic or multi-view stitched image covering the entire warehouse scene for global analysis. Through spatial mapping algorithms, images from each camera are uniformly converted into target locations within the warehouse map coordinate system, facilitating subsequent use by the behavior trajectory analysis module.

[0032] The system also needs to be configured with a video data caching mechanism to backtrack failed detection frames or conflicting analysis frames, and to use a sliding time window mechanism to save the most recent 15 seconds of image frame data for use in backtracking and locating abnormal events. Through the above deployment and processing, this implementation method can achieve accurate identification and stable tracking of people and vehicles in the working environment, providing a solid data foundation for the behavior recognition module.

[0033] Furthermore, to prevent information loss or keyframe omission during image acquisition, the system introduces an edge caching mechanism to asynchronously back up image frames. If the main channel camera experiences image quality degradation issues such as occlusion, blurring, or defocusing, the system can automatically call upon images from the suboptimal angle camera and perform image compensation by combining the target's predicted position with depth matching methods. For situations where the recognition results for the same object differ across multiple cameras, the system uses a Kalman filter-based fusion algorithm to fit and correct its position, ensuring that target recognition and trajectory continuity remain unaffected.

[0034] Meanwhile, the video acquisition system maintains a real-time connection with the cloud data center, uploading sampled image frames hourly for continuous model training and correction updates, enabling the model to continuously evolve in response to new changes in cigarette storage operations. The model periodically evaluates its recognition accuracy and automatically adjusts the threshold based on the actual false alarm rate, improving system stability and adaptability. In this embodiment, the entire image acquisition and target detection system significantly improves the perception quality of the pre-unsafe behavior recognition stage through multi-view and multi-model collaborative fusion, laying a solid technical foundation for building an intelligent recognition and accurate early warning system.

[0035] Implementation Method Two: This implementation method, based on trajectory data after image target detection, employs advanced sequence modeling technology to identify dynamic behavioral changes in personnel and equipment through time series learning, with a particular focus on identifying high-risk behavioral patterns and determining risk levels. By constructing a behavioral feature library and risk level model, it achieves dynamic analysis and risk classification of operational behaviors, providing technical support for graded early warning decision-making.

[0036] The system first constructs a vector sequence of parameters such as coordinates, velocity, orientation changes, and acceleration for each target over several consecutive frames (recommended to be no less than 5 seconds). After inputting these sequences into an LSTM model, the system uses model training to identify whether the target exhibits abnormal behavior patterns, such as sudden changes in speed, abrupt changes in direction, or frequent head-turning within a short distance.

[0037] Simultaneously, the system establishes a high-risk behavior template library, including typical high-risk behaviors such as rapid personnel retreat, personnel crossing at an angle greater than 90 degrees with forklifts, and personnel entering the forklift reversing path. These are all modeled as standard behavior templates using time-series features. The identification module matches the target sequence with the behavior templates in real time, using cosine similarity or dynamic time warping (DTW) algorithms to determine the degree of matching. When the similarity of a sequence exceeds a risk threshold, the behavior is considered to have occurred.

[0038] In addition, to accurately assess risk levels, the system introduces a multi-factor scoring mechanism: for example, an intersection angle greater than 60 degrees earns 5 points, a distance less than 0.5 meters earns 3 points, and an additional 2 points are awarded if the incident occurred in a high-frequency accident area. The final score is mapped to three risk levels: high, medium, and low. Each identified behavior is assigned a risk level label and enters the early warning decision-making module for subsequent response.

[0039] To further enhance the system's adaptability in real-world scenarios, the behavior recognition module can be configured with a dynamic learning mode. Based on changes in behavioral characteristics across different shifts and time periods, the system is periodically guided to re-label behavior samples and update model parameters using incremental training, ensuring stable long-term application performance. Simultaneously, the system can integrate a scene recognition module to automatically identify the current work area scene type (such as stacking areas, forklift aisles, loading and unloading areas, etc.), defining specific behavior recognition templates and risk thresholds for different areas. This enables dynamic switching of behavior recognition logic, improving the system's adaptability to various scenarios.

[0040] After the behavior recognition results are output, the system will automatically cross-reference historical behavior records with the worker's historical safety record. If a worker is found to have repeatedly exhibited medium- to high-risk behaviors with time intervals below a set threshold, the system will classify the worker's behavior as "suspicious high-frequency risk behavior" and report it to the warehouse safety management backend to support managers in making personalized intervention decisions. Simultaneously, the system also supports outputting the recognition results to a knowledge graph platform, serving as an important data source for subsequent warehouse safety situation assessments and operational optimization analyses.

[0041] Implementation Method 3: This implementation method focuses on modeling and predicting the dynamic changes in spatial risks in warehousing operation sites. It constructs a refined spatial grid structure and combines a Markov state transition model with historical data analysis to dynamically assess the risk evolution trends in each area. The system adjusts the grid state based on behavior recognition results and uses the risk change prediction results to drive real-time early warning responses, thereby achieving dynamic intervention and proactive prevention and control.

[0042] After completing behavior recognition, in order to link behavioral risks with the warehouse space, the system needs to construct a spatial risk assessment framework. This implementation uses a regular grid structure to physically divide the entire warehouse area, with each grid being a 1m x 1m square unit, initially set to low risk. Based on historical incident data, the system assigns an initial risk factor to each grid; for example, a high-risk area has an initial value of 0.6, a normal area 0.2, and an extremely low-risk area 0.

[0043] The system updates the status of each grid in real time, based on dynamic factors including target density, target category, and target behavior level within the current grid. For example, if a forklift speeds through a grid and personnel enter, the system will immediately overlay a high-risk behavior factor, update the grid status to "high-risk," and trigger a visual heatmap display highlighted in red.

[0044] To predict the development trend of risk areas, the system introduces a Markov chain model for state prediction. Let the current state of each grid be S_t, and the state at the next time step be controlled by the transition probability matrix P. The system constructs the transition probability by statistically analyzing the transition frequency between each state using historical trajectory data. For example, the probability of a grid transitioning from a "medium-risk" state to a "high-risk" state is 0.35, and the probability of transitioning from a "low-risk" state to a "medium-risk" state is 0.6.

[0045] Regarding the early warning and feedback mechanism, this implementation not only provides an audible and visual alarm device, but also integrates IoT smart tags and wearable devices (such as smart safety helmets, vibration bracelets, etc.). When a worker enters a high-risk grid area and engages in suspected unsafe behavior, the system can send vibration or voice prompts to the worker's device via Bluetooth or UWB signals, prompting the worker to evacuate immediately or adjust their work posture.

[0046] Meanwhile, the management backend will collect heat map information from the entire area in real time and dynamically update the visualization interface based on the results of the predictive model. Managers can clearly view the concentrated risk distribution areas and their evolution trends on a large screen, and conduct personnel evacuation or task rescheduling. All risk events, trajectory changes, and early warning response records will be logged every minute and stored long-term in the security audit database for various purposes such as later review, security scoring, and management assessment, helping warehouse security move from passive prevention to proactive governance.

[0047] This invention effectively overcomes the problem of traditional systems struggling to identify dynamic interactive behaviors in complex warehousing scenarios by introducing a deep learning multi-object detection model and a temporal behavior modeling algorithm. The system can not only identify the trajectory information of personnel and equipment such as forklifts in real time, but also capture high-risk behavioral characteristics such as temporary personnel turning back or reversing direction, and achieve early intervention by combining behavioral intent analysis. Compared with existing early warning systems based on static rules, this invention significantly improves the accuracy and timeliness of behavior recognition, especially demonstrating stronger robustness in situations involving multiple object intersections and visual blind spots.

[0048] This invention constructs a spatial grid risk mapping model and introduces a behavioral risk scoring mechanism. This allows for real-time quantification of the risk level of each unsafe behavior, categorized into high, medium, and low levels. Simultaneously, a prediction mechanism based on Markov state transitions and time weighting is employed to model the risk evolution trend, enabling the early warning mechanism to go beyond reactive responses and predict potential hazards in advance. This tiered early warning mechanism ensures differentiated responses to different risky behaviors, improving the efficiency of the management system in handling safety hazards and the rationality of resource allocation.

[0049] This invention achieves comprehensive intelligent monitoring of the safety status of personnel and equipment in cigarette storage environments by constructing a closed-loop system encompassing image acquisition, behavior analysis, and intelligent early warning. The system is highly automated and adaptable, automatically adjusting its early warning strategy based on changes in the on-site environment and simultaneously triggering early warning information through multiple terminals (such as audio-visual equipment, a back-end platform, and mobile devices), ensuring "immediate alert upon detection." This intelligent linkage mechanism significantly reduces problems such as human error and missed reports, and delayed responses, fundamentally improving the inherent safety level and intelligent management capabilities of warehousing operations.

[0050] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0051] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

[0052] It should be noted that, in this document, the use of relational terms such as "first" and "second" is merely for distinguishing one entity or operation from another, and does not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0053] It should be understood that in the various embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0054] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0055] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0056] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0057] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0058] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0059] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

Claims

1. A method for intelligent identification and graded early warning of unsafe behaviors in cigarette storage environments, characterized in that, Includes the following steps: Multiple video acquisition devices are deployed in the cigarette storage area, fixedly installed at different angles and heights, to collect real-time image data of personnel and equipment in the area. The collected image data is processed for multi-target detection. A trained deep learning model is used to identify the position and motion trajectory of personnel, forklifts and mobile devices, and the recognition results are output in a structured manner. Establish a behavior recognition module based on temporal feature analysis to perform correlation analysis on target trajectories in different time periods and identify high-risk behavior patterns including personnel turning back, sudden stops, and reverse crossing. A spatial risk mapping model is constructed to divide the warehouse area into regional grids, label the risk weight of each area, and compare the relative positional relationship and behavioral status of personnel and equipment in the grid in real time. Based on the preset hazard level determination rules, the identified unsafe behaviors are classified into levels according to behavior category, risk level and location of occurrence, and corresponding warning level labels are generated; For different levels of unsafe behavior, warning information is sent to on-site safety personnel and the management system through audible and visual alarms, mobile push notifications, and background pop-ups to achieve tiered intervention.

2. The intelligent identification and graded early warning method for unsafe behaviors in cigarette storage environments according to claim 1, characterized in that, The video acquisition device is an industrial-grade camera with infrared night vision capability and high frame rate function, with a frame rate of no less than 60 frames / second and a resolution of no less than 1080p, used to ensure clear and stable image acquisition of personnel and equipment in low light and high-speed motion conditions. When deploying the data acquisition device, the warehouse movement line is used as a reference, and the installation angle is controlled between 30 and 60 degrees to cover the high-frequency areas of forklift operations and high-risk activity areas of personnel. Overlapping coverage is also applied to the blind spots of the camera's field of view to achieve seamless continuity and regional complementarity of image data.

3. The intelligent identification and graded early warning method for unsafe behaviors in cigarette storage environments according to claim 1, characterized in that, The deep learning model adopts a multi-label classification structure with ResNet50 as the backbone network and introduces an attention mechanism module to enhance local features of human body posture, including the enhanced extraction of head orientation, hand movements, and walking posture details. Furthermore, the recognition model is retrained on a real-world cigarette storage environment dataset through transfer learning to adapt to application scenarios with complex backgrounds including stacked shelves, light reflection, and smoke interference, thereby improving the model's recognition accuracy and robustness in real-world storage environments.

4. The intelligent identification and graded early warning method for unsafe behaviors in cigarette storage environments according to claim 1, characterized in that, The behavior recognition module adopts a sequence behavior modeling structure based on LSTM and combines a time window sliding strategy to extract a set of trajectory points of people and equipment within 5 consecutive seconds, with a sampling frequency of 10 times per second to construct dynamic trajectory sequence data; The system is trained and classified using multidimensional features such as trajectory change amplitude, velocity vector direction change rate, and path intersection density. It is used to identify whether there is an intention to turn back, cross, or stop suddenly, and to perform logical reasoning to determine whether it belongs to the defined "high-risk behavior template library".

5. The intelligent identification and graded early warning method for unsafe behaviors in cigarette storage environments according to claim 1, characterized in that, The spatial risk mapping model adopts a regular grid structure, dividing the storage area into 1-meter × 1-meter square units. Each square is assigned a basic risk factor. The value of the basic factor is calculated based on historical accident density, equipment activity frequency and personnel density statistics. It also dynamically combines the current behavior status and interaction scenario to adjust the risk level in each square in real time. The regional heat map is updated every 60 seconds and the distribution of high-risk areas is displayed in real time with color gradients to guide people to avoid danger.

6. The intelligent identification and graded early warning method for unsafe behaviors in cigarette storage environments according to claim 1, characterized in that, The space risk level update mechanism adopts a prediction method based on a Markov chain state transition model combined with a dynamic risk gain function, specifically including the following steps: Construct a set of states, let the set of spatial risk states be . Each status corresponds to the risk level of one grid in the storage area, starting from the lowest level. To the highest level , This is the current state; The statistical state transition probability is calculated by extracting the number of state transitions from historical trajectory data and then calculating the state transition probability matrix, where each element represents the transition from state... Transition to state The probability is calculated using the following expression: In the formula, From historical data, from the state Transition to state Number of times, From historical data, from the state Transition to state Number of times, These are elements in the static state transition probability matrix, representing the transition from state... Transition to state The probability of; A risk triggering factor is introduced, and its reinforcement coefficient for state transition is calculated based on the average current behavior risk score within each grid. The calculation expression is as follows: In the formula, It is the first All spatial grids corresponding to each state at time t Average behavioral risk score It is the maximum behavioral risk score threshold. It is the current time. The risk amplification coefficient; The dynamic state transition matrix is ​​calculated to construct a dynamically corrected state transition matrix, which is used to reflect the risk-inducing effect. The calculation expression is as follows: In the formula, These are elements in the dynamically corrected state transition matrix. It is the Kronecker function; The spatial risk state distribution for the next time step is updated using the state vector, calculated as follows: In the formula, It is the state of a certain spatial grid region in the next time step. The probability, It is the entire dynamically corrected state transition matrix. ; If the risk level corresponding to the updated state of any grid exceeds the set threshold ,Right now If the risk is high, a high-level early warning response will be triggered for that grid area, and the area will be marked as a high-risk zone in the risk heat map.

7. The intelligent identification and graded early warning method for unsafe behaviors in cigarette storage environments according to claim 1, characterized in that, The warning level label is divided into three levels: high risk, medium risk, and low risk. The risk level is jointly assessed based on the location of the unsafe behavior, the number of people, the speed of the equipment, and historical risk records. High-risk behaviors include people entering restricted areas and coming into close proximity to forklifts; medium-risk behaviors include not wearing safety equipment or speeding; and low-risk behaviors include slightly crossing the line or making a brief stop. By constructing a multi-factor decision tree model, node conditions are determined for each type of behavior, thereby determining the corresponding warning level. The priority of sending warning information is sorted in order of level.

8. The intelligent identification and graded early warning method for unsafe behaviors in cigarette storage environments according to claim 1, characterized in that, Behavioral risk assessment employs a weighted dynamic scoring mechanism, with the following specific steps: Collect behavioral characteristic indicators and calculate the raw risk score. The calculation expression is as follows: In the formula, It is the rate of change of the velocity vector. It is the duration of the behavior. It is a relatively close distance. It is the frequency of crossover. It is a behavior category code. It is the original risk score. It is the weighting coefficient of the rate of change of the velocity vector. It is a weighting coefficient for the duration of the behavior. It is a weighting coefficient based on relative proximity. It is the weighting coefficient for crossover frequency. These are the weighting coefficients for behavior category encoding. satisfy ; Introducing a scenario adjustment factor, and weighting it by combining the current area's historical risk density and equipment density, the calculation expression is as follows: In the formula, It is a scene adjustment factor. It is the density of historical risks. It is equipment density. and These are all adjustment coefficients, used to control historical risk density. With equipment density Its influence in overall risk compensation, and meets the following requirements. ; The final risk score is output, calculated using the following formula: In the formula, It is the final risk score, used for subsequent warning level determination.

Images