A waterfront landscape belt safety early warning system based on behavior recognition

Abstract

This invention relates to the field of waterfront landscape belt safety early warning technology, specifically disclosing a waterfront landscape belt safety early warning system based on behavior recognition. The system includes a data acquisition module, comprising an environmental sensing unit, a visual perception unit, and a wireless positioning unit. The environmental sensing unit is used to collect environmental data of the waterfront landscape belt, while the visual perception unit and the wireless positioning unit are both used to collect visitor target perception data and transmit the collected data to the corresponding associated modules. This invention achieves graded and precise early warning intervention based on fused risk and behavioral information. This solution systematically solves the core defects of existing technologies, such as delayed, generalized, and untargeted early warning methods, as well as the disconnect between environmental risks and personnel behavior, significantly improving the intelligence level and proactive early warning capability of waterfront landscape belt safety management.

Description

Technical Field

[0001] This invention relates to the field of waterfront landscape safety early warning technology, specifically a waterfront landscape safety early warning system based on behavior recognition. Background Technology

[0002] As an important component of the urban public space system, waterfront landscape belts not only meet the daily needs of citizens for leisure, recreation, sightseeing, and social interaction, but also play an irreplaceable role in enhancing the city's image, improving the ecological environment, and promoting the vitality of waterfront areas. These areas typically include various open spaces such as waterfront platforms, riverside walkways, and scenic boardwalks. Due to their waterfront characteristics, they are highly favored by the public, with dense crowds and diverse activities on weekdays and holidays.

[0003] In traditional safety management practices, waterfront areas primarily rely on basic methods such as setting up fixed warning signs and arranging regular patrols. With the development of technology, environmental sensing devices such as water level sensors and meteorological monitoring stations have been gradually introduced to collect basic parameters such as water level and wind speed, and attempts have been made to set fixed thresholds to trigger audible and visual alarms. Simultaneously, some locations have also deployed video surveillance systems to monitor pedestrian flow and detect simple boundary violations in specific restricted areas. However, these traditional or rudimentary technological management methods have significant limitations: environmental monitoring can only provide point-based or single-factor data, making it difficult to integrate multi-source information such as tides, waves, and weather to dynamically and accurately assess and spatially present changes in the risk level of the entire waterfront area; video surveillance is often limited to detecting the presence of people and making rough location judgments, failing to deeply identify and understand specific behavioral patterns of visitors. This results in existing early warning methods often being delayed and generalized, making it difficult to achieve accurate, timely, and differentiated early warning interventions for the potential dangerous behaviors of specific individuals in high-risk areas under complex and dynamic risk environments. The intelligence and proactivity of safety management are severely lacking. Summary of the Invention

[0004] The purpose of this invention is to provide a waterfront landscape safety early warning system based on behavior recognition, and to solve the following technical problems:

[0005] The question is how to achieve intelligent perception and collaborative early warning of dynamic environmental risks and tourist behavior in waterfront areas.

[0006] The objective of this invention can be achieved through the following technical solutions:

[0007] A behavior recognition-based safety early warning system for waterfront landscape zones, the system comprising:

[0008] The data acquisition module includes an environmental sensing unit, a visual perception unit, and a wireless positioning unit. The environmental sensing unit is used to collect environmental data of the waterfront landscape belt, and the visual perception unit and the wireless positioning unit are both used to collect visitor target perception data and transmit the collected data to the corresponding associated modules.

[0009] The environmental risk analysis module is used to receive hydrological and meteorological data transmitted by the data acquisition module, calculate and output the dynamic risk field of the waterfront area based on real-time and predicted environmental data, and the dynamic risk field divides the waterfront area into areas with different risk levels.

[0010] The behavior state recognition module is used to receive tourist target perception data transmitted by the data acquisition module and dynamic risk field information output by the environmental risk analysis module. Based on the risk level area defined by the dynamic risk field, it prioritizes to start behavior recognition analysis for tourist targets in high-risk areas and outputs a structured state description vector containing tourist behavior category and attention state level.

[0011] The early warning decision module is used to generate early warning instructions of different levels based on the risk level area output by the environmental risk analysis module and the structured state description vector output by the behavior state identification module, and to control the execution terminals deployed in the landscape belt to perform intervention operations.

[0012] Furthermore, the environmental risk analysis module includes:

[0013] The regional division unit is used to dynamically divide the waterfront area of ​​the waterfront landscape belt into core risk zone, extended warning zone and safe zone based on real-time and predicted data of tides, waves and wind speed, combined with a preset geographic information model.

[0014] The core risk area is the area that will be submerged by the tide or directly impacted by large waves in the current or predicted short term.

[0015] The extended early warning zone is the area adjacent to the core risk zone that is expected to be affected by the spread of risk within a set future timeframe.

[0016] The safe zone is an area that has no direct hydrological risk in the current and foreseeable future period;

[0017] The core risk zone, extended early warning zone, and safe zone each correspond to different perception parameter configurations. The area division unit synchronously outputs perception parameter adjustment instructions for each area to drive the data acquisition module to dynamically adjust its data acquisition strategy.

[0018] Furthermore, the behavior state recognition module includes:

[0019] The target screening unit is used to screen targets located in the core risk area and the extended warning area from all perceived tourist targets based on the regional division results of the dynamic risk field, and mark the targets as priority identification objects;

[0020] The deep recognition unit is configured to initiate machine learning-based behavior recognition analysis on the priority recognition object based on the attention state determination model. The behavior recognition analysis includes individual deep behavior recognition and group situation analysis.

[0021] The attention state determination model is used at least to quantify the target's attention distraction index using the following formula:

[0022]

[0023] in, The Distraction Index; For the first The identification results of predefined attention-distraction behaviors; For the first Weight coefficients corresponding to class behaviors; The total number of behavior categories. ; The normalized population density value for the local area where the target is located; The duration of the target under a pre-defined dangerous orientation; The preset standard tracking duration; The angle between the target body's orientation and the direction of the water; For A function with independent variables used to quantify the orientation risk; , , , These are the adjustment coefficients for each item.

[0024] Furthermore, the depth recognition unit is further configured as follows:

[0025] For all targets located in the core risk area, continuous behavior identification and status tracking are conducted;

[0026] For targets located in the extended warning zone, the identification frequency is dynamically adjusted based on the initial judgment of their behavior and movement trend, with an increased identification frequency for targets moving toward the core risk zone or remaining stationary.

[0027] Furthermore, the early warning decision module includes:

[0028] The comprehensive risk assessment unit receives the risk level of the target's location and the corresponding attention distraction index, and obtains the target's comprehensive risk score through the following coupled calculation formula:

[0029]

[0030] in, For comprehensive risk scoring; This is the basic hydrological risk value of the target's location calculated based on the environmental risk analysis module. This is the predicted rate of expansion of the risk area towards the target location; The attention distraction index is calculated by the behavior state recognition module. , These are the weighting coefficients for the basic hydrological risk value and the expansion rate, respectively; This is the risk amplification factor for the attention distraction factor.

[0031] Furthermore, the early warning decision module, based on the early warning decision model, also includes:

[0032] The strategy execution unit is configured to perform the action based on the comprehensive risk score. Based on the given numerical range, implement a tiered intervention strategy:

[0033] when When the value is in the first range, a level one warning is triggered, and an area broadcast reminder is made through nearby fixed audible and visual alarm terminals;

[0034] when When the value is in the second range, a level two warning is triggered, and mobile active guidance terminals are dispatched to provide close-range, directional audio-visual warnings to specific targets;

[0035] when When the value is in the third range, a level 3 warning is triggered. While executing the level 2 warning, an alarm message containing the precise location of the target and the scene image is pushed to the nearest personnel collaborative handling terminal.

[0036] Wherein, the lower limit of the third numerical interval is higher than that of the second numerical interval, and the lower limit of the second numerical interval is higher than that of the first numerical interval;

[0037] The policy execution unit is further configured as follows:

[0038] After the warning is triggered, the intervention effect data is monitored and recorded, and the intervention effect data is fed back to the preset data feedback interface;

[0039] Among them, differentiated optimization processing priorities are set for the intervention effect data of different levels of early warning.

[0040] Furthermore, the system also includes:

[0041] The data feedback optimization module is used to receive the early warning intervention effect data transmitted through the preset data feedback interface, form a training sample set, and optimize the parameters of the attention state determination model in the behavior state recognition module and the early warning decision model in the early warning decision module based on the training sample set.

[0042] Furthermore, the steps for optimizing the parameters of the early warning decision model include:

[0043] Establish an early warning effectiveness evaluation function, which is based on a comprehensive risk score. With the final safety outcome The degree of matching is calculated, and the final security result is determined. A quantitative value defined based on the final safety status of tourists;

[0044] By minimizing the loss value of the early warning effect evaluation function, the weight coefficients in the comprehensive risk score calculation formula are iteratively updated. , and risk amplification factor .

[0045] Furthermore, the step of optimizing the parameters of the attention state determination model includes: adjusting the behavioral weight coefficients in the attention distraction index calculation formula. Optimization;

[0046] During the optimization process, specific behavior categories are calculated. With the final safety outcome Statistical correlation between And according to the following formula, the weighting coefficients Make dynamic adjustments:

[0047]

[0048] in, and These are the weight coefficients before and after optimization, respectively; The learning rate; For the first Normalized correlation degree of class behavior; This refers to the system's expected safety value predicted based on the current parameters, where the current parameters are the behavioral weight coefficients before optimization. .

[0049] The beneficial effects of this invention are:

[0050] (1) This invention achieves synchronous perception of environmental parameters and tourist information by deploying a data acquisition module; overcomes the limitations of traditional single-point and static monitoring by using an environmental risk analysis module, and can generate a dynamically changing spatial risk field by integrating multi-source environmental information; overcomes the shortcomings of existing video surveillance which can only perform presence detection or boundary judgment by using a behavior state recognition module, and can deeply identify tourist behavior patterns and attention states, and associate them with the dynamic risk field; finally, the early warning decision module realizes graded and precise early warning intervention based on the fused risk and behavior information. This solution systematically solves the core defects of existing technologies, such as the lagging, generalized, and untargeted early warning methods, as well as the disconnect between environmental risks and personnel behavior states, and significantly improves the intelligent level and proactive early warning capability of waterfront landscape belt safety management.

[0051] (2) This invention collects actual effect data of early warning intervention, constructs early warning effect evaluation function and safety result correlation analysis, and iteratively optimizes the parameters of the decision model and the behavior weight coefficients in the behavior recognition model, so that the system can continuously learn from actual operation experience, automatically calibrate the quantitative evaluation standards of environmental risk and personnel behavior risk, thereby continuously improving the accuracy and reliability of early warning, realizing the evolution from static rule system to dynamic intelligent system, and ensuring long-term effective safety protection capabilities. Attached Figure Description

[0052] The invention will now be further described with reference to the accompanying drawings.

[0053] Figure 1 This is a schematic diagram of a waterfront landscape safety early warning system based on behavior recognition proposed in this invention. Detailed Implementation

[0054] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0055] Please see Figure 1 As shown, in one embodiment, a waterfront landscape safety early warning system based on behavior recognition is provided, the system comprising:

[0056] The data acquisition module includes an environmental sensing unit, a visual perception unit, and a wireless positioning unit. The environmental sensing unit is used to collect hydrological and meteorological environmental data of the waterfront landscape belt. The visual perception unit is used to collect visual perception data of tourists through images or videos. The wireless positioning unit is used to obtain the location information of tourists and transmit the collected data to the corresponding associated module.

[0057] The environmental risk analysis module is used to receive hydrological and meteorological data transmitted by the data acquisition module, and calculate and output a dynamic risk field of the safety status of the waterfront area based on real-time and predicted environmental data and a preset environmental risk model. The dynamic risk field dynamically divides the waterfront area into different risk level zones in space.

[0058] The behavior state recognition module is used to receive tourist target perception data transmitted by the data acquisition module and dynamic risk field information output by the environmental risk analysis module. Based on the risk level area defined by the dynamic risk field, the priority of the tourist target to be analyzed is determined, and behavior recognition analysis is initiated for tourist targets in high-risk areas first. The behavior recognition analysis includes at least the determination of individual tourist behavior categories and the assessment of tourist attention state, and outputs a structured state description vector containing tourist behavior category and attention state level.

[0059] The early warning decision module is used to assess the comprehensive risk of specific tourist targets based on the risk level area output by the environmental risk analysis module and the structured state description vector output by the behavior state recognition module. Based on this, it generates early warning instructions of different levels and controls various execution terminals deployed in the landscape belt to perform differentiated safety operations ranging from broadcast warnings to targeted interventions.

[0060] Through the above technical solutions, this embodiment provides a waterfront landscape belt safety early warning system based on behavior recognition. The system achieves synchronous perception of environmental parameters and visitor information by deploying a data acquisition module; overcomes the limitations of traditional single-point, static monitoring through an environmental risk analysis module, and can generate a dynamically changing spatial risk field by integrating multi-source environmental information; and overcomes the shortcomings of existing video surveillance, which can only perform presence detection or boundary judgment, through a behavior state recognition module, which can deeply identify visitor behavior patterns and attention states and correlate them with the dynamic risk field; finally, the early warning decision module realizes graded and precise early warning intervention based on the fused risk and behavior information. This solution systematically solves the core defects of existing technologies, such as lagging, generalized, and untargeted early warning methods, as well as the disconnect between environmental risks and personnel behavior states, and significantly improves the intelligent level and proactive early warning capability of waterfront landscape belt safety management.

[0061] In one embodiment, the environmental risk analysis module includes:

[0062] The regional division unit is used to dynamically divide the waterfront area of ​​the waterfront landscape belt into core risk zone, extended warning zone and safe zone based on real-time and predicted data of tides, waves and wind speed, combined with a preset geographic information model that includes topography, shoreline and facility layout, through spatial overlay and risk propagation algorithms.

[0063] The core risk area refers to the area that is currently submerged by tides or directly impacted by large waves, based on current hydrological and meteorological conditions, or the area that, based on forecast data, is expected to experience the above situation within a short time window (e.g., within the next 5-15 minutes). This area represents the highest immediate risk.

[0064] The extended warning zone refers to the area that is directly adjacent to the core risk zone in space. Although there is no direct danger in this area at the present moment, according to the risk contagion model, it is highly likely to be affected by the expansion of the core risk zone in a set long time window in the future (e.g., within the next 15-60 minutes).

[0065] The safe zone refers to an area that, based on all environmental parameters, poses no direct hydrological or meteorological risk at the present moment and in the foreseeable future.

[0066] The core risk zone, extended early warning zone, and safe zone each correspond to different perception parameter configurations. After the area division unit completes the division, it will synchronously generate and send the corresponding perception parameter adjustment instructions for each area to the data acquisition module. For example, the instructions may require increasing the frame rate and resolution of video surveillance in the core risk zone, or increasing the scanning frequency of wireless positioning signals in the extended early warning zone, thereby driving the data acquisition module to dynamically optimize its data acquisition strategy according to the risk distribution and realize intelligent scheduling of computing resources.

[0067] The behavior state recognition module includes:

[0068] The target screening unit is used to screen out targets located in the core risk area and the extended warning area from the tourist targets sensed in real time by the data acquisition module based on the regional division results of the dynamic risk field. These screened targets will be given a higher processing priority and marked as priority identification objects for subsequent in-depth analysis unit to focus on processing.

[0069] The deep recognition unit is configured to initiate behavior recognition analysis based on machine learning models (such as convolutional neural networks and temporal behavior analysis models) on the priority recognition object based on the attention state determination model. The behavior recognition analysis includes individual deep behavior recognition and group situation analysis. Individual deep behavior recognition refers to recognizing the specific actions of individual tourists, such as running, falling, sitting still for a long time, leaning on railings, etc. Group situation analysis refers to analyzing the group characteristics such as the degree of crowd gathering and the consistency of movement speed in the local area around the target.

[0070] The attention state determination model is used at least to quantify the target's attention distraction index using the following formula. :

[0071]

[0072] in, The Distraction Index; For the first The identification results of predefined distracting behaviors can be obtained by using a behavior recognition model to detect the i-th predefined distracting behavior (such as operating a mobile phone, talking to others for a long time, walking with your head down, etc.) and then binarizing the result (e.g., "1" indicates occurrence, "0" indicates non-occurrence). For the first The weight coefficients corresponding to different behaviors are weights set based on experience or historical data, and are used to measure the degree of contribution of different distracting behaviors to the overall distraction. This refers to the total number of behavior categories, that is, the total number of attention-distracting behavior categories predefined by the system. ; The normalized crowd density value of the local area where the target is located is the ratio of the instantaneous number of tourists in the local area where the target is located (such as within a radius of 3 meters) to the maximum capacity of the area, reflecting the potential interference of crowd congestion on individual attention. The duration of the target under a preset dangerous orientation, which is generally defined as the safe orientation of facing away from or to the side of the water. The preset standard tracking duration is used for normalization. The time base is set based on experience, for example, 60 seconds; The angle between the target body's orientation and the direction it points towards the water is obtained through visual positioning or sensor fusion technology. For A function with independent variables used to quantify the differences in risk arising from different orientations, for example, can be defined as when... When the angle exceeds a certain threshold (such as 90 degrees), the function value increases, indicating a higher risk when facing away from the water. , , , These are the adjustment coefficients for each item, which are usually set during system initialization based on scenario optimization, and can be further optimized later using feedback data.

[0073] The depth recognition unit is further configured as follows:

[0074] For all targets located in the core risk area, conduct continuous and high-frequency behavior identification and status tracking to ensure that no risk changes are overlooked;

[0075] For targets located in the extended warning zone, the identification frequency is dynamically adjusted based on the initial judgment of their behavior and movement trend (such as calculating the movement vector through positioning data). The identification frequency is increased for targets moving towards the core risk zone or those that are stationary, allowing for closer monitoring. For targets that are clearly moving away from the safe zone, the identification frequency can be appropriately reduced to save system resources.

[0076] Through the above technical solutions, this embodiment provides a refined, adaptive, and quantifiable core analysis method for safety early warning of waterfront landscape zones. The method achieves precise allocation of computing resources by dynamically dividing risk areas and intelligently filtering monitoring targets accordingly. By constructing an attention distraction index calculation formula that integrates individual behavior, group situation, and multi-dimensional environmental factors, the judgment of tourists' risk status is elevated from subjective experience to an objective quantitative level. This method not only responds to dynamic environmental changes but also achieves a deep perception of tourists' inherent risk status, thereby systematically improving the timeliness, accuracy, and intelligence level of early warning.

[0077] In one embodiment, the early warning decision module includes:

[0078] The comprehensive risk assessment unit receives the risk level of the target's location and the corresponding attention distraction index, and obtains the target's comprehensive risk score through the following coupled calculation formula:

[0079]

[0080] in, The comprehensive risk score is used to determine the overall level of danger faced by the target. The higher the score, the higher the overall level of danger faced by the target. This score is the core basis for triggering a graded early warning. This is a scalar value calculated by the environmental risk analysis module based on the specific geographical location of the target and combined with hydrological data such as real-time tides, wave heights, and water flow speeds. It reflects the inherent environmental hazard level of the location itself, excluding human behavior factors, and is often obtained through hydrological model simulation or mapping of historical disaster data. This is a predicted rate of expansion of the risk area towards the target location, derived by the environmental risk analysis module based on historical data and real-time trend predictions. It represents the predicted speed at which the risk area (such as the boundary of the core risk area) where the target is located spreads or approaches the target location. This value reflects the dynamic nature and urgency of the hazard. When the target is located within the risk area... It can take the value of zero or a value representing the intensity of the internal hazard; The attention distraction index is calculated by the behavior state recognition module. , These are the weighting coefficients for the basic hydrological risk value and the expansion rate, respectively. They are typically initially set during system deployment based on the characteristics of the specific waterfront area (such as terrain steepness and mainstream risk type). For example, if the area is prone to sudden surges, V may be assigned a higher weighting. ; This is the risk amplification coefficient for the attention distraction factor, a coefficient greater than 0, used to adjust the attention distraction index. The amplifying effect on the final overall risk occurs when tourists' attention is distracted ( When the value is large, its ability to cope with sudden risks decreases. Therefore, even when facing the same external environmental risks, the actual comprehensive risks it faces should be magnified. The value is obtained through historical accident data analysis or system feedback learning, and it quantifies the coupling strength of danger between behavioral factors and environmental factors.

[0081] The early warning decision module, based on the early warning decision model, also includes:

[0082] The strategy execution unit is configured to perform the action based on the comprehensive risk score. Based on the given numerical range, implement a tiered intervention strategy:

[0083] when When the value is in the first range, a level one warning is triggered, and a regional, non-targeted safety broadcast reminder is given through nearby fixed sound and light alarm terminals (such as loudspeakers and warning light poles). This is suitable for scenarios where risks are initially apparent and need to be widely communicated.

[0084] when When the value is in the second range, a level 2 warning is triggered, and mobile active guidance terminals (such as patrol robots and drones) or steerable fixed-point devices are dispatched to the vicinity of the target to provide close-range, directional audio-visual warnings. The intervention is more targeted and aims to attract the attention of specific high-risk individuals.

[0085] when When the value is in the third range, a level 3 warning is triggered. While executing the level 2 warning, an alarm message containing the precise location coordinates of the target and real-time scene images will be immediately pushed to the management backend and the nearest personnel collaborative handling terminal (such as security personnel handheld terminals or duty room systems) to request manual intervention and emergency handling. This is applicable to the highest level of urgent danger.

[0086] Wherein, the lower limit of the third numerical interval is higher than that of the second numerical interval, and the lower limit of the second numerical interval is higher than that of the first numerical interval;

[0087] The policy execution unit is further configured as follows:

[0088] After triggering any level of warning and executing intervention, monitor and record intervention effect data (such as whether the target leaves the risk area, whether the behavior changes, warning response time, etc.), and feed the intervention effect data back to the preset data feedback interface;

[0089] Among them, for intervention effect data of different levels of early warning, different optimization processing priorities can be set. For example, the feedback data of the Level 3 early warning involves manual handling and may have a higher priority for model calibration.

[0090] Through the above technical solution, this embodiment provides a multi-level, quantifiable, and feedback-mechanism-enabled intelligent early warning decision-making and execution method. The method couples basic environmental risks, the dynamic expansion speed of risks, and the attention state of personnel, and uses a comprehensive risk scoring formula to achieve accurate quantitative assessment of risks. Then, based on the scoring range, it strictly implements a progressive graded intervention strategy from broad reminders and targeted warnings to manual collaborative handling, ensuring precise matching between response measures and risk levels. This not only solves the problems of traditional early warning methods being singular, lagging, and lacking specificity, but also lays the foundation for continuous adaptive optimization of the system by introducing intervention effect feedback and priority setting, thereby significantly improving the initiative, accuracy, and overall effectiveness of waterfront safety management.

[0091] In one embodiment, the system further includes:

[0092] The data feedback optimization module is used to receive early warning intervention effect data (e.g., changes in target behavior, location movement, and whether a safety accident has occurred before and after the early warning is triggered) transmitted through a preset data feedback interface, form a training sample set, and optimize the parameters of the attention state determination model in the behavior state recognition module and the early warning decision model in the early warning decision module based on the training sample set.

[0093] The steps for optimizing the parameters of the early warning decision model include:

[0094] Establish an early warning effectiveness evaluation function, which is based on a comprehensive risk score. With the final safety outcome The degree of matching is calculated, and the early warning effect evaluation function generally adopts the mean square error function. The smaller the value of the function, the more accurate the prediction of the risk score R on the actual safety result S. The final safety result... The quantitative value is defined based on the final safety status of tourists. For example, it can be defined as follows: if tourists leave safely after the warning and no accident occurs, then S=1 (safe); if a minor danger occurs or manual intervention is required, then S=0.5 (normal); if a safety accident occurs, then S=0 (dangerous). The value is obtained by annotating the monitoring video or the manual reporting records after the event.

[0095] Optimization algorithms such as gradient descent are employed to minimize the loss value of the early warning effect evaluation function. During the iteration process, the weight coefficients in the comprehensive risk score calculation formula are automatically adjusted and updated. , and risk amplification factor .

[0096] The step of optimizing the parameters of the attention state determination model includes adjusting the behavioral weight coefficients in the attention distraction index calculation formula. Optimization;

[0097] During the optimization process, specific behavior categories are calculated. With the final safety outcome Statistical correlation between For example, correlation coefficients or other methods can be used to calculate it. The larger the value, the stronger the positive association between this type of behavior and unsafe outcomes; its weight... Theoretically, it should be larger;

[0098] Then, according to the following formula, the weighting coefficients are... Make dynamic adjustments:

[0099]

[0100] in, and These are the weight coefficients before and after optimization, respectively; The learning rate is a pre-defined small constant (usually...). (e.g., 0.01) is used to control the step size of each parameter update, prevent oscillation, and ensure the stability of the learning process; For the first Normalized correlation of class behaviors, where It is the correlation between the behavior and the safety outcome. It is the maximum correlation among all current behavior categories; This refers to the expected safety value predicted by the system based on the current parameters before optimization; the current parameters are the behavioral weight coefficients before optimization. This can be understood as the safe outcome that should be achieved after assessing the current situation using the old model.

[0101] Through the above technical solution, this embodiment provides a method for adaptive optimization based on data feedback. The method collects actual effect data of early warning intervention, constructs an early warning effect evaluation function and a safety result correlation analysis, and iteratively optimizes the parameters of the decision model and the behavior weight coefficients in the behavior recognition model, respectively. This enables the system to continuously learn from actual operating experience, automatically calibrate the quantitative assessment standards for environmental risks and personnel behavior risks, thereby continuously improving the accuracy and reliability of early warning, realizing the evolution from a static rule system to a dynamic intelligent system, and ensuring long-term effective safety protection capabilities.

[0102] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A behavior recognition-based safety warning system for waterfront landscape belt, characterized in that, The system includes: The data acquisition module includes an environmental sensing unit, a visual perception unit, and a wireless positioning unit. The environmental sensing unit is used to collect environmental data of the waterfront landscape belt, and the visual perception unit and the wireless positioning unit are both used to collect visitor target perception data and transmit the collected data to the corresponding associated modules. The environmental risk analysis module is used to receive hydrological and meteorological data transmitted by the data acquisition module, calculate and output the dynamic risk field of the waterfront area based on real-time and predicted environmental data, and the dynamic risk field divides the waterfront area into areas with different risk levels. The behavior state recognition module is used to receive tourist target perception data transmitted by the data acquisition module and dynamic risk field information output by the environmental risk analysis module. Based on the risk level area defined by the dynamic risk field, it prioritizes to start behavior recognition analysis for tourist targets in high-risk areas and outputs a structured state description vector containing tourist behavior category and attention state level. The early warning decision module is used to generate early warning instructions of different levels based on the risk level area output by the environmental risk analysis module and the structured state description vector output by the behavior state identification module, and to control the execution terminals deployed in the landscape belt to perform intervention operations.

2. The behavior recognition-based safety warning system for waterfront landscape belt according to claim 1, characterized in that, The environmental risk analysis module includes: The regional division unit is used to dynamically divide the waterfront area of ​​the waterfront landscape belt into core risk zone, extended warning zone and safe zone based on real-time and predicted data of tides, waves and wind speed, combined with a preset geographic information model. The core risk area is the area that will be submerged by the tide or directly impacted by large waves in the current or predicted short term. The extended early warning zone is the area adjacent to the core risk zone that is expected to be affected by the spread of risk within a set future timeframe. The safe zone is an area that has no direct hydrological risk in the current and foreseeable future period; The core risk zone, extended early warning zone, and safe zone each correspond to different perception parameter configurations. The area division unit synchronously outputs perception parameter adjustment instructions for each area to drive the data acquisition module to dynamically adjust its data acquisition strategy.

3. The behavior recognition based safety warning system for waterfront landscape belt according to claim 2, characterized in that, The behavior state recognition module includes: The target screening unit is used to screen targets located in the core risk area and the extended warning area from all perceived tourist targets based on the regional division results of the dynamic risk field, and mark the targets as priority identification objects; The deep recognition unit is configured to initiate machine learning-based behavior recognition analysis on the priority recognition object based on the attention state determination model. The behavior recognition analysis includes individual deep behavior recognition and group situation analysis. The attention state determination model is used at least to quantify the target's attention distraction index using the following formula: wherein, is an attention distraction index; is a recognition result of the attention distraction behavior of the class; is a weight coefficient corresponding to the behavior of the class; is a total number of behavior classes, ; is a normalized crowd density value of the local area where the target is located; is a duration of the target in a preset dangerous orientation; is a preset standard tracking duration; is an included angle between the body orientation of the target and the direction of the water area; is a function for quantifying the orientation risk with as the independent variable; , , , are adjustment coefficients of each item.

4. The behavior recognition based safety warning system for waterfront landscape belt according to claim 3, characterized in that, The depth recognition unit is further configured as follows: For all targets located in the core risk area, continuous behavior identification and status tracking are conducted; For targets located in the extended warning zone, the identification frequency is dynamically adjusted based on the initial judgment of their behavior and movement trend, with an increased identification frequency for targets moving toward the core risk zone or remaining stationary.

5. The behavior recognition based safety warning system for waterfront landscape belt according to claim 4, characterized in that, The early warning decision module includes: The comprehensive risk assessment unit receives the risk level of the target's location and the corresponding attention distraction index, and obtains the target's comprehensive risk score through the following coupled calculation formula: wherein, is a comprehensive risk score; is a basic hydrological risk value of a location where the target is located, calculated according to the environmental risk analysis module; is a prediction value of an expansion speed of the risk area in the direction of the target location; is an attention distraction index calculated by the behavior state recognition module; , are weight coefficients of the basic hydrological risk value and the expansion speed, respectively; is a risk amplification coefficient of the attention distraction factor.

6. The behavior recognition based safety warning system for waterfront landscape belt according to claim 5, characterized in that, The early warning decision module, based on the early warning decision model, also includes: a policy execution unit configured to execute a hierarchical intervention policy depending on the composite risk score the numerical interval in which the policy execution unit is configured to execute a hierarchical intervention policy: When When in the first numerical interval, trigger a first-level early warning, through the nearby fixed sound and light alarm terminal for regional broadcast reminder; When When in the second numerical interval, a secondary early warning is triggered, and a mobile active guidance terminal is dispatched to conduct close-range and directional sound and light warning on the specific target. When When in the third numerical interval, a third-level early warning is triggered, and the alarm information containing the target accurate position and the scene picture is pushed to the nearest personnel cooperative treatment terminal while the second-level early warning is executed. Wherein, the lower limit of the third numerical interval is higher than that of the second numerical interval, and the lower limit of the second numerical interval is higher than that of the first numerical interval; The policy execution unit is further configured as follows: After the warning is triggered, the intervention effect data is monitored and recorded, and the intervention effect data is fed back to the preset data feedback interface; Among them, differentiated optimization processing priorities are set for the intervention effect data of different levels of early warning.

7. The behavior recognition based safety warning system for waterfront landscape belt according to claim 6, characterized in that, The system also includes: The data feedback optimization module is used to receive the early warning intervention effect data transmitted through the preset data feedback interface, form a training sample set, and optimize the parameters of the attention state determination model in the behavior state recognition module and the early warning decision model in the early warning decision module based on the training sample set.

8. The behavior recognition based safety warning system for waterfront landscape belt according to claim 7, characterized in that, The steps for optimizing the parameters of the early warning decision model include: establishing an early warning effectiveness evaluation function, the early warning effectiveness evaluation function being based on the composite risk score with the degree of match with the final safety outcome , wherein the final safety outcome is a quantified value defined according to the final safety status of the tourist; The weight coefficient in the comprehensive risk score calculation formula is iteratively updated by minimizing the loss value of the early warning effect evaluation function , and the risk amplification coefficient .

9. The behavior recognition based safety warning system for waterfront landscape belt according to claim 8, characterized in that, The step of performing parameter optimization on the attention state determination model comprises: optimization of a behavior weight coefficient in the attention distraction index calculation formula ​ During the optimization process, the statistical correlation between the specific behavior category and the final safety result is calculated , and the weight coefficient is dynamically adjusted according to the following formula: wherein, and are the weight coefficients after and before optimization, respectively; is the learning rate; is the first normalized correlation degree of the class behavior; is the safety expectation value predicted by the system based on the current parameters, i.e. the behavior weight coefficients before optimization .

Images