A smart city public safety management system based on video surveillance

Abstract

The present invention discloses a smart city public safety management system based on video surveillance, which relates to the field of public safety management technology. It collects raw video data streams and real-time road data sets D in real time through an API interface, generates a time series feature matrix TSM and a road congestion coefficient Ck, and improves data integration efficiency. The behavior feature extraction module identifies group behavior characteristics through target detection and optical flow algorithm, and encrypts transmission to ensure privacy. The risk prediction module calculates the cross-regional risk index RRI based on federated learning, triggers a response mechanism, and accurately warns of risks. The dynamic decision-making module combines real-time data to generate a resource scheduling plan and selects appropriate paths and callable resources. By combining the execution feedback module and the drive iterative optimization module, a dynamic adjustment algorithm is realized to form a closed-loop management, which solves the problems of data fragmentation, response lag and model rigidity in traditional systems, and significantly improves the real-time, accuracy and scenario adaptability of public safety management.

Description

Technical Field

[0001] The present invention relates to the technical field of public safety management, and in particular to a smart city public safety management system based on video surveillance. Background Art

[0002] With the deepening development of smart cities, video surveillance networks have become the core infrastructure for urban security management, covering transportation hubs, commercial areas, and community settings, becoming the core perception layer for public safety management. With the increasing affluence of people in public places, the flow of people is increasing daily. Efficiently managing crowd safety, effectively integrating cross-regional video collection and analysis, and building effective public safety prediction models and management systems have become urgent tasks.

[0003] Current smart city management systems that integrate video surveillance data primarily focus on video capture and simple behavior recognition in a single area, such as detecting pedestrians or vehicles running red lights. First, risk prediction relies on manual experience and static rules, lacking quantitative modeling for group behavior, such as movement speed, direction, disorder, and abnormal behavior index, leading to delayed response decisions. Second, existing systems lack global analytical capabilities for regional risk prediction and management. Cross-regional camera data cannot be integrated, resulting in a lack of early warning of crowd movement directions in other areas when a pedestrian safety hazard arises in one location. Finally, resource scheduling is rigid, and emergency response plans cannot be dynamically adjusted based on real-time risks, often leading to delayed regional responses or redundant resources in low-risk areas. These shortcomings have prevented cities from developing a comprehensive management framework encompassing "data integration → dynamic modeling → resource scheduling → closed-loop optimization," hindering further improvements in the efficiency of predicting and preventing pedestrian safety hazards in cities. Summary of the Invention

[0004] In response to the deficiencies of the existing technology, the present invention provides a smart city public safety management system based on video surveillance, which solves the problems mentioned in the background technology.

[0005] To achieve the above objectives, the present invention is implemented through the following technical solutions: a smart city public safety management system based on video surveillance, characterized by comprising a data acquisition and analysis module, a behavior feature extraction module, a risk prediction module, a dynamic decision module, an execution feedback module and an iterative optimization module;

[0006] The data acquisition and analysis module collects the original video data stream and the real-time road data set D through the API interface, obtains the time series feature matrix TSM by integrating the crowd density CD and the individual movement speed vector V in the original video data stream, and obtains the road congestion coefficient Ck by analyzing the real-time road data set D;

[0007] The behavior feature extraction module analyzes the time series feature matrix TSM at each local node, extracts the crowd density distribution coefficient CDM, the movement direction consistency coefficient MDC, and the abnormal behavior coefficient ABFV, generates a homomorphically encrypted real-time behavior feature data set FV, and transmits it to the central node in encrypted form along with the road congestion coefficient Ck.

[0008] The risk prediction module receives the encrypted real-time behavior feature data set FV from each node. Based on the data in the real-time behavior feature data set FV, it establishes a cross-regional risk prediction algorithm through federated learning, calculates the cross-regional risk index RRI, compares it with the preset risk threshold range T, and implements the security management plan;

[0009] The dynamic decision-making module establishes a dynamic resource scheduling algorithm based on the cross-regional risk index RRI and the road congestion coefficient Ck and generates a resource scheduling plan and sends it to the duty unit;

[0010] The execution feedback module quantifies the execution effect of the resource scheduling solution, obtains feedback effect evaluation and generates a feedback dataset F;

[0011] The iterative optimization module optimizes the cross-regional risk prediction algorithm and dynamic resource scheduling algorithm based on the feedback dataset F.

[0012] Preferably, the data acquisition and analysis module includes a data stream acquisition unit and a data analysis unit;

[0013] The data stream acquisition unit uses APIs to call communication protocols compatible with different terminals, access public video surveillance terminals in various areas of the city, and obtain original video streams;

[0014] By calling the urban traffic management platform and vehicle GPS data through the API, the maximum traffic flow max (Tf), real-time traffic flow RTf, historical average vehicle speed Hv and real-time average vehicle speed Rv of the roads within the monitoring area are obtained to obtain the real-time road data set D.

[0015] Preferably, the data analysis unit obtains the time series feature matrix TSM and the road congestion coefficient Ck by analyzing the original video stream and the implementation road data set D;

[0016] Use the YOLOv8 target detection algorithm to identify the bounding box Bi of people in the original video stream frame by frame, count the number of people per unit area, calculate the crowd density CD in the video area, and use the improved optical flow algorithm to track the displacement of people and calculate the individual movement velocity vector V;

[0017] Integrate the crowd density CD and individual movement speed vector V in the original video stream to obtain the temporal feature matrix TSM of the video area;

[0018] Based on the analysis of the real-time road data set D, the road congestion coefficient Ck is obtained. The analysis expression of the road congestion coefficient Ck is as follows:

[0019] ;

[0020] Where η represents the contribution weight coefficient of the ratio of the real-time traffic flow RTf to the maximum traffic flow max(Tf), and ι represents the contribution weight coefficient of the ratio of the real-time average vehicle speed Rv to the historical average vehicle speed Hv.

[0021] Preferably, the behavior feature extraction module includes a behavior feature analysis unit and a data processing unit;

[0022] The behavior feature analysis unit extracts and encrypts the behavior features of the time series feature matrix TSM at each local node, and converts the time series feature matrix TSM into the crowd density distribution coefficient CDM, the moving direction consistency coefficient MDC, and the abnormal behavior coefficient ABFV;

[0023] The crowd density distribution coefficient CDM analysis expression is as follows:

[0024] ;

[0025] Where (x, y) represents the video pixel coordinates, (xi, yi) represents the detected human coordinates of the i-th person, N represents the total number of people detected, π represents the constant 3.14, σ represents the preset Gaussian kernel bandwidth, and e represents a natural constant;

[0026] By analyzing the individual moving velocity vector V, the individual moving direction angle θ is obtained. The moving direction consistency coefficient MDC analysis expression is as follows:

[0027] ;

[0028] Where θi represents the moving direction angle of the i-th person, θj represents the moving direction angle of the j-th person, and j≠i;

[0029] The number of people squatting and lying down in the surveillance video is analyzed based on the person bounding box Bi. The number of people standing still and running in the video is analyzed based on the individual movement velocity vector V. The abnormal behavior coefficient ABFV is obtained based on the ratio of the number of people squatting, lying down, standing still, and running in the surveillance video to the total number of people detected N.

[0030] Among them, the detection standards for squatting and lying down are as follows:

[0031] If the current height of the person bounding box Bi is less than or equal to 0.5 times the original height of the person bounding box Bi, and the current width of the person bounding box Bi is greater than or equal to 1.25 times the original width of the person bounding box Bi, then the current detection target is determined to be in a squatting state;

[0032] If the current height of the person bounding box Bi is greater than 0.5 times the original height of the person bounding box Bi, and the current width of the person bounding box Bi is less than 1.25 times the original width of the person bounding box Bi, then it is determined that the current detection target is in a squatting state;

[0033] If the current height of the person's bounding box Bi is less than or equal to 0.25 times the original height of the person's bounding box Bi, and the current width of the person's bounding box Bi is greater than or equal to 1.5 times the original width of the person's bounding box Bi, then the current detection target is determined to be in a lying state;

[0034] If the current height of the person bounding box Bi is greater than 0.25 times the original height of the person bounding box Bi, and the current width of the person bounding box Bi is less than 1.5 times the original width of the person bounding box Bi, then it is determined that the current detection target is not in a lying state;

[0035] The standards for stationary and running behavior detection are as follows:

[0036] If the individual moving speed vector V = 0m / s, it is determined that the current detection target is in a stationary state;

[0037] If the individual moving speed vector V≠0m / s, it is determined that the current detection target is not in a stationary state;

[0038] If the individual moving speed vector V>2.5m / s, it is determined that the current detection target is in a running state;

[0039] If the individual moving speed vector V≤2.5m / s, it is determined that the current detection target is not in a running state.

[0040] Preferably, the data processing unit generates a real-time behavior feature data set FV by vertically merging the encrypted crowd density distribution coefficient CDM, the moving direction consistency coefficient MDC and the abnormal behavior coefficient ABFV, homomorphically encrypts the obtained real-time behavior feature data set FV and the road congestion coefficient Ck, and transmits them to the central node in ciphertext form.

[0041] Preferably, the risk prediction module includes a risk modeling unit and an execution unit;

[0042] The risk modeling unit receives the encrypted real-time behavior feature data set FV from each node, designs a cross-region risk prediction algorithm based on the data in the real-time behavior feature data set FV through federated learning, and obtains the cross-region risk index RRIk of region k;

[0043] The cross-regional risk prediction algorithm expression is as follows:

[0044] ;

[0045] Wherein, k represents the region, RRIk represents the cross-regional risk index of region k, max(CDMk) represents the maximum value of the population density distribution coefficient CDM in region k, a represents the set density sensitivity coefficient, b represents the set curve density threshold, α represents the contribution weight coefficient of the maximum value of the population density distribution coefficient max(CDMk) in region k, MDCk represents the moving direction consistency coefficient of region k, β represents the contribution weight coefficient of the moving direction consistency coefficient MDCk of region k, ABFVk represents the abnormal behavior coefficient of region k, and γ represents the contribution weight coefficient of the abnormal behavior coefficient ABFVk of region k.

[0046] Preferably, the execution unit compares the cross-region risk index RRIk of region k with the risk threshold range T, and executes a graded response plan according to the comparison result;

[0047] If the cross-region risk index RRIk of region k is less than the risk threshold Tmin, only the current data is recorded as historical reference data;

[0048] If the risk threshold Tmin ≤ the cross-regional risk index RRIk of region k < the risk threshold Tmax, a risk warning is triggered and pushed to the duty units in the jurisdiction of region k and public electronic screens in nearby areas, advising people and vehicles that have not entered region k to choose routes with caution;

[0049] If the cross-region risk index RRIk of region k is greater than or equal to the risk threshold Tmax, a risk warning is triggered and a prompt is generated on public electronic screens in nearby areas, warning people and vehicles that have not entered region k to avoid entering region k. The dynamic resource scheduling algorithm is also triggered to automatically generate a resource scheduling plan and send it to the duty unit in the jurisdiction.

[0050] The risk threshold Tmin represents the lower limit value in the risk threshold range T, and the risk threshold Tmax represents the upper limit value in the risk threshold range T.

[0051] Preferably, the dynamic decision module includes a dynamic resource scheduling decision unit;

[0052] The dynamic resource scheduling decision unit designs a dynamic resource scheduling algorithm based on the cross-region risk index RRIk and road congestion coefficient Ck of region k and generates the resource scheduling optimization value PD;

[0053] By calling the API from the urban traffic management platform, we obtain the paths n from the duty units in the jurisdiction of area k to area k, and integrate the paths n to form a real-time candidate path set p;

[0054] The dynamic resource scheduling algorithm expression is as follows:

[0055] ;

[0056] Where argmin represents the parameter for finding the minimum value of the objective function, pn represents the nth path in the candidate path set p, Spnk represents the Manhattan distance S of the nth path in the candidate path set p to area k, Rpnm represents the Manhattan distance R from the available resource unit m on the nth path in the candidate path set p to area k, t represents the time difference from the start of the risk warning to the current moment, δ represents the path efficiency contribution weight coefficient of the nth path in the candidate path set p to area k, w represents the contribution weight coefficient of the Manhattan distance R from the available resource unit m on the nth path in the candidate path set p to area k, and μ represents the contribution weight coefficient of the time difference t from the start of the risk warning to the current moment;

[0057] The parameters when the resource scheduling preference value PD is minimum are exported and a resource scheduling plan is generated. The plan content includes the nth path in the selected candidate path set p and the available resource unit m called from the nth path in the selected candidate path set p. The resource scheduling plan is sent to the on-duty units within the jurisdiction.

[0058] Preferably, the execution feedback module includes a feedback data collection unit;

[0059] The feedback data collection unit generates a feedback data set F by collecting the execution feedback data of the resource scheduling plan to quantify the execution effect of the resource scheduling plan;

[0060] Among them, the feedback data set F includes response time RT, resource utilization RC and risk prediction accuracy SR;

[0061] The response time RT represents the time difference between the available resource unit m and the time difference between the resource scheduling plan sent to the arrival site based on the timestamp obtained by the GPS device;

[0062] Resource utilization rate RC represents the ratio of the actual number of resources on site to the predicted number of required resources;

[0063] The risk prediction accuracy rate SR represents the ratio of the number of correct risk predictions to the total number of risk predictions.

[0064] Preferably, the iterative optimization module includes an iterative optimization unit;

[0065] The iterative optimization unit dynamically optimizes the cross-regional risk prediction algorithm and dynamic resource scheduling algorithm based on the feedback data set F;

[0066] The cross-regional risk prediction algorithm optimization scheme is as follows:

[0067] Based on the risk prediction accuracy rate SR in the feedback dataset F, if the risk prediction accuracy rate SR is ≥ 95%, the prediction result is judged to be reasonable, and there is no need to dynamically adjust the contribution weight coefficients α, β, and γ in the cross-regional risk prediction algorithm;

[0068] If the risk prediction accuracy rate SR is less than 95%, the prediction result is judged to be unreasonable, and the contribution weight coefficients α, β and γ in the cross-regional risk prediction algorithm are dynamically adjusted;

[0069] The optimization scheme of dynamic resource scheduling algorithm is as follows:

[0070] Based on the response time RT and resource utilization RC in the feedback dataset F;

[0071] Optimization direction 1: If the single response time RT is ≤ 15 minutes, the nth path in the selected candidate path set p is determined to be reasonable, and there is no need to dynamically adjust the contribution weight coefficients δ and w in the dynamic resource scheduling algorithm;

[0072] If the single response time RT is greater than 15 minutes, the nth path in the selected candidate path set p is determined to be unreasonable, and the contribution weight coefficients δ and w in the dynamic resource scheduling algorithm are dynamically adjusted;

[0073] Optimization direction 2: If the resource utilization RC is ≥ 80% for three consecutive times, the available resource unit m is determined to be reasonably selected, and there is no need to dynamically adjust the contribution weight coefficients δ and w in the dynamic resource scheduling algorithm;

[0074] If the resource utilization rate RC is less than 80% for three consecutive times, it is determined that the selection of the available resource unit m is unreasonable, and the contribution weight coefficients δ and w in the dynamic resource scheduling algorithm are dynamically adjusted.

[0075] The present invention provides a smart city public safety management system based on video surveillance, which has the following beneficial effects:

[0076] (1) When the system is running, it accesses the time monitoring terminals, traffic management platforms and vehicle GPS data in various areas of the city through APIs, obtains raw video streams, road traffic flow and vehicle speed data in real time, constructs the time series feature matrix TSM and the road congestion coefficient Ck, integrates and analyzes multi-source data, and uses homomorphic encryption technology to ensure the security of sensitive information transmission, providing multi-dimensional reliable data for subsequent analysis, solving the data island problem and data security risks of traditional systems. Based on the time series feature matrix TSM, the YOLOv8 target detection algorithm and the improved optical flow algorithm are used to identify the person bounding box Bi frame by frame and track individual displacement, obtain the crowd density distribution coefficient CDM, the movement direction consistency coefficient MDC, and the abnormal behavior coefficient ABFV, and generate a real-time behavior feature data set FV, solving the problem of the lack of quantification of group behavior status in traditional methods. These data provide a key basis for risk prediction and resource scheduling, realize the deep collaboration between data collection and analysis and processing, and enhance the overall stability of the system.

[0077] (2) By receiving the encrypted real-time behavior feature data set FV from each node, a risk prediction algorithm is designed based on federated learning and the real-time behavior feature data set FV. This algorithm takes into account the combined influence of the maximum population density distribution coefficient max (CDMk) in region k, the movement direction consistency coefficient MDCk in region k, and the abnormal behavior coefficient ABFVk in region k, and generates a cross-region risk index RRIk. Through a hierarchical response mechanism, accurate risk prevention and control is achieved, avoiding problems after the response caused by manual experience. By combining the cross-region risk index RRIk and the road congestion coefficient Ck, a dynamic source scheduling algorithm is designed. By comprehensively considering the efficiency of each path n in the candidate path set p and the Manhattan distance R from the available resource unit m on the nth path in the candidate path set p to region k, a resource scheduling plan is obtained to avoid resource redundancy in low-risk areas. The federated learning mechanism realizes the combination of the cross-region risk prediction algorithm and the dynamic resource scheduling algorithm, breaking through the problems of global analysis capabilities, regional response delays, and resource redundancy in low-risk areas caused by data silos in traditional systems, and improving cross-regional collaboration efficiency.

[0078] (3) By collecting the execution feedback data of the resource scheduling plan, a feedback data set F is generated to quantify the execution effect of the resource scheduling plan. Based on the global analysis of the feedback data set F, the system can automatically identify decision deviations and algorithm optimization trends, and optimize the risk prediction algorithm and resource scheduling algorithm through an adaptive learning mechanism. Based on the response time RT and resource utilization rate RC, the contribution weight coefficients of the path and resource unit selection are corrected to improve the real-time and accuracy of resource scheduling. In response to the risk prediction error, the contribution weight coefficients α, β and γ in the regional risk prediction algorithm are dynamically balanced to improve the robustness of the algorithm. This closed-loop mechanism not only avoids the long-term performance degradation caused by static rules in traditional systems, but also significantly improves the prevention and control efficiency and long-term stability through continuous data-driven optimization. BRIEF DESCRIPTION OF THE DRAWINGS

[0079] Figure 1 This is a schematic diagram of a smart city public safety management system based on video surveillance according to the present invention;

[0080] Figure 2 This is a schematic diagram of a public safety management resource planning process flow chart of a smart city public safety management system based on video surveillance according to the present invention;

[0081] Figure 3 Schematic diagram of data acquisition, processing and flow. DETAILED DESCRIPTION

[0082] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention. Example 1

[0083] The present invention provides a smart city public security management system based on video surveillance, please refer to Figure 1 and Figure 3 , including data collection and analysis module, behavior feature extraction module, risk prediction module, dynamic decision module, execution feedback module and iterative optimization module;

[0084] The data acquisition and analysis module collects the original video data stream and the real-time road data set D in real time through the API interface, obtains the time series feature matrix TSM by integrating the crowd density CD and the individual movement speed vector V in the original video data stream, and obtains the road congestion coefficient Ck by analyzing the real-time road data set D;

[0085] The behavior feature extraction module analyzes the time series feature matrix TSM at each local node, extracts the crowd density distribution coefficient CDM, the movement direction consistency coefficient MDC, and the abnormal behavior coefficient ABFV, generates a homomorphically encrypted real-time behavior feature data set FV, and transmits it to the central node in encrypted form along with the road congestion coefficient Ck.

[0086] The risk prediction module receives the encrypted real-time behavior feature data set FV from each node. Based on the data in the real-time behavior feature data set FV, it establishes a cross-regional risk prediction algorithm through federated learning, calculates the cross-regional risk index RRI, compares it with the preset risk threshold range T, and implements the security management plan;

[0087] The dynamic decision-making module establishes a dynamic resource scheduling algorithm based on the cross-regional risk index RRI and the road congestion coefficient Ck and generates a resource scheduling plan and sends it to the duty unit;

[0088] The execution feedback module quantifies the execution effect of the resource scheduling solution, obtains feedback effect evaluation and generates a feedback dataset F;

[0089] The iterative optimization module optimizes the cross-regional risk prediction algorithm and dynamic resource scheduling algorithm based on the feedback dataset F.

[0090] In this embodiment, real-time access to data from video surveillance terminals, traffic management platforms, and vehicle GPS in various urban areas is achieved through an API interface. This integrates multi-dimensional data, including video streams, road traffic volume, and vehicle speed, to generate a time series feature matrix (TSM) and a road congestion coefficient (Ck). This solves the traditional data silo problem. Homomorphic encryption technology ensures data transmission security and mitigates the risk of data leakage. The behavioral feature extraction module utilizes target detection and optical path algorithms to identify individual bounding boxes (Bi) frame by frame and track individual displacements. This accurately extracts a real-time behavioral feature data set (FV), which is encrypted and transmitted to a central node. This overcomes the lag in response to complex behavior detection often associated with traditional methods. The risk prediction module integrates encrypted data from multiple regions within a federated learning framework, dynamically models the cross-regional risk index (RRI), and triggers a graded response based on a preset threshold range (T). This addresses the response lag inherent in traditional systems, which rely on manual experience and static rules. The dynamic decision-making module combines the cross-regional risk index (RRIk) and the road congestion coefficient (Ck) for region k to generate a resource scheduling plan, allocating routes and on-duty units in real time. This optimizes emergency response efficiency and avoids the bottlenecks of rigid plans and low resource utilization associated with traditional systems. The execution feedback module quantifies response time (RT), resource utilization (RC), and risk prediction accuracy (SR). This generates a feedback dataset (F) that drives the iterative optimization module to dynamically adjust the algorithm's contribution weight coefficients, forming a closed-loop management mechanism of "data collection and analysis → risk prediction → resource scheduling → feedback optimization." This addresses the long-term performance degradation caused by the lack of adaptive optimization in traditional systems. Through modular collaboration and closed-loop data flow, a public safety management system has been established that combines privacy, real-time capabilities, and adaptability. This addresses the pain points of traditional systems, such as rigid resource scheduling and the inability to dynamically adjust emergency plans based on real-time risks, leading to delayed regional responses or redundant resources in low-risk areas. This provides a scalable technical paradigm for smart city security management. Example 2

[0091] This embodiment is explained in Example 1, please refer to Figure 1 and Figure 3 ,Specifically: the data acquisition and analysis module includes a data stream acquisition unit and a data ,analysis unit;

[0092] The data stream acquisition unit uses APIs to call communication protocols compatible with different terminals, access public video surveillance terminals in various areas of the city, and obtain original video streams;

[0093] By calling the urban traffic management platform and vehicle GPS data through the API, the maximum traffic flow max (Tf), real-time traffic flow RTf, historical average vehicle speed Hv and real-time average vehicle speed Rv of the roads within the monitoring area are obtained to obtain the real-time road data set D.

[0094] The data analysis unit obtains the time series feature matrix TSM and the road congestion coefficient Ck by analyzing the original video stream and the implementation road data set D;

[0095] Use the YOLOv8 target detection algorithm to identify the bounding box Bi of people in the original video stream frame by frame, count the number of people per unit area, calculate the crowd density CD in the video area, and use the improved optical flow algorithm to track the displacement of people and calculate the individual movement velocity vector V;

[0096] Integrate the crowd density CD and individual movement speed vector V in the original video stream to obtain the temporal feature matrix TSM of the video area;

[0097] Based on the analysis of the real-time road data set D, the road congestion coefficient Ck is obtained. The analysis expression of the road congestion coefficient Ck is as follows:

[0098] ;

[0099] Where η represents the contribution weight coefficient of the ratio of the real-time traffic flow RTf to the maximum traffic flow max(Tf), and ι represents the contribution weight coefficient of the ratio of the real-time average vehicle speed Rv to the historical average vehicle speed Hv.

[0100] The purpose of this formula is to quantify the current traffic complexity through the ratio of real-time traffic flow RTf to the maximum road traffic flow max (Tf), and to reflect the road traffic efficiency by combining the degree of deviation between the real-time average vehicle speed Rv and the historical average speed Hv, thereby providing a real-time road status basis for dynamic resource scheduling.

[0101] The behavior feature extraction module includes a behavior feature analysis unit and a data processing unit;

[0102] The behavior feature analysis unit extracts and encrypts the behavior features of the time series feature matrix TSM at each local node, and converts the time series feature matrix TSM into the crowd density distribution coefficient CDM, the moving direction consistency coefficient MDC, and the abnormal behavior coefficient ABFV;

[0103] The crowd density distribution coefficient CDM analysis expression is as follows:

[0104] ;

[0105] Where (x, y) represents the video pixel coordinates, (xi, yi) represents the body coordinates of the i-th person detected, N represents the total number of people detected, π represents the constant 3.14, σ represents the preset Gaussian kernel bandwidth, which is a value set based on experience and experimental rules, and e represents a natural constant;

[0106] By analyzing the individual moving velocity vector V, the individual moving direction angle θ is obtained. The moving direction consistency coefficient MDC analysis expression is as follows:

[0107] ;

[0108] Where θi represents the moving direction angle of the i-th person, θj represents the moving direction angle of the j-th person, and j≠i;

[0109] This formula achieves its purpose: by dynamically adjusting the Gaussian kernel bandwidth σ, it adapts to changes in crowd density in different scenarios, accurately quantifies the spatial distribution of people in the video area, quantifies the consistency of the group's movement direction by calculating the cosine similarity of the movement direction angles of all individuals, introduces the difference calculation between all individuals, and comprehensively evaluates the coordination of group movement, providing key input for the calculation of the cross-regional risk index (RRI), thereby enhancing the accuracy of risk prediction.

[0110] The number of people squatting and lying down in the surveillance video is analyzed based on the person bounding box Bi. The number of people standing still and running in the video is analyzed based on the individual movement velocity vector V. The abnormal behavior coefficient ABFV is obtained based on the ratio of the number of people squatting, lying down, standing still, and running in the surveillance video to the total number of people detected N.

[0111] Among them, the detection standards for squatting and lying down are as follows:

[0112] If the current height of the person bounding box Bi is less than or equal to 0.5 times the original height of the person bounding box Bi, and the current width of the person bounding box Bi is greater than or equal to 1.25 times the original width of the person bounding box Bi, then the current detection target is determined to be in a squatting state;

[0113] If the current height of the person bounding box Bi is greater than 0.5 times the original height of the person bounding box Bi, and the current width of the person bounding box Bi is less than 1.25 times the original width of the person bounding box Bi, then it is determined that the current detection target is in a squatting state;

[0114] If the current height of the person's bounding box Bi is less than or equal to 0.25 times the original height of the person's bounding box Bi, and the current width of the person's bounding box Bi is greater than or equal to 1.5 times the original width of the person's bounding box Bi, then the current detection target is determined to be in a lying state;

[0115] If the current height of the person bounding box Bi is greater than 0.25 times the original height of the person bounding box Bi, and the current width of the person bounding box Bi is less than 1.5 times the original width of the person bounding box Bi, then it is determined that the current detection target is not in a lying state;

[0116] The standards for stationary and running behavior detection are as follows:

[0117] If the individual moving speed vector V = 0m / s, it is determined that the current detection target is in a stationary state;

[0118] If the individual moving speed vector V≠0m / s, it is determined that the current detection target is not in a stationary state;

[0119] If the individual moving speed vector V>2.5m / s, it is determined that the current detection target is in a running state;

[0120] If the individual moving speed vector V≤2.5m / s, it is determined that the current detection target is not in a running state.

[0121] The data processing unit generates a real-time behavior feature data set FV by vertically merging the encrypted crowd density distribution coefficient CDM, moving direction consistency coefficient MDC and abnormal behavior coefficient ABFV, homomorphically encrypts the obtained real-time behavior feature data set FV and road congestion coefficient Ck, and transmits them to the central node in ciphertext form.

[0122] This embodiment uses an API interface to access a multi-terminal traffic protocol, providing real-time access to video surveillance terminals, education and management platforms, and vehicle GPS data in various urban areas. It simultaneously acquires raw video streams, maximum road traffic flow (max (Tf), real-time traffic flow (RTf), real-time vehicle average speed (Rv), and historical vehicle average speed (Hv), generating a real-time road data set D. This addresses the data fragmentation problem caused by information silos in traditional systems. The YOLOv8 algorithm is used to identify person bounding boxes (Bi) frame by frame for the collected data. Crowd density (CD) is calculated, and individual movement velocity vectors (V) are tracked. Combined with an improved optical flow algorithm, a time series feature matrix (TSM) is generated, and the road congestion coefficient (Ck) is dynamically calculated, improving the real-time and accuracy of traffic status analysis. The behavioral feature analysis unit calculates the crowd density distribution coefficient (CDM) using a Gaussian kernel function and quantifies the movement direction consistency coefficient (MDC) using cosine similarity. It detects squatting and lying down based on changes in the height and width of the person bounding box (Bi), and detects standing and running based on the individual movement velocity vectors (V). This generates an abnormal behavior coefficient (ABFV), addressing the lag in response to complex behavior detection in traditional detection methods. The data processing unit vertically merges and homomorphically encrypts the crowd density distribution coefficient (CDM), the movement direction consistency coefficient (MDC), and the abnormal behavior coefficient (ABFV), generating a real-time behavioral feature data set (FV) that is encrypted and transmitted to the central node. Through localized processing and privacy protection mechanisms, this eliminates the data leakage risks associated with centralized computing in traditional systems. This deep fusion of multi-source data and refined behavioral modeling provides high-precision input for risk prediction and resource scheduling, addressing the decision-making lag and error accumulation issues inherent in traditional systems due to data fragmentation, crude analysis, and privacy risks, significantly improving the reliability and adaptability of public safety management. Example 3

[0123] This embodiment is explained in Example 2, please refer to Figure 1 and Figure 2 ,Specifically: the risk prediction module includes a risk modeling unit and an ,execution unit;

[0124] The risk modeling unit receives the encrypted real-time behavior feature data set FV from each node, designs a cross-region risk prediction algorithm based on the data in the real-time behavior feature data set FV through federated learning, and obtains the cross-region risk index RRIk of region k;

[0125] The cross-regional risk prediction algorithm expression is as follows:

[0126] ;

[0127] Wherein, k represents the region, RRIk represents the cross-regional risk index of region k, max(CDMk) represents the maximum value of the population density distribution coefficient CDM in region k, a represents the set density sensitivity coefficient, b represents the set curve density threshold, α represents the contribution weight coefficient of the maximum value of the population density distribution coefficient max(CDMk) in region k, MDCk represents the moving direction consistency coefficient of region k, β represents the contribution weight coefficient of the moving direction consistency coefficient MDCk of region k, ABFVk represents the abnormal behavior coefficient of region k, and γ represents the contribution weight coefficient of the abnormal behavior coefficient ABFVk of region k.

[0128] This formula achieves its purpose: through the logical function The maximum value of the crowd density distribution coefficient (CDMk) is mapped to the interval [0,1] to avoid the sensitivity of the linear model to extreme density changes. Combined with the directional consistency coefficient MDCk and the abnormal behavior coefficient ABFVk, the risk index of area k is accurately quantified, providing a scientific basis for the graded response mechanism.

[0129] The execution unit compares the cross-region risk index RRIk of region k with the risk threshold range T and implements a graded response plan based on the comparison results;

[0130] If the cross-region risk index RRIk of region k is less than the risk threshold Tmin, only the current data is recorded as historical reference data;

[0131] If the risk threshold Tmin ≤ the cross-regional risk index RRIk of region k < the risk threshold Tmax, a risk warning is triggered and pushed to the duty units in the jurisdiction of region k and public electronic screens in nearby areas, advising people and vehicles that have not entered region k to choose routes with caution;

[0132] If the cross-region risk index RRIk of region k is greater than or equal to the risk threshold Tmax, a risk warning is triggered and a prompt is generated on public electronic screens in nearby areas, warning people and vehicles that have not entered region k to avoid entering region k. The dynamic resource scheduling algorithm is also triggered to automatically generate a resource scheduling plan and send it to the duty unit in the jurisdiction.

[0133] The risk threshold Tmin represents the lower limit value in the risk threshold range T, and the risk threshold Tmax represents the upper limit value in the risk threshold range T.

[0134] In this embodiment, a federated learning framework receives encrypted real-time behavioral feature data sets FV from each node. Based on this data, a cross-regional risk prediction algorithm is designed to calculate the cross-regional risk index (RRI). This cross-regional risk prediction algorithm considers the influence of the maximum population density distribution coefficient (max(CDMk), the movement direction consistency coefficient (MDCk), and the abnormal behavior coefficient (ABFVk) within region k. Using a set density sensitivity coefficient a, a curve density threshold (b), and respective contribution weights (α, β, and γ), a nonlinear mapping of population density is performed, enhancing the model's sensitivity to high-density areas. A comprehensive assessment is then performed to determine the cross-regional risk index (RRIk) for region k. This comprehensive analysis method not only quantifies the degree of crowd concentration within a specific area, but also measures the consistency of group movement and the frequency of abnormal behavior, providing comprehensive data support for risk assessment. The execution unit then compares the calculated cross-regional risk index (RRIk) with a preset risk threshold range (T) to trigger a corresponding tiered response plan, such as recording historical reference data, issuing warning information, or generating an emergency resource scheduling plan. In this way, the system not only rapidly responds to potential security threats but also enables precise prevention and control while protecting privacy, effectively avoiding the delayed response issues associated with traditional reliance on manual experience and static rules. Furthermore, this approach enhances the ability to foresee and respond to complex public safety incidents, enabling city managers to take the most appropriate measures immediately, improving the efficiency of public resource utilization and the overall level of public safety management. This mechanism places particular emphasis on the deep integration and intelligent analysis of data from multiple dimensions, laying a solid foundation for building a smarter and safer urban environment. Example 4

[0135] This embodiment is explained in Example 3, please refer to Figure 1 and Figure 2 ,Specifically: the dynamic decision module includes a dynamic resource scheduling ,decision unit;

[0136] The dynamic resource scheduling decision unit designs a dynamic resource scheduling algorithm based on the cross-region risk index RRIk and road congestion coefficient Ck of region k and generates the resource scheduling optimization value PD;

[0137] By calling the API from the urban traffic management platform, we obtain the paths n from the duty units in the jurisdiction of area k to area k, and integrate the paths n to form a real-time candidate path set p;

[0138] The dynamic resource scheduling algorithm expression is as follows:

[0139] ;

[0140] Where argmin represents the parameter for finding the minimum value of the objective function, pn represents the nth path in the candidate path set p, and Spnk represents the Manhattan distance S from the nth path in the candidate path set p to the area k. represents the path efficiency of the nth path in the candidate path set p, Rpnm represents the Manhattan distance R from the available resource unit m to the region k on the nth path in the candidate path set p, t represents the time difference from the start of the risk warning to the current moment, δ represents the path efficiency contribution weight coefficient of the nth path in the candidate path set p to the region k, w represents the contribution weight coefficient of the Manhattan distance R from the available resource unit m to the region k on the nth path in the candidate path set p, μ represents the contribution weight coefficient of the time difference t from the start of the risk warning to the current moment, e μt The time factor of the cross-regional risk index RRIk of region k;

[0141] The parameters when the resource scheduling preference value PD is minimum are exported and a resource scheduling plan is generated. The plan content includes the nth path in the selected candidate path set p and the available resource unit m called from the nth path in the selected candidate path set p. The resource scheduling plan is sent to the on-duty units within the jurisdiction.

[0142] In this embodiment, the dynamic resource scheduling decision unit designs a dynamic resource scheduling algorithm based on the cross-regional risk index RRIk and road congestion coefficient Ck of region k. By calling the actual path data of the urban traffic management platform, a candidate path set p is generated, and the resource scheduling preference value PD of each path is calculated. The parameter when the resource scheduling preference value PD takes the minimum value is selected as the resource scheduling solution. By implementing the integration of the cross-regional risk index RRIk and road congestion coefficient Ck of region k, the algorithm reduces the negative impact of high congestion or other high-risk areas on path selection, and introduces the time factor e μt, dynamically enhancing sensitivity to response timeliness and ensuring that resource scheduling priorities automatically adjust over time. This mechanism breaks through the limitations of traditional reliance on static path planning or a single distance metric, achieving multi-dimensional real-time integration of path efficiency, resource accessibility, and time urgency, significantly shortening emergency response time and improving resource utilization. By dynamically optimizing the optimal path and available resource units m in the candidate path set p, the system solves the problems of resource redundancy in low-risk areas or response delays in high-risk areas caused by the rigidity of traditional plans, providing a quantifiable and reusable technical path for the precise handling of urban public safety incidents, and further enhancing the agility and scenario adaptability of smart city management. Example 5

[0143] This embodiment is explained in Example 4. Please refer to Figure 1 and Figure 2 ,Specifically: the execution feedback module includes a feedback data ,collection unit;

[0144] The feedback data collection unit generates a feedback data set F by collecting the execution feedback data of the resource scheduling plan to quantify the execution effect of the resource scheduling plan;

[0145] Among them, the feedback data set F includes response time RT, resource utilization RC and risk prediction accuracy SR;

[0146] The response time RT represents the time difference between the available resource unit m and the time difference between the resource scheduling plan sent to the arrival site based on the timestamp obtained by the GPS device;

[0147] Resource utilization rate RC represents the ratio of the actual number of resources on site to the predicted number of required resources;

[0148] The risk prediction accuracy rate SR represents the ratio of the number of correct risk predictions to the total number of risk predictions.

[0149] The iterative optimization module includes an iterative optimization unit;

[0150] The iterative optimization unit dynamically optimizes the cross-regional risk prediction algorithm and dynamic resource scheduling algorithm based on the feedback data set F;

[0151] The cross-regional risk prediction algorithm optimization scheme is as follows:

[0152] Based on the risk prediction accuracy rate SR in the feedback dataset F, if the risk prediction accuracy rate SR is ≥ 95%, the prediction result is judged to be reasonable, and there is no need to dynamically adjust the contribution weight coefficients α, β, and γ in the cross-regional risk prediction algorithm;

[0153] If the risk prediction accuracy rate SR is less than 95%, the prediction result is judged to be unreasonable, and the contribution weight coefficients α, β and γ in the cross-regional risk prediction algorithm are dynamically adjusted;

[0154] The optimization scheme of dynamic resource scheduling algorithm is as follows:

[0155] Based on the response time RT and resource utilization RC in the feedback dataset F;

[0156] Optimization direction 1: If the single response time RT is ≤ 15 minutes, the nth path in the selected candidate path set p is determined to be reasonable, and there is no need to dynamically adjust the contribution weight coefficients δ and w in the dynamic resource scheduling algorithm;

[0157] If the single response time RT is greater than 15 minutes, the nth path in the selected candidate path set p is determined to be unreasonable, and the contribution weight coefficients δ and w in the dynamic resource scheduling algorithm are dynamically adjusted;

[0158] Optimization direction 2: If the resource utilization RC is ≥ 80% for three consecutive times, the available resource unit m is determined to be reasonably selected, and there is no need to dynamically adjust the contribution weight coefficients δ and w in the dynamic resource scheduling algorithm;

[0159] If the resource utilization rate RC is less than 80% for three consecutive times, it is determined that the selection of the available resource unit m is unreasonable, and the contribution weight coefficients δ and w in the dynamic resource scheduling algorithm are dynamically adjusted.

[0160] In this embodiment, by introducing an execution feedback module and an iterative optimization module, the system's adaptive adjustment capabilities are enhanced while significantly increasing its long-term operational stability and flexibility. The feedback data collection unit in the execution feedback module collects system response time (RT), resource utilization (RC), and risk prediction accuracy (SR), providing solid data support for subsequent optimization decisions. This feedback mechanism ensures that the system is always operating efficiently and can be dynamically adjusted when needed. Simultaneously, the iterative optimization module, through the iterative optimization unit, continuously adjusts and optimizes the core algorithms in the system based on the collected feedback data. To address the emergency response needs of different scenarios, the optimization unit adjusts the dynamic resource scheduling strategy based on real-time feedback, ensuring that resources are accurately and efficiently allocated in the face of emergencies and avoiding resource waste or shortages. Furthermore, for cross-regional risk prediction optimization, the optimization unit analyzes historical data and real-time feedback to adjust the parameters of the risk assessment model, improving prediction accuracy and response speed. Through this closed-loop feedback and iterative optimization mechanism, the system not only responds and adapts quickly to emergencies but also continuously improves decision-making accuracy and resource allocation efficiency over the long term, thereby ensuring the efficiency and reliability of the entire public safety management system in diverse environments and conditions. Overall, the introduction of this solution has significantly improved the system's dynamic adaptability and decision-making intelligence, making public safety management more accurate, efficient and sustainable.

[0161] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A smart city public safety management system based on video surveillance, characterized by: It includes data collection and analysis module, behavior feature extraction module, risk prediction module, dynamic decision module, execution feedback module and iterative optimization module; The data acquisition and analysis module collects the original video data stream and the real-time road data set D in real time through the API interface, obtains the time series feature matrix TSM by integrating the crowd density CD and the individual movement speed vector V in the original video data stream, and obtains the road congestion coefficient Ck by analyzing the real-time road data set D; The behavior feature extraction module analyzes the time series feature matrix TSM at each local node, extracts the crowd density distribution coefficient CDM, the movement direction consistency coefficient MDC, and the abnormal behavior coefficient ABFV, generates a homomorphically encrypted real-time behavior feature data set FV, and transmits it to the central node in encrypted form along with the road congestion coefficient Ck. The risk prediction module receives the encrypted real-time behavior feature data set FV from each node. Based on the data in the real-time behavior feature data set FV, it establishes a cross-regional risk prediction algorithm through federated learning, calculates the cross-regional risk index RRI, compares it with the preset risk threshold range T, and implements the security management plan; The dynamic decision-making module establishes a dynamic resource scheduling algorithm based on the cross-regional risk index RRI and the road congestion coefficient Ck and generates a resource scheduling plan and sends it to the duty unit; The execution feedback module quantifies the execution effect of the resource scheduling solution, obtains feedback effect evaluation and generates a feedback dataset F; The iterative optimization module optimizes the cross-regional risk prediction algorithm and dynamic resource scheduling algorithm based on the feedback dataset F.

2. A smart city public safety management system based on video surveillance according to claim 1, characterized in that: The data acquisition and analysis module includes a data stream acquisition unit and a data analysis unit; The data stream acquisition unit uses APIs to call communication protocols compatible with different terminals, access public video surveillance terminals in various areas of the city, and obtain original video streams; By calling the urban traffic management platform and vehicle GPS data through the API, the maximum traffic flow max (Tf), real-time traffic flow RTf, historical average vehicle speed Hv and real-time average vehicle speed Rv of the roads within the monitoring area are obtained to obtain the real-time road data set D.

3. A smart city public safety management system based on video surveillance according to claim 2, characterized in that: The data analysis unit obtains the time series feature matrix TSM and the road congestion coefficient Ck by analyzing the original video stream and the implementation road data set D; Use the YOLOv8 target detection algorithm to identify the bounding box Bi of people in the original video stream frame by frame, count the number of people per unit area, calculate the crowd density CD in the video area, and use the improved optical flow algorithm to track the displacement of people and calculate the individual movement velocity vector V; Integrate the crowd density CD and individual movement speed vector V in the original video stream to obtain the temporal feature matrix TSM of the video area; Based on the analysis of the real-time road data set D, the road congestion coefficient Ck is obtained. The analysis expression of the road congestion coefficient Ck is as follows: ； Where η represents the contribution weight coefficient of the ratio of the real-time traffic flow RTf to the maximum traffic flow max(Tf), and ι represents the contribution weight coefficient of the ratio of the real-time average vehicle speed Rv to the historical average vehicle speed Hv.

4. A smart city public safety management system based on video surveillance according to claim 3, characterized in that: The behavior feature extraction module includes a behavior feature analysis unit and a data processing unit; The behavior feature analysis unit extracts and encrypts the behavior features of the time series feature matrix TSM at each local node, and converts the time series feature matrix TSM into the crowd density distribution coefficient CDM, the moving direction consistency coefficient MDC, and the abnormal behavior coefficient ABFV; The crowd density distribution coefficient CDM analysis expression is as follows: ； Where (x, y) represents the video pixel coordinates, (xi, yi) represents the detected human coordinates of the i-th person, N represents the total number of people detected, π represents the constant 3.14, σ represents the preset Gaussian kernel bandwidth, and e represents a natural constant; By analyzing the individual moving velocity vector V, the individual moving direction angle θ is obtained. The moving direction consistency coefficient MDC analysis expression is as follows: ； Where θi represents the moving direction angle of the i-th person, θj represents the moving direction angle of the j-th person, and j≠i; The number of people squatting and lying down in the surveillance video is analyzed based on the person bounding box Bi. The number of people standing still and running in the video is analyzed based on the individual movement velocity vector V. The abnormal behavior coefficient ABFV is obtained based on the ratio of the number of people squatting, lying down, standing still, and running in the surveillance video to the total number of people detected N. Among them, the detection standards for squatting and lying down are as follows: If the current height of the person bounding box Bi is less than or equal to 0.5 times the original height of the person bounding box Bi, and the current width of the person bounding box Bi is greater than or equal to 1.25 times the original width of the person bounding box Bi, then the current detection target is determined to be in a squatting state; If the current height of the person bounding box Bi is greater than 0.5 times the original height of the person bounding box Bi, and the current width of the person bounding box Bi is less than 1.25 times the original width of the person bounding box Bi, then it is determined that the current detection target is in a squatting state; If the current height of the person's bounding box Bi is less than or equal to 0.25 times the original height of the person's bounding box Bi, and the current width of the person's bounding box Bi is greater than or equal to 1.5 times the original width of the person's bounding box Bi, then the current detection target is determined to be in a lying state; If the current height of the person bounding box Bi is greater than 0.25 times the original height of the person bounding box Bi, and the current width of the person bounding box Bi is less than 1.5 times the original width of the person bounding box Bi, then it is determined that the current detection target is not in a lying state; The standards for stationary and running behavior detection are as follows: If the individual moving speed vector V = 0m / s, it is determined that the current detection target is in a stationary state; If the individual moving speed vector V≠0m / s, it is determined that the current detection target is not in a stationary state; If the individual moving speed vector V>2.5m / s, it is determined that the current detection target is in a running state; If the individual moving speed vector V≤2.5m / s, it is determined that the current detection target is not in a running state.

5. The smart city public safety management system based on video surveillance according to claim 4 is characterized in that: The data processing unit generates a real-time behavior feature data set FV by vertically merging the encrypted crowd density distribution coefficient CDM, moving direction consistency coefficient MDC and abnormal behavior coefficient ABFV, homomorphically encrypts the obtained real-time behavior feature data set FV and road congestion coefficient Ck, and transmits them to the central node in ciphertext form.

6. The smart city public safety management system based on video surveillance according to claim 5, characterized in that: The risk prediction module includes a risk modeling unit and an execution unit; The risk modeling unit receives the encrypted real-time behavior feature data set FV from each node, designs a cross-region risk prediction algorithm based on the data in the real-time behavior feature data set FV through federated learning, and obtains the cross-region risk index RRIk of region k; The cross-regional risk prediction algorithm expression is as follows: ； Wherein, k represents the region, RRIk represents the cross-regional risk index of region k, max(CDMk) represents the maximum value of the population density distribution coefficient CDM in region k, a represents the set density sensitivity coefficient, b represents the set curve density threshold, α represents the contribution weight coefficient of the maximum value of the population density distribution coefficient max(CDMk) in region k, MDCk represents the moving direction consistency coefficient of region k, β represents the contribution weight coefficient of the moving direction consistency coefficient MDCk of region k, ABFVk represents the abnormal behavior coefficient of region k, and γ represents the contribution weight coefficient of the abnormal behavior coefficient ABFVk of region k.

7. A smart city public safety management system based on video surveillance according to claim 6, characterized in that: The execution unit compares the cross-region risk index RRIk of region k with the risk threshold range T and implements a graded response plan based on the comparison results; If the cross-region risk index RRIk of region k is less than the risk threshold Tmin, only the current data is recorded as historical reference data; If the risk threshold Tmin ≤ the cross-regional risk index RRIk of region k < the risk threshold Tmax, a risk warning is triggered and pushed to the duty units in the jurisdiction of region k and public electronic screens in nearby areas, advising people and vehicles that have not entered region k to choose routes with caution; If the cross-region risk index RRIk of region k is greater than or equal to the risk threshold Tmax, a risk warning is triggered and a prompt is generated on public electronic screens in nearby areas, warning people and vehicles that have not entered region k to avoid entering region k. The dynamic resource scheduling algorithm is also triggered to automatically generate a resource scheduling plan and send it to the duty unit in the jurisdiction. The risk threshold Tmin represents the lower limit value in the risk threshold range T, and the risk threshold Tmax represents the upper limit value in the risk threshold range T.

8. The smart city public safety management system based on video surveillance according to claim 7, characterized in that: The dynamic decision module includes a dynamic resource scheduling decision unit; The dynamic resource scheduling decision unit designs a dynamic resource scheduling algorithm based on the cross-region risk index RRIk and road congestion coefficient Ck of region k and generates the resource scheduling optimization value PD; By calling the API from the urban traffic management platform, we obtain the paths n from the duty units in the jurisdiction of area k to area k, and integrate the paths n to form a real-time candidate path set p; The dynamic resource scheduling algorithm expression is as follows: ； Where argmin represents the parameter for finding the minimum value of the objective function, pn represents the nth path in the candidate path set p, Spnk represents the Manhattan distance S of the nth path in the candidate path set p to area k, Rpnm represents the Manhattan distance R from the available resource unit m on the nth path in the candidate path set p to area k, t represents the time difference from the start of the risk warning to the current moment, δ represents the path efficiency contribution weight coefficient of the nth path in the candidate path set p to area k, w represents the contribution weight coefficient of the Manhattan distance R from the available resource unit m on the nth path in the candidate path set p to area k, and μ represents the contribution weight coefficient of the time difference t from the start of the risk warning to the current moment; The parameters when the resource scheduling preference value PD is minimum are exported and a resource scheduling plan is generated. The plan content includes the nth path in the selected candidate path set p and the available resource unit m called from the nth path in the selected candidate path set p. The resource scheduling plan is sent to the on-duty units within the jurisdiction.

9. The smart city public safety management system based on video surveillance according to claim 8, characterized in that: The execution feedback module includes a feedback data collection unit; The feedback data collection unit generates a feedback data set F by collecting the execution feedback data of the resource scheduling plan to quantify the execution effect of the resource scheduling plan; Among them, the feedback data set F includes response time RT, resource utilization RC and risk prediction accuracy SR; The response time RT represents the time difference between the available resource unit m and the time difference between the resource scheduling plan sent to the arrival site based on the timestamp obtained by the GPS device; Resource utilization rate RC represents the ratio of the actual number of resources on site to the predicted number of required resources; The risk prediction accuracy rate SR represents the ratio of the number of correct risk predictions to the total number of risk predictions.

10. The smart city public safety management system based on video surveillance according to claim 8, characterized in that: The iterative optimization module includes an iterative optimization unit; The iterative optimization unit dynamically optimizes the cross-regional risk prediction algorithm and dynamic resource scheduling algorithm based on the feedback data set F; The cross-regional risk prediction algorithm optimization scheme is as follows: Based on the risk prediction accuracy rate SR in the feedback dataset F, if the risk prediction accuracy rate SR is ≥ 95%, the prediction result is judged to be reasonable, and there is no need to dynamically adjust the contribution weight coefficients α, β, and γ in the cross-regional risk prediction algorithm; If the risk prediction accuracy rate SR is less than 95%, the prediction result is judged to be unreasonable, and the contribution weight coefficients α, β and γ in the cross-regional risk prediction algorithm are dynamically adjusted; The optimization scheme of dynamic resource scheduling algorithm is as follows: Based on the response time RT and resource utilization RC in the feedback dataset F; Optimization direction 1: If the single response time RT is ≤ 15 minutes, the nth path in the selected candidate path set p is determined to be reasonable, and there is no need to dynamically adjust the contribution weight coefficients δ and w in the dynamic resource scheduling algorithm; If the single response time RT is greater than 15 minutes, the nth path in the selected candidate path set p is determined to be unreasonable, and the contribution weight coefficients δ and w in the dynamic resource scheduling algorithm are dynamically adjusted; Optimization direction 2: If the resource utilization RC is ≥ 80% for three consecutive times, the available resource unit m is determined to be reasonably selected, and there is no need to dynamically adjust the contribution weight coefficients δ and w in the dynamic resource scheduling algorithm; If the resource utilization rate RC is less than 80% for three consecutive times, it is determined that the selection of the available resource unit m is unreasonable, and the contribution weight coefficients δ and w in the dynamic resource scheduling algorithm are dynamically adjusted.

Images