Linkage management method for patrol and security monitoring of unmanned aerial vehicle

Abstract

The invention discloses an unmanned aerial vehicle patrol and security monitoring linkage management method. According to the system, urban geographic information data and key monitoring area coordinates are acquired, hierarchical grid division is realized by using a quadtree spatial index algorithm, and a patrol task distribution basic data structure is established; solving an approximate optimal solution of the patrol path of the unmanned aerial vehicle by adopting an improved greedy strategy, and performing optimization adjustment according to technical parameters and weather conditions of the unmanned aerial vehicle; the unmanned aerial vehicle collects video data in real time through high-definition camera equipment in the patrol process, target detection is conducted through a deep learning model, and potential safety hazards such as personnel gathering and vehicle abnormity are recognized; detecting violent conflicts and illegal intrusion behaviors in combination with a time sequence analysis algorithm, and generating an abnormal event report; cooperative work of the unmanned aerial vehicle, the ground robot and the command center is realized; and a patrol scheme is dynamically adjusted according to meteorological conditions, and robot scheduling is optimized through energy consumption management.

Description

A method for integrated management of drone patrols and security monitoring Technical Field

[0001] This invention relates to the field of intelligent security technology, specifically to a method for the coordinated management of drone patrols and security monitoring, and particularly to a method for achieving air-ground collaborative intelligent patrol and monitoring of key urban areas using computer technology, graph theory algorithms, artificial intelligence, and multi-objective optimization technology. Background Technology

[0002] Drone patrol and security monitoring technology is an important component of modern urban security and prevention systems, widely used in public safety, traffic management, emergency rescue, and other fields. With the acceleration of urbanization and the continuous improvement of security demands, traditional manual patrols and fixed monitoring points can no longer meet the needs of large-scale, all-weather, and intelligent security.

[0003] Currently, the main technologies include single-drone patrol systems and ground robot patrol systems. Single-drone patrol systems conduct aerial monitoring through preset paths, offering advantages such as wide coverage and high mobility; while ground robot patrol systems conduct ground patrols through fixed routes or simple scheduling algorithms, featuring long endurance and large payload.

[0004] The most relevant existing technology is the air-ground cooperative patrol system based on a fixed algorithm. Its working principle involves triggering a ground equipment response when an anomaly is detected by aerial surveillance, based on preset scheduling rules. This system employs traditional greedy or heuristic algorithms for path planning and allocates tasks using simple distance-first or time-first strategies.

[0005] However, existing technologies have the following key problems: First, traditional algorithms have excessive computational complexity when dealing with large-scale urban area patrol scheduling, making it difficult to respond quickly in real-time environments; second, there is a lack of efficient approximation algorithms for ground equipment scheduling under multiple constraints, resulting in lengthy response times and affecting security effectiveness. Summary of the Invention

[0006] The purpose of this invention is to provide a method for the coordinated management of drone patrol and security monitoring, which aims to solve the technical problems of high computational complexity and low efficiency of ground equipment scheduling under multiple constraints in the existing technology for large-scale urban area patrol scheduling, and to achieve efficient air-ground collaborative intelligent patrol and security monitoring.

[0007] To achieve the above objectives, this invention provides a method for the coordinated management of drone patrols and security monitoring, comprising: acquiring urban geographic information data and coordinates of key monitoring areas; dividing the key monitoring areas into hierarchical grids according to security level and terrain features using a quadtree spatial indexing algorithm; establishing an adjacency matrix between grid cells; and generating a basic data structure for patrol task allocation containing spatial relationships and attribute information of all grid cells; using the basic data structure for patrol task allocation, transforming the grid coverage problem into a minimum dominance set problem in graph theory; employing an improved greedy strategy to select the location points that cover the most unvisited grids in each iteration and introducing random perturbations to avoid local optima, achieving an O(n log n) time complexity. Within an n-time complexity, a patrol path scheme with an approximation ratio of ln(Δ)+1 is solved and optimized, outputting an aerial patrol scheduling scheme containing patrol order, estimated arrival time, and dwell time, where n is the number of grids and Δ is the maximum grid degree. Based on the aerial patrol scheduling scheme, real-time video streams are collected according to the predetermined path. A pre-trained deep learning model is used to detect and identify personnel gatherings, vehicle anomalies, and abandoned items. A temporal analysis algorithm is used to perform behavioral analysis on continuous frame images to detect abnormal behavior patterns and calculate the confidence score of abnormal events, generating an abnormal event report containing event type, location, confidence score, and urgency level. Based on the abnormal event report, a multi-constraint optimization model is established to abstract the ground robot scheduling problem into a vehicle path problem with a time window. A (2+ε) approximation algorithm based on Lagrange relaxation is used with dynamic programming techniques to achieve a quasi-polynomial time complexity of O(n^2). 1 / ε The system solves for an approximate optimal solution and dynamically adjusts the weight coefficients according to the urgency level of the abnormal event. It outputs a ground response plan that includes a robot allocation scheme, a travel route, and an estimated arrival time, where ε is a preset accuracy parameter. Based on the ground response plan, a three-level early warning system is established to automatically trigger the corresponding level of linkage response according to the confidence score and urgency level of the abnormal event report.

[0008] Preferably, the process involves acquiring urban geographic information data and coordinates of key monitoring areas, using a quadtree spatial indexing algorithm to hierarchically divide the key monitoring areas into grids based on security levels and terrain features, establishing an adjacency matrix between grid units, and generating a basic data structure for patrol task allocation containing spatial relationships and attribute information of all grid units. This includes: acquiring urban geographic information data and coordinates of key monitoring areas; determining the initial grid division level based on security levels and terrain features; generating an initial grid set; recursively subdividing each grid using a quadtree spatial indexing algorithm based on the initial grid set, assigning each grid unit location coordinates, security level weight, and terrain complexity coefficient to obtain a hierarchical grid division result; and calculating the distance and access difficulty between each grid unit based on the hierarchical grid division result, establishing an adjacency matrix between grid units, and obtaining the basic data structure for patrol task allocation.

[0009] Preferably, the attribute information of the grid cell includes location coordinates, security level weight, and terrain complexity coefficient; the adjacency matrix records the distance and passage difficulty between each grid cell.

[0010] Preferably, using the patrol task allocation basic data structure, the grid coverage problem is transformed into a minimum dominance set problem in graph theory. An improved greedy strategy is adopted to select the location point that can cover the most unvisited grids in each iteration and introduce random perturbation to avoid local optima. A patrol path scheme with an approximate ratio of ln(Δ)+1 is solved in O(n log n) time complexity and optimized. The output is an aerial patrol scheduling scheme containing patrol order, expected arrival time, and dwell time. This includes: using the patrol task allocation basic data structure to model the grid coverage problem as a minimum dominance set problem in graph theory, determining the objective function and constraints, and obtaining an optimization model; based on the optimization model, using an improved greedy strategy to calculate the number of unvisited grids that each candidate location point can cover in each iteration and selecting the location point with the most coverage, introducing a random perturbation mechanism to avoid getting trapped in local optima, and solving the problem in O(n log n) time complexity to obtain an approximate ratio of ln(Δ)+1 patrol path scheme and optimizing it. Solving within an n) time complexity yields an initial patrol path scheme with an approximation ratio of ln(Δ)+1; based on the initial patrol path scheme, UAV endurance parameters and real-time weather condition data are obtained, and the initial patrol path scheme is optimized through constraint condition verification and path adjustment algorithms to obtain the aerial patrol scheduling scheme.

[0011] Preferably, based on the aerial patrol scheduling scheme, real-time video streams are collected along a predetermined path. A pre-trained deep learning model is used to detect and identify gatherings of people, abnormal vehicles, and abandoned items. A temporal analysis algorithm is employed to perform behavioral analysis on consecutive frames to detect abnormal behavior patterns and calculate the confidence score of the abnormal event. An abnormal event report is then generated, containing the event type, location, confidence score, and urgency level. This includes: based on the aerial patrol scheduling scheme, controlling a high-definition camera mounted on a drone to capture real-time video stream data along a predetermined path; based on the real-time video stream data, a pre-trained deep learning model is used to perform target detection on the video frames, identifying gatherings of people, abnormal vehicles, and abandoned items, obtaining target detection results; based on the target detection results, a temporal analysis algorithm is used to perform behavioral pattern analysis on consecutive frames to detect violent conflicts and illegal intrusions, and the confidence score of the abnormal event is calculated by combining environmental context information such as time, location, and weather, resulting in the abnormal event report.

[0012] Preferably, the pre-trained deep learning model is an improved version of YOLO v5; the abnormal behavior patterns include violent conflicts and illegal intrusions; and the environmental context information includes time, location, and weather.

[0013] Preferably, based on the abnormal event report, a multi-constraint optimization model is established to abstract the ground robot scheduling problem into a vehicle routing problem with a time window. A (2+ε) approximation algorithm based on Lagrange relaxation is employed, using dynamic programming techniques to achieve a quasi-polynomial time complexity of O(n^2). 1 / ε Within a given timeframe, an approximate optimal solution is found, and weighting coefficients are dynamically adjusted based on the urgency level of the abnormal event. The resulting ground response plan includes a robot allocation scheme, travel route, and estimated arrival time. This includes: based on the abnormal event report, obtaining robot quantity limits, battery life, maximum payload, and road conditions; establishing a multi-constraint optimization model for the vehicle routing problem with a time window to obtain a set of constraints; and based on this set of constraints, employing a (2+ε) approximation algorithm based on Lagrange relaxation, and utilizing dynamic programming techniques to achieve a quasi-polynomial time complexity of O(n^2). 1 / ε Solve within the range to obtain the initial scheduling scheme; based on the initial scheduling scheme, dynamically adjust the weight coefficient of each abnormal event according to the urgency level in the abnormal event report, giving priority to ensuring a rapid response to high-priority events, and obtain the ground response scheme.

[0014] Preferably, based on the ground response plan, a three - level early warning system is established. According to the confidence score and the level of urgency of the abnormal event report, corresponding - level linkage responses are automatically triggered. Through a unified dispatching platform, three - party collaboration of continuous tracking of drones, rapid arrival of ground robots at the scene, and real - time monitoring by the command center is achieved, and a real - time communication link is established. The progress of disposal is pushed to relevant responsible persons through mobile terminals, forming a full - process closed - loop management from abnormal event discovery to disposal completion, including: based on the ground response plan, establishing a three - level early warning system of blue early warning, yellow early warning, and red early warning, determining the early warning level according to the confidence score and the level of urgency of the abnormal event report, and automatically triggering the corresponding - level linkage response mechanism; based on the linkage response mechanism, sending instructions through a unified dispatching platform to achieve three - party collaboration of continuous tracking of abnormal targets by drones, rapid arrival of ground robots at the scene according to the ground response plan, and real - time monitoring of the disposal process by the command center, establishing a real - time communication link among aerial devices, ground devices, and the command center, and obtaining collaborative disposal status data; based on the collaborative disposal status data, pushing real - time disposal progress information to relevant responsible persons through mobile terminals, and obtaining full - process closed - loop management from abnormal event discovery to disposal completion.

[0015] Preferably, the method further includes: based on the aerial patrol scheduling plan, obtaining real - time meteorological information including wind speed, rainfall, and visibility, evaluating the impact of weather on the flight safety of drones through a Bayesian network model, recalculating the patrol path using a dynamic weight adjustment algorithm, and automatically triggering an alternate path or a temporary landing procedure when adverse weather conditions are detected, obtaining a dynamically adjusted aerial patrol scheduling plan.

[0016] Preferably, the method further includes: based on the ground response plan, establishing a robot battery status monitoring system to continuously track the power consumption and remaining working time of each ground robot, dynamically adjusting task allocation through a load - balancing algorithm, automatically transferring the tasks of a robot with insufficient power to nearby devices with sufficient power when a certain robot has insufficient power, and planning the optimal return path to the charging station, obtaining a ground response plan with optimized energy consumption.

[0017] The beneficial effects of the present invention are as follows: 1. For the first time, a nearly linear time - complexity algorithm is applied to the optimization of the patrol paths of drones in large - scale areas, significantly improving the computational efficiency, making it possible to complete path planning within the time complexity of O(n log n), and effectively solving the problem of excessively high computational complexity of traditional algorithms in large - scale area scheduling; 2. Innovatively, a (2 + ε) approximation algorithm is used to solve the multi - constraint scheduling problem of ground robots, within quasipolynomial time O(n 1 / ε1) Achieving high-quality approximate solutions within a given timeframe balances computational efficiency with solution quality, effectively addressing the problem of low ground equipment scheduling efficiency under multiple constraints in existing technologies; 2) Constructing a multi-level linkage early warning mechanism based on confidence assessment, automatically triggering different levels of early warning responses according to the type, confidence level, and urgency of abnormal events, improving the system's accuracy in identifying abnormal events and its response speed; 3) Realizing integrated air-ground collaboration between UAV patrols and ground robot responses, forming a closed-loop management system from anomaly detection to event handling, significantly enhancing the intelligence level of urban security monitoring and its ability to respond to emergencies. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of this disclosure, the accompanying drawings used in the embodiments will be briefly described below. These drawings are incorporated in and constitute a part of this specification. They illustrate embodiments conforming to this disclosure and, together with the specification, serve to explain the technical solutions of this disclosure. It should be understood that the following drawings only show some embodiments of this disclosure and should not be considered as limiting the scope. Those skilled in the art can obtain other related drawings based on these drawings without creative effort.

[0019] Figure 1 is a flowchart illustrating the linkage management method of drone patrol and security monitoring according to the present invention; Figure 2 is a schematic diagram illustrating the grid division of key urban areas in the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of this disclosure clearer, the technical solutions of the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this disclosure, and not all embodiments. The components of the embodiments of this disclosure described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0021] Therefore, the following detailed description of the embodiments of this disclosure provided in the accompanying drawings is not intended to limit the scope of the claimed disclosure, but merely to illustrate selected embodiments of the disclosure. All other embodiments obtained by those skilled in the art based on the embodiments of this disclosure without inventive effort are within the scope of this disclosure.

[0022] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0023] In this document, the term "and / or" merely describes a relationship, indicating that three relationships can exist. For example, A and / or B can represent three cases: A alone, A and B simultaneously, and B alone. Furthermore, the term "at least one" in this document means any combination of at least two of any one or more elements. For example, including at least one of A, B, and C can mean including any one or more elements selected from the set consisting of A, B, and C.

[0024] As shown in Figure 1, this invention provides a method for the coordinated management of drone patrols and security monitoring, including: acquiring urban geographic information data and coordinates of key monitoring areas, and dividing the area into patrol grids; calculating the optimal patrol path for the drones; performing anomaly detection in video images; calculating the ground robot response scheme; and realizing multi-level coordinated early warning response. The specific implementation methods of this invention will be described in detail below. Examples

[0025] As shown in Figure 2, urban geographic information data and coordinates of key monitoring areas are obtained. The key monitoring areas are then divided into hierarchical grids according to security level and terrain features using a quadtree spatial indexing algorithm. An adjacency matrix between grid units is established, and a basic data structure for patrol task allocation containing spatial relationships and attribute information of all grid units is generated.

[0026] Step S1: First, obtain urban geographic information data and coordinates of key monitoring areas. Urban geographic information data includes data on building distribution, road network, and terrain elevation; coordinates of key monitoring areas include the geographical location information of key locations such as commercial centers, transportation hubs, and government agencies.

[0027] Taking a central urban area as an example, obtain 3D building model data, road network topology data, and latitude and longitude coordinates of key monitoring areas within that area.

[0028] Step S2: Determine the initial grid division level based on the security level and terrain features, and generate the initial grid set. The security level refers to the importance and security risk of the area, usually divided into high, medium, and low levels; terrain features refer to the terrain complexity of the area, such as flat land, hills, or densely populated areas of high-rise buildings.

[0029] For example, for commercial center areas, the security level is set to "high", the terrain feature is "dense high-rise building area", and the initial grid division level is set to level 4; while for ordinary residential areas, the security level is set to "medium", the terrain feature is "flat land", and the initial grid division level is set to level 2.

[0030] Step S3: Based on the initial grid set, recursively subdivide each grid using the quadtree spatial indexing algorithm, assigning each grid cell location coordinates, security level weights, and terrain complexity coefficients to obtain the hierarchical grid division result.

[0031] The quadtree algorithm recursively divides the planar space into four quadrants, and each quadrant can be further subdivided as needed.

[0032] In the implementation process, areas with higher security levels or greater terrain complexity are divided into finer-grained grids. For example, a 10km × 10km urban area may first be divided into four 5km × 5km primary grids. Then, based on security levels and terrain features, some primary grids may be further subdivided into multiple secondary grids, with the finest granularity potentially reaching 100m × 100m grid units. Each grid unit is assigned the following attributes: Location coordinates: latitude, longitude, and altitude of the grid center point; Security level weight: assigned a weight value of 1-10 based on importance, with higher importance resulting in a larger weight; Terrain complexity coefficient: reflects the density and height differences of terrain obstacles, ranging from 0.1 to 2.0. Step S4: Based on the hierarchical grid division results, the distance and access difficulty between each grid unit are calculated, an adjacency matrix between grid units is established, and the basic data structure for patrol task allocation is obtained.

[0033] The adjacency matrix records the connectivity parameters between any two grid cells, including: straight-line distance: the Euclidean distance between the center points of the two grid cells; and travel difficulty: a travel cost coefficient considering factors such as height difference and obstacle density. For example, the element M[i,j] in the adjacency matrix M represents the connectivity parameters from grid i to grid j, including the distance value and the travel difficulty value. When there is an insurmountable obstacle between two grid cells, the travel difficulty value is set to infinity, indicating that the two grid cells cannot travel directly together.

[0034] The final generated patrol task allocation basic data structure is a composite data structure containing grid cell attributes and adjacency relationships, providing basic data support for subsequent path planning and task scheduling.

[0035] Example 1 details how to acquire urban geographic information data and coordinates of key monitoring areas, and the process of hierarchical grid division using a quadtree spatial indexing algorithm. First, the system needs to acquire complete urban geographic information data, including the 3D distribution of buildings, the topology of the road network, and basic geographic information such as terrain elevation. Simultaneously, the system also needs to acquire the precise coordinates of key monitoring areas, which typically include densely populated commercial centers, transportation hubs, government agencies, and critical infrastructure—locations requiring special security protection. Taking a typical city center as an example, the system acquires 3D model data of all buildings within that area, the complete road network topology, and the latitude and longitude coordinates of each key monitoring location.

[0036] After acquiring basic data, the system determines the initial grid division level based on the security level and terrain features, generating an initial grid set. The security level reflects the importance and potential security risks of an area, typically categorized into high, medium, and low levels, each corresponding to different patrol frequencies and monitoring density requirements. Terrain features indicate the complexity of the area's terrain, such as flat, open areas, hilly areas, or areas with dense high-rise buildings. Different terrain features significantly impact the drone's patrol path and field of view. For example, for high-risk and densely populated areas like commercial centers, the system sets its security level to "high." Considering that commercial centers are typically high-rise buildings with complex terrain, the system sets a higher initial grid division level (e.g., level 4). Conversely, for ordinary residential areas, the system might set the security level to "medium," the terrain feature to "flat," and the initial grid division level accordingly to a lower level (e.g., level 2).

[0037] Next, based on the initial grid set, the system recursively subdivides each grid using a quadtree spatial indexing algorithm. The quadtree algorithm is a classic spatial data indexing method that recursively divides a two-dimensional plane into four equally sized quadrants. Each quadrant can be further subdivided as needed, forming a hierarchical spatial partitioning structure. In this system, the quadtree algorithm subdivides the grid according to different security levels and terrain features. For areas with high security levels or complex terrain, the system performs finer-grained grid partitioning to ensure more precise coverage; while for areas with low security levels and simple terrain, larger grid cells are used to reduce computation and improve patrol efficiency. In practical applications, a typical urban area (e.g., 10 square kilometers) might first be divided into four 2.5 square kilometer level-one grids, and then further subdivided to different degrees according to the characteristics of each area, ultimately forming a hierarchical structure composed of grid cells ranging in size from hundreds to tens of meters.

[0038] During the grid generation process, the system assigns key attribute information to each grid cell, including precise location coordinates (longitude, latitude, and altitude), security level weight, and terrain complexity coefficient. Location coordinates define the spatial position of the grid; security level weights are typically integer values ​​from 1 to 10, with higher values ​​indicating greater security importance and higher priority in subsequent patrol path planning; the terrain complexity coefficient reflects the density and height differences of terrain obstacles in the area, usually between 0.1 and 2.0, with higher values ​​indicating more complex terrain, greater difficulty for drone patrols, and the need for more careful path planning.

[0039] Finally, based on the hierarchical grid division results, the system calculates the distance and access difficulty between each grid cell, establishes an adjacency matrix between grid cells, and generates the basic data structure for patrol task allocation. The adjacency matrix is ​​a two-dimensional table recording the connectivity parameters between any two grid cells, mainly including straight-line distance and access difficulty. Straight-line distance is the Euclidean distance between the center points of two grid cells, directly affecting the UAV's flight time; access difficulty considers factors such as altitude difference and obstacle density, and is a comprehensive access cost coefficient that directly affects the UAV's energy consumption and safety. When there are insurmountable obstacles (such as tall buildings or restricted flight zones) between two grid cells, the system sets the access difficulty value to infinity, indicating that the two grid cells cannot pass directly and a detour is required. Through this adjacency matrix, the system can quickly query the connectivity between any two grid cells, laying the foundation for subsequent path planning.

[0040] The basic data structure for patrol task allocation is a composite data structure that integrates the attribute information and adjacency matrix of grid cells, comprehensively describing the spatial characteristics and connectivity of the entire monitoring area. This data structure is the core foundation for subsequent UAV patrol path planning and task scheduling, supporting efficient spatial queries and path calculations, ensuring that the entire system can intelligently allocate security monitoring tasks based on regional characteristics. Example

[0041] By utilizing the basic data structure for patrol task allocation, the grid coverage problem is transformed into the minimum dominance set problem in graph theory. An improved greedy strategy is adopted to select the location point that can cover the most unvisited grids in each iteration and introduce randomization perturbation to avoid local optima. The patrol path scheme with an approximation ratio of ln(Δ)+1 is solved in O(n log n) time complexity and optimized. The output is an aerial patrol scheduling scheme that includes patrol order, expected arrival time and dwell time.

[0042] Step S1: Utilizing the basic data structure for patrol task allocation, the grid coverage problem is modeled as a minimum dominating set problem in graph theory. The objective function and constraints are determined to obtain the optimization model. In graph theory, a dominating set is a subset D⊆V of vertices in a graph G=(V,E), such that any vertex in the graph is either in D or adjacent to at least one vertex in D. The minimum dominating set problem aims to find a dominating set with the smallest size. In this invention, grid cells are considered vertices in the graph. If two grid cells are adjacent or one grid cell can observe another grid cell, an edge is established between these two vertices. The goal is to find the minimum number of grid cells that can be used as observation points for the UAV, ensuring that all grid cells are covered.

[0043] The objective function is set as: minimize ∑(x_i) for i ∈ V, where x_i is a binary variable representing whether to select grid i as an observation point. The constraint condition is: x_i + ∑(x_j) ≥ 1 for all i ∈ V and j ∈ N(i), where N(i) represents the set of grids adjacent to grid i. Step S2: Based on the optimization model, an improved greedy strategy is adopted to calculate the number of unvisited grids that each candidate location point can cover in each iteration and select the location point with the most coverage. A randomized perturbation mechanism is introduced to avoid getting trapped in local optima. The solution is obtained in O(n log n) time complexity, and the initial patrol path scheme with an approximation ratio of ln(Δ)+1 is obtained. The specific algorithm implementation is as follows: 1) Initialization: Mark all grids as "uncovered" and set an empty observation point set S. 2) Greedy iteration: For each candidate grid i, calculate the number of unvisited grids it can cover c_i. Use a max-heap data structure to maintain the candidate grids and their coverage numbers. Select the top element of the heap (the grid with the most coverage) as a new observation point. Select a suboptimal solution with a certain probability p (e.g., 0.1) to introduce random perturbation. Add the newly selected observation point to set S and mark the grids it covers as "covered". Update the coverage number of the affected grids and adjust the heap. 3) Termination condition: Stop iteration when all grids are covered or the predetermined upper limit of the number of observation points is reached. Since a max-heap is used to maintain the candidate grids, the operation time for finding the maximum coverage number is O(log n), and a total of O(n) operations are required. Therefore, the total time complexity is O(n log n). It can be proven that the algorithm can obtain a solution with an approximate ratio of ln(Δ) + 1, where Δ is the maximum degree in the grid graph.

[0044] Step S3: Based on the initial patrol path plan, obtain the UAV endurance parameters and real-time weather condition data. Optimize the initial patrol path plan through constraint checks and path adjustment algorithms to obtain the aerial patrol scheduling plan. Constraints include: UAV endurance time limit: typically 30-120 minutes; UAV flight speed: typically 10-20 m / s; Weather influencing factors: the impact of wind speed, rainfall, etc. on flight. The path adjustment algorithm first checks whether the initial patrol path meets the above constraints. If not, the following adjustments are made: 1) Path segmentation: Divide the long path exceeding the endurance capacity into multiple short path segments. 2) Path smoothing: Use cubic spline interpolation to make path turns smoother and reduce energy consumption. 3) Time allocation: Allocate dwell time according to the grid security level weight, with longer dwell time in important areas. The final output aerial patrol scheduling plan includes the following information: Patrol order: A list of visit order for observation points. Estimated arrival time: The estimated arrival timestamp for each observation point. Dwell time: The dwell time at each observation point. Flight path: A detailed sequence of flight path coordinates connecting each observation point. Coverage area: A list of grid cells that can be monitored from each observation point.

[0045] Example 2 details how to utilize the patrol task allocation basic data structure to calculate the optimal patrol path for a UAV using an efficient approximation algorithm. First, the system uses the previously constructed patrol task allocation basic data structure to model the grid coverage problem as a minimum dominance set problem in graph theory. In this model, each grid cell is considered a vertex in the graph. If one grid is observable from another (i.e., two grids are mutually visible within the UAV's observation range), an edge is established between these two vertices. The goal of the minimum dominance set problem is to find the minimum number of vertices (grids) such that all other vertices in the graph are adjacent to these selected vertices. In the context of UAV patrol, this means finding the minimum number of observation points so that all monitored grid areas are covered.

[0046] The objective function of this optimization model is to minimize the number of selected observation points, while the constraint is to ensure that each grid cell is either an observation point itself or adjacent to at least one observation point (i.e., covered). In practical applications, the system also considers the security level weight of each grid cell, giving higher priority to covering areas with higher security levels. This modeling transforms patrol path planning into a structured combinatorial optimization problem that can be solved using efficient algorithms.

[0047] Based on the above optimization model, the system employs an improved greedy strategy to find the near-optimal patrol path. Traditional greedy algorithms select the location point that covers the most uncovered grid cells in each iteration, but are prone to getting trapped in local optima. This system improves the greedy strategy in two ways: firstly, it uses a max-heap data structure to maintain candidate grid cells and their coverage counts, significantly improving the efficiency of finding the location with the maximum coverage count; secondly, it introduces a randomized perturbation mechanism to select a suboptimal solution with a certain probability (usually 0.1-0.2), effectively preventing the algorithm from getting trapped in local optima.

[0048] The improved greedy algorithm is implemented as follows: First, the system marks all grids to be monitored as "uncovered" and initializes an empty set of observation points. In each iteration, the system calculates the number of unvisited grids that each candidate location can cover and maintains this information using a max-heap data structure, reducing the time complexity of finding the location with the most covered grids to O(log n). The system typically selects the top element of the heap (i.e., the location with the most covered grids) as the new observation point, but it may select a suboptimal solution with a certain probability to introduce random perturbation. After selecting a new observation point, the system adds it to the observation point set and marks all grids it covers as "covered". Then, the coverage count of the affected candidate locations is updated, and the heap structure is adjusted accordingly. The algorithm iterates this process until all grids to be monitored are covered, or the predetermined upper limit for the number of observation points is reached.

[0049] This improved greedy algorithm offers a significant advantage in time complexity. Due to the use of a max-heap data structure, the time complexity of each operation to find the maximum coverage is O(log n), and a total of O(n) such operations are required, resulting in an overall time complexity of O(n log n), where n is the number of grid cells. This is much lower than the O(n²) time complexity of the traditional algorithm, enabling the system to handle patrol path planning for large-scale grid regions in real-time environments. Theoretical analysis and practical tests both show that the algorithm can obtain a solution with an approximate ratio of ln(Δ) + 1, where Δ is the maximum degree in the grid graph (i.e., the maximum number of other grid cells a grid cell can be adjacent to).

[0050] After obtaining the initial patrol path plan, the system also needs to consider the actual operational constraints of the UAV and optimize and adjust the path. The system first acquires various technical parameters of the UAV and real-time weather condition data, including the UAV's endurance (usually ranging from 30 to 120 minutes), flight speed (usually 10-20 meters per second), and the impact of weather factors (such as wind speed, rainfall, etc.) on flight. Then, it verifies whether the initial path meets these realistic limitations through constraint checks.

[0051] If the initial path does not meet the constraints, the system will make a series of adjustments. First, path segmentation: long paths exceeding the drone's endurance are divided into multiple shorter segments, ensuring each segment remains within the drone's range. Second, path smoothing: techniques such as cubic spline interpolation are used to make path turns smoother, reducing energy consumption and flight time. Finally, time allocation: the dwell time of the drone at each observation point is rationally allocated according to the security level weight of each grid, ensuring longer monitoring time for important areas.

[0052] After these optimizations and adjustments, the system finally outputs a complete aerial patrol scheduling plan, including a detailed patrol sequence (i.e., the order in which observation points are visited), the estimated arrival time at each observation point, the duration of stay at each observation point, the detailed flight path coordinate sequence connecting each observation point, and a list of grid cells that can be monitored from each observation point. This scheduling plan comprehensively considers factors such as the importance of regional security, terrain complexity, UAV performance limitations, and weather effects, making it a patrol execution plan that meets both monitoring requirements and is practically feasible.

[0053] Through the detailed implementation of the two embodiments described above, the system successfully transforms complex urban areas into structured grid data and calculates near-optimal drone patrol paths using efficient algorithms. This method considers both the security importance and terrain features of the area, while also taking into account algorithm efficiency and solution quality, providing strong technical support for intelligent security monitoring in large-scale urban areas.

[0054] Example 3 is based on an aerial patrol scheduling scheme. Real-time video streams are collected according to a predetermined path. A pre-trained deep learning model is used to detect and identify crowds, abnormal vehicles, and abandoned items. A time-series analysis algorithm is used to perform behavioral analysis on continuous frame images to detect abnormal behavior patterns and calculate the confidence score of abnormal events. An abnormal event report is generated, which includes the event type, location, confidence score, and urgency level.

[0055] Step S1: Based on the aerial patrol scheduling scheme, control the high-definition camera on the drone to shoot in real time along a predetermined path, collecting real-time video stream data. The drone performs patrol missions according to the prescribed path and schedule of the aforementioned aerial patrol scheduling scheme, collecting video data through the onboard high-definition camera during flight. The camera adopts gimbal stabilization technology to ensure stable and clear images during flight. The main parameters of the camera include: resolution: 4K ultra-high definition (3840×2160 pixels); frame rate: 25-30 frames / second; viewing angle: adjustable 90°-120° wide-angle; zoom capability: 10-30x optical zoom. The video stream data is transmitted in real time to the backend processing server via 4G / 5G network, and is also backed up locally on the drone. In areas with unstable network, the system will automatically reduce the transmission resolution to ensure real-time performance, and will retransmit high-definition data after the network is restored.

[0056] Step S2: Based on real-time video stream data, a pre-trained deep learning model is used to perform target detection on video frames, identifying crowds, vehicle anomalies, and abandoned items, obtaining target detection results. The system uses an improved YOLO v5 deep learning model for target detection. This model is improved on the original YOLO v5 as follows: the backbone network uses EfficientNet-B4 instead of CSPDarknet53 to improve feature extraction capabilities; an attention mechanism (CBAM module) is introduced to enhance the detection capability of small targets; multi-scale feature fusion technology is used to improve adaptability to targets of different sizes; a loss function customized for security scenarios balances accuracy and recall. The target types that this model can detect include: crowds: crowds of 5 or more people; vehicle anomalies: illegal parking, driving against traffic, speeding, and other abnormal behaviors; abandoned items: luggage, packages, and other items remaining in the same location for more than a predetermined time. The model output includes three key pieces of information: target category, location coordinates (bounding box), and detection confidence.

[0057] Step S3: Based on the target detection results, a temporal analysis algorithm is used to perform behavioral pattern analysis on continuous frame images to detect violent conflicts and illegal intrusions. The confidence score of the abnormal event is calculated by combining environmental context information such as time, location and weather, and an abnormal event report is obtained. The temporal analysis algorithm analyzes target behavior patterns based on continuous multi-frame image data, mainly including: Optical Flow: analyzing target motion trajectory and velocity; Pose Estimation: analyzing human posture to determine whether violent behaviors such as fighting exist; Temporal-Spatial Correlation: analyzing the interaction relationship between multiple targets. Abnormal behavior patterns that the system can identify include: Violent conflict: fighting, robbery, and other violent behaviors; Illegal intrusion: climbing over walls, breaking down doors, and other intrusion behaviors. The fusion of environmental context information considers the following factors: Time factor: distinguishing between day / night, weekday / holiday, and other time features; Location factor: considering area type (commercial area / residential area) and historical police data; Weather factor: the impact of visibility and lighting conditions on detection. The confidence score calculation uses a Bayesian network model to comprehensively consider the target detection confidence, behavior recognition results, and environmental context information, and outputs an abnormal event confidence score between 0 and 1. The urgency level is divided into the following categories based on confidence scores: Low (blue): 0.3-0.5; Medium (yellow): 0.5-0.8; High (red): 0.8-1.0. The final generated anomaly report includes the following information: Event type: gathering of people / vehicle anomaly / item left behind / violent conflict / illegal intrusion; Location: latitude and longitude coordinates and corresponding area description; Confidence score: a value between 0 and 1; Urgency level: Low (blue) / Medium (yellow) / High (red); Event description: template-generated natural language description; Image evidence: keyframe images and video clip URLs. Example 3 details how the system performs target detection and anomaly identification based on an aerial patrol scheduling scheme and advanced video analysis technology. First, the UAV performs patrol missions according to the path and schedule specified in the aforementioned aerial patrol scheduling scheme, collecting real-time video data through its onboard high-definition camera during flight. These cameras are not ordinary devices, but high-end camera systems designed specifically for security monitoring. They employ gimbal stabilization technology to ensure stable and clear images during flight, acquiring high-quality images even under slight turbulence or wind interference. These cameras feature 4K ultra-high-definition resolution (3840×2160 pixels), capture 25-30 frames per second, and are equipped with an adjustable wide-angle lens of 90°-120° and 10-30x optical zoom, enabling them to clearly capture ground details while maintaining a high cruising altitude.

[0058] After video data is acquired, it is transmitted in real time to the backend processing server via high-speed 4G or 5G networks, while a backup is stored locally on the drone to prevent data transmission interruption. The system employs an intelligent transmission strategy: transmitting complete high-definition images when network conditions are good, automatically reducing the transmission resolution in areas with unstable networks to ensure real-time performance, and automatically retransmitting high-definition data once the network recovers. This dual-protection mechanism ensures the integrity and availability of video data, providing a reliable data foundation for subsequent video analysis.

[0059] Upon receiving video data, the system uses a pre-trained deep learning model to perform target detection on the video frames, identifying potential security risks such as crowds, abnormal vehicles, and leftover items. This system does not use the standard YOLO v5 model, but a specially modified version optimized for security monitoring scenarios. Specific improvements include: replacing the original CSPDarknet53 backbone network with the more efficient EfficientNet-B4, significantly improving feature extraction capabilities; introducing an attention mechanism (CBAM module) to enhance the detection of small targets, crucial for identifying small objects on the ground from high altitudes; employing multi-scale feature fusion technology to improve adaptability to targets of different sizes; and customizing a special loss function for security scenarios to achieve a better balance between accuracy and recall, reducing false positives and false negatives.

[0060] This improved deep learning model can accurately detect various security-related target types, including crowds (defined as groups of 5 or more people), abnormal vehicle behavior (such as illegal parking, driving against traffic, speeding, etc.), and abandoned items (such as luggage, packages, etc., remaining in the same location for more than a predetermined time). The model output includes three key pieces of information: target category (indicating the type of target detected), location coordinates (precisely marking the target's position in the image using bounding boxes), and detection confidence (representing the model's degree of confidence in the detection result, typically a value between 0 and 1). This information provides the foundational data for subsequent behavior analysis.

[0061] Simply detecting a target is insufficient to determine the existence of a security threat; the system also needs to analyze the target's behavioral patterns. Based on the target detection results, the system employs temporal analysis algorithms to analyze behavioral patterns in consecutive frames of images, detecting abnormal behaviors such as violent conflicts and illegal intrusions. The temporal analysis algorithm integrates multiple computer vision techniques: optical flow is used to analyze the target's trajectory and speed, detecting abnormal rapid movement or sudden stops; pose estimation is used to analyze human posture to determine if violent behaviors such as fighting exist; and temporal-spatial correlation analysis is used to study the interaction relationships between multiple targets, identifying potential group conflicts or collaborative behaviors.

[0062] The system can identify typical abnormal behavior patterns including violent conflicts (such as fights, robberies, and other violent acts) and illegal intrusions (such as scaling walls, breaking down doors, and other intrusion behaviors). To improve the accuracy of identification, the system also integrates environmental context information, considering time factors (distinguishing between day / night, weekdays / holidays, etc.), location factors (considering area type such as commercial / residential areas, historical police incident data, etc.), and weather factors (such as the impact of visibility and lighting conditions on detection). For example, gatherings of people in commercial areas late at night are more likely to be identified as abnormal than during the day; while gatherings of people near schools during dismissal time are considered normal phenomena.

[0063] After behavioral analysis, the system uses a Bayesian network model to calculate the confidence score of the anomalous event. This model comprehensively considers the target detection confidence, behavior recognition results, and environmental context information, outputting an anomalous event confidence score between 0 and 1. For example, if the system detects a group of people gathering at the bank entrance, identifies possible violent postures, and considers that the bank is not open during business hours, the comprehensive analysis of these three pieces of information may result in a high confidence score of 0.85. Based on this confidence score, the system classifies the urgency level into three levels: events with a confidence score between 0.3 and 0.5 are marked as low-level (blue) emergencies; events with a confidence score between 0.5 and 0.8 are marked as medium-level (yellow) emergencies; and events with a confidence score between 0.8 and 1.0 are marked as high-level (red) emergencies.

[0064] Ultimately, the system generates an anomaly report containing rich information, including the event type (clearly specifying whether it's a gathering of people, vehicle malfunction, abandoned items, violent conflict, or illegal intrusion), location (precise latitude and longitude coordinates and corresponding area description, such as "northeast corner of Financial Street commercial district"), confidence score (a value between 0 and 1, indicating the degree of certainty in the event assessment), urgency level (low / medium / high, corresponding to blue / yellow / red alert levels respectively), event description (a natural language description automatically generated by the system based on a template, such as "Approximately 7 people were detected gathering in the northeast corner of Financial Street commercial district, with suspected violent behavior, confidence level 0.85, immediate action recommended"), and image evidence (keyframe images and access links to relevant video clips). This detailed anomaly report provides comprehensive contextual information for subsequent ground response, enabling ground robots to take targeted actions.

[0065] Example 4 establishes a multi-constraint optimization model based on abnormal event reporting. The ground robot scheduling problem is abstracted into a vehicle routing problem with a time window. A (2+ε) approximation algorithm based on Lagrange relaxation is used, employing dynamic programming techniques to achieve a quasi-polynomial time complexity of O(n^2). 1 / ε The system solves for an approximate optimal solution and dynamically adjusts the weight coefficients according to the urgency level of the abnormal event, outputting a ground response plan that includes robot allocation scheme, travel route, and estimated arrival time.

[0066] Step S1: Based on the abnormal event report, obtain the robot quantity limit, battery life, maximum load, and road traffic conditions to establish a multi-constraint optimization model for the vehicle routing problem with time windows, thus obtaining the set of constraints. The vehicle routing problem with time windows (VRPTW) is a classic combinatorial optimization problem. Its core is to plan a set of vehicle paths that minimize the total cost while satisfying various constraints. In this invention, vehicles correspond to ground robots, and customer points correspond to the locations where abnormal events occur.

[0067] The system first obtains the following resource constraints: Robot quantity limit: Total number of available ground robots n_robots; Battery life: Maximum working time of each robot t_max; Maximum load: Maximum load capacity of the robot w_max; Road traffic conditions: Road network data including traffic congestion index and road construction information. Then, it sets time window constraints for each abnormal event: Earliest arrival time e_i: Event reporting time; Latest arrival time l_i: Deadline set based on urgency level: High (red): Reporting time + 10 minutes; Medium (yellow): Reporting time + 20 minutes; Low (blue): Reporting time + 30 minutes. Multiple appointments established... The bundle optimization model comprises the following main components: Decision variables: x_{i,j,k}: indicates whether robot k travels from location i to location j; y_{i,k}: indicates whether robot k processes event; it_{i,k}: indicates the time it takes for robot k to reach location i. Objective function: Minimize the weighted sum of total response time and total travel distance. Constraints: Each event must be processed by one and only one robot; each robot's route must start from the origin and return; the time for the robot to reach each event location must be within a specified time window; the robot's working time cannot exceed the battery's endurance; the robot's workload cannot exceed its maximum load capacity. Step S2: Based on the constraint set, a (2+ε) approximation algorithm based on Lagrange relaxation is adopted, and dynamic programming is used to achieve a quasi-polynomial time O(n^2) time complexity. 1 / ε The initial scheduling scheme is obtained by solving the problem within a certain time. Since VRPTW is an NP-hard problem, it is difficult to obtain an exact solution in a reasonable time for large-scale instances. Therefore, this invention employs an efficient approximation algorithm. Lagrange relaxation is a method that transforms difficult-to-handle constraints into penalty terms of the objective function using Lagrange multipliers, thus decomposing the original problem into more easily solvable subproblems.

[0068] The main steps of the algorithm are as follows: 1) Problem relaxation: Relax the hard constraint that "each event must be handled by one and only one robot" into the objective function using the Lagrange multiplier λ_i. 2) Sub-problem solving: Decompose the relaxed problem into n_robots single-robot path planning sub-problems. 3) Dynamic programming solution: For each sub-problem, construct a state-space network and apply the dynamic programming algorithm to solve it. State definition: (S,i,t) represents the visited set S, current position i, and current time t. State transition: f(S,i,t) = min{f(S-{j},i,t-t_{i,j}) + c_{i,j}} 4) State space reduction: Discretize the continuous time using ε-grid technology, with a time precision of (L / n)·ε, where L is the time span. 5) Solution integration: Iteratively update the Lagrange multiplier λ_i using the sub-gradient method, and combine it with a greedy allocation strategy to integrate the sub-problem solutions into a feasible solution to the original problem. This algorithm has a quasi-polynomial time complexity of O(n^2).1 / ε Within a given range, a solution with an approximate ratio of (2+ε) can be obtained, where n is the number of events and ε is a preset precision parameter. In practical applications, ε is usually set to a decimal between 0.1 and 0.5; a smaller ε value will produce a more accurate solution but will take longer to compute.

[0069] Step S3: Based on the initial scheduling scheme, dynamically adjust the weight coefficients of each abnormal event according to the urgency level in the abnormal event report, prioritizing the rapid response of high-priority events to obtain the ground response scheme. The system dynamically sets the weight coefficients according to the urgency level of the abnormal events: High (red): weight coefficient w_high = 3.0; Medium (yellow): weight coefficient w_medium = 2.0; Low (blue): weight coefficient w_low = 1.0. These weight coefficients directly affect the importance of each event in the objective function; events with higher weights are given priority in the solution process. The system adopts the following optimization strategies: 1) Redefine the objective function: ∑w_i·t_i, where w_i is the weight coefficient of event i and t_i is the response time. 2) Apply simulated annealing algorithm to locally optimize the initial solution: generate a new solution by randomly swapping the processing order of two events or transferring the event from one robot to another. Decide whether to accept the new solution based on the degree of improvement of the objective function and the current "temperature". Gradually reduce the "temperature" parameter so that the algorithm gradually converges to the local optimum. 3) Conflict detection and adjustment: check whether high-priority events are responded to in a timely manner, and manually adjust the arrangement of low-priority events if necessary. The final output ground response scheme includes the following: Robot allocation scheme: list of events that each robot is responsible for handling. Travel route: detailed travel path and turning points of each robot. Estimated arrival time: estimated time for the robot to arrive at each event location. Task description: handling guidelines and precautions for each event. Collaboration arrangement: division of labor for multiple robots to collaboratively handle the same event. Example 4 details how the system calculates the optimal scheduling scheme for ground robots based on abnormal event reports and through an efficient multi-constraint optimization algorithm. First, based on abnormal event reports, the system obtains the constraints on robot resources, including the total number of available ground robots, the battery life of each robot, its maximum payload capacity, and the current road traffic conditions. These constraints constitute the resource limitations of the scheduling problem. For example, the system may have 8 ground patrol robots available for scheduling, each with a battery life of approximately 4-6 hours and a maximum payload of 20 kg. The system also acquires road network data containing real-time traffic congestion and road construction information.

[0070] Next, the system sets time window constraints for each exception event that needs to be handled. The start time of the time window is the time the event was reported, while the end time is dynamically set according to the urgency of the event: the processing deadline for high-urgency (red) events is within 10 minutes after the reporting time, for medium-urgency (yellow) events it is within 20 minutes, and for low-urgency (blue) events it is within 30 minutes. This differentiated time window setting ensures that high-urgency events receive a more timely response, reflecting the system's priority handling mechanism.

[0071] Based on the aforementioned resource and time window constraints, the system establishes a complex multi-constraint optimization model, abstracting the ground robot scheduling problem into a vehicle routing problem with a time window (VRPTW). In this model, decision variables include: which robot is responsible for handling which event (represented by the binary variable x_{i,j,k} indicating whether robot k travels from location i to location j), and the arrival time of each robot (represented by the continuous variable t_{i,k} indicating the time robot k takes to arrive at location i). The model's objective function is to minimize the weighted sum of total response time and total travel distance, with higher weight given to the response time of high-urgency events. Constraints include: each event must be handled by one and only one robot; each robot's route must start from the origin and end in a return path; the robot's arrival time at each event location must be within the specified time window; the robot's working time cannot exceed its battery life; and the robot's workload cannot exceed its maximum payload capacity.

[0072] Since VRPTW is an NP-hard problem, it is difficult to find the exact optimal solution within a reasonable time for large-scale instances. Therefore, the system adopts a (2+ε) approximation algorithm based on Lagrange relaxation to solve this problem. Lagrange relaxation is a method that transforms difficult-to-handle constraints into penalty terms in the objective function using Lagrange multipliers, decomposing the original problem into more easily solvable subproblems. In this system, the hard constraint that "each event must be handled by one and only one robot" is first relaxed into the objective function using Lagrange multipliers λ_i, transforming it into a penalty for solutions that fail to satisfy this constraint. In this way, the originally complex problem is decomposed into multiple independent single-robot path planning subproblems, each considering how a robot can handle as many events as possible within its capabilities.

[0073] For these sub-problems, the system employs a dynamic programming algorithm for solution. Specifically, the system constructs a state-space network, where each state is defined as a triple (S,i,t), representing the set of currently visited events S, the current position i, and the current time t. The state transition equation describes the cost of transitioning from one state to another, i.e., the time and distance cost required to travel from the current position to the next event location. However, since time is a continuous variable, the state space becomes extremely large, making efficient solution difficult. Therefore, the system uses an ε-grid technique to discretize the continuous time, with a time precision set to (L / n)·ε, where L is the entire scheduling time span, n is the number of events, and ε is a preset precision parameter. This discretization technique reduces the state space size to a manageable range, but at the cost of introducing an approximation error of ε.

[0074] After solving multiple subproblems, the system iteratively updates the Lagrange multipliers λ_i using the subgradient method and combines this with a greedy allocation strategy to integrate the solutions to the subproblems into a feasible solution to the original problem. This method can achieve a quasi-polynomial time complexity of O(n^2). 1 / ε The approximate solution ratio is (2+ε), where n is the number of events and ε is a preset precision parameter. In practical applications, ε is usually set to a decimal between 0.1 and 0.5. Smaller ε values ​​produce more accurate solutions but take longer to compute. A balance between accuracy and computational efficiency needs to be struck based on the actual requirements.

[0075] After obtaining the initial scheduling plan, the system needs to adjust the priorities based on the urgency of the events. The system assigns different weight coefficients to events of different urgency levels: high-level (red) urgency events have a weight of 3.0, medium-level (yellow) urgency events have a weight of 2.0, and low-level (blue) urgency events have a weight of 1.0. These weights directly affect the importance of each event in the objective function; events with higher weights are given priority in the solution process and are more likely to achieve faster response times.

[0076] The system further employs simulated annealing to locally optimize the initial solution, attempting to improve the solution quality by swapping the order of event processing or transferring events from one robot to another. Simulated annealing is a heuristic optimization method that allows the algorithm to accept a certain degree of inferior solutions during the search process, in order to escape local optima. Specifically, the algorithm randomly generates new solutions (such as swapping the processing order of two events) and decides whether to accept this new solution based on the degree of improvement in the objective function and the current "temperature" parameter. In the early stages of the algorithm, the temperature is higher, making it easier to accept inferior solutions and increasing the diversity of the search space; as the algorithm progresses, the temperature gradually decreases, and the algorithm becomes more "picky," tending to accept only improved solutions, eventually converging to a local optimum.

[0077] In addition, the system performs conflict detection and adjustment, checking whether high-priority events can be responded to in a timely manner. If the expected response time of a high-priority event is found to exceed its time window requirement, the system will manually adjust the scheduling of low-priority events, and may even temporarily suspend some low-priority events to ensure that high-priority events are processed in a timely manner.

[0078] Ultimately, the system outputs a complete ground response plan, including a detailed robot allocation plan (clearly indicating which events each robot is responsible for handling), a precise route (including detailed waypoints and turning navigation instructions), the estimated arrival time of each event (for real-time monitoring and adjustment), handling guidelines and precautions for each event (such as whether special equipment or handling techniques are required), and multi-robot collaboration arrangements (when a single robot cannot handle a complex event, the system will arrange multiple robots to work together, with a clear division of labor).

[0079] This ground response solution comprehensively considers factors such as the urgency of the event, robot capability limitations, and road conditions. Through an efficient approximation algorithm, it achieves high-quality scheduling results within an acceptable timeframe, ensuring that ground robots can respond optimally to anomalies detected by drones, achieving seamless air-ground collaboration. Particularly for high-urgency events, the system can guarantee the deployment of ground robots to the scene in a very short time, significantly improving the response speed and handling efficiency of the security system.

[0080] Example 5 establishes a three-level early warning system based on the ground response scheme. The system automatically triggers the corresponding level of linkage response according to the confidence score and urgency level of the abnormal event report. Through a unified scheduling platform, it realizes the three-party collaboration of continuous UAV tracking, rapid on-site arrival of ground robots and real-time monitoring by the command center, and establishes a real-time communication link. The system pushes the handling progress to relevant responsible persons through mobile terminals, forming a closed-loop management of the entire process from the discovery of abnormality to the completion of handling.

[0081] Step S1: Based on the ground response plan, establish a three-tiered early warning system of blue, yellow, and red alerts. The alert level is determined according to the confidence score and urgency level of the abnormal event report, and the corresponding linkage response mechanism is automatically triggered. The specific settings of the three-tiered early warning system are as follows: Blue Alert (General Abnormality): Triggering condition: Confidence 0.3-0.5, Low urgency response measures: Dispatch a single robot to investigate, drones maintain regular patrols. Notification scope: Regional security personnel. Yellow Alert (Significant Abnormality): Triggering condition: Confidence 0.5-0.8, Medium urgency response measures: Dispatch 1-2 robots to handle the situation, drones adjust their patrol paths for observation. Notification scope: Regional security personnel and security supervisors. Red Alert (Major Abnormality): Triggering condition: Confidence 0.8-1.0, High urgency response measures: Dispatch multiple robots to the scene urgently, drones hover for fixed-point monitoring. Notification scope: Regional security personnel, security supervisors, and relevant department heads. The system automatically calculates the alert level based on the confidence score and urgency level in the abnormal event report and triggers the corresponding linkage response mechanism. The warning level is calculated using a weighted average method: Warning Index = 0.7 × Confidence Level + 0.3 × Urgency Level Warning Index < 0.5: Blue Warning 0.5 ≤ Warning Index < 0.8: Yellow Warning Warning Index ≥ 0.8: Red Warning Step S2: Based on the linkage response mechanism, instructions are sent through the unified dispatch platform to achieve three-way collaboration: continuous tracking of abnormal targets by UAVs, rapid deployment of ground robots to the scene according to the ground response plan, and real-time monitoring of the handling process by the command center. A real-time communication link is established between aerial equipment, ground equipment, and the command center to obtain collaborative handling status data. The unified dispatch platform is the core control center of the entire system, responsible for coordinating the collaborative work of various equipment.The platform has the following functional modules: Command distribution module: generates and issues task commands to various devices according to the warning level; Real-time positioning module: tracks the location information of all devices and displays it visually; Video aggregation module: integrates video streams from UAVs and ground robots and performs intelligent analysis; Communication management module: ensures real-time communication and data transmission between devices; Status monitoring module: monitors the working status and task execution of each device. The three-party collaborative working mode is as follows: 1) Continuous UAV tracking: After receiving an abnormal event report, the nearest UAV adjusts its flight path to the event location, uses computer vision technology to lock and track the abnormal target, adjusts its hovering altitude and monitoring range according to the warning level, and transmits high-definition video streams to the command center in real time. 2) Ground robots rush to the scene: select the best route to the event location according to the ground response plan, continuously receive route adjustment instructions from the command center en route, and handle the situation according to the preset procedure after arriving at the scene, upload the scene situation, and wait for further instructions. 3) Command Center Real-Time Monitoring: A large screen simultaneously displays video streams from drones and ground robots, providing real-time analysis of the situation and adjusting response strategies. When necessary, it contacts human security personnel for support. The entire incident handling process is recorded. Real-time communication links employ a multi-level redundancy design to ensure communication reliability: Main Channel: 5G private network communication, providing high-bandwidth video transmission; Backup Channel: 4G public network communication, ensuring uninterrupted basic communication; Emergency Channel: Self-organizing Mesh network to address communication needs in extreme situations. Collaborative Handling Status Data includes the following information: Device Location Data: Real-time GPS coordinates of each device; Video Image Data: Compressed video streams and keyframe images; Event Status Data: Handling stage and result evaluation of the abnormal event; Environmental Data: Environmental parameters such as on-site temperature, humidity, and lighting. Step S3: Based on the collaborative handling status data, real-time progress information is pushed to relevant responsible persons via mobile terminals, achieving closed-loop management of the entire process from anomaly discovery to completion. A dedicated mobile application is developed and installed on the smartphones or tablets of security personnel and managers to receive and display real-time progress information.The mobile application has the following functions: Real-time notification: Push notifications to remind users of new abnormal events; Map display: Displays the location of the event and the location of the handling equipment on an electronic map; Video viewing: View real-time video transmitted by drones and ground robots; Status tracking: Displays the current stage of event handling and the estimated completion time; Instruction feedback: Allows users to send simple instructions such as "urgent processing" or "request assistance"; The system provides differentiated information content for users with different roles: Frontline security personnel: Receive specific instructions and on-site situation; Security supervisor: View the overall situation and resource allocation; Department head: Obtain summary reports and reminders of major events; Full-process closed-loop management realizes complete tracking of event handling, including the following stages: 1) Anomaly detection: Drone patrols detect anomalies and generate event reports; 2) Task allocation: The system automatically allocates ground robots for response; 3) On-site handling: Equipment coordinates to handle the abnormal event; 4) Result confirmation: The system confirms whether the abnormal event has been resolved; 5) Summary report: Generates an event summary including a timeline, measures taken, and effect evaluation. The status changes of each stage are pushed to the mobile terminals of relevant responsible persons in real time, ensuring that all participants can keep abreast of the latest situation. The system will also automatically generate a complete log of the handling process, including timestamps, operation records and video clips, for subsequent event analysis and security strategy optimization.

[0082] Example 5 details how the system establishes a three-tiered early warning system and achieves closed-loop management throughout the entire process based on a ground response plan. First, the system establishes a three-tiered early warning system consisting of blue, yellow, and red alerts, with different levels corresponding to different response measures and notification scopes. A blue alert represents a general abnormal situation, typically triggered when the confidence level of the abnormal event is between 0.3 and 0.5 and the urgency level is low. When the system triggers a blue alert, the response measures are relatively mild, usually only dispatching a single ground robot to check the situation, while drones maintain their regular patrol routes, only paying slight attention to the abnormal area. Regarding the notification scope, a blue alert only notifies the security personnel directly responsible for the area, avoiding alerting higher-level management personnel and preventing resource waste and unnecessary panic.

[0083] A yellow alert indicates a significant anomaly, typically triggered when the confidence level of the anomaly is between 0.5 and 0.8 and the urgency is moderate. At this time, the system's response measures are significantly strengthened. Generally, 1-2 ground robots are dispatched to handle the situation, and drones will adjust their original patrol routes, increasing the frequency and duration of observation of the abnormal area to provide real-time aerial support for the ground robots. In terms of notification scope, a yellow alert not only notifies the directly responsible security personnel but also the security supervisor, informing them of the situation and allowing them to prepare possible manpower support.

[0084] A red alert signifies a major anomaly and is the highest level of warning, typically triggered when the confidence level of the anomaly exceeds 0.8 and the urgency is high. At this point, the system initiates a comprehensive response, dispatching multiple ground robots to the scene, while drones completely change their original patrol mission, becoming fixed-point hovering surveillance units above the anomaly location, providing continuous aerial monitoring. In terms of notification scope, a red alert simultaneously notifies regional security personnel, security supervisors, and relevant department heads, ensuring that all levels of personnel are promptly informed and can make appropriate decisions.

[0085] The system determines the warning level through a scientific calculation method, rather than simply relying on a single indicator. Specifically, the system uses a weighted average to calculate the warning index: Warning Index = 0.7 × Confidence Level + 0.3 × Urgency Level. The confidence level is derived from the abnormal event detection results in the previous embodiment, while the urgency level is comprehensively assessed based on factors such as event type, location, and time. A blue warning is triggered when the warning index is less than 0.5, a yellow warning when it is between 0.5 and 0.8, and a red warning when it is greater than or equal to 0.8. This calculation method considers both the reliability of the system's detection results and the urgency of the event itself, thus more accurately reflecting the true severity of the event.

[0086] After determining the warning level and triggering the corresponding warning, the system coordinates the collaborative work of various devices through a unified scheduling platform. The unified scheduling platform is the central nervous system of the entire system, composed of multiple functional modules, including an instruction distribution module (generating and issuing task instructions to each device based on the warning level), a real-time positioning module (tracking the location of all devices and visualizing it on an electronic map), a video aggregation module (integrating video streams from drones and ground robots and performing intelligent analysis), a communication management module (ensuring real-time communication and data transmission between devices), and a status monitoring module (monitoring the working status and task execution of each device). These modules work together to ensure that the entire system can respond to abnormal events in a coordinated manner.

[0087] In this three-way collaborative working model, drones, ground robots, and the command center each perform their respective functions, forming an organic whole. Upon receiving a report of an anomaly, the drone immediately adjusts its flight path to the incident location, using advanced computer vision technology to lock onto and track the abnormal target. Depending on the warning level, the drone adjusts its hovering altitude and monitoring range; for example, during a red alert, it may lower its flight altitude to obtain a clearer view. The drone transmits high-definition video streams to the command center in real time throughout the entire process, providing intuitive visual evidence for decision-making.

[0088] The ground robot then selects the optimal route to the event location according to the ground response plan calculated in the previous embodiment. During its journey, the robot continuously receives route adjustment instructions from the command center to respond to changes in road conditions or updates to the event's location. Upon arrival at the scene, the ground robot performs appropriate actions according to preset procedures, such as issuing warnings and dispersing crowds, providing voice prompts, and monitoring the environment, while simultaneously uploading the scene information and awaiting further instructions. In complex situations, multiple ground robots may work collaboratively to complete the task.

[0089] As the brain of the entire system, the command center simultaneously displays video streams transmitted from drones and ground robots on a large screen, providing real-time updates on the situation. Command center operators analyze this real-time information to predict developments and adjust response strategies, contacting human security personnel for support when necessary to ensure the proper handling of any incidents. The command center is also responsible for recording the entire incident handling process, providing detailed data support for subsequent incident analysis and system optimization.

[0090] To ensure the reliability of the three-way collaborative operation, the system is designed with multi-level redundant real-time communication links. The main channel uses 5G private network communication, providing high-bandwidth, low-latency video transmission capabilities; the backup channel uses 4G public network communication, automatically switching in case of problems with the main channel to ensure uninterrupted basic communication; and the emergency channel uses self-organizing mesh network technology to maintain basic communication needs even in extreme situations (such as large-scale power outages or natural disasters). This multi-level redundancy design greatly improves the system's anti-interference capability and reliability, ensuring smooth information exchange in various complex environments.

[0091] The collaborative work among the three parties generates a large amount of collaborative response status data, including real-time location data of each device (GPS coordinates, altitude, orientation, etc.), video image data (compressed video streams and keyframe images), event status data (response stage, progress, result evaluation, etc.), and environmental data (parameters such as temperature, humidity, and light intensity). This data is used for real-time command and decision-making, and is also saved for subsequent analysis and system improvement.

[0092] To ensure that relevant personnel are kept informed of the progress of the incident, a dedicated mobile application has been developed and installed on the smartphones or tablets of security personnel and managers. This application offers a wealth of features, including real-time notifications (pushing messages to alert users to new incidents), map display (visually showing the location of the incident and the location of the response equipment on an electronic map), video viewing (real-time viewing of on-site video footage), status tracking (displaying the current stage of the incident response and the estimated completion time), and instruction feedback (allowing users to send simple instructions such as "urgent processing" or "request assistance").

[0093] The system provides differentiated information content for users with different roles. Frontline security personnel primarily receive specific instructions and detailed on-site information for immediate action; security supervisors can view the overall situation and resource allocation for better coordination; department heads mainly receive overview reports and alerts to major events, helping them grasp the big picture and intervene in decision-making when necessary. This role-based information distribution ensures that everyone receives information appropriate to their responsibilities, avoiding information overload while guaranteeing the timely delivery of critical information.

[0094] Through the aforementioned series of measures, the system achieves closed-loop management of the entire process from anomaly detection to resolution, specifically including five stages: anomaly detection (drone patrols detect anomalies and generate event reports), task allocation (the system automatically assigns ground robots to respond), on-site handling (equipment collaboratively handles the anomaly), result confirmation (the system confirms whether the anomaly has been resolved), and summary report (generating an event summary including a timeline, measures taken, and effectiveness evaluation). Status changes at each stage are pushed to the mobile terminals of relevant personnel in real time, ensuring all participants are promptly informed of the latest situation. The system also automatically generates a complete log of the handling process, including timestamps, operation records, and video clips, for subsequent event analysis and security strategy optimization.

[0095] Example 6 is based on an aerial patrol scheduling scheme. Real-time meteorological information, including wind speed, rainfall and visibility, is obtained. The impact of weather on UAV flight safety is evaluated by a Bayesian network model. The patrol path is recalculated by a dynamic weight adjustment algorithm. When adverse weather conditions are detected, an alternative path or temporary landing procedure is automatically triggered to obtain a dynamically adjusted aerial patrol scheduling scheme.

[0096] Step S1: First, the system acquires real-time meteorological information through an interface, including parameters such as wind speed, rainfall, visibility, temperature, and lightning activity. Meteorological data sources include: regional weather forecast data provided by the meteorological bureau's public API; localized precise measurements provided by micro-meteorological stations distributed throughout the city; and flight altitude meteorological data collected by the drone's onboard meteorological sensors. Step S2: The system uses a Bayesian network model to assess the impact of weather on drone flight safety. A Bayesian network is a probabilistic graphical model that can represent the conditional dependencies between random variables. In this embodiment, the established Bayesian network includes the following nodes: Input nodes: wind speed, rainfall, visibility, temperature, lightning activity; Intermediate nodes: mechanical stability, sensor performance, communication quality, battery performance; Output node: flight safety level (safe / warning / dangerous). The conditional probability table of the Bayesian network is trained using expert knowledge and historical data to achieve accurate inference of the flight safety level. For example, when the wind speed exceeds level 8 (20.8-24.4 m / s), the mechanical stability is extremely low, and the flight safety level is "dangerous"; when the visibility is less than 100 meters but other conditions are normal, the flight safety level is "warning".

[0097] Step S3: Based on the flight safety level assessment results, the system recalculates the patrol path using a dynamic weight adjustment algorithm. The adjustment strategies are as follows: Safety Level (Green): Maintain the original patrol path. Warning Level (Yellow): Increase the weight coefficient of high-risk areas (×1.5), reduce flight altitude in open areas (reduced by 20%), reduce dwell time (reduced by 30%), increase navigation margin between waypoints. Hazard Level (Red): Automatically trigger alternative path planning, prioritize safe landing points, reduce coverage area, focus on core areas, and cancel the mission and return to base if necessary. Step S4: When adverse weather conditions are detected, the system automatically triggers an alternative path or temporary landing procedure. The criteria for determining adverse weather conditions include: Gale warning: wind speed exceeding 12 m / s; Heavy rainfall: rainfall exceeding 20 mm / h; Low visibility: visibility less than 50 meters; Thunderstorm activity: lightning activity detected within a 3 km radius. Response measures include: 1) Alternate path activation: Switching to a preset alternate patrol path for severe weather, reducing coverage area, concentrating on the core area, lowering flight altitude to reduce wind impact, increasing navigation point density, and improving positioning accuracy. 2) Temporary landing procedure: When weather conditions deteriorate to the point of endangering flight safety, selecting the nearest suitable location from the preset safe landing point database, executing a sequence of actions to control descent speed, and sending landing position coordinates to the command center to await weather improvement or manual recovery. Step S5: The system integrates the above processing results and outputs a dynamically adjusted aerial patrol scheduling plan, including: the adjusted patrol path coordinate sequence, new arrival time and stay duration, weather response strategy, marking of backup landing point coordinates, and explanation of the reasons for the adjustment. For example, when thunderstorm activity is detected in the northeastern area of ​​the city, the system will automatically adjust the path of the drones originally scheduled to patrol in that area to avoid the thunderstorm area, increase the patrol density in other areas, and calculate temporary landing points in advance to ensure safe landing if the weather deteriorates further.

[0098] Example 6 details how the system dynamically adjusts its drone patrol plan based on weather conditions. First, the system acquires real-time weather information through multiple channels, including key parameters such as wind speed, rainfall, visibility, temperature, and lightning activity. This meteorological data comes from diverse sources, including regional weather forecasts provided by meteorological bureaus via public APIs, precise local measurements from micro-weather stations distributed throughout the city, and real-time flight altitude meteorological data collected by the drone's onboard sensors. This multi-channel data acquisition strategy ensures the comprehensiveness and accuracy of the meteorological information, providing a reliable basis for subsequent flight safety assessments.

[0099] After acquiring meteorological information, the system uses a Bayesian network model to assess the impact of weather on UAV flight safety. A Bayesian network is a probabilistic graphical model that can represent the conditional dependencies between random variables, making it particularly suitable for handling complex decision-making problems with uncertainty. In this system, the established Bayesian network consists of three layers of nodes: input nodes (including raw meteorological parameters such as wind speed, rainfall, visibility, temperature, and lightning activity), intermediate nodes (including system functions affected by weather, such as mechanical stability, sensor performance, communication quality, and battery performance), and output nodes (flight safety levels, categorized into safe, warning, and dangerous levels).

[0100] This Bayesian network, trained with extensive historical data and expert knowledge, forms a comprehensive conditional probability table capable of accurately inferring flight safety levels under different weather conditions. For example, when wind speeds exceed level 8 (approximately 20.8-24.4 m / s), the system determines that mechanical stability is extremely low, classifying the flight safety level as "dangerous." Conversely, when visibility is below 100 meters but other conditions are normal, the system classifies the flight safety level as "warning." This probability-based inference method is more flexible and accurate than simple threshold-based judgments, better able to handle complex and ever-changing weather environments.

[0101] Based on the flight safety level assessment results, the system recalculates the patrol path using a dynamic weight adjustment algorithm. Different adjustment strategies are adopted for different safety levels. When the safety level is "Safe" (green), the system maintains the original patrol path and executes the predetermined patrol plan normally. When the safety level is "Warning" (yellow), the system makes a series of adjustments: first, it increases the weight coefficient of high-risk areas (usually multiplied by 1.5) to ensure these areas still receive sufficient attention; second, it reduces the flight altitude (usually reduced by 20%) to reduce the impact of wind on flight; then, it reduces the dwell time (usually reduced by 30%) to increase patrol speed and reduce the time the drone is exposed to adverse weather conditions; finally, it increases the navigation margin between waypoints to reserve more buffer space to cope with emergencies.

[0102] When the safety level reaches "dangerous" (red), the system will take more aggressive measures: First, it will automatically trigger pre-designed backup path plans, which are usually more conservative and avoid open areas and high-altitude areas; prioritize safe landing points to ensure a rapid and safe landing if necessary; reduce the coverage area and focus on monitoring core and important areas; in extreme cases, it may even cancel the current mission and instruct the drone to return to base to ensure equipment safety.

[0103] The system also sets a series of criteria for determining adverse weather conditions, and will automatically trigger alternative paths or temporary landing procedures when these conditions are detected. These criteria include: gale warnings with wind speeds exceeding 12 m / s, heavy rainfall exceeding 20 mm / h, low visibility conditions with visibility less than 50 meters, and detection of lightning activity within a 3 km radius. Once these conditions are triggered, the system will immediately initiate response measures.

[0104] When an alternate route is activated, the system switches to a pre-defined patrol route specifically designed for severe weather. These routes are characterized by a reduced coverage area, focusing on core areas to minimize unnecessary long-distance flights; lower flight altitudes, typically flying close to buildings to reduce wind impact; and increased navigation point density to improve positioning accuracy and ensure accurate navigation even in poor visibility conditions. For temporary landing procedures, when weather conditions deteriorate to the point of directly jeopardizing flight safety, the system selects the nearest and suitable location from a pre-defined safe landing point database and executes a carefully designed sequence of actions to control descent speed, ensuring a smooth landing for the UAV. Simultaneously, the system sends the landing location coordinates to the command center for subsequent recovery and continuously monitors weather conditions, automatically resuming the mission or awaiting manual recovery once conditions improve.

[0105] Finally, the system integrates all processing results and outputs a dynamically adjusted aerial patrol scheduling plan, including the adjusted patrol path coordinate sequence, updated arrival and stay duration estimates, clear weather response strategy markers, alternative landing point coordinates, and detailed explanations of the adjustment reasons. For example, when the system detects thunderstorm activity in the northeastern part of the city, it will automatically adjust the paths of drones originally scheduled to patrol that area to avoid the thunderstorm zone, while appropriately increasing the patrol density in other areas to maintain overall coverage, and pre-calculating temporary landing points to ensure safe landing if the weather worsens.

[0106] This dynamic adjustment mechanism based on real-time meteorological data significantly improves the safety and reliability of the UAV patrol system, enabling it to maintain basic functionality under various adverse weather conditions while maximizing equipment safety and preventing unnecessary losses. This function is particularly important in regions with variable monsoons and frequent extreme weather events, serving as a key guarantee for the system's long-term stable operation.

[0107] Example 7, based on the ground response scheme, establishes a robot battery status monitoring system to track the power consumption and remaining working time of ground robots in real time. The system dynamically adjusts task allocation through a load balancing algorithm. When a robot's power is insufficient, its tasks are automatically transferred to nearby devices with sufficient power, and the optimal return path to the charging station is planned, resulting in an energy-optimized ground response scheme.

[0108] Step S1: First, the system establishes a robot battery status monitoring system to track the power consumption and remaining working time of robots in various locations in real time. The monitoring system includes the following components: Battery Management System (BMS): Monitors the battery status of each robot in real time, including: current battery percentage, state of health (SOH), discharge rate and current consumption, battery temperature, and energy consumption prediction model: Predicts the remaining working time based on historical data and current tasks, taking into account the impact of road conditions (slope, road conditions), the impact of load weight, the impact of ambient temperature, and the impact of battery aging. Status Reporting System: Reports battery status data to the central system periodically (every 30 seconds). Step S2: The system dynamically adjusts task allocation through a load balancing algorithm. The load balancing algorithm is based on a multi-factor scoring model, which comprehensively considers the following factors: Battery status: current battery percentage and estimated remaining working time; Task load: number of assigned tasks and estimated completion time; Geographic location: distance from task location and charging station; Equipment capability: performance differences between different robot models. The algorithm implementation process is as follows: 1) Calculate the comprehensive load index of each robot: Load index = w1 × (1 - battery percentage) + w2 × number of tasks + w3 × distance coefficient, where w1, w2, and w3 are weight parameters that are adjusted according to the actual situation. 2) Sort the load indices to identify robots with excessively high or low loads. 3) Balance the load of each robot through task transfer to optimize overall resource utilization. Step S3: When a robot is detected to have insufficient battery power, the system automatically transfers its tasks to nearby devices with sufficient battery power. The system manages battery levels using the following thresholds: Warning threshold: 30% battery, triggering task transfer assessment; Emergency threshold: 15% battery, forcing task transfer; Return threshold: 10% battery, canceling all tasks and returning to the charging station. The task transfer process is as follows: 1) Identify robots with insufficient battery and their incomplete task list. 2) Search for available robots with sufficient battery (>50%) in the nearby area. 3) Calculate the task transfer cost (additional travel distance and time). 4) Select the option with the lowest transfer cost and perform task reassignment. 5) Send updated task instructions to the relevant robots. 6) Update the task allocation status within the system. Step S4: The system plans the optimal charging station return path for robots with insufficient battery. The selection of the optimal charging station considers the following factors: Distance factor: straight-line distance and actual path length to the charging station; Battery factor: whether the current battery level can support reaching the charging station; Queuing situation: current usage status of the charging station and waiting queue; Charging speed: differences in charging power between different charging stations. Path planning uses the A* algorithm combined with an energy consumption model to generate the energy-optimal return path. Unlike the traditional shortest path, the energy-optimal path tends to choose flat, unobstructed roads, even if these paths are slightly longer.

[0109] Step S5: The system integrates the above processing results and outputs an energy-optimized ground response plan, including: an adjusted robot task allocation table, task priorities and execution order, charging plans and schedules, path planning for each robot, and estimated task completion time. For example, when the system detects that robot 3's battery level has dropped to 25% and it is expected that it cannot complete the remaining two tasks, it will automatically evaluate other nearby robots. If it finds that robot 5 has sufficient battery (75%) and only has one low-priority task, the system will transfer robot 3's task to robot 5, and at the same time plan a path for robot 3 to the nearest charging station, ensuring that the overall response efficiency is not affected by battery limitations.

[0110] This invention is not limited to the preferred embodiments described above. Any changes or modifications made by any person based on the teachings of this invention fall within the scope of protection claimed by this invention. The scope of protection of this invention is defined in the appended claims and their equivalents.

[0111] Example 7 details how the system optimizes the scheduling scheme of ground robots through energy management. First, the system establishes a complete robot battery status monitoring system to track the power consumption and remaining working time of each ground robot in real time. This monitoring system consists of several core components, with the Battery Management System (BMS) being the most fundamental. It is responsible for real-time monitoring of key battery parameters for each robot, including the current battery percentage (reflecting remaining power), state of health (SOH) (reflecting battery aging), discharge rate and current consumption (reflecting the current energy consumption rate), and battery temperature (affecting battery efficiency and safety). These parameters are collected in real time by sensors built into the robots, providing first-hand data for energy management.

[0112] In addition to basic battery parameter monitoring, the system includes a sophisticated energy consumption prediction model to forecast the robot's remaining working time based on historical data and current task characteristics. This model comprehensively considers multiple influencing factors: road conditions (including slope and road type) directly affect the robot's energy consumption; for example, energy consumption increases significantly when climbing slopes. Load weight is also a key factor; the heavier the load, the higher the energy consumption. Ambient temperature has a significant impact on battery performance, especially in extremely low temperatures where battery efficiency drops drastically. Battery aging also affects its capacity and discharge characteristics. By integrating these factors, the model can provide a more accurate prediction of remaining working time than simply calculating the battery percentage.

[0113] The third component of the monitoring system is the status reporting system, which is responsible for periodically (usually every 30 seconds) reporting battery status data to the central system. This periodic reporting mechanism ensures that the central system can keep abreast of the energy status of all ground robots, providing a basis for subsequent mission adjustments. The reporting frequency is a carefully balanced result, taking into account both real-time requirements and avoiding the waste of resources caused by overly frequent communication.

[0114] Based on data collected by the battery status monitoring system, the system dynamically adjusts task allocation through a load balancing algorithm. This algorithm is based on a multi-factor scoring model, comprehensively considering multiple aspects such as battery status (current battery percentage and estimated remaining working time), task load (number of assigned tasks and estimated completion time), geographical location (distance from task locations and charging stations), and equipment capabilities (performance differences between different robot models). The algorithm first calculates the comprehensive load index for each robot, which is a weighted formula: Load Index = w1 × (1 - Battery Percentage) + w2 × Number of Tasks + w3 × Distance Coefficient, where w1, w2, and w3 are weight parameters that are dynamically adjusted according to actual operating conditions.

[0115] After calculating the load index, the system sorts all robots by index value, identifying robots with excessively high loads (top 20% of index values) and robots with excessively low loads (bottom 20% of index values). Then, the system attempts to transfer some tasks from overloaded robots to underloaded robots through task transfer, balancing the workload of each robot and optimizing overall resource utilization efficiency. This load balancing not only considers power consumption but also comprehensively takes into account task volume and geographical location, representing a comprehensive optimization strategy.

[0116] The system sets three key thresholds for battery management: a warning threshold is triggered when the robot's battery level drops to 30%, at which point the system begins to assess whether a task transfer is necessary; an emergency threshold is triggered when the battery level drops to 15%, forcing a task transfer to ensure the robot can complete its current task and return safely; and a return-to-home threshold is triggered when the battery level drops to 10%, canceling all unfinished tasks for the robot and instructing it to immediately return to the nearest charging station. This tiered threshold design ensures that the system can take appropriate measures based on the battery status, avoiding premature task interruption that wastes resources and delays that cause the robot to run out of power.

[0117] When a robot's battery is detected to be low, the system initiates a task transfer process. First, it identifies the robot with low battery and its list of unfinished tasks. Then, it searches the nearby area for available robots with sufficient battery power (typically requiring more than 50% remaining). The system calculates the transfer cost for each possible task transfer option, primarily considering additional travel distance and time. Among all feasible options, the system selects the option with the lowest total cost to perform the task reallocation and sends updated task instructions to the relevant robots. The entire process is fully automated, requiring no manual intervention, significantly improving the system's response speed and scheduling efficiency.

[0118] For robots that need to return to recharge, the system plans the optimal return path to the charging station. The selection of the optimal charging station is not simply based on distance, but rather considers multiple factors: in addition to the straight-line distance and actual path length, the system also assesses whether the current battery level can support reaching the charging station (avoiding running out of power en route), the current usage status and waiting queues of the charging station (avoiding long wait times), and the differences in charging power among different charging stations (prioritizing fast-charging stations). Path planning uses the A* algorithm combined with an energy consumption model to generate the most energy-efficient return path, not just the shortest distance. This energy-optimized path tends to choose flat, unobstructed roads, even if these paths are slightly longer, because steep slopes and congested sections significantly increase energy consumption.

[0119] After the above processing, the system outputs an energy-optimized ground response plan, including an adjusted robot task allocation table, updated task priorities and execution order, detailed charging plans and schedules, optimized path planning for each robot, and adjusted estimated task completion times. This plan considers both task requirements and energy constraints, making it a comprehensive and optimal execution solution.

[0120] For a concrete example, when the system detects that robot #3's battery level has dropped to 25%, and the energy consumption prediction model determines that its remaining power is insufficient to complete the two assigned tasks, the system will automatically scan for other nearby robots. If it finds that robot #5 has sufficient power (75%) and currently only has one low-priority task, the system will transfer robot #3's tasks to robot #5, while simultaneously planning a path for robot #3 to the nearest charging station. During path planning, the system will avoid steep slopes and congested areas to ensure robot #3 can safely reach the charging station. After charging is complete, robot #3 will rejoin the task queue and accept new task assignments.

[0121] Through this dynamic energy management and task adjustment mechanism, the system can maximize the working time of ground robots, avoid task interruptions due to power issues, and ensure the continuous and effective operation of the entire security monitoring network. Simultaneously, this mechanism also helps extend battery life, reduce unnecessary charge-discharge cycles, lower equipment maintenance costs, and improve the overall system's economy and sustainability. This function is particularly valuable in large-scale deployments and 24 / 7 uninterrupted operation scenarios, serving as a crucial guarantee for the system's long-term stable operation.

[0122] Finally, it should be noted that the above-described embodiments are merely specific implementations of this disclosure, used to illustrate the technical solutions of this disclosure, and not to limit it. The protection scope of this disclosure is not limited thereto. Although this disclosure has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments, or make equivalent substitutions for some of the technical features, within the scope of the technology disclosed in this disclosure; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this disclosure, and should all be covered within the protection scope of this disclosure. Therefore, the protection scope of this disclosure should be determined by the protection scope of the claims.

Claims

1. A method for the coordinated management of drone patrols and security monitoring, characterized in that... This includes: acquiring urban geographic information data and coordinates of key monitoring areas; using a quadtree spatial indexing algorithm to hierarchically divide the key monitoring areas into grids according to security level and terrain features; establishing an adjacency matrix between grid cells; and generating a basic data structure for patrol task allocation containing spatial relationships and attribute information of all grid cells. Using this basic data structure, the grid coverage problem is transformed into a minimum dominance set problem in graph theory. An improved greedy strategy is adopted to select the location point that covers the most unvisited grids in each iteration and introduces random perturbation to avoid local optima, achieving an O(n log n) time complexity. Within an n-time complexity, a patrol path scheme with an approximation ratio of ln(Δ)+1 is solved and optimized, outputting an aerial patrol scheduling scheme containing patrol order, estimated arrival time, and dwell time, where n is the number of grids and Δ is the maximum grid degree. Based on the aerial patrol scheduling scheme, real-time video streams are collected according to the predetermined path. A pre-trained deep learning model is used to detect and identify personnel gatherings, vehicle anomalies, and abandoned items. A temporal analysis algorithm is used to perform behavioral analysis on consecutive frames to detect abnormal behavior patterns and calculate the confidence score of abnormal events, generating an abnormal event report containing event type, location, confidence score, and urgency level. Based on the abnormal event report, a multi-constraint optimization model is established to abstract the ground robot scheduling problem into a vehicle path problem with a time window. A (2+ε) approximation algorithm based on Lagrange relaxation is used with dynamic programming techniques to achieve a quasi-polynomial time complexity of O(n^2). 1 / ε The system solves for an approximate optimal solution and dynamically adjusts the weight coefficients according to the urgency level of the abnormal event. It outputs a ground response plan that includes a robot allocation scheme, a travel route, and an estimated arrival time, where ε is a preset accuracy parameter. Based on the ground response plan, a three-level early warning system is established to automatically trigger the corresponding level of linkage response according to the confidence score and urgency level of the abnormal event report.

2. The method according to claim 1, characterized in that... The process involves acquiring urban geographic information data and coordinates of key monitoring areas. Using a quadtree spatial indexing algorithm, the key monitoring areas are hierarchically divided into grids based on security levels and terrain features. An adjacency matrix between grid units is established, generating a basic data structure for patrol task allocation containing spatial relationships and attribute information of all grid units. This includes: acquiring urban geographic information data and coordinates of key monitoring areas; determining the initial grid division level based on security levels and terrain features; generating an initial grid set; recursively subdividing each grid using a quadtree spatial indexing algorithm, assigning location coordinates, security level weights, and terrain complexity coefficients to each grid unit to obtain a hierarchical grid division result; and calculating the distance and access difficulty between each grid unit based on the hierarchical grid division result, establishing an adjacency matrix between grid units, and obtaining the basic data structure for patrol task allocation.

3. The method according to claim 2, characterized in that... The attribute information of the grid cell includes location coordinates, security level weight, and terrain complexity coefficient; the adjacency matrix records the distance and passage difficulty between each grid cell.

4. The method according to claim 1, characterized in that... Using the aforementioned patrol task allocation basic data structure, the grid coverage problem is transformed into a minimum dominance set problem in graph theory. An improved greedy strategy is adopted to select the location point that can cover the most unvisited grids in each iteration and introduce random perturbation to avoid local optima. A patrol path scheme with an approximate ratio of ln(Δ) + 1 is solved in O(n log n) time complexity and optimized. The output includes an aerial patrol scheduling scheme containing patrol order, estimated arrival time, and dwell time. The process includes: using the aforementioned patrol task allocation basic data structure to model the grid coverage problem as a minimum dominance set problem in graph theory, determining the objective function and constraints, and obtaining an optimization model; based on the optimization model, using an improved greedy strategy to calculate the number of unvisited grids that each candidate location point can cover in each iteration and selecting the location point with the most coverage, introducing a random perturbation mechanism to avoid getting trapped in local optima, and solving the problem in O(n log n) time complexity to obtain an approximate ratio of ln(Δ) + 1 patrol path scheme. Solving within an n) time complexity yields an initial patrol path scheme with an approximation ratio of ln(Δ)+1; based on the initial patrol path scheme, UAV endurance parameters and real-time weather condition data are obtained, and the initial patrol path scheme is optimized through constraint condition verification and path adjustment algorithms to obtain the aerial patrol scheduling scheme.

5. The method according to claim 1, characterized in that... , Based on the above air patrol scheduling plan, collect real-time video streams according to a predetermined path, perform object detection through a pre-trained deep learning model to identify personnel gatherings, vehicle anomalies, and item leftovers, use a time series analysis algorithm to analyze the behavior of consecutive frame images to detect abnormal behavior patterns and calculate the confidence score of abnormal events, and generate an abnormal event report including event type, occurrence location, confidence score, and emergency level, including: Based on the above air patrol scheduling plan, control the high-definition camera carried by the drone to take real-time shots according to a predetermined path, and collect real-time video stream data; Based on the real-time video stream data, perform object detection on video frames through a pre-trained deep learning model to identify personnel gatherings, vehicle anomalies, and item leftovers, and obtain the object detection results; Based on the object detection results, use a time series analysis algorithm to analyze the behavior patterns of consecutive frame images to detect violent conflicts and illegal intrusions, and calculate the confidence score of abnormal events in combination with environmental context information such as time, location, and weather, and obtain the above abnormal event report.

6. The method according to claim 5, characterized in that... , The pre-trained deep learning model is an improved version of YOLO v5; The abnormal behavior patterns include violent conflicts and illegal intrusions; The environmental context information includes time, location, and weather.

7. The method according to claim 1, characterized in that... Based on the aforementioned abnormal event reports, a multi-constraint optimization model is established to abstract the ground robot scheduling problem into a vehicle pathing problem with a time window. A (2+ε) approximation algorithm based on Lagrange relaxation is employed, utilizing dynamic programming techniques to achieve a quasi-polynomial time complexity of O(n^2). 1 / ε Within a given timeframe, an approximate optimal solution is found, and weighting coefficients are dynamically adjusted based on the urgency level of the abnormal event. The resulting ground response plan includes a robot allocation scheme, travel route, and estimated arrival time. This includes: based on the abnormal event report, obtaining robot quantity limits, battery life, maximum payload, and road conditions; establishing a multi-constraint optimization model for the vehicle routing problem with a time window to obtain a set of constraints; and based on this set of constraints, employing a (2+ε) approximation algorithm based on Lagrange relaxation, and utilizing dynamic programming techniques to achieve a near-polynomial time complexity of O(n^2). 1 / ε Solve within the range to obtain the initial scheduling scheme; based on the initial scheduling scheme, dynamically adjust the weight coefficients of each abnormal event according to the urgency level in the abnormal event report, giving priority to ensuring a rapid response to high-priority events, and obtain the ground response scheme.

8. The method according to claim 1, characterized in that... , Based on the above ground response plan, establish a three-level early warning system to automatically trigger corresponding-level linkage responses according to the confidence score and emergency level of the abnormal event report, achieve three-party collaboration of continuous drone tracking, rapid arrival of ground robots at the scene, and real-time monitoring by the command center through a unified dispatching platform, and establish a real-time communication link, and push the disposal progress to relevant responsible persons through a mobile terminal to form a full-process closed-loop management from abnormal discovery to disposal completion, including: Based on the above ground response plan, establish a three-level early warning system of blue warning, yellow warning, and red warning, determine the warning level according to the confidence score and emergency level of the abnormal event report, and automatically trigger the corresponding-level linkage response mechanism; Based on the linkage response mechanism, send instructions through a unified dispatching platform to achieve three-party collaboration of continuous drone tracking of abnormal targets, rapid arrival of ground robots at the scene according to the above ground response plan, and real-time monitoring of the disposal process by the command center, establish a real-time communication link between air equipment, ground equipment, and the command center, and obtain collaborative disposal status data; Based on the collaborative disposal status data, push real-time disposal progress information to relevant responsible persons through a mobile terminal to obtain full-process closed-loop management from abnormal discovery to disposal completion.

9. The method according to claim 1, characterized in that... , It also includes: Based on the above air patrol scheduling plan, obtain real-time meteorological information including wind speed, rainfall, and visibility, evaluate the impact of weather on the flight safety of the drone through a Bayesian network model, use a dynamic weight adjustment algorithm to recalculate the patrol path, and automatically trigger an alternative path or a temporary landing procedure when adverse weather conditions are detected to obtain a dynamically adjusted air patrol scheduling plan.

10. The method according to claim 1, characterized in that... It also includes: based on the ground response scheme, establishing a robot battery status monitoring system to track the power consumption and remaining working time of ground robots in real time, dynamically adjusting task allocation through a load balancing algorithm, automatically transferring the task to a nearby device with sufficient power when a robot's power is insufficient, and planning the optimal return path to the charging station, thus obtaining an energy-optimized ground response scheme.

Images