An Adaptive Video Inspection and Scheduling Method Based on Spatial Risk Awareness and Related Components
By constructing a spatial topology model and setting differentiated frequencies, the problems of insufficient blind spot risk assessment and suboptimal resource allocation in video inspection systems were solved, realizing full-domain risk assessment and dynamic scheduling, and improving the accuracy and rationality of inspections.
Patent Information
- Application Number
- CN202611311726.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-27
- Publication Date
- 2026-11-17
AI Technical Summary
Existing video inspection systems cannot achieve comprehensive risk assessment, have blind spots in monitoring and insufficient risk assessment in areas obstructed by equipment, lack adaptability in scheduling strategies, have suboptimal resource allocation, and lack interpretability in output results.
By constructing a spatial topology model with associated weights, dividing logical regions and performing video anomaly detection, generating risk scores, setting differentiated inspection frequencies, and updating the scheduling plan based on feedback data, dynamic adaptive adjustment is achieved.
It effectively fills the gaps in risk assessment for monitoring blind spots, optimizes resource allocation, improves the accuracy and rationality of inspections, enhances scheduling effectiveness, reduces computing power waste, and provides interpretable risk assessments.
Smart Images

Figure CN123123545A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of artificial intelligence technology, and in particular to an adaptive video inspection and scheduling method and related components based on spatial risk perception. Background Technology
[0002] Current video inspection systems generally adopt a fixed-cycle task configuration architecture that combines cameras with algorithms. Maintenance personnel pre-bind corresponding inspection algorithms to each camera location, limit the inspection time period, and set a fixed inspection frequency. Based on preset parameters, the system periodically performs intelligent analysis on the monitoring video stream, video frames, or captured screenshots. Once an abnormal situation is identified on site, it triggers and outputs corresponding alarm information. It is the mainstream technical solution for safety inspection in various scenarios such as parks, workshops, and warehouses.
[0003] However, in practical engineering applications, the deployment conditions of each monitoring point vary significantly. The installation location of cameras, field of view angles, on-site obstruction, network bandwidth, edge device computing power reserves, and algorithm deployment costs all differ, making it difficult to achieve a unified and standardized configuration. To ensure overall inspection coverage, traditional technical solutions typically employ a crude optimization approach of uniformly increasing the inspection frequency across all points, or rely entirely on the manual experience of maintenance personnel, only encrypting inspection parameters for a few key areas. This maintenance configuration model lacks risk adaptive adjustment capabilities, cannot dynamically optimize inspection strategies based on real-time on-site risk conditions, and struggles to conduct effective risk assessments for areas not covered by cameras, resulting in significant shortcomings in the accuracy and rationality of overall inspections.
[0004] Specifically, existing video inspection technologies suffer from numerous substantial deficiencies. Current inspection systems consistently build their logic around single camera locations or preset inspection targets, failing to establish corresponding risk-layering modeling systems for complete physical spaces such as parks, workshops, and warehouses. This prevents integrated, comprehensive risk assessment of the overall spatial operation, resulting in significant limitations in inspection analysis dimensions. For blind spots, areas obstructed by equipment, camera field-of-view boundaries, and temporarily unobservable areas caused by equipment malfunctions or temporary obstructions, existing technologies lack effective risk inference and compensation mechanisms. This leads to various blind spot risks being consistently underestimated and missed, creating inspection security vulnerabilities.
[0005] Meanwhile, traditional multi-camera collaborative scheduling and PTZ camera scheduling strategies primarily focus on basic functions such as full coverage of the monitoring field of view, reasonable allocation of communication resources, and accurate target acquisition. They fail to deeply integrate spatial topological correlation information such as on-site business flow processes, personnel and vehicle traffic paths, and historical anomaly propagation patterns. Consequently, the business adaptability and risk-specificity of the scheduling strategies are insufficient. The fixed-frequency mechanical inspection mode causes low-risk, stable areas to continuously occupy and consume edge and cloud computing resources, resulting in a waste of computing resources. Conversely, for areas where anomalies have occurred, surrounding areas spatially related to the anomaly areas, and various monitoring blind spots, it is impossible to promptly activate encrypted inspection mechanisms, which can easily lead to delayed handling of anomalies and the continued spread of risks.
[0006] Furthermore, existing technologies do not consider the cost differences of various execution resources. The resource costs, computing power consumption, and response efficiency of inspection execution methods such as edge lightweight models, cloud-based high-precision models, PTZ camera turning scheduling, and manual on-site verification differ significantly. However, traditional systems lack multi-level resource switching and intelligent scheduling mechanisms based on risk-reward comparisons, failing to achieve optimal resource allocation. The output of existing inspection systems is limited to a single format, only displaying abnormal alarm information and inspection task execution results. It cannot intuitively demonstrate risk propagation paths, risk contribution sources, and the core basis of intelligent scheduling strategies, resulting in poor interpretability of the technical solutions. Simultaneously, the spatial relationships built into the system are mostly factory-preset static configuration parameters. They cannot rely on real business data such as the co-occurrence patterns of abnormal events on-site, the accuracy of inspection scheduling, and feedback results from manual handling to continuously iterate and correct spatial topology weights. This makes it difficult to adapt to complex and ever-changing on-site inspection scenarios, resulting in a low overall level of intelligence and adaptability. Summary of the Invention
[0007] This invention provides an adaptive video inspection and scheduling method, system, computer equipment, and storage medium based on spatial risk perception, aiming to enhance the effectiveness of video inspection and scheduling.
[0008] In a first aspect, embodiments of the present invention provide an adaptive video inspection scheduling method based on spatial risk awareness, comprising: Obtain the physical space to be inspected, divide the physical space to be inspected into multiple logical regions, and construct a spatial topology model containing association weights for the multiple logical regions. The multiple logical regions are divided into a first region that can be covered by inspection and a second region that is a blind spot for inspection. Video anomaly detection is performed on the first region to obtain the corresponding anomaly observation results. A first risk score for the first region is generated based on the abnormal observation results and the spatial topology model, and a second risk score for the second region is extrapolated based on the first risk score and the spatial topology model. Based on the computing power budget constraint, the inspection frequency is set for each logical region by combining the first risk score and the second risk score. Based on the inspection frequency, an inspection task scheduling plan is generated, and the inspection task scheduling plan is sent to the corresponding inspection equipment. The system acquires feedback data from the inspection equipment and updates the spatial topology model, the first risk score, and the second risk score based on the feedback data. Then, based on a preset update mode, it generates a new inspection task scheduling plan using the updated spatial topology model, the first risk score, and the second risk score.
[0009] Secondly, embodiments of the present invention provide an adaptive video patrol scheduling system based on spatial risk awareness, comprising: A spatial partitioning unit is used to obtain the physical space to be inspected, divide the physical space to be inspected into multiple logical regions, and construct a spatial topology model containing association weights for the multiple logical regions. The region division unit is used to divide the multiple logical regions into a first region that can be covered by inspection and a second region that is a blind spot of inspection, and to perform video anomaly perception on the first region to obtain the corresponding anomaly observation results. The scoring generation unit is used to generate a first risk score for the first region based on the abnormal observation results and the spatial topology model, and extrapolate a second risk score for the second region based on the first risk score and the spatial topology model. A frequency setting unit is used to set the inspection frequency for each of the logical regions based on computing power budget constraints and in combination with the first risk score and the second risk score. The task generation unit is used to generate an inspection task scheduling plan based on the inspection frequency and to send the inspection task scheduling plan to the corresponding inspection equipment. The feedback update unit is used to acquire feedback data from the inspection equipment, update the spatial topology model and the first risk score and the second risk score based on the feedback data, and then generate a new inspection task scheduling plan based on the preset update mode using the updated spatial topology model, the first risk score and the second risk score.
[0010] Thirdly, embodiments of the present invention provide a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the adaptive video inspection and scheduling method based on spatial risk perception as described in the first aspect.
[0011] Fourthly, embodiments of the present invention provide a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the adaptive video inspection and scheduling method based on spatial risk perception as described in the first aspect.
[0012] This invention provides an adaptive video inspection scheduling method, system, computer device, and storage medium based on spatial risk perception. The method includes: acquiring a physical space to be inspected; dividing the physical space to be inspected into multiple logical regions; and constructing a spatial topology model containing correlation weights for the multiple logical regions; dividing the multiple logical regions into a first region that can be covered by inspection and a second region that is a blind spot for inspection; performing video anomaly perception on the first region to obtain corresponding anomaly observation results; generating a first risk score for the first region based on the anomaly observation results and the spatial topology model; and extrapolating a second risk score for the second region based on the first risk score and the spatial topology model; setting an inspection frequency for each logical region based on computing power budget constraints and combining the first risk score and the second risk score; generating an inspection task scheduling plan based on the inspection frequency and issuing the inspection task scheduling plan to the corresponding inspection equipment; acquiring feedback data from the inspection equipment; updating the spatial topology model, the first risk score, and the second risk score based on the feedback data; and then generating a new inspection task scheduling plan based on a preset update mode using the updated spatial topology model, the first risk score, and the second risk score. This invention achieves integrated risk assessment across the entire domain by constructing a spatial topology model with associated weights. By extrapolating blind spot risk scores based on anomaly perception results from observable areas and spatial topology relationships, it effectively fills risk assessment gaps in monitoring blind spots and prevents security vulnerabilities during inspections. Simultaneously, by combining risk scores with computational constraints to set differentiated inspection frequencies and continuously updating the spatial topology model and risk scores based on inspection feedback, it enables dynamic adaptive adjustment of the inspection scheduling plan. This ensures inspection intensity in high-risk areas and blind spot-related areas while preventing excessive consumption of computational resources in low-risk areas, thus optimizing the allocation of computational resources. This effectively improves the accuracy and rationality of video inspections, enhances the overall video inspection scheduling effect, and solves the problems of wasted computational resources caused by fixed-frequency inspections in traditional video inspections, the inability to assess camera blind spot risks, the inability to automatically encrypt inspections of abnormal areas, the lack of interpretability in scheduling strategies, and the high cost of manual configuration. Attached Figure Description
[0013] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0014] Figure 1 A flowchart illustrating an adaptive video inspection and scheduling method based on spatial risk perception, provided in an embodiment of the present invention; Figure 2 A schematic block diagram of an adaptive video inspection and scheduling system based on spatial risk perception provided in an embodiment of the present invention; Figure 3 A schematic diagram of the overall system architecture of an adaptive video inspection and scheduling method based on spatial risk awareness provided in an embodiment of the present invention; Figure 4 A schematic diagram of the perception-scoring-feedback-scheduling closed-loop process of an adaptive video inspection and scheduling method based on spatial risk perception provided in an embodiment of the present invention; Figure 5 A schematic diagram of the logical region, camera field of view mapping, and blind spot marking of an adaptive video inspection and scheduling method based on spatial risk perception provided in an embodiment of the present invention; Figure 6 A schematic diagram illustrating the spatial topology risk transmission and blind zone uncertainty compensation of an adaptive video inspection and scheduling method based on spatial risk perception provided in an embodiment of the present invention; Figure 7 A schematic diagram illustrating non-uniform inspection frequency optimization and dynamic penalty value scheduling of an adaptive video inspection scheduling method based on spatial risk perception provided in an embodiment of the present invention; Figure 8 This is a schematic diagram illustrating the multi-level execution resource switching of an adaptive video inspection scheduling method based on spatial risk awareness, provided in an embodiment of the present invention. Figure 9 A schematic diagram of topology weight self-correction for an adaptive video inspection scheduling method based on spatial risk perception provided in an embodiment of the present invention; Figure 10 This is a schematic diagram of the mode rescheduling process of an adaptive video inspection scheduling method based on spatial risk perception provided in an embodiment of the present invention. Detailed Implementation
[0015] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0016] It should be understood that, when used in this specification and the appended claims, the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.
[0017] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.
[0018] It should also be further understood that the term "and / or" as used in this specification and the appended claims refers to any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.
[0019] Please see below. Figure 1 This invention provides an adaptive video inspection and scheduling method based on spatial risk perception, specifically including steps S101 to S106.
[0020] Step S101: Obtain the physical space to be inspected, divide the physical space to be inspected into multiple logical regions, and construct a spatial topology model containing association weights for the multiple logical regions. Step S102: Divide the multiple logical regions into a first region that can be covered by inspection and a second region that is a blind spot for inspection, and perform video anomaly perception on the first region to obtain the corresponding anomaly observation results. Step S103: Generate a first risk score for the first region based on the abnormal observation results and the spatial topology model, and extrapolate a second risk score for the second region based on the first risk score and the spatial topology model. Step S104: Based on the computing power budget constraint, and in combination with the first risk score and the second risk score, set the inspection frequency for each logical region; Step S105: Generate an inspection task scheduling plan based on the inspection frequency, and send the inspection task scheduling plan to the corresponding inspection equipment. Step S106: Obtain feedback data from the inspection equipment, update the spatial topology model, the first risk score, and the second risk score based on the feedback data, and then generate a new inspection task scheduling plan based on the preset update mode using the updated spatial topology model, the first risk score, and the second risk score.
[0021] In this embodiment, the physical space to be inspected is first acquired and the spatial logical regions are divided, and a spatial topology model with associated weights is constructed. Then, based on the camera coverage capability, each logical region is divided into an observable first region and a second region with blind spots. Video anomaly perception is performed on the first region to obtain anomaly observation results. Subsequently, based on the anomaly observation results and the spatial topology model, the direct risk score calculation for the first region and the blind spot risk extrapolation score calculation for the second region are completed respectively. Then, under the condition of computing power constraints, the differentiated inspection frequency of each logical region is adaptively configured in combination with the global risk score. Then, a complete inspection task scheduling plan is generated based on the optimized inspection frequency and sent to the corresponding inspection equipment. Finally, the equipment inspection feedback data is collected, the spatial topology model weights are dynamically corrected and the global risk score is updated. A new inspection scheduling plan is generated through iterative optimization using a preset update mode, realizing the adaptive operation of the entire process of video inspection, including spatial perception, risk quantification, frequency optimization, intelligent scheduling and dynamic self-correction.
[0022] This embodiment achieves integrated risk assessment across the entire domain by constructing a spatial topology model with associated weights. By combining anomaly detection results from observable areas with spatial topology relationships to extrapolate blind spot risk scores, it effectively fills the risk assessment gaps in monitoring blind spots and prevents security vulnerabilities during inspections. Simultaneously, by combining risk scores with computing power constraints to set differentiated inspection frequencies, and by continuously updating the spatial topology model and risk scores based on inspection feedback, it enables dynamic adaptive adjustment of the inspection scheduling plan. This ensures inspection intensity in high-risk areas and blind spot-related areas while preventing excessive consumption of computing resources in low-risk areas, thus optimizing the allocation of computing resources. This effectively improves the accuracy and rationality of video inspections, enhances the overall video inspection scheduling effect, and solves the problems of wasted computing power due to fixed-frequency inspections in traditional video inspections, inability to assess camera blind spot risks, inability to automatically encrypt inspections of abnormal areas, lack of interpretability in scheduling strategies, and high manual configuration costs.
[0023] In particular, the adaptive video inspection scheduling method provided in this embodiment can be widely applied to various intelligent video inspection scenarios with multiple camera locations, interconnected spatial structures, related business processes, and limited computing resources, such as factory parks, warehouse sites, energy stations, commercial complexes, public buildings, transportation hubs, data centers, and industrial pipe corridors. It has strong scenario adaptability and engineering implementation value.
[0024] Combination Figure 3 Based on the adaptive video inspection scheduling method provided in this embodiment, a corresponding scheduling system can be built. The overall architecture of this scheduling system, from top to bottom, consists of a camera access layer, an edge perception layer, a spatial risk layer, a scheduling optimization layer, and an execution feedback layer. Specifically, the camera access layer handles the access of video streams and image data from fixed cameras and PTZ cameras; the edge perception layer performs anomaly detection based on the accessed video data and outputs model version and task execution cost information; the spatial risk layer manages logical regions, maintains topology weights, and performs blind zone risk extrapolation calculations, and pushes risk data to the interpretation output module; the scheduling optimization layer optimizes inspection frequency, calculates dynamic penalty values, and makes multi-level execution resource switching decisions, and allocates execution resources to the cloud / edge collaborative system; the execution feedback layer is responsible for issuing inspection tasks, managing manual verification processes, and performing topology weight self-correction. The cloud and edge collaborative system is equipped with lightweight models, high-precision models, and cloud-based verification capabilities, and after execution, it sends operational feedback back to the interpretation output module. The data storage module persistently stores spatial models, camera logs, topology weights, risk scores, inspection plans, and various anomaly records. The interpretation and output module provides interpretable information such as risk propagation paths, risk contribution ratios, scheduling decision bases, and risk heatmaps. Within the architecture, each layer forms a complete closed loop based on anomaly observation results, regional risk scores, inspection scheduling plans, and task execution feedback, enabling adaptive video inspection with seamless, interconnected operation.
[0025] Combination Figure 4 The adaptive video inspection scheduling method provided in this embodiment can also form a complete closed loop, which includes four core stages: perception, scoring, scheduling, and feedback. Specifically, the perception stage performs video anomaly detection, outputs anomaly observation information, and transmits it to the scoring stage. The scoring stage completes multi-source risk score calculation, driving a periodic update process based on a recalculation mechanism, iteratively updating the risk score and inspection plan according to a fixed time window, and distributing the risk score to the scheduling stage. The scheduling stage optimizes the inspection frequency and generates an inspection task plan, supporting two types of update triggering methods: periodic updates and event-triggered updates. When there is a sudden increase in anomalies, blind spot risk exceeding limits, changes in resource status, or manual feedback, scheduling recalculation is initiated immediately. The scheduling stage outputs the task execution results to the feedback stage, which summarizes the anomaly verification results, execution costs, and manual handling information, updates the model based on this information, and sends it back to the perception stage. The newly generated scheduling plan can adjust the sampling frequency, algorithm level, and inspection coverage for the next cycle. The continuous linkage of each stage forms a complete continuous closed loop, realizing the dynamic adaptive iteration of the video inspection strategy.
[0026] In one embodiment, step S101 includes: Obtain inspection parameters for the physical space to be inspected; wherein, the inspection parameters include spatial information and camera information; The physical space to be inspected is divided according to the spatial information to obtain multiple logical regions; Obtain the association information between the multiple logical regions, and set association weights for the associated logical regions based on the association information; A spatial topology model is constructed for the logical region by combining the associated information and associated weights.
[0027] In this embodiment, when dividing the physical space to be inspected, the first step is to collect the floor plan, floor plan, and area division information of the physical space to be inspected. Simultaneously, the camera installation log and various camera parameters are retrieved as the basic data for space division and model construction. Two division methods can be used in the logical area division process: one is to uniformly divide the entire physical space into grid areas of fixed size; the other is to combine the actual business scenario on site and divide the physical space into functional logical areas such as entrances / exits, access roads, warehouses, loading / unloading areas, production workshops, hazardous materials storage areas, fire exits, and equipment rooms according to business units, achieving refined and scenario-based decomposition of the physical space. In other words, the resulting logical areas can be regular grids, or they can be floors, rooms, production lines, roads, fire zones, or business areas.
[0028] After the logical regions are divided, a spatial topology model is constructed based on the association information of each logical region. In this embodiment, the spatial topology model can be defined as T=(G,E,W), where G is the set of nodes composed of all logical regions, E is the set of association edges between different logical regions, and W is the set of association weights corresponding to each association relationship. The association weight w_jm is used to characterize the risk transmission intensity of the source logical region j to the target logical region m.
[0029] The assignment of this correlation weight can be determined by comprehensively considering multiple core factors to fully align with spatial physical characteristics and operational patterns. Specifically, it can be based on factors such as geographical distance and physical proximity between regions, spatial connectivity relationships including doors and windows, passageways, vehicular roads, staircases and elevators, and utility tunnel structures; on-site process flows; standard operating procedures (SOPs); logistics transportation routes; upstream and downstream sequential relationships of personnel's routine movement paths; dynamic connectivity data formed by pedestrian, vehicle, and logistics flow and equipment operating status; co-occurrence frequency, consecutive occurrence frequency, and risk propagation sequence patterns of historical anomalies in each region; and comprehensive judgment based on factors such as the safety level of each region, hazard type, business importance, and potential impact range of accidents. In practical applications, spatial topology weights can be determined by rule configuration, statistical models, machine learning models, or a combination of the above methods.
[0030] In a preferred embodiment, the initial association weight of logical regions can be jointly determined by combining rule configuration with historical data statistics to ensure the rationality and accuracy of weight assignment. For example, for two logical regions that are physically adjacent and have connecting structures such as doors or passages, a higher initial association weight can be set; for regions that are only geographically adjacent but have no connecting structures separating them, a relatively lower initial association weight can be set; for related regions that belong to the upstream and downstream of the same work process, a directional association weight can be set according to the process transmission characteristics to accurately match the directional propagation pattern of risks.
[0031] In one embodiment, step S102 includes: The inspection coverage area and blind spot are determined based on the camera information. The logical area is divided into a first area and a second area using the covered area and the inspection blind spot. Anomaly identification is performed using cameras within the first area to obtain anomaly observation results; wherein, the anomaly observation results include anomaly type, anomaly confidence level, anomaly quantity, and anomaly time.
[0032] This embodiment determines the effective shooting coverage of a single camera based on its installation position, shooting orientation, field of view, focal length, effective recognition distance, and on-site occlusion information. The obtained coverage is then projected onto pre-defined logical regions, establishing a mapping matrix M between the camera's field of view and the logical regions. In this matrix, M(i,k) characterizes the spatial coverage status, coverage ratio, or actual coverage quality of the i-th camera over the k-th logical region. For PTZ cameras, a mapping relationship can be established between gimbal attitude, preset position, focal length parameters, and each logical region. This allows the PTZ camera to match different sets of logical regions when switching to different preset positions, achieving precise regional correspondence for variable field of view.
[0033] For any logical area, if there is no corresponding camera coverage, the camera coverage ratio does not reach the preset threshold, or the area observation quality does not meet the preset standard due to factors such as camera offline, scene occlusion, and low light at night, such logical areas can be uniformly marked as inspection blind spots, weak coverage areas, or temporarily unobservable areas, thereby completing the accurate division between the first and second areas.
[0034] like Figure 5 As shown, the space to be inspected is divided into multiple independent logical regions. Figure 5It includes fixed cameras C1 and C2, and a PTZ camera. Each camera has a corresponding field of view. A coverage relationship is established between the cameras and logical areas, and the coverage degree of each logical area is quantified by a mapping matrix M(i,k). Areas without direct camera coverage or with camera coverage below a preset threshold are marked as blind spots (e.g., ...). Figure 5 The system includes a B3 blind zone (where video perception data cannot be directly obtained), which automatically enters the risk extrapolation calculation process. PTZ cameras can be rotated to preset positions to enable blind zone supplementary inspections. Through spatial division, camera field-of-view matching, and blind zone marking, a full-domain spatial inspection model is completed, providing a spatial basis for subsequent risk assessment and inspection scheduling.
[0035] After the area is divided, based on the available camera equipment within the first area, anomaly identification and analysis are performed on real-time video streams, video frames, or captured screenshots, and corresponding anomaly observation results are output. The obtained anomaly observation results can cover multi-dimensional information, specifically including various anomaly types such as personnel gathering, area intrusion, abnormal vehicle parking, smoke and fire hazards, fire lane obstruction, equipment malfunction, and violations of safety helmet and work uniform wearing regulations. It also includes anomaly confidence level, number of anomaly targets, anomaly duration or number of consecutive frames, anomaly density or anomaly frequency per unit time, along with associated camera identifiers, logical area identifiers, timestamp information, algorithm model version, and execution resource type, forming a complete anomaly observation data system.
[0036] The corresponding real-time anomaly observation components can be calculated using the following formula: O_k(t)=sum_q alpha_q * conf_q(t) * sev_q * dur_q(t); Where q represents the anomaly type, alpha_q represents the anomaly type weight, conf_q(t) represents the confidence level, sev_q represents the severity weight, and dur_q(t) represents the duration or number of frames. Through multi-parameter coupled calculation, the real-time anomaly risk level of each logical region can be accurately quantified.
[0037] In one embodiment, step S103 includes: Obtain historical inspection data for the first area; wherein, the historical inspection data includes historical inspection trends, basic business risks, and ongoing status; A first risk score for the first region is generated by combining the aforementioned anomaly observation results, historical inspection data, and spatial topology model. Based on the spatial topology model, a target first region associated with the second region is obtained, and a second risk score for the second region is derived based on the association weight between the target first region and the second region, the abnormal observation results of the target first region, and the preset blind zone uncertainty compensation term.
[0038] This embodiment maintains a risk score for each logical region at a corresponding time. This risk score can be used to quantitatively characterize the probability of anomalies occurring, the overall risk level, and the inspection priority or normalized risk value of the corresponding logical region at the current time. In specific implementation, for logical regions that can be inspected and covered, the risk score can be comprehensively calculated by combining real-time anomaly observation results, historical anomaly change trends, spatial topology risk transmission effects, regional business basic risks, and multiple dimensions of anomaly persistence status. The corresponding risk score calculation formula is as follows: P_k(t)=β1*O_k(t)+β2*H_k(t)+β3*sum_j w_jk*O_j(t)+β4*B_k+β5* D_k(t); Wherein, O_k(t) is the real-time anomaly observation component of the current logical region, H_k(t) is the historical anomaly trend component of the region, sum_j w_jk * O_j(t) is the risk transmission component based on spatial topology, B_k is the inherent risk component corresponding to the basic risk or business importance of the region, D_k(t) is the state component corresponding to the anomaly duration, anomaly density or anomaly diffusion speed, and β1 to β5 are the weight parameters corresponding to each dimension component.
[0039] In practical applications, risk scores follow a dynamic update mechanism. For example, when a high-confidence anomaly occurs in a region, the risk score for that region can be adjusted upwards accordingly. Conversely, if no anomalies are detected in a region for several consecutive inspection cycles, the risk score can be gradually adjusted downwards according to a preset time decay mechanism to avoid the problem of fixed risk scores. The specific update formula is as follows: P_k(t+1)=rho*P_k(t)+(1-rho)*B_k; Among them, rho is a preset time decay coefficient, which is used to adjust the decay rate of the risk score and ensure that the risk score fits the real-time risk situation of the region.
[0040] For blind spots where cameras are not covered, have insufficient coverage, or are temporarily unobservable, risk extrapolation and quantitative assessment can be performed using a spatial topology model. Specifically, the set of known first-order regions with topological associations to the current blind spot can be retrieved. This is combined with abnormal observation results from each associated known region, their corresponding association weights, and a blind spot uncertainty compensation term is introduced to calculate the real-time risk score for the blind spot. The specific calculation formula is as follows: P_m(t)=sum_{j in Neighbor(m)}w_jm*O_j(t)+epsilon_m(t); Wherein, P_m(t) is the risk score of blind region m at time t, w_{jm} is the risk transmission correlation weight of known logical region j pointing to blind region m, O_j(t) is the real-time abnormal observation result associated with known logical region j, and epsilon_m(t) is the blind region uncertainty compensation term.
[0041] In practical applications, the aforementioned blind zone uncertainty compensation item can be determined by combining multiple scenario factors to fully adapt to the complex observation characteristics of the blind zone. Specifically, it can refer to the duration of the blind zone not being directly observed, the actual coverage ratio and coverage quality of the camera, the consistency and conflict degree of abnormal observation results in each associated known area, the inherent historical risks and business importance level of the blind zone, the observation failure state caused by camera offline, physical obstruction, low light at night, network anomalies, etc., as well as special working conditions such as missing manual verification, verification timeout, and inability to verify on-site.
[0042] In the preferred embodiment, the value of the blind zone uncertainty compensation term can be dynamically and adaptively adjusted. The compensation intensity gradually increases as the duration of no direct observation in the blind zone increases. After the completion of PTZ fixed-point inspection, manual on-site verification, or the restoration of direct observation capability in the blind zone, the compensation intensity is reduced accordingly, effectively solving the problem of long-term underestimation of risk in the blind zone due to the lack of direct abnormal data support in the traditional scheme.
[0043] In another implementation, the blind spot risk score can be further optimized by combining it with a Bayesian update mechanism, thereby improving the accuracy of risk projection. The specific update formula is as follows: P_m(t | O)=eta*P_m(t)*P(O | m); Wherein, P_m(t) is the prior risk score of the blind zone, P(O | m) is the conditional probability of corresponding abnormal observation results appearing in adjacent related areas when there is a risk in the blind zone, and eta is the normalization coefficient used to normalize the range of risk score values.
[0044] Of course, in other embodiments, the blind zone uncertainty compensation can also be determined by time decay models, learning models, etc.
[0045] Furthermore, during the calculation of risk scores for each region and blind spot, interpretable risk propagation path information can be generated simultaneously, thereby enabling the visualization and traceability of risk source tracing and transmission logic. This type of interpretable information can integrate multiple dimensions such as the source logical region, the target logical region or blind spot to be evaluated, the corresponding anomaly type and anomaly observation value, the topological association edges and association weights between regions, the contribution value and contribution ratio of each source risk, the risk transmission level and transmission direction, and the subsequent inspection and scheduling actions triggered by this risk score.
[0046] For example, when an abnormal gathering of personnel is detected in the loading and unloading area, complete risk transmission and tracing information can be output, clearly identifying the loading and unloading area as the risk source area. The risk is transmitted to the warehouse entrance through the topological association edge of the logistics channel, and further spread to the blind area inside the warehouse. At the same time, the risk contribution ratio of each link is quantified. Among them, the original anomaly in the loading and unloading area contributes 38% to the risk score of the blind area inside the warehouse, the contribution ratio of the associated anomaly at the warehouse entrance is 24%, and the contribution ratio of the uncertainty compensation item for the blind area is 20%.
[0047] The aforementioned explainable risk information can be applied to front-end risk heatmap visualization, inspection and scheduling process auditing, manual targeted verification, anomaly and false alarm identification, and iterative correction of spatial topology weights to continuously optimize the accuracy of risk assessment and inspection and scheduling. Alternatively, it can be stored solely in backend logs or scheduling audit records.
[0048] like Figure 6 As shown, each observable logical area outputs anomaly observation values O, such as O=0.82 for the loading and unloading area, O=0.55 for the warehouse entrance, and O=0.20 for the fire escape. Logical areas are connected by topological edges with association weights w, such as w=0.35 between the loading and unloading area and the fire escape, and w=0.62 between the loading and unloading area and the warehouse entrance, thus establishing a risk transmission relationship. Anomalies generated in the observable areas are transmitted along these topological edges to adjacent areas and blind spots. The blind spot inside the warehouse, being an area without direct observation capability, has a risk score P_m(t) calculated jointly by the risk transmission term and the uncertainty compensation term epsilon_m(t). The risk transmission term is the sum of the products of the anomalies in all associated observable areas and their corresponding topological weights; the uncertainty compensation term comprehensively considers various factors such as the duration of no observation in the blind spot, camera coverage quality, observation conflicts, equipment offline status, historical risks, and missing information from manual verification. Once an anomaly occurs at a known location, the risk is transmitted to adjacent areas and blind spots via spatial topology links. The final risk value of the blind spot is then integrated with the risk transmission results and uncertainty compensation to complete the quantitative extrapolation assessment of the blind spot risk.
[0049] In one embodiment, step S104 includes: Obtain the computational cost of a single inspection task, the importance of the region, and the corresponding risk score for each logical region. Under the constraints of the computing power budget, a non-uniform inspection frequency is generated for each logical region based on the computational cost of the single inspection task, the importance of the region, and the risk score.
[0050] Combination Figure 7 To achieve optimal allocation of computing resources across the entire domain, this embodiment constructs an inspection frequency optimization model based on a non-uniform sampling mechanism. This model replaces the traditional fixed-frequency inspection mode with differentiated inspection frequencies, adapting to the risk differences and resource constraints of each logical region. Here, the overall computing budget is set as C_{total}, the inspection frequency corresponding to logical region g_k is f_k, the computational cost of a single inspection task is cost_k, the region importance coefficient is S_k, and the risk score at the current moment is P_k(t).
[0051] This embodiment takes maximizing the risk-weighted anomaly capture revenue as the core optimization objective, while also considering global computing power budget constraints and upper and lower limits of single-region inspection frequency constraints, and constructs a corresponding optimization solution model, the specific expression of which is as follows: Maximize sum_{k=1}^n S_k*P_k(t)*(1 - exp(-lambda * f_k)); subject to sum_{k=1}^n f_k*cost_k <= C_total; f_min,k <= f_k <= f_max,k; Wherein, 1-exp(-lambda*f_k) represents the nonlinear mapping relationship between inspection frequency and anomaly capture probability, lambda is the anomaly arrival rate or sampling benefit parameter, and f_min,k and f_max,k are the preset minimum and maximum inspection frequency boundary constraints for each logical region, which can effectively avoid resource waste caused by excessively high inspection frequency or missed detection caused by excessively low frequency.
[0052] Furthermore, under the premise of satisfying all boundary constraints, the theoretical optimal inspection frequency of each logic region can be obtained by using the Lagrange multiplier method. The specific solution formula is as follows: f_k*=(1 / lambda)*ln((S_k*P_k(t)*lambda) / (mu*cost_k)); Here, *mu* represents the Lagrange multiplier obtained by solving the global computing power budget constraint. According to this theoretical formula, the optimized inspection frequency is positively correlated with the regional risk score and regional importance, and negatively correlated with the computational cost of a single inspection task. This allows for differentiated configuration effects, such as intensive inspections in high-risk, high-importance areas and reduced-frequency inspections in low-risk, high-computing-cost areas. For example, the inspection frequency for high-risk areas is 5 seconds, for medium-risk areas it is 30 seconds, and for low-risk areas it is 5 minutes.
[0053] In other embodiments, heuristic optimization, integer programming, dynamic programming, or reinforcement learning strategies may also be considered to solve for non-uniform inspection frequencies.
[0054] After obtaining the theoretically optimal inspection frequency for each region, the theoretical frequency parameters can be converted into actual task time slots that can be implemented. At the same time, a secondary adaptation scheduling is completed by combining the actual working conditions such as the carrying capacity of on-site equipment, the number of concurrent processing operations of the algorithm model, network transmission bandwidth, and the latest execution time limit of the task, so as to ensure that the inspection task can be implemented stably and efficiently.
[0055] In one embodiment, step S105 includes: Obtain the inspection requirements of each logical region, and obtain the inspection risk and benefit of the corresponding logical region based on the inspection requirements; By combining the risk-reward ratio, regional importance, and corresponding risk score of each logical region, the task priority of each logical region is calculated. The inspection task scheduling plan is generated by combining the inspection frequency and task priority; Select the corresponding execution resources according to the inspection task scheduling plan, and issue the inspection task scheduling plan to the execution resources.
[0056] In this embodiment, after obtaining the inspection requirements of each logical area, the various inspection requirements can first be converted into a complete set of candidate tasks. Each candidate task can include target area, corresponding acquisition camera, selected algorithm type, matching execution resources, estimated task time, estimated computational cost, latest execution time limit of the task, and multi-dimensional information on the risk and benefit of the task.
[0057] Then, the task priority of each logical region can be calculated based on the multi-dimensional information contained in each candidate task. The priority calculation formula is as follows: Priority_k=(S_k*P_k(t)*Gain_k) / cost_k; Among them, Gain_k represents the anomaly capture benefit, blind zone uncertainty reduction benefit, or risk confirmation benefit that can be obtained by executing the current task. The priority is quantified by combining business value and resource consumption.
[0058] Furthermore, in combination Figure 7 Furthermore, dynamic penalty values can be introduced as scheduling constraint parameters. The dynamic penalty value is negatively correlated with the regional risk score, that is, the higher the risk, the lower the penalty. It is positively correlated with task execution cost, resource contention conflict degree, task queuing waiting time, and screen repetition coverage area. Tasks with higher expected return level, higher risk score, and more return per unit of computing power correspond to lower penalty values, which can prioritize the allocation of executable time slots.
[0059] Furthermore, this embodiment also incorporates real-time regional risk scoring, task execution costs, risk-reward ratios per unit of computing power, and real-time operational status of various resources to achieve adaptive switching and scheduling of different types of execution resources. For example... Figure 8 As shown, in practical applications, the execution resources include: lightweight edge models adapted to low-risk routine inspection scenarios; high-precision edge models adapted to medium- and high-risk areas or scenarios requiring high recognition accuracy; cloud-based models and large cloud-based models adapted to high-risk review, complex anomaly recognition, and multi-dimensional comprehensive image analysis scenarios; PTZ camera preset position turning scheduling methods adapted to blind spot supplementary inspection and temporary encrypted inspection scenarios in high-risk areas; multi-algorithm serial review processes used to reduce false alarms and verify complex events; and manually generated verification work orders for scenarios with no directly observable blind spots, insufficient model recognition confidence, and handling of major high-risk events. Specifically, the execution resources may include edge boxes, camera built-in computing power, cloud services, PTZ devices, mobile inspection robots, or manual work order systems.
[0060] For example, when the blind spot risk score exceeds the first preset threshold, the inspection frequency of adjacent fixed cameras can be increased; when the blind spot risk score rises to the second preset threshold (the first preset threshold is less than the second preset threshold), the PTZ camera is activated to rotate to the matching preset position to complete the supplementary inspection; if the PTZ device still cannot cover the target area or the abnormal state still cannot be verified, a manual verification work order is generated and sent for processing.
[0061] In specific embodiments, low-risk resource release judgment conditions adapted to the global scheduling logic can also be set: for example, when no abnormalities are detected in a single logical region and all its topologically related regions for multiple consecutive inspection cycles, and the region risk score is lower than the preset release threshold, the blind zone uncertainty value meets the standard, and the manual handling feedback determines that the region risk has been eliminated or no longer needs to be inspected, multiple resource optimization operations can be executed simultaneously, including reducing the current region inspection frequency, downgrading the high-precision identification model to a lightweight edge model, terminating cloud-based verification tasks, and reducing the number of PTZ devices temporarily jumping the queue for inspection, thereby releasing idle edge computing power, cloud computing power, and idle task time slots, and reallocating all recovered resources to high-risk regions, observation blind zones, or tasks with long waiting times.
[0062] Relying on this resource release logic, redundant computing resources can be automatically recovered after the regional risk level drops, avoiding the loss of ineffective computing power caused by continuous high-frequency inspections, and achieving dynamic and balanced allocation of computing resources across the entire region.
[0063] In one embodiment, step S106 includes: The feedback data is generated based on the inspection results of the inspection equipment. The association weights of the spatial topology model are self-corrected based on the feedback data; The first and second risk scores are updated using a self-correcting spatial topology model. Based on the updated spatial topology model, the first risk score, and the second risk score, the inspection task scheduling plan is recalculated and updated using either a periodic rescheduling mode or an event-triggered rescheduling mode.
[0064] This embodiment relies on various feedback data during the inspection process to complete the adaptive correction of spatial topology weights. The feedback data can include multi-dimensional operational information, such as the historical frequency of co-occurrence of anomalies in different areas, the propagation time sequence of anomalies in adjacent areas or downstream areas of business processes, the hit rate of scheduling tasks triggered by each topology-related edge on real anomalies, the final confirmation results of manual on-site handling, the verification results of false anomaly reports, and the actual anomaly discovery results after blind spot supplementary inspections and PTZ special inspections.
[0065] Combination Figure 9In the specific correction process, if an anomaly occurs in source region a, and related anomalies frequently occur in the target associated region b within a preset time window, the association weight w_ab between the two regions can be adaptively increased. If the scheduling task triggered based on the association weight w_ab fails to hit a valid anomaly for a long period, or the corresponding inspection result is manually verified as a false alarm, the corresponding association weight w_ab can be adaptively decreased. Through the above-mentioned topology weight self-correction mechanism, the actual risk propagation patterns on site can be continuously learned, eliminating excessive reliance on manual static configuration parameters and making the spatial topology model more consistent with the actual on-site inspection scenario.
[0066] In addition, combined Figure 10 This embodiment also features a dual-mode rescheduling mechanism combining periodic and event-triggered methods to adapt to the dual needs of routine inspection updates and rapid response to emergencies. The periodic rescheduling, according to a preset fixed time window, periodically updates the overall risk score, inspection frequency for each area, and inspection task plan, ensuring the stability and continuity of the overall inspection strategy.
[0067] Event-triggered rescheduling immediately initiates risk score recalculation and inspection task plan update when a preset trigger event is detected. Triggering scenarios can include anomaly density, anomaly confidence, or anomaly duration exceeding preset thresholds; blind zone risk score or blind zone uncertainty compensation value exceeding the standard; high-risk area inspection waiting time exceeding the limit; changes in resource conditions such as edge computing power, network bandwidth, cloud services, and device online status; device anomalies such as PTZ camera occupancy, preset position call failure, and camera obstruction failure; as well as various dynamic scenarios such as manual handling feedback, false alarm and real anomaly confirmation, and low-risk resource release conditions being met.
[0068] The dual-mode rescheduling mechanism in this embodiment works together to achieve routine periodic iterative optimization of inspection strategies, ensuring the stable operation of the overall inspection system, and to make instantaneous responses to sudden risks and hidden dangers and dynamic changes in equipment resources, greatly improving the real-time performance, flexibility and adaptability of inspection scheduling.
[0069] In one specific embodiment, taking the factory park inspection scenario as an example, the overall physical space of the park is functionally divided into multiple logical areas such as loading and unloading area, warehouse entrance, warehouse interior, production workshop, hazardous materials temporary storage area, and fire passage, and a mapping relationship between the field of view of each camera and the corresponding logical area is established to complete the full-domain spatial inspection model.
[0070] When cameras deployed in the loading and unloading area detect abnormal behaviors such as vehicles lingering for extended periods or people gathering, the video perception system can accurately output multi-dimensional observation information, including the anomaly type, identification confidence level, number of abnormal targets, and duration of the anomaly. Based on this anomaly observation result, the regional risk score of the loading and unloading area can be adjusted in real time. Furthermore, relying on a pre-built spatial topology model and combining the topological association weights between different areas, the local risks in the loading and unloading area can be adaptively transmitted to related logical areas such as warehouse entrances, warehouse interiors, and fire exits, achieving a comprehensive, interconnected risk assessment.
[0071] For blind spots within warehouses not directly covered by cameras, a blind spot risk extrapolation mechanism can be used. This mechanism combines real-time anomaly observation data from loading / unloading areas and warehouse entrances to match the topological association weights between these two types of areas and the blind spots. An uncertainty compensation term for the blind spots is then added to accurately calculate the real-time risk score for these blind spots. Simultaneously, an interpretable risk propagation path is output, clearly indicating the risk source, the risk contribution percentage of each related area, and the transmission level, thus achieving traceability and interpretability in blind spot risk assessment.
[0072] After updating the risk scores for all logical areas across the entire domain, the differentiated non-uniform inspection frequency for each area is calculated by combining the business importance of each area, the computational cost of a single inspection, and the overall global computing power budget, thus achieving optimal allocation of computing resources. Based on the optimized inspection frequency and real-time task priority, the inspection frequency for high-risk areas such as loading and unloading areas and warehouse entrances is automatically increased. At the same time, PTZ cameras are scheduled to rotate to corresponding preset positions to conduct special intensive supplementary inspections of blind spots inside the warehouse, filling the gaps in blind spot inspections.
[0073] If the PTZ inspection results verify that there are no real anomalies inside the warehouse, and the risk status of each associated area remains stable over subsequent inspection cycles, a low-risk resource release mechanism can be triggered. This mechanism automatically reduces the redundant inspection frequency of each area, reclaims idle computing power and task time slots, and re-optimizes their allocation. Simultaneously, based on the complete anomaly observation, scheduling execution, and result verification feedback, the topological association weights between logical areas are adaptively corrected, and the risk transmission accuracy of the spatial topology model is continuously iterated and optimized, achieving dynamic self-optimization of the inspection strategy and spatial model.
[0074] In summary, compared with existing fixed-frequency, location-centralized traditional video inspection solutions, the adaptive video inspection scheduling method provided in this embodiment has multi-dimensional technical advantages, which can comprehensively improve the full-range, accuracy, and economy of intelligent video inspection.
[0075] Specifically, this embodiment, by constructing a global spatial topology model and introducing a blind zone uncertainty compensation mechanism, overcomes the limitations of traditional solutions that only conduct risk assessments for camera locations. It can effectively extrapolate risks in areas with weak coverage, no coverage, obstructions, and temporarily unobservable blind zones, significantly improving the overall physical space's risk perception capability and eliminating the problems of underestimating and missing risks in inspection blind zones. Based on a risk-scoring-driven non-uniform inspection frequency optimization mechanism, it can achieve dynamic and intensified inspections of high-risk areas, anomaly-related areas, and blind zones, while appropriately reducing the frequency of inspections in low-risk areas, effectively improving the accuracy of global anomaly capture and the targeting of inspections.
[0076] This embodiment can also automatically release redundant computing resources in low-risk scenarios, avoiding the ineffective computing power loss caused by the traditional fixed high-frequency inspection mode. It can accurately allocate limited edge computing power and cloud computing power resources to high-risk areas and emergency inspection tasks, greatly improving the utilization efficiency of computing resources. Through the event-triggered rescheduling mechanism, the scheduling plan can be recalculated in real time in sudden scenarios such as abnormal density surges, blind spot risk exceeding limits, equipment and network resource status anomalies, and manual feedback updates. Compared with single-cycle scheduling, it has a faster response speed and is suitable for complex and ever-changing on-site inspection conditions.
[0077] Meanwhile, this embodiment establishes a multi-level heterogeneous execution resource adaptive switching system, which can dynamically match different execution methods such as lightweight edge algorithms, high-precision identification models, cloud verification, PTZ mobile supplementary inspection, and manual verification based on real-time risk and benefit, achieving precise matching of resource capabilities and inspection needs. By outputting interpretable information such as risk propagation paths, risk contribution sources, and transmission levels, the entire process of risk assessment and scheduling decisions is traceable and auditable, providing effective basis for manual verification, false alarm identification, and scheduling optimization.
[0078] In addition, this embodiment has long-term self-iterative optimization capabilities. It can continuously correct the spatial topology weights based on historical anomaly co-occurrence patterns, scheduling task hit status, manual handling and verification feedback, so that the risk transmission model can continuously fit the real risk propagation characteristics on site. This enables the overall inspection system to adapt to changes in on-site business, scenario iterations and risk evolution patterns in the long term, and continuously ensure stable and optimal inspection results.
[0079] Figure 2 A schematic block diagram of an adaptive video inspection and scheduling system 200 based on spatial risk perception provided in an embodiment of the present invention. The system 200 includes: The spatial partitioning unit 201 is used to acquire the physical space to be inspected, divide the physical space to be inspected into multiple logical regions, and construct a spatial topology model containing association weights for the multiple logical regions. The region division unit 202 is used to divide the multiple logical regions into a first region that can be covered by inspection and a second region that is a blind spot of inspection, and to perform video anomaly perception on the first region to obtain the corresponding anomaly observation results. The scoring generation unit 203 is used to generate a first risk score for the first region based on the abnormal observation results and the spatial topology model, and extrapolate a second risk score for the second region based on the first risk score and the spatial topology model. The frequency setting unit 204 is used to set the inspection frequency for each of the logical regions based on the computing power budget constraint and in combination with the first risk score and the second risk score. The task generation unit 205 is used to generate an inspection task scheduling plan based on the inspection frequency and to send the inspection task scheduling plan to the corresponding inspection equipment. The feedback update unit 206 is used to acquire feedback data from the inspection equipment, update the spatial topology model and the first risk score and the second risk score based on the feedback data, and then generate a new inspection task scheduling plan based on the preset update mode using the updated spatial topology model, the first risk score and the second risk score.
[0080] In one embodiment, the spatial partitioning unit 201 includes: The parameter acquisition unit is used to acquire the inspection parameters of the physical space to be inspected; wherein, the inspection parameters include spatial information and camera information; The inspection division unit is used to divide the physical space to be inspected into multiple logical regions based on the spatial information. The weight setting unit is used to obtain the correlation information between the multiple logical regions and set the correlation weights for the associated logical regions according to the correlation information. A topology building unit is used to construct a spatial topology model of the logical region by combining the association information and association weights.
[0081] In one embodiment, the region division unit 202 includes: A coverage determination unit is used to determine the inspection coverage area and inspection blind spot based on the camera information; A blind spot division unit is used to divide the logical area into a first area and a second area using the coverable area and the inspection blind spot. An anomaly identification unit is used to identify anomalies using cameras in the first area and obtain anomaly observation results; wherein, the anomaly observation results include anomaly type, anomaly confidence level, anomaly quantity, and anomaly time.
[0082] In one embodiment, the rating generation unit 203 includes: The historical data acquisition unit is used to acquire historical inspection data of the first area; wherein, the historical inspection data includes historical inspection trends, basic business risks, and ongoing status. The first generation unit is used to generate a first risk score for the first area by combining the abnormal observation results, historical inspection data and spatial topology model. The second generation unit is used to obtain a target first region that is associated with the second region based on the spatial topology model, and to derive and generate a second risk score for the second region based on the association weight between the target first region and the second region, the abnormal observation results of the target first region, and the preset blind zone uncertainty compensation term.
[0083] In one embodiment, the frequency setting unit 204 includes: The region acquisition unit is used to acquire the single inspection task calculation cost, region importance and corresponding risk score of each logical region. A non-uniform generation unit is used to generate a non-uniform inspection frequency for each logical region based on the calculation cost of the single inspection task, the importance of the region, and the risk score, under the constraints of the computing power budget.
[0084] In one embodiment, the task generation unit 205 includes: The requirement acquisition unit is used to acquire the inspection requirements of each of the logical regions, and to acquire the inspection risk and benefit of the corresponding logical region based on the inspection requirements. The priority calculation unit is used to calculate the task priority of each logical region by combining the risk-reward, regional importance and corresponding risk score of each logical region. The plan generation unit is used to generate the inspection task scheduling plan by combining the inspection frequency and task priority. The selection and distribution unit is used to select the corresponding execution resource according to the inspection task scheduling plan, and distribute the inspection task scheduling plan to the execution resource.
[0085] In one embodiment, the feedback update unit 206 includes: A feedback setting unit is used to generate the feedback data based on the inspection results of the inspection equipment. A self-calibration unit is used to self-calibrate the association weights of the spatial topology model based on the feedback data; The scoring update unit is used to update the first risk score and the second risk score using the self-correcting spatial topology model. The recalculation and update unit is used to recalculate and update the inspection task scheduling plan based on the updated spatial topology model, the first risk score, and the second risk score, through either a periodic rescheduling mode or an event-triggered rescheduling mode.
[0086] Since the embodiments of the apparatus and the embodiments of the method correspond to each other, please refer to the description of the embodiments of the method for the embodiments of the apparatus, which will not be repeated here.
[0087] This invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed, can perform the steps provided in the above embodiments. The storage medium may include various media capable of storing program code, such as a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.
[0088] This invention also provides a computer device, which may include a memory and a processor. The memory stores a computer program, and when the processor calls the computer program in the memory, it can implement the steps provided in the above embodiments. Of course, the computer device may also include various network interfaces, power supplies, and other components.
[0089] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to in the method section. It should be noted that those skilled in the art can make various improvements and modifications to this application without departing from the principles of this application, and these improvements and modifications also fall within the protection scope of the claims of this application.
[0090] It should also be noted that, in this specification, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.
Claims
1. An adaptive video inspection and scheduling method based on spatial risk perception, characterized in that, include: Obtain the physical space to be inspected, divide the physical space to be inspected into multiple logical regions, and construct a spatial topology model containing association weights for the multiple logical regions. The multiple logical regions are divided into a first region that can be covered by inspection and a second region that is a blind spot for inspection. Video anomaly detection is performed on the first region to obtain the corresponding anomaly observation results. A first risk score for the first region is generated based on the abnormal observation results and the spatial topology model, and a second risk score for the second region is extrapolated based on the first risk score and the spatial topology model. Based on the computing power budget constraint, the inspection frequency is set for each logical region by combining the first risk score and the second risk score. Based on the inspection frequency, an inspection task scheduling plan is generated, and the inspection task scheduling plan is sent to the corresponding inspection equipment. The system acquires feedback data from the inspection equipment and updates the spatial topology model, the first risk score, and the second risk score based on the feedback data. Then, based on a preset update mode, it generates a new inspection task scheduling plan using the updated spatial topology model, the first risk score, and the second risk score.
2. The adaptive video inspection and scheduling method based on spatial risk perception according to claim 1, characterized in that, The step of acquiring the physical space to be inspected, dividing the physical space to be inspected into multiple logical regions, and constructing a spatial topology model including association weights for the multiple logical regions includes: Obtain inspection parameters for the physical space to be inspected; wherein, the inspection parameters include spatial information and camera information; The physical space to be inspected is divided according to the spatial information to obtain multiple logical regions; Obtain the association information between the multiple logical regions, and set association weights for the associated logical regions based on the association information; A spatial topology model is constructed for the logical region by combining the associated information and associated weights.
3. The adaptive video inspection and scheduling method based on spatial risk perception according to claim 2, characterized in that, The process of dividing the multiple logical regions into a first region that can be covered by inspection and a second region that is a blind spot for inspection, and performing video anomaly detection on the first region to obtain corresponding anomaly observation results, includes: The inspection coverage area and blind spot are determined based on the camera information. The logical area is divided into a first area and a second area using the covered area and the inspection blind spot. Anomaly identification is performed using cameras within the first area to obtain anomaly observation results; wherein, the anomaly observation results include anomaly type, anomaly confidence level, anomaly quantity, and anomaly time.
4. The adaptive video inspection and scheduling method based on spatial risk perception according to claim 1, characterized in that, The process of generating a first risk score for the first region based on the abnormal observation results and the spatial topology model, and extrapolating a second risk score for the second region based on the first risk score and the spatial topology model, includes: Obtain historical inspection data for the first area; wherein, the historical inspection data includes historical inspection trends, basic business risks, and ongoing status; A first risk score for the first region is generated by combining the aforementioned anomaly observation results, historical inspection data, and spatial topology model. Based on the spatial topology model, a target first region associated with the second region is obtained, and a second risk score for the second region is derived based on the association weight between the target first region and the second region, the abnormal observation results of the target first region, and the preset blind zone uncertainty compensation term.
5. The adaptive video inspection and scheduling method based on spatial risk perception according to claim 1, characterized in that, The step of setting the inspection frequency for each logical region based on computing power budget constraints and combining the first risk score and the second risk score includes: Obtain the computational cost of a single inspection task, the importance of the region, and the corresponding risk score for each logical region. Under the constraints of the computing power budget, a non-uniform inspection frequency is generated for each logical region based on the computational cost of the single inspection task, the importance of the region, and the risk score.
6. The adaptive video inspection and scheduling method based on spatial risk perception according to claim 5, characterized in that, The step of generating an inspection task scheduling plan based on the inspection frequency and issuing the inspection task scheduling plan to the corresponding inspection equipment includes: Obtain the inspection requirements of each logical region, and obtain the inspection risk and benefit of the corresponding logical region based on the inspection requirements; By combining the risk-reward ratio, regional importance, and corresponding risk score of each logical region, the task priority of each logical region is calculated. The inspection task scheduling plan is generated by combining the inspection frequency and task priority; Select the corresponding execution resources according to the inspection task scheduling plan, and issue the inspection task scheduling plan to the execution resources.
7. The adaptive video inspection and scheduling method based on spatial risk perception according to claim 1, characterized in that, The process of acquiring feedback data from the inspection equipment, updating the spatial topology model and the first and second risk scores based on the feedback data, and then generating a new inspection task scheduling plan using the updated spatial topology model, the first risk score, and the second risk score based on a preset update mode includes: The feedback data is generated based on the inspection results of the inspection equipment. The association weights of the spatial topology model are self-corrected based on the feedback data; The first and second risk scores are updated using a self-correcting spatial topology model. Based on the updated spatial topology model, the first risk score, and the second risk score, the inspection task scheduling plan is recalculated and updated using either a periodic rescheduling mode or an event-triggered rescheduling mode.
8. An adaptive video inspection and scheduling system based on spatial risk perception, characterized in that, include: A spatial partitioning unit is used to obtain the physical space to be inspected, divide the physical space to be inspected into multiple logical regions, and construct a spatial topology model containing association weights for the multiple logical regions. The region division unit is used to divide the multiple logical regions into a first region that can be covered by inspection and a second region that is a blind spot of inspection, and to perform video anomaly perception on the first region to obtain the corresponding anomaly observation results. The scoring generation unit is used to generate a first risk score for the first region based on the abnormal observation results and the spatial topology model, and extrapolate a second risk score for the second region based on the first risk score and the spatial topology model. A frequency setting unit is used to set the inspection frequency for each of the logical regions based on computing power budget constraints and in combination with the first risk score and the second risk score. The task generation unit is used to generate an inspection task scheduling plan based on the inspection frequency and to send the inspection task scheduling plan to the corresponding inspection equipment. The feedback update unit is used to acquire feedback data from the inspection equipment, update the spatial topology model and the first risk score and the second risk score based on the feedback data, and then generate a new inspection task scheduling plan based on the preset update mode using the updated spatial topology model, the first risk score and the second risk score.
9. A computer device, characterized in that, The method includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the adaptive video inspection scheduling method based on spatial risk perception as described in any one of claims 1 to 7.
10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the adaptive video inspection scheduling method based on spatial risk perception as described in any one of claims 1 to 7.
