A multi-camera differentiated polling video recognition method based on scene grading and dynamic weight

Abstract

The application relates to the computer technical field, in particular to a multi-camera differential polling video recognition method based on scene grading and dynamic weight, aiming to solve the problems of lacking of differential priority division, fixed polling strategy, low resource utilization efficiency and lacking of self-adjusting ability in the prior art. The method comprises: camera network initialization, obtaining the deployment position, coverage area, resolution and frame rate parameters of each camera. Through the establishment of a three-dimensional scene grading evaluation system and a reinforcement learning driven adaptive mechanism, the application realizes accurate investment and flexible allocation of recognition resources in high-value monitoring areas, so that the system can dynamically adjust resource allocation according to real-time situation changes, the polling frequency of high-priority areas is greatly improved, and the recognition accuracy is continuously optimized with running time.

Description

Technical Field

[0001] This application belongs to the field of computer technology, specifically relating to a multi-camera differentiated polling video recognition method based on scene classification and dynamic weights. Background Technology

[0002] With the rapid development of video surveillance technology and the continuous advancement of smart city construction, multi-camera systems have been widely used in numerous fields such as intelligent transportation, security monitoring, and industrial inspection. These systems deploy a large number of camera devices to monitor target areas in real time and use video recognition technology to intelligently analyze the collected image data, thereby achieving multiple functions such as traffic flow monitoring, abnormal event detection, and personnel behavior recognition. In practical application scenarios, a single monitoring area often requires the deployment of multiple cameras to achieve comprehensive coverage, while large-scale monitoring networks may contain hundreds or even thousands of camera nodes. How to efficiently organize and manage the massive video stream data collected by these camera devices has become one of the core technical challenges currently facing the video surveillance field.

[0003] Among them, polling recognition technology using multiple camera video streams plays a crucial role in video surveillance systems as a key means to optimize computing resource allocation and improve recognition efficiency. This technology coordinates the access order and recognition task allocation of multiple cameras, enabling the system to achieve balanced coverage and efficient processing of multiple video sources under the constraint of limited computing resources. With the increasing complexity of application scenarios and the growing diversity of application requirements, traditional uniform polling methods are no longer sufficient to meet the needs of prioritizing monitoring and differentiated processing of key areas in practical applications. Therefore, there is an urgent need to introduce more intelligent polling strategies to improve the overall system performance.

[0004] Existing technologies have conducted some research and proposed corresponding solutions for polling recognition of multi-camera video streams. For example, the patent technology with publication number CN121861462A acquires video stream data collected by each camera and stores it in a video resource pool. When the polling time is reached, it retrieves unpollated video stream data from the video resource pool and polls multiple different video recognition algorithms according to a preset first time interval. However, the polling strategy in this solution adopts a uniform polling method with a fixed time interval, failing to differentiate processing according to the actual needs and importance of different monitoring scenarios. This makes it difficult to achieve optimal resource allocation and maximize recognition efficiency in complex and ever-changing real-world application scenarios. In addition, this solution only implements polling of different video recognition algorithms without addressing the priority division and dynamic weight adjustment of different cameras, which has obvious limitations. For example, the patent technology with publication number CN115482667A establishes multiple polling queues by grouping multiple image acquisition devices, sequentially extracting video data from each group and performing recognition to achieve coarse-grained assessment of traffic conditions in sparse road networks. However, although this scheme groups the cameras, the grouping strategy is relatively fixed and fails to flexibly adjust according to dynamic factors such as the importance of road areas, time characteristics, and sudden events. Furthermore, this scheme uses a uniform polling method, with each group of cameras being accessed at the same frequency, lacking a differentiated attention mechanism for key areas and high-priority scenes, making it difficult to meet the needs of key road sections and congested areas in actual traffic management. The various shortcomings of the aforementioned existing technical solutions indicate that current multi-camera video recognition polling methods still have significant technical deficiencies in terms of scene adaptability, resource optimization allocation, and recognition efficiency improvement. There is an urgent need to propose a multi-camera differentiated polling video recognition method that can achieve scene classification and dynamic weight adjustment to meet the pressing needs of practical applications for differentiated polling of cameras of different priorities, increasing the recognition frequency and response speed of key areas, optimizing computational resource allocation, and improving the overall system recognition efficiency and monitoring effect. Summary of the Invention

[0005] The purpose of this invention is to provide a multi-camera differentiated polling video recognition method based on scene classification and dynamic weights, which can effectively solve the problems in the background art mentioned above.

[0006] To achieve the above objectives, the technical solution adopted by this invention is as follows: a multi-camera differentiated polling video recognition method based on scene classification and dynamic weights, comprising the following specific steps: Step 1: Camera network initialization, obtaining the deployment location, coverage area, resolution parameters, and frame rate parameters of each camera in the multi-camera system, and establishing a complete list of camera node information; Step 2: Scene classification evaluation, based on the coverage area characteristics of each camera, performing geographical importance scoring, time period feature scoring, and event frequency scoring, calculating a comprehensive scene priority score, and classifying the cameras into three levels: high priority, medium priority, and low priority; Step 3: Dynamic weights The process involves several steps: Step 1: Recalculation. Based on the scene priority level, combined with real-time situational data and environmental correction factors, the dynamic polling weight value of each camera is calculated. Step 4: Differentiated Polling Scheduling. Based on the dynamic weight values, a priority polling queue is constructed according to a weighted fair scheduling algorithm to achieve differentiated polling access to cameras with different priorities. Step 5: Video Recognition Execution. Video recognition processing, such as target detection, anomaly recognition, or feature extraction, is performed on the video streams collected by the currently polled cameras. Step 6: Adaptive Feedback Update. Based on the video recognition processing results and system operating status indicators, the scene classification score and dynamic weight parameters are dynamically updated to form a closed-loop feedback adjustment mechanism.

[0007] Preferably, the specific method for scene classification assessment in step 2 includes: a geographic importance score is generated by spatial analysis and calculation based on the distance between the camera's monitoring coverage area and important facilities, and its correlation with main roads and transportation hubs; a time period characteristic score is generated by statistical analysis of historical data to obtain the distribution patterns of traffic flow and pedestrian flow in each time period; an event frequency score is generated by evaluating the frequency of abnormal events detected by each camera in the past statistical period; and the comprehensive scene priority score is calculated by weighted summation, with the weights of each assessment dimension set as preset geographic importance weights, time period characteristic weights, and event frequency weights.

[0008] Preferably, the specific method for calculating the dynamic weight in step 3 includes: the basic weight is determined according to the priority level of the camera in step 2, and each priority camera is assigned a corresponding preset basic weight value; the time decay factor is calculated based on the time elapsed since the last polling, using an exponential form of a preset decay coefficient, with greater decay over longer periods; the environmental correction factor is dynamically adjusted based on real-time situational data and environmental monitoring data, with corresponding preset correction factors set for different scenario situations; and the final dynamic weight value is the product of the basic weight, the time decay factor, and the environmental correction factor.

[0009] Preferably, the specific method of differentiated polling scheduling in step 4 includes: the priority polling queue is constructed according to the weighted fair scheduling algorithm, and the high-priority queue, medium-priority queue and low-priority queue maintain their respective camera node lists and the most recent polling timestamp; each time a polling selection is performed, the time difference between the current time and the most recent polling time of each camera is calculated, and the priority score is calculated in combination with its dynamic weight value, and the camera with the highest priority score is selected for polling access; the polling interval of cameras with different priorities is adjusted according to a preset ratio, the basic polling interval of high-priority cameras is shortened according to a first preset ratio, the basic polling interval of medium-priority cameras remains at a standard value, and the basic polling interval of low-priority cameras is extended according to a second preset ratio.

[0010] Preferably, the specific method for performing video recognition in step 5 includes: real-time acquisition and preprocessing of video stream data from polled cameras, including image denoising, brightness correction, and size normalization; target detection using a deep learning model to identify and locate targets such as vehicles, pedestrians, and non-motorized vehicles in the video stream; anomaly detection performing behavior analysis on detected targets to identify abnormal event types such as driving in the wrong direction, intrusion, loitering, and leaving objects behind; feature extraction performing structured feature extraction on video frames to generate structured data containing target category, location, size, and confidence level; and storing the recognition results in a recognition result database after associating the recognition results with camera identifiers, timestamps, and recognition types.

[0011] Preferably, the adaptive feedback update in step 6 includes the following specific methods: the system continuously monitors operational status indicators such as recognition accuracy, false negative rate, false positive rate, and response latency; calculates the correction coefficient for scene classification based on the operational status indicators; automatically increases the scene priority score of the corresponding camera when the recognition accuracy decreases, and appropriately decreases the weight value when the false positive rate increases; optimizes the weight adjustment strategy using a reinforcement learning algorithm, with recognition accuracy as a positive reward and false positive rate as a negative penalty, to calculate the optimal weight adjustment parameters; the corrected scene classification and weight parameters take effect in the next polling cycle, forming a continuously optimized closed-loop feedback mechanism.

[0012] Preferably, it also includes a recognition result storage and query module. The recognition result storage module uses a time-series database to store recognition result data, retains historical records for a preset time period, supports multi-condition combination queries based on camera identifier, time range, and recognition type, and returns recognition statistical summary information for the corresponding time period.

[0013] Preferably, it also includes an abnormal event alarm module, which generates corresponding alarm information based on the abnormal event type and severity identified in step 5. The abnormal event types include safety accidents, group gatherings, abandoned items, suspicious loitering, etc. The alarm priority is calculated based on the severity of the abnormality and the camera weight. The higher the severity of the abnormality, the higher the priority of the alarm. After a high-priority alarm is triggered, the next round of polling weight of the corresponding camera is automatically increased and the polling interval is shortened to achieve rapid response to key areas.

[0014] Preferably, the system also includes a dynamic computing resource allocation module, which dynamically adjusts the computing resource quota allocated to each camera's video recognition task based on the computing load of the video recognition task and the system resource usage. Resource monitoring indicators include CPU utilization, memory usage, and GPU load. The load balancing algorithm performs weighted allocation based on the computing needs and resource utilization of each camera's current task to ensure that high-priority cameras receive sufficient computing resources. When the resource utilization exceeds a preset threshold, a resource reclamation mechanism is automatically triggered to prioritize the reclamation of computing resources allocated to low-priority cameras and inactive tasks.

[0015] Preferably, it also includes a historical data statistics module, which uses a sliding window mechanism to count key indicators such as the recognition frequency, number of abnormal events, and false alarm rate of each camera within a preset statistical period; the statistical results are used to dynamically adjust the scene classification score and weight parameters of each camera, and cameras with higher recognition frequency are automatically given higher priority weights; the historical data is also used to generate monitoring status reports for each area, providing data support for the optimized deployment of the video surveillance network.

[0016] Preferably, the quality control of the recognition results adopts a confidence filtering mechanism. Recognition results with a confidence level greater than a preset confidence threshold are determined as valid recognition results and processed normally. Recognition results with a confidence level less than the preset confidence threshold are marked as low confidence results and their influence coefficient is reduced during weight calculation. Low confidence results are also used to trigger camera status detection. When a single camera generates more than a preset number of low confidence results in a short time window, a device fault warning message is generated.

[0017] Compared with the prior art, the present invention has the following beneficial effects:

[0018] 1. Precise scene classification capability: By establishing a three-dimensional evaluation system based on geographical importance, time period characteristics, and event frequency, the system achieves refined classification of camera surveillance scenes, breaking through the limitations of existing technologies that lack differentiated priority classification mechanisms, and significantly increasing the investment of identification resources in high-value surveillance areas.

[0019] 2. Flexible dynamic weight adjustment: Through the synergistic effect of time decay factor and environmental correction factor, the polling weight is dynamically adjusted in real time, breaking through the limitation of fixed polling strategy in existing technologies, and enabling the system to flexibly adjust resource allocation according to the real-time situational changes of the monitoring scenario.

[0020] 3. Efficient differentiated polling: By constructing a priority polling queue through a weighted fair scheduling algorithm, differentiated polling access for cameras with different priorities is realized. This overcomes the limitation of low resource utilization efficiency caused by uniform polling in existing technologies, and significantly improves the polling frequency of high-priority areas.

[0021] 4. Intelligent adaptive optimization: Through a closed-loop feedback mechanism driven by reinforcement learning algorithms, the scene classification score and weight parameters are continuously optimized, breaking through the limitations of existing technologies that lack self-regulation capabilities, and enabling the system's recognition accuracy to continuously improve over time. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the overall technical architecture of the multi-camera differentiated polling video recognition method based on scene classification and dynamic weights proposed in this invention.

[0023] Figure 2 This is a schematic diagram illustrating the core principle framework of dynamic weight calculation and differentiated round-robin scheduling in this invention. Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] Specific implementation examples are given below.

[0026] This invention proposes a multi-camera differentiated polling video recognition method based on scene classification and dynamic weights. This method establishes a complete camera node information management system, constructs a three-dimensional evaluation system for scene classification, calculates dynamic weights using multiple factors, achieves differentiated polling based on weighted fair scheduling, performs deep recognition processing on the video stream, and achieves adaptive optimization through reinforcement learning-driven closed-loop feedback, forming a complete intelligent video surveillance polling recognition system. The core of this method lies in breaking the limitations of traditional uniform polling. Through refined scene perception and dynamic resource allocation, high-value monitoring areas receive more recognition resources and a higher polling frequency, thereby significantly improving the overall system's monitoring efficiency and response capability.

[0027] Step 1, camera network initialization, is the foundational step of this method. Its main objective is to establish a complete and accurate list of camera node information, providing a data foundation for subsequent scene classification assessment and differentiated polling scheduling. In this step, the system first acquires the deployment location information of each camera in the multi-camera system. This location information includes the precise latitude and longitude coordinates of the camera in the geographic coordinate system. The location information is acquired through methods such as marking on engineering drawings, on-site surveying, or positioning based on satellite positioning systems. The location accuracy is required to be within the meter range to ensure the accuracy of subsequent spatial analysis and calculations. Simultaneously, the system acquires the coverage area information of each camera. This coverage area is calculated and determined using the camera's focal length parameters, pitch angle parameters, installation height parameters, and the terrain features of the actual monitoring scene. The coverage area is digitally represented in the form of a polygonal geofence, and the vertex coordinates of the polygon are determined using spatial triangulation methods.

[0028] Regarding resolution parameter acquisition, the system acquires the effective pixel resolution of the video streams output by each camera. Common resolution specifications include Full HD resolution (1920×1080), 2K resolution (2560×1440), 4K resolution (3840×2160), and HD resolution (1280×720). Resolution parameters directly determine the clarity and detail retention of the video image, directly impacting the accuracy of subsequent target detection and feature extraction. As for frame rate parameters, the system acquires the video output frame rate configuration of each camera. Typical frame rate configurations include 15, 25, 30, and 60 frames per second. Frame rate parameters determine the temporal sampling density of the video stream. High frame rate cameras can provide smoother video and richer temporal information, but require more computing and network bandwidth resources for data processing and transmission.

[0029] The camera node information list uses a hybrid storage structure combining hash tables and linked lists. The key is a unique identifier for the camera, formed by combining the device number and channel number in the format "DEV-XXXX-CH-Y", where "DEV-XXXX" represents the device number and "CH-Y" represents the channel number. The linked list nodes store detailed attribute information for the cameras, including location coordinates, a polygon vertex array for the coverage area, resolution objects, frame rate integer values, installation timestamps, and last online timestamps. The system also maintains a camera status flag array to record the online, fault, and maintenance status of each camera in real time. Status information is updated periodically using a heartbeat detection mechanism. The default heartbeat detection period is set to 30 seconds; if three consecutive heartbeats time out, the camera is considered offline.

[0030] Step 2, scene classification assessment, is the core of this method for achieving differentiated services. It establishes a three-dimensional assessment system based on geographical importance, time-period characteristics, and event frequency to comprehensively quantify and score the monitoring scenes of each camera. Based on the scoring results, cameras are divided into three priority levels: high priority, medium priority, and low priority. This provides a priority classification basis for subsequent dynamic weight calculation and differentiated polling scheduling. The geographical importance scoring module first constructs a spatial correlation analysis model between the camera coverage area and important facilities. Important facility categories include government agency locations, schools, hospitals, financial institutions, commercial centers, transportation hubs, large residential communities, industrial plants, and hazardous materials storage facilities. The weight coefficients for each type of important facility in the geographical importance assessment are preset based on their social influence and security sensitivity. For example, the weight coefficient for government agency locations is set to 1.0, the weight coefficients for hospitals and schools are set to 0.9, and the weight coefficient for commercial centers is set to 0.8.

[0031] The calculation process for geographic importance score first determines the coordinates of the geometric center point of each camera's coverage area. Then, it calculates the spatial distance between this center point and the locations of each important facility. The distance calculation uses a great circle distance formula corrected for the Earth's curvature to ensure accuracy. The distance attenuation function uses a negative exponential form; the smaller the distance value, the higher the importance score after attenuation. The specific calculation formula is: the importance score equals the sum of the product of the weight coefficients of each important facility and the distance attenuation factor. The distance attenuation factor is defined as exp(-d / τ), where d represents the spatial distance from the center of the camera's coverage area to the important facility in meters, and τ represents the attenuation scale parameter, defaulted to 500 meters, meaning the importance score attenuates to approximately 36.8% of its original value every 500 meters. When multiple important facilities exist within the coverage area of ​​a single camera, the importance scores of each facility are superimposed during the calculation.

[0032] In terms of the correlation analysis between main roads and transportation hubs, the system first loads vectorized road network map data, which includes road level information, road length information, and road direction information. For each camera coverage area, the system calculates the degree of intersection between that area and roads of each level. The degree of intersection is quantitatively represented by the number of spatial intersections between the coverage area and the road centerline and the total length of the intersecting line segments. The more intersections and the longer the total length of the intersecting line segments, the stronger the camera's monitoring coverage of road traffic conditions, and the higher the corresponding geographical importance score. The transportation hub correlation analysis performs spatial buffer analysis on traffic nodes such as highway entrances and exits, urban expressway ramps, railway level crossings, airport terminals, and train stations. The buffer radius is set differently according to the level of the transportation hub: 2000 meters for level 1 transportation hubs, 1000 meters for level 2 transportation hubs, and 500 meters for level 3 transportation hubs.

[0033] The time-period feature scoring module establishes vehicle and pedestrian flow distribution models for each time period based on historical data statistical analysis, constructing a 24-hour time-period importance scoring curve for each camera. The historical data collection period defaults to the most recent 30 days, with data sources including intelligent analysis results of the camera's historical video stream and data from external traffic flow sensors. Vehicle flow statistics employ a vehicle detection algorithm combining inter-frame difference and optical flow methods, with detection results aggregated and statistically analyzed hourly to generate the mean and standard deviation of vehicle flow for each hourly segment. Pedestrian flow statistics utilize a deep learning-based pedestrian detection model, with a detection sensitivity set to 0.7 to balance the false negative and false positive rates. Pedestrian flow statistics are also aggregated hourly.

[0034] The construction of the time-period characteristic scoring curve first divides the day into 24 time periods, each lasting one hour. Then, standardized traffic flow scores for each time period are calculated based on historical statistical data. The standardization method uses min-max normalization to map the traffic flow values ​​of each time period to the interval between 0 and 1. Considering the concentrated traffic pressure characteristics during morning and evening peak hours, the system introduces a time-period amplification factor after standardization. The amplification factor is set to 1.3 for the morning peak period (7:00-9:00), 1.4 for the evening peak period (17:00-19:00), and 0.6 for the midday off-peak period (11:00-13:00). The final output of the time-period characteristic score is the score value corresponding to the current time. This value is associated with the date feature, distinguishing between weekdays and holidays. On holidays, the scores for each time period are uniformly adjusted downward to reflect the differences in traffic patterns.

[0035] The event frequency scoring module counts the frequency of abnormal events detected by each camera within a preset statistical period. Abnormal event types include traffic accidents, traffic violations, unusual gatherings, abandoned items, loitering, intrusion, wrong-way driving, and illegal parking. The default statistical period is 7 days. The statistical results are updated in real-time using a sliding window mechanism with a 1-hour step size, meaning that the earliest event record from the previous hour is removed and the latest hour's record is added every hour. The event frequency score is calculated as follows: the score equals the weighted sum of the counts of each type of abnormal event and its corresponding event weight. The weight coefficients for each type of abnormal event are preset based on the severity and urgency of the event. Severe events such as traffic accidents and intrusions have a weight coefficient of 2.0; general events such as wrong-way driving and illegal parking have a weight coefficient of 1.0; and minor events such as short-term loitering have a weight coefficient of 0.5.

[0036] The comprehensive scenario priority score is calculated using a weighted summation method, linearly combining the geographical importance score, time period characteristic score, and event frequency score according to preset weights. The weight parameters are determined through expert scoring and the analytic hierarchy process (AHP): geographical importance weight is set to 0.35, time period characteristic weight to 0.25, and event frequency weight to 0.40. The highest event frequency weight reflects the invention's emphasis on the patterns of historical abnormal events; historically high-incidence areas should receive more attention. After the comprehensive score calculation, the system determines the thresholds for the three priorities based on the score distribution. The classification method uses the natural breakpoint grading method, dividing the score sequence into three levels with the smallest internal differences and the largest inter-group differences. The high-priority threshold is the top 33% quantile of the score sequence, the medium-priority threshold is the 33% to 66% quantile range, and the low-priority threshold is the bottom 33% quantile.

[0037] Step 3: Dynamic Weight Calculation. Based on the priority levels determined in Step 2, and combined with real-time situational data and environmental correction factors, the dynamic polling weight value of each camera at the current moment is calculated, providing a quantitative basis for differentiated polling scheduling. The determination of the base weight directly depends on the priority levels divided in Step 2. The base weight of high-priority cameras is set to 10.0, the base weight of medium-priority cameras is set to 5.0, and the base weight of low-priority cameras is set to 2.0. The base weight value reflects the baseline proportion of cameras of different priorities in polling resource allocation. The base weight of high-priority cameras is twice that of medium-priority cameras and five times that of low-priority cameras.

[0038] The time decay factor is used to prioritize cameras that haven't been polled for a long time, preventing low-priority cameras from experiencing service starvation due to prolonged lack of service. The time decay factor is calculated using an exponential decay model, mathematically expressed as: the time decay factor equals 1 plus a decay coefficient multiplied by an exponential function of the time interval since the last polling, where the decay coefficient is set to 0.1, and the time unit is seconds. When the time interval since the last polling for a camera is 60 seconds, the time decay factor is calculated as 1 plus 0.1 multiplied by the square of 60 divided by 1000, resulting in 1.36, indicating that the camera's priority increases by 36% after 60 seconds. The time decay factor is set to an upper limit of 3.0 to prevent cameras that haven't been polled for a long time from having excessively high weights and interfering with the normal polling rhythm.

[0039] The environmental correction factor is dynamically adjusted based on real-time situational data and environmental monitoring data. Situational data sources include meteorological monitoring data, real-time traffic flow data, security levels for major events, and early warning information for emergencies. Meteorological monitoring data includes parameters such as temperature, humidity, visibility, and precipitation. When visibility is below 500 meters or precipitation reaches moderate to heavy rain levels, the environmental correction factor for all cameras is uniformly reduced by 0.2 to reflect the negative impact of severe weather on video recognition performance. Real-time traffic flow data is obtained by acquiring vehicle speed and congestion indices for each road segment. When the congestion index for a road segment exceeds a preset threshold, the environmental correction factor for the associated cameras on that segment is increased by 0.15. Security levels for major events are divided into four levels, corresponding to correction factors of 1.3, 1.2, 1.1, and 1.0, respectively; the higher the security level, the larger the corresponding correction factor.

[0040] The final dynamic weight value is calculated as the product of the base weight, the time decay factor, and the environment correction factor. The formula is: Dynamic weight equals base weight multiplied by the time decay factor multiplied by the environment correction factor. Taking a high-priority camera as an example, with a base weight of 10.0, a time decay factor of 1.36, and an environment correction factor of 1.15, its dynamic weight is calculated as 10.0 multiplied by 1.36 multiplied by 1.15, resulting in 15.64. Compared to its base weight of 10.0, this camera's polling weight in the current situation has increased by 56.4%, demonstrating the dynamic weight mechanism's responsiveness to real-time situations.

[0041] Step 4, Differentiated Round-Robin Scheduling, is based on the dynamic weight values ​​calculated in Step 3. A priority round-robin queue is constructed according to the weighted fair scheduling algorithm to achieve differentiated round-robin access for cameras of different priorities, ensuring that high-priority cameras receive a higher polling frequency and faster response time. The priority round-robin queue adopts a three-level queue structure: a high-priority queue, a medium-priority queue, and a low-priority queue. Each queue maintains a doubly linked list structure of camera nodes and a recent polling timestamp for each node. The recent polling timestamp records the absolute time point of each camera's most recent polling access, with millisecond precision, and is used to calculate the time difference and priority score.

[0042] During each round-robin selection, the system first calculates the current available round-robin quota for each queue. The quota calculation uses a weighted round-robin algorithm, with the formula: queue quota equals the sum of the dynamic weights of all cameras within the queue multiplied by a priority adjustment coefficient. The adjustment coefficient for high-priority queues is set to 1.5, for medium-priority queues to 1.0, and for low-priority queues to 0.6. After quota calculation, the system performs round-robin selection sequentially according to the order of high-priority queues, medium-priority queues, and low-priority queues. Cameras within the same queue are selected based on their priority scores, from highest to lowest.

[0043] The priority score is calculated by combining the camera's dynamic weight value and the time difference since the last polling. The mathematical formula is: Priority score equals dynamic weight multiplied by a time difference correction factor, where the time difference correction factor is defined as the ratio of the time difference to the basic polling interval. When a camera's dynamic weight is 15.64, the time difference since the last polling is 90 seconds, and the basic polling interval is 30 seconds, its priority score is calculated as 15.64 multiplied by 3.0, resulting in 46.92. The system selects the camera with the highest priority score at the current moment for polling, and after each polling visit, updates the camera's most recent polling timestamp to the current moment.

[0044] The polling intervals for cameras of different priorities are adjusted according to a differentiated strategy. The basic polling interval for high-priority cameras is shortened from the standard value of 30 seconds to 18 seconds, an adjustment ratio of 0.6 times. The basic polling interval for medium-priority cameras remains at the standard value of 30 seconds, while the basic polling interval for low-priority cameras is extended from the standard value of 30 seconds to 48 seconds, an adjustment ratio of 1.6 times. This differentiated setting of polling intervals ensures that high-priority cameras are polled significantly more times per unit time than low-priority cameras. Assuming the system runs for 1 hour, the theoretical number of polling times for high-priority cameras is 200, for medium-priority cameras it is 120, and for low-priority cameras it is 75.

[0045] Step 5, Video Recognition, performs video recognition processing on the video streams captured by the currently polled cameras, including target detection, anomaly detection, and feature extraction, generating structured recognition result data. Video stream data is acquired in real-time via RTSP or RTMP protocols to establish a media session with the cameras and obtain real-time H.264 or H.265 encoded video streams. The video stream data first enters the preprocessing module for data cleaning, including image denoising, brightness correction, and size normalization. Image denoising uses a spatiotemporal bilateral filtering algorithm with a filter kernel radius of 3 pixels, a spatial standard deviation of 50, and a grayscale standard deviation of 20, effectively removing video compression noise and sensor noise while preserving edge details. Brightness correction uses histogram equalization to adaptively stretch the image grayscale histogram, resulting in a more uniform brightness distribution. Size normalization adjusts video frames captured by cameras of different resolutions to a standard resolution of 832×480 to meet the input requirements of subsequent deep learning models.

[0046] The object detection module employs a deep convolutional neural network-based model. The model architecture is a hybrid of a two-stage detector and a single-stage detector, with a ResNet50 backbone network and a feature pyramid structure for multi-scale feature fusion. During training, the model utilizes a large-scale labeled dataset for transfer learning, including over 500,000 labeled images covering more than 50 object categories, such as vehicles, pedestrians, non-motorized vehicles, animals, and other targets. A confidence threshold of 0.5 is set for object detection; detection boxes below this threshold are filtered out and discarded. A non-maximum suppression (NMS) cross-union (CUI) threshold of 0.5 is also set to remove overlapping detections. Detection results are output in a structured data format, including object category identifiers, bounding box coordinates, confidence scores, and object size estimates.

[0047] The anomaly detection module performs temporal behavior analysis on target detection results to identify various anomaly event types. The anomaly event type identification algorithm employs a dual-stream architecture combining a Long Short-Term Memory (LSTM) network and a spatiotemporal graph convolutional network. Spatial stream processing handles the spatial feature representation of a single frame image, while temporal stream processing handles the temporal dynamic features of a multi-frame image sequence. Reverse driving event recognition determines whether the target is moving in the wrong direction by analyzing the consistency between the direction vector of the target's trajectory and the permitted road direction. When the angle between the target's movement direction and the permitted direction exceeds 135 degrees, it is considered reverse driving. Intrusion event recognition uses a defined electronic fence area. When a target is detected entering the fenced area from the outside, an intrusion alarm is triggered. The fence area can be set in polygonal or rectangular formats. Loitering event recognition tracks the target's dwell time and movement trajectory within a specific area. When the dwell time exceeds a preset threshold and the movement trajectory exhibits repetitive characteristics, it is considered loitering. The dwell time threshold is differentiated based on the area type: 60 seconds for sensitive areas and 180 seconds for general areas. The abandoned object event recognition analyzes the changes in inter-frame differences and triggers an abandoned object alarm when a new stationary object is detected and there is no one watching it for more than a set time. The default value of the set time is 30 seconds.

[0048] The feature extraction module performs structured feature extraction on video frames, generating structured data containing target category, location, size, and confidence score. Feature extraction employs a multi-task learning network, simultaneously outputting target classification features, bounding box regression features, and attribute regression features. The classification features are 512-dimensional and mapped to the probability distribution of each target category through a fully connected layer. Location features include the target's center point coordinates and width and height dimensions in the image; the coordinate values ​​are normalized to the 0-1 range. The confidence score features directly reuse the confidence score output by the target detection module. The structured data is stored using Protocol Buffer serialization, offering high encoding compression rates and cross-platform compatibility.

[0049] The associated storage step of the recognition results assembles the recognition results with camera identifiers, timestamps, and recognition types. Camera identifiers use the standard format identifier established in step 1, and timestamps are accurate to the millisecond level, using the UTC time standard to ensure time consistency across multiple cameras. The recognition type field records the specific type of recognition, including categories such as regular target detection, anomaly event recognition, and quality inspection. The assembled recognition results are stored in the recognition result database via a database write interface. The database uses the time-series database InfluxDB, which is specifically optimized for writing and querying time-series data, supporting efficient data aggregation queries and time-range retrieval.

[0050] Step 6, Adaptive Feedback Update, dynamically updates scene classification scores and dynamic weight parameters based on video recognition processing results and system operation status indicators, forming a closed-loop feedback adjustment mechanism to achieve continuous self-optimization of the system. The system operation status monitoring module continuously collects and calculates various performance indicators, including recognition accuracy, false negative rate, false positive rate, and response latency. Recognition accuracy is calculated using a sampling verification method, randomly selecting a certain proportion of samples from historical recognition results for manual annotation and comparison. The default sampling proportion is 5%, and the calculation formula is: accuracy equals the number of correctly recognized items divided by the total number of samples. The false negative rate is calculated based on the ratio of the number of events discovered manually but not recognized by the system to the total number of events. The false positive rate is calculated based on the ratio of the number of events that trigger system alarms but are manually confirmed as normal to the total number of alarms. The response latency indicator records the end-to-end delay time from the occurrence of an abnormal event to the system completing the recognition and alarm process, including video stream transmission delay, video decoding delay, model inference delay, and result processing delay.

[0051] The scene classification correction coefficient is calculated based on the deviation of the operating status index from the baseline value. When the recognition accuracy is lower than 90% of the baseline value, it indicates that the recognition difficulty of the scene corresponding to the camera has increased. The system automatically increases the scene priority score of the camera, and the correction magnitude is proportional to the accuracy deviation; for every 1 percentage point decrease in deviation, the score increases by 0.5 points. When the false alarm rate is higher than the baseline value by 5%, the system also increases the scene priority score of the corresponding camera, with the correction magnitude being twice the false alarm rate deviation. When the false alarm rate is higher than the baseline value by 10%, the system appropriately reduces the weight value of the camera to reduce its occupation of system resources, with the correction magnitude being 1.5 times the false alarm rate deviation. Upper and lower limits are set for each correction coefficient, and the score adjustment range is limited to within ±20%.

[0052] The reinforcement learning optimization algorithm employs a deep Q-network architecture, using recognition accuracy as a positive reward signal and false alarm rate as a negative penalty signal. It learns the optimal weight adjustment strategy through continuous interaction with the environment. The reinforcement learning environment simulator outputs expected system performance feedback based on the current weight configuration and scene state. The reward function is defined as recognition accuracy multiplied by the weight coefficient minus the false alarm rate multiplied by the penalty coefficient. The agent policy network uses a deep neural network architecture. The input layer dimension is the sum of the scene classification parameters and the running state index dimension, while the output layer dimension is the weight adjustment action space, which includes three discrete actions: weight increase, weight decrease, and weight hold. The training process employs an experience replay mechanism and a target network stabilization technique, updating the target network parameters every 100 training epochs to ensure the stability of the learning process.

[0053] The revised scenario classification scores and weight parameters are written to the system configuration storage area via the parameter update interface. The update operation employs transactional processing to ensure the atomicity and consistency of parameter changes. The updated parameters take effect in the next polling cycle. The polling scheduling module reads the latest weight parameters from the configuration storage area at the beginning of each polling cycle. The closed-loop feedback mechanism operates on a default cycle of 10 minutes, meaning parameter evaluation and updates are performed every 10 minutes. This long-cycle feedback adjustment avoids the impact of frequent parameter fluctuations on system stability.

[0054] The recognition result storage and query module uses a time-series database to store the recognition results data. InfluxDB is chosen as the core storage engine, specifically optimized for time-series workloads and capable of supporting write throughput of millions of data points per second. The database employs a distributed cluster deployment mode, with a data replica count of 3 to ensure data reliability and high service availability. The data retention policy sets a retention period of 90 days, automatically archiving historical data older than 90 days to cold storage nodes to reduce storage costs. The database table structure uses a modeling approach that separates tag columns and field columns. Tag columns include query dimension fields such as camera identifier, recognition type, and event type, while field columns include numerical fields such as recognition result details, confidence scores, and associated metadata.

[0055] Multi-condition combined queries support flexible data retrieval using Structured Query Language (SCL). Query conditions include any combination of three main dimensions: camera identifier, time range, and recognition type. Query results are returned as JSON structured data, containing the total number of records matched, aggregated statistical values ​​for each field, and an array of time-series data points. Statistical summaries include the total number of recognitions for each camera within a specified time period, statistics on abnormal event classifications, trends in recognition accuracy, and percentiles of response latency. The query interface uses a RESTful API design, supporting GET requests for data queries and POST requests for submitting complex query conditions, with an interface response time controlled within 100 milliseconds.

[0056] The abnormal event alarm module generates corresponding alarm information based on the abnormal event type and severity identified in step 5. The alarm information generation adopts an event-driven asynchronous processing architecture. The rules for mapping abnormal types to alarm levels include: safety accident anomalies, including traffic accidents, fire alarms, and equipment failures, are mapped to alarm level 1, with lower alarm level values ​​indicating higher priority; crowd gathering anomalies, including excessive crowd density and panic stampede warnings, are mapped to alarm level 2; abandoned items anomalies, including suspicious abandoned items and unattended luggage, are mapped to alarm level 3; and suspicious loitering anomalies, including targets staying in sensitive areas for extended periods and abnormal trajectory tracking, are mapped to alarm level 4.

[0057] The comprehensive calculation formula for alarm priority is: alarm priority equals the base priority of the anomaly severity multiplied by the camera weight coefficient. The base priority of the anomaly severity is determined according to the alarm level: level 1 alarms have a base priority of 100, level 2 alarms have a base priority of 80, level 3 alarms have a base priority of 60, and level 4 alarms have a base priority of 40. The camera weight coefficient is normalized based on the dynamic weight calculated in step 3, with the normalization benchmark being the maximum value of the dynamic weights of all cameras in the system. After a high-priority alarm is triggered, the system automatically performs two linked operations: the first is to increase the corresponding camera's polling weight in the next round by 50% of the base weight; the second is to shorten the polling interval of the camera by 40% of the base interval. The default duration of the linked adjustment is 10 minutes, after which the normal weight configuration is automatically restored.

[0058] The dynamic resource allocation module dynamically adjusts the computing resource quota allocated to each camera's video recognition task based on the computational load and system resource usage. This module is crucial for ensuring the effective execution of the differentiated polling strategy. System resource monitoring metrics include three dimensions: CPU utilization, memory usage, and GPU load. The monitoring data collection frequency is set to once per second, and the collected results are smoothed using an exponentially weighted moving average method to eliminate instantaneous fluctuations. The alarm threshold for resource utilization is set at 80%. When any metric exceeds the alarm threshold, the system triggers a load balancing process.

[0059] The load balancing algorithm uses a priority-based weighted least connections algorithm. The calculation formula is: task resource quota equals task priority coefficient multiplied by the total available resources divided by the total number of currently active tasks. The task priority coefficient is determined according to the camera's priority level: high-priority tasks have a coefficient of 1.5, medium-priority tasks have a coefficient of 1.0, and low-priority tasks have a coefficient of 0.5. The total available resources are calculated by subtracting the reserved resources from the total system resources. The reserved resources are set at 20% of the total resources to handle sudden loads. The criteria for determining an active task are that the task is running and the time since the last scheduling does not exceed 60 seconds.

[0060] The resource reclamation mechanism is triggered when the overall system resource utilization exceeds 90%. The reclamation strategy adopts a reverse priority order, prioritizing the reclamation of computing resources allocated to low-priority cameras, with a reclamation amount of 30% of the current quota for low-priority tasks. Inactive tasks are defined as those where more than 300 seconds have passed since the last video recognition processing and there is no pending data in the current task queue; the reclamation amount for inactive tasks is 100% of their total quota. The resource reclamation operation adopts a gradual execution method, with each reclamation not exceeding 10% of the total system resources to avoid interrupting running tasks.

[0061] The historical data statistics module employs a sliding window mechanism to collect key indicators such as recognition frequency, number of abnormal events, and false alarm rate for each camera within a preset statistical period. The statistical results guide scene classification scoring and dynamic adjustment of weight parameters. The sliding window mechanism utilizes a circular buffer data structure, with a default window size of 24 hours. The buffer is divided into 24 equal-length sub-windows, each corresponding to a one-hour data storage interval. Data writing uses an overwrite mode; new data automatically overwrites the old data in the corresponding time window, ensuring that the statistical results reflect the data characteristics of the most recent 24 hours.

[0062] The calculation of statistical indicators includes three dimensions: identification frequency statistics, abnormal event classification statistics, and false alarm rate statistics. Identification frequency statistics calculate the total number of identifications by each camera within each sliding window, on an hourly basis, and also calculate the moving average of the total number of identifications within the sliding window to reflect the changing trend of the identification volume. Abnormal event classification statistics count the number of various abnormal events detected by each camera within the sliding window, categorized by event type. The statistical results are stored in the form of a nested mapping table, where the first-level key is the camera identifier, the second-level key is the abnormality type, and the value is the cumulative count for that type. False alarm rate statistics, based on alarm confirmation result data, calculate the ratio of the number of false alarms generated by each camera within the sliding window to the total number of alarms.

[0063] The monitoring situation reports are generated on a scheduled basis, with the default scheduling cycle set to 2:00 AM daily. Report content includes a workload distribution map for each area, a spatiotemporal heatmap of abnormal events, a performance ranking table for each camera, and a system resource usage trend chart. Reports are output in both PDF and HTML formats. PDF is used for archiving and printing, while HTML is used for web page display and interactive queries. Report data is synchronously exported to a data warehouse for long-term archiving. The data warehouse uses a distributed columnar storage engine, supporting online analysis, processing, querying, and large-scale historical data mining and analysis.

[0064] The quality control of the recognition results employs a confidence filtering mechanism to improve the reliability of the results. The confidence threshold is configured differently based on the recognition type: 0.7 for regular target detection and 0.85 for anomaly event recognition. Recognition results with a confidence level greater than the threshold are considered valid and proceed to subsequent processing and database storage. Recognition results with a confidence level less than the threshold are marked as low-confidence results, and their influence coefficient is automatically reduced during weight calculation. The influence coefficient is calculated as: influence coefficient equals confidence level divided by the threshold. When the confidence level is 0.5 and the threshold is 0.7, the influence coefficient is calculated to be 0.714.

[0065] The continuous accumulation of low-confidence results triggers camera status detection. When a single camera generates more than a preset threshold number of low-confidence results within a short time window, the system generates a device fault warning. The default short time window is 10 minutes, and the default threshold is 5 times. The generation of the device fault warning includes three fault type determinations: a significant decrease in the number of target detection boxes in the low-confidence results indicates camera occlusion or lens contamination; an offset of the detection box position beyond the normal range indicates camera angle offset; and an overall decrease in confidence but a normal detection count indicates abnormal lighting conditions. The fault warning information is pushed to the operation and maintenance management system via an alarm interface, and maintenance personnel perform on-site inspections and troubleshooting based on the warning information.

[0066] The technical solution of this invention constructs a complete intelligent multi-camera video recognition polling system through the coordinated operation of the aforementioned six core steps and multiple functional modules. This system overcomes the limitations of traditional uniform polling strategies by employing refined 3D scene hierarchical evaluation, a real-time dynamic weight adjustment mechanism, an intelligent weighted fair scheduling algorithm, and a reinforcement learning-based adaptive optimization strategy. This achieves efficient allocation of video recognition resources towards high-value monitoring areas. Practical application verification shows that this method can significantly improve the monitoring frequency and response speed of high-priority areas while maintaining an optimized balance in overall system resource utilization efficiency, providing an advanced technical solution for multi-camera video recognition applications in fields such as intelligent transportation, security monitoring, and industrial inspection.

[0067] Example 2

[0068] This embodiment serves as an alternative technical solution to Embodiment 1, focusing on the specific implementation details and parameter configurations of the method of the present invention applied in a large-scale park or complex setting. Assume a smart park has deployed 200 surveillance cameras, covering various scene types such as main roads, parking lots, office buildings, production workshops, and warehouse areas. The park management needs to achieve efficient monitoring of key areas and rapid response to abnormal events.

[0069] In this application scenario, the camera network initialization phase first involves grouping and labeling the 200 cameras according to their functional areas. The grouping rules are set based on the actual functional zoning of the park, including six groups: core production area, general production area, office area, living area, public facilities area, and border control area. The number of cameras in each group is dynamically configured according to the actual deployment; for example, the core production area group might have 30 cameras, the general production area group 60 cameras, the office area group 50 cameras, the living area group 20 cameras, the public facilities area group 25 cameras, and the border control area group 15 cameras. The mapping relationship between group information and camera labels is stored in a group configuration table for subsequent batch management and statistical analysis.

[0070] The scenario-based assessment requires parameter customization to meet the specific needs of the park. Regarding geographical importance scoring, key facilities within the park include the central control room, power distribution room, hazardous materials warehouse, data center, and fire control center. The importance coefficients for these facilities are increased to reflect the high standards of safety management within the park. The importance coefficient for the hazardous materials warehouse is set at 1.2, and for the data center at 1.15. In terms of time-of-day characteristics scoring, the park's work schedule significantly impacts the scoring curve. The score for the workday (8:00 AM to 6:00 PM) is generally higher than that for the nighttime, while the weekend scoring curve differs significantly from weekdays. The score for the midday (12:00 PM to 1:00 PM) is relatively lower due to staff being away from the park. Regarding event frequency scoring, historical data from the park indicates that unauthorized intrusions and missing persons are the main anomaly types at night. The weight coefficients for these two types of events are increased compared to typical scenario settings.

[0071] Dynamic weighting calculations incorporate a park-specific environmental correction factor in this scenario. The production safety status factor is adjusted based on real-time data from the park's safety management system. When the park issues a safety production warning, the environmental correction factor for all cameras associated with production areas is increased by 0.2. The visibility correction factor is adaptively adjusted for nighttime and foggy weather conditions, with a uniform nighttime visibility correction factor set at 0.85. The personnel density factor is adjusted based on the real-time number of people entering the park as counted by the park's access control system. When the number of people entering the park exceeds 70% of the park's maximum capacity, the environmental correction factor for cameras in key areas is increased by 0.1 to reflect the increased monitoring pressure caused by dense crowds.

[0072] Differentiated polling scheduling in this scenario needs to address the scheduling efficiency issue of a large number of cameras. The system adopts a hierarchical scheduling architecture. The first layer is group-level scheduling, which allocates polling quotas between groups according to group priorities. The second layer is intra-group scheduling, which selects cameras within a group based on their dynamic weights. The determination of group priorities comprehensively considers the number of cameras within a group, the average weight within the group, and the service completion rate of the group in the previous cycle. Intra-group scheduling adopts a pipelined processing mode, parallelizing the polling selection operation of cameras within the group to improve scheduling efficiency.

[0073] The video recognition system incorporates a specific model tailored to the park's characteristics. A safety equipment detection model is added to the production area to automatically detect whether personnel entering the area are wearing safety helmets, protective glasses, and gloves, with a confidence threshold of 0.8. The vehicle recognition model includes a license plate recognition submodule to automatically extract and verify license plate numbers of vehicles entering and leaving the park. The smoke and fire detection model employs a dual-band infrared thermal imaging and visible light image fusion analysis scheme, enabling early detection of fire hazards and the issuance of warnings.

[0074] In this scenario, the adaptive feedback update employs a multi-objective optimization algorithm to balance performance metrics across multiple dimensions. The optimization objectives include maximizing recognition accuracy, minimizing false alarm rate, minimizing response latency, and optimizing resource utilization. The multi-objective optimization uses a weighted summation method to linearly combine multiple objective functions, with weight allocation dynamically adjusted based on the preferences of the park management. The historical data statistics module adds production safety indicator statistics functions, tracking the number of violations, safety equipment wearing rates, and equipment anomaly detection frequencies in each production area, providing data support for park safety management.

[0075] Example 3

[0076] This embodiment, as an alternative technical solution, focuses on illustrating the cross-regional collaborative implementation of the method of this invention in the context of smart city construction. Assume a smart city project needs to integrate video surveillance resources from multiple administrative regions within a city to achieve cross-regional collaborative monitoring of key areas and coordinated response to emergencies.

[0077] In this application scenario, the initial phase of the camera network involves establishing a unified cross-regional coding system. This system employs a six-segment structure: the first segment is the area code (two digits representing the district / county); the second segment is the street code (two digits representing the street office); the third segment is the community code (two digits representing the community); the fourth segment is the camera type (one letter representing the monitoring point type); the fifth segment is the camera serial number (three digits representing the serial number within that type); and the sixth segment is the channel number (one digit representing the device channel number). This unified coding system ensures unique identification and hierarchical attribution lookup for cameras throughout the city.

[0078] The scenario-based hierarchical assessment requires the establishment of a cross-regional importance assessment model. The regional importance dimension includes three primary indicators: political impact, economic impact, and social attention. The political impact dimension assessment criteria consider whether the target area is a venue for important political activities, a location for diplomatic missions or consulates, or a concentrated area of ​​government agencies. The economic impact dimension assessment criteria consider the target area's commercial density, average annual passenger flow, and economic output density. The social attention dimension assessment criteria consider the frequency of historical public opinion events, the number of resident complaints and reports, and media exposure in the target area. The scores for the three dimensions are weighted using the analytic hierarchy process (AHP), with the political impact dimension weighted at 0.4, the economic impact dimension at 0.35, and the social attention dimension at 0.25.

[0079] Dynamic weight calculation incorporates city-level situational factors in this application scenario. The large-scale event support factor activates during major exhibitions, sporting events, or cultural performances held in the city, with different correction coefficients set based on the event's scale and impact. The severe weather warning factor activates when meteorological departments issue warnings for heavy rain, heavy snow, or strong winds, increasing the weight of relevant cameras in areas prone to meteorological disasters. The major event linkage factor activates when a sudden event occurs, causing a linked increase in the weight of cameras in the event's location and surrounding areas; the range of this increase is determined based on the event type and its radius of influence.

[0080] Differentiated round-robin scheduling in this application scenario requires cross-regional load balancing. The system deployment adopts a cloud-edge collaborative architecture, with the central cloud platform responsible for cross-regional task scheduling and global resource allocation, while edge computing nodes handle local video stream preprocessing and initial recognition. The round-robin scheduling strategy employs a two-level weight calculation: edge nodes calculate local dynamic weights for edge-side priority scheduling, while the central cloud platform calculates global dynamic weights for cross-regional collaborative scheduling. The calculation of the global dynamic weights adds a regional coordination factor to the local dynamic weights, and this regional coordination factor is dynamically adjusted based on the real-time monitoring pressure and service needs of each region.

[0081] In this application scenario, video recognition has been enhanced with cross-regional trajectory tracking. When a target moves across regions, the system uses pedestrian re-identification and vehicle ReID technologies to achieve cross-camera identity matching and trajectory stitching. The trajectory tracking results are stored in a distributed graph database for spatiotemporal relationship analysis. Abnormal event identification results are synchronized in real time to the city's central command center, which then coordinates resources from surrounding areas based on the event's location and type.

[0082] In this application scenario, adaptive feedback updates employ a federated learning framework for model optimization. Edge nodes in each region train their models on local data, uploading only the model gradient parameters to the central platform for aggregation. The aggregated model parameters are then distributed back to the edge nodes for updates. This federated learning framework ensures data privacy and security across regions while enabling collaborative optimization of model parameters. The historical data statistics module adds city-level situational analysis capabilities, statistically analyzing the distribution of monitoring activity across regions, the spatiotemporal patterns of abnormal events, and cross-regional population flow patterns, providing decision support for urban planning and emergency management.

[0083] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A multi-camera differentiated polling video recognition method based on scene classification and dynamic weights, characterized in that, Includes the following steps: Step 1: Camera network initialization. Obtain the deployment location, coverage area, resolution parameters, and frame rate parameters of each camera in the multi-camera system, and establish a complete list of camera node information. Step 2: Scene classification and evaluation. Based on the coverage area characteristics of each camera, perform geographical importance scoring, time period feature scoring, and event frequency scoring, calculate a comprehensive scene priority score, and classify the cameras into three levels: high priority, medium priority, and low priority. Step 3: Dynamic weight calculation. Based on the scene priority levels, combined with real-time situational data and environmental correction factors, calculate the dynamic polling weight value of each camera. Step 4, Differentiated Polling Scheduling: Based on the dynamic weight values, a priority polling queue is constructed according to the weighted fair scheduling algorithm to achieve differentiated polling access to cameras with different priorities; Step 5, Video Recognition Execution: Target detection, anomaly recognition, or feature extraction processing is performed on the video streams collected by the currently polled cameras; Step 6, Adaptive Feedback Update: Based on the video recognition processing results and system operating status indicators, the scene classification score and dynamic weight parameters are dynamically updated to form a closed-loop feedback adjustment mechanism.

2. The multi-camera differentiated polling video recognition method based on scene classification and dynamic weights according to claim 1, characterized in that, The specific methods for scene classification assessment in step 2 include: Geographical importance scoring is generated by spatial analysis and calculation based on the distance between the camera's monitoring coverage area and important facilities, and its correlation with main roads and transportation hubs; Time period characteristic scoring is generated by statistical analysis of historical data to obtain the distribution patterns of traffic and pedestrian flow in each time period; Event frequency scoring is generated by assessing the frequency of abnormal events detected by each camera in the past statistical period; The comprehensive scene priority score is calculated by weighted summation, with the weights of each assessment dimension set as preset geographical importance weight, time period characteristic weight, and event frequency weight.

3. The multi-camera differentiated polling video recognition method based on scene hierarchical and dynamic weights according to claim 1, characterized in that, The specific method for calculating the dynamic weight in step 3 includes: the basic weight is determined according to the priority level of the camera in step 2, and each priority camera is assigned a corresponding preset basic weight value; the time decay factor is calculated based on the time elapsed since the last polling, using the exponential form of the preset decay coefficient, with greater decay as time increases; the environmental correction factor is dynamically adjusted based on real-time situational data and environmental monitoring data, and corresponding preset correction factors are set for different scenario situations; the final dynamic weight value is the product of the basic weight, the time decay factor, and the environmental correction factor.

4. The multi-camera differentiated polling video recognition method based on scene hierarchical and dynamic weights according to claim 1, characterized in that, The specific method of differentiated polling scheduling in step 4 includes: the priority polling queue is constructed according to the weighted fair scheduling algorithm, and the high-priority queue, medium-priority queue and low-priority queue maintain their own camera node list and the most recent polling timestamp respectively; each time a polling selection is performed, the time difference between the current time and the most recent polling time of each camera is calculated, and the priority score is calculated in combination with its dynamic weight value, and the camera with the highest priority score is selected for polling access; the polling interval of cameras with different priorities is adjusted according to a preset ratio, the basic polling interval of high-priority cameras is shortened according to the first preset ratio, the basic polling interval of medium-priority cameras remains at the standard value, and the basic polling interval of low-priority cameras is extended according to the second preset ratio.

5. The multi-camera differentiated polling video recognition method based on scene classification and dynamic weights according to claim 1, characterized in that, The specific methods for performing video recognition in step 5 include: real-time acquisition and preprocessing of video stream data from polled cameras, including image denoising, brightness correction, and size normalization; target detection using a deep learning model to identify and locate vehicles, pedestrians, and non-motorized vehicles in the video stream; anomaly detection performing behavior analysis on detected targets to identify abnormal event types such as driving in the wrong direction, intrusion, loitering, and leaving objects behind; feature extraction performing structured feature extraction on video frames to generate structured data containing target category, location, size, and confidence level; and storing the recognition results in the recognition result database after associating the recognition results with camera identifiers, timestamps, and recognition types.

6. The multi-camera differentiated polling video recognition method based on scene hierarchical and dynamic weights according to claim 1, characterized in that, The specific method of adaptive feedback update in step 6 includes: the system continuously monitors the recognition accuracy, false negative rate, false positive rate, and response delay operation status indicators; calculates the correction coefficient of scene classification based on the operation status indicators; automatically increases the scene priority score of the corresponding camera when the recognition accuracy decreases, and appropriately decreases the weight value when the false positive rate increases; optimizes the weight adjustment strategy using reinforcement learning algorithm, with recognition accuracy as a positive reward and false positive rate as a negative penalty, and calculates the optimal weight adjustment parameters; the corrected scene classification and weight parameters take effect in the next polling cycle, forming a closed-loop feedback mechanism for continuous optimization.

7. The multi-camera differentiated polling video recognition method based on scene hierarchical and dynamic weights according to claim 1, characterized in that, It also includes a recognition result storage and query module. The recognition result storage module uses a time-series database to store recognition result data, retains historical records for a preset time period, and supports multi-condition combined queries based on camera identifier, time range, and recognition type. The query results return recognition statistical summary information for the corresponding time period.

8. The multi-camera differentiated polling video recognition method based on scene classification and dynamic weights according to claim 1, characterized in that, It also includes an abnormal event alarm module, which generates corresponding alarm information based on the abnormal event type and severity identified in step 5. The abnormal event types include safety accidents, group gatherings, abandoned items, and suspicious loitering. Alarm priority is calculated based on the severity of the anomaly and the camera weight. The more severe the anomaly, the higher the priority of the alarm. After a high-priority alarm is triggered, the corresponding camera's polling weight in the next round is automatically increased and the polling interval is shortened to achieve rapid response to key areas.

9. The multi-camera differentiated polling video recognition method based on scene hierarchical and dynamic weights according to claim 1, characterized in that, It also includes a dynamic computing resource allocation module, which dynamically adjusts the computing resource quota allocated to each camera video recognition task based on the computing load of the video recognition task and the system resource usage; the resource monitoring indicators include CPU utilization, memory usage and graphics processor load. The load balancing algorithm allocates resources based on the current task's computational requirements and resource utilization of each camera, ensuring that high-priority cameras receive sufficient computational resources. When the resource utilization exceeds a preset threshold, a resource reclamation mechanism is automatically triggered, prioritizing the reclamation of computational resources allocated to low-priority cameras and inactive tasks.

10. The multi-camera differentiated polling video recognition method based on scene classification and dynamic weights according to claim 1, characterized in that, It also includes a historical data statistics module, which uses a sliding window mechanism to count key indicators such as the recognition frequency, number of abnormal events, and false alarm rate of each camera within a preset statistical period. The statistical results are used to dynamically adjust the scene classification scores and weight parameters of each camera, and cameras with higher identification frequency are automatically given higher priority weights; historical data is also used to generate monitoring status reports for each area, providing data support for the optimized deployment of the video surveillance network.

Images