Distributed video monitoring integration method based on edge computing

Abstract

The invention discloses a distributed video monitoring integration method based on edge computing, and relates to the technical field of intelligent monitoring, and the method comprises the following steps: screening key nodes and related nodes, carrying out the first analysis and the second analysis, obtaining a time feature and a space feature, and further obtaining a first accident degree and an accident level; and constructing a first influence model and sending a first early warning signal. According to the invention, through two-stage topology of key nodes and related nodes mined by historical data, the AI reasoning model, the cache capacity and the communication bandwidth can be pre-deployed to a potential accident area in advance, non-differentiated redundancy calculation of the whole network and direct embedding of a mapping model from a first accident degree to an accident level into an edge ARM-NPU are avoided, local closed-loop decision is realized, and the reliability of the whole network is improved. And the edge node regularly returns a new accident sample to the center for incremental training, and the weight of the key node is dynamically updated once every 24 hours, so that the same-frequency resonance of the model and the scene is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent monitoring, and in particular to a distributed video monitoring integration method based on edge computing. Background Art

[0002] In recent years, distributed video surveillance integration technology has been rapidly evolving towards edge intelligence, system collaboration, data fusion, and active defense. It has formed a relatively complete technical system and industrial application path, integrating video, fiber optic vibration sensing, and millimeter-wave radar to achieve high-precision intrusion detection in rain, snow, fog, and haze, build virtual monitoring scenes, and overlay environmental parameters such as temperature and electromagnetic fields in real time to form three-dimensional situational awareness. NPU / AI chips are integrated in cameras or edge nodes to realize local face detection, behavior recognition, and other calculations. Only structured data is uploaded, significantly reducing latency.

[0003] Currently, a Chinese invention patent with publication number CN119963823A discloses a real-time integrated multi-video stream multi-level analysis and monitoring system. This method introduces a warning area demarcation module and a coupling analysis module, enabling the system to focus on monitoring and analyzing key areas, improving the accuracy and efficiency of anomaly detection. Secondly, the Kalman filter algorithm is used to construct a time series model to achieve effective coupling analysis of micro-motion characteristics and background changes, thereby more accurately assessing target abnormal behavior. Through a multi-level priority alarm mechanism, the system can automatically adjust the alarm strategy based on different abnormal behaviors, improving the intelligence level of the monitoring system. However, the related technology does not perform multi-level screening of each road section based on the correlation order of time and space, which is not conducive to the rapid identification of the main contradictions. It is also not conducive to the simplicity of data during edge computing. The current road data is not planned and warned based on historical data, which is not conducive to driving safety and the foresight of traffic restrictions, and has certain limitations. Summary of the Invention

[0004] The technical problem solved by the present invention is that the relevant technology does not perform multi-level screening of each road section according to the correlation order of time and space, which is not conducive to the rapid identification of the main contradictions, and is not conducive to the simplicity of data during edge computing. It does not plan and warn the current road data based on historical data, which is not conducive to driving safety and the foresight of traffic restrictions, and has certain limitations.

[0005] To solve the above technical problems, the present invention provides the following technical solution: a distributed video surveillance integration method based on edge computing, comprising the following steps: Step S100, screening key nodes and related nodes based on historical monitoring data, performing a first analysis and a second analysis on the historical monitoring data of the key nodes and the historical monitoring data of the related nodes to obtain temporal features and spatial features; Step S200, obtaining a first accident severity based on the time characteristics and the spatial characteristics, and obtaining an accident grade based on the first accident severity; Step S300: Perform a third analysis on the historical monitoring data and the accident level to obtain a first impact model, and send a first warning signal according to the first impact model.

[0006] As a preferred solution of the distributed video surveillance integration method based on edge computing described in the present invention, wherein: the historical monitoring data includes historical time, historical natural data and historical road data; The historical time is expressed as a historical date; The historical natural data include visibility and ground friction; The historical road data includes historical road sections, historical traffic flow, historical average speed, historical minimum vehicle spacing and corresponding historical license plate numbers. The historical traffic flow is represented by the traffic flow within the first historical time period, the historical average speed is represented by the average of the average speeds of each vehicle on the road within the first historical time period, the historical minimum vehicle spacing is represented by the minimum value of the distance between the front and rear of any adjacent vehicles in the same lane on the road within the first historical time period, and the historical license plate numbers are represented by the license plate numbers of two or more vehicles belonging to the historical minimum vehicle spacing.

[0007] As a preferred solution of the distributed video surveillance integration method based on edge computing described in the present invention, before obtaining historical road data, a road segment cutting method is configured, and the road segment cutting method includes: Obtain the historical road length, set the first length as the segmentation distance, segment the historical road according to the first length to obtain N historical road segments, and number the historical road segments from either end point of the road. The numbers are natural numbers and are distributed from 1 to N.

[0008] As a preferred solution of the distributed video surveillance integration method based on edge computing described in the present invention, the method of screening key nodes and related nodes based on historical monitoring data includes: Select any historical road section, obtain the historical minimum vehicle spacing corresponding to the historical road section, perform a first comparison between the historical minimum vehicle spacing and 0, and when the historical minimum vehicle spacing is greater than 0, jump to the next first historical time period of the historical road section and repeat the first comparison process until the historical minimum vehicle spacing is less than or equal to 0, obtain the corresponding historical license plate number, count a first number of historical license plate numbers, traverse the first number of the historical road section in the first historical time period in a first period, and calculate a first sum value of the first number, where the first period is represented as one year; Traverse each historical road section, calculate the first sum value of each historical road section, sort each of the sum values ​​in descending order, select the top three first sum values, obtain the historical road sections corresponding to the top three first sum values, set the historical road sections corresponding to the top three first sum values ​​as key nodes, and set the previous historical road section of the key node as a related node in order of numbering; The key nodes represent the top three road sections with the highest number of traffic accidents in the first historical time period within the first cycle, and the related nodes represent the road sections before the key nodes where attention should be paid to controlling road data.

[0009] As a preferred solution of the distributed video surveillance integration method based on edge computing described in the present invention, wherein: the time feature includes a first change feature and a second change feature, and a first change feature of a first quantity of key nodes in each first historical time period within a first cycle, and a second change feature of the first quantity of key nodes in each first historical time period on the same day are obtained through a first analysis; Performing a first analysis and a second analysis on the historical monitoring data of key nodes and related nodes to obtain temporal features and spatial features; Select any key node and any historical time period, obtain the first number of the key node in the same first historical time period within the first cycle, set the first value as the first number threshold, the first value is expressed as the average value of the first number of the historical time period, compare each first number with the first value, when the first number is greater than or equal to the first value, retain the corresponding date, when the first number is less than the first value, delete the corresponding date, traverse the retained dates corresponding to each historical time period, count the number of times the retained date appears, sort the number of times the retained date appears in descending order, set the first number as the number threshold, compare the number of times the retained date appears with the first number, when the number of times the retained date appears is greater than or equal to the first number, set the retained date as an accident-prone date, and set the accident-prone date as the first change feature; When the number of times a reserved date appears is less than the first number, the reserved date is deleted and the next reserved date is jumped to.

[0010] As a preferred solution of the distributed video surveillance integration method based on edge computing described in the present invention, the first quantity of the key nodes in each first historical time period on the same day is obtained, the first difference of adjacent first quantities is calculated in chronological order, each first difference is traversed, and the first historical time period corresponding to the first difference with the largest value is selected, the first historical time period corresponding to each first difference with the largest value in the first cycle is traversed, the number of occurrences of each first historical time period is counted, the number of occurrences of each first historical time period is sorted in descending order, the first historical time period corresponding to the largest number of occurrences is selected, recorded as the accident-prone time period, and the accident-prone time period is set as the second change feature.

[0011] As a preferred solution of the distributed video surveillance integration method based on edge computing described in the present invention, wherein: the third change feature between the first number of relevant nodes and the first number of key nodes in different first historical time periods on different days in the first cycle is obtained through the second analysis; Obtain any time point in the first cycle, obtain the first number of any key node and the first number of corresponding related nodes, calculate the second difference between the first number of key nodes and the first number of corresponding related nodes, traverse each second difference in any day, calculate the average value of the second difference, traverse the average value of the second difference of each day in the first cycle, calculate the average value of the average value of the second difference, record the average value of the average value of the second difference as the number of associated accidents, and set the number of associated accidents as the third change feature.

[0012] As a preferred solution of the distributed video surveillance integration method based on edge computing described in the present invention, wherein: the accident level is obtained according to the first accident degree, and the first demarcation value is obtained according to the historical monitoring data; The accident levels include the first, second, and third levels, where the severity of the accident represented by the first, second, and third levels gradually increases. The method for obtaining the first accident level based on time characteristics and spatial characteristics includes: Counting the number of days of the first change feature as a second quantity, counting the total duration of the second change feature, converting the total duration into days as a third quantity, recording the number of accidents associated with the third associated quantity as a fourth quantity, calculating a second sum of the second quantity, the third quantity, and the fourth quantity, and setting the second sum as the first accident degree; Setting the second value and the third value as a second sum threshold, wherein the second value is less than the third value, and comparing the second sum with the second sum threshold; When the second sum is less than or equal to the second value, the accident level is set to the first level; when the second sum is greater than the second value and less than or equal to the third value, the accident level is set to the second level; when the second sum is greater than the third value, the accident level is set to the third level.

[0013] As a preferred solution of the distributed video surveillance integration method based on edge computing described in the present invention, the method for constructing the first impact model includes: Obtain the visibility, ground friction, historical traffic volume, historical average speed corresponding to any key node and the historical average speed of the corresponding related nodes, take the first accident degree as the dependent variable, and visibility, ground friction, historical traffic volume and historical average speed as independent variables, conduct a multiple regression analysis, and obtain the first impact model.

[0014] As a preferred solution of the distributed video surveillance integration method based on edge computing described in the present invention, the method of sending the first warning signal according to the first impact model includes: Obtaining the current road section on which any vehicle is traveling. If the current road section does not belong to a key node, no first warning signal is sent. If the current road section belongs to a key node, a first warning signal is sent. The first warning signal includes a final accident probability and a warning decision. Obtaining current visibility, current ground friction, current traffic volume, current average speed, and current average speed of relevant nodes to obtain a current first accident severity, obtaining a current accident level based on the current first accident severity, obtaining an accident level corresponding to a key node, subtracting an upper limit of the first accident severity corresponding to the accident level from the first accident severity to obtain a third difference, calculating a ratio of the third difference to an upper limit of the underestimated level, and setting the ratio of the third difference to the upper limit of the underestimated level as the accident probability; Get the current date and the first current time period. If the current date is a time feature or the first current time period, weight the accident probability with 0.1. If the current date is a time feature and the first current time period, weight the accident probability with 0.2 to obtain the final accident probability. The early warning decision includes: setting the expected accident level, obtaining expected parameters based on the first impact model, the expected parameters including expected traffic flow, expected average speed and the expected average speed of the corresponding related nodes, and according to the expected parameters, not restricting or limiting the current traffic flow, not restricting or reducing the current average speed, and not restricting or reducing the current average speed of the related nodes until the expected parameters are met, wherein the traffic flow of the key node or the related node is limited when it is greater than the expected average speed, and is not restricted when the current traffic flow is less than or equal to the expected average speed.

[0015] The beneficial effects of the present invention are as follows: through the two-level topology of key nodes and related nodes mined from historical data, AI reasoning models, cache capacity and communication bandwidth can be pre-deployed to potential accident areas in advance to avoid indiscriminate redundant calculations across the entire network. The first mapping model from accident degree to accident level is directly embedded in the edge ARM-NPU to achieve local closed-loop decision-making without the need to transmit back to the cloud. The edge nodes regularly transmit new accident samples back to the center for incremental training, and the weights of key nodes are dynamically updated every 24 hours to achieve synchronous resonance between the model and the scenario. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 A schematic diagram of the basic flow of a distributed video surveillance integration method based on edge computing provided by one embodiment of the present invention. DETAILED DESCRIPTION

[0017] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the drawings. It is obvious that the described embodiments are only part of the embodiments of the present invention, but not all of the embodiments.

[0018] Example, see Figure 1 , as an embodiment of the present invention, provides a distributed video surveillance integration method based on edge computing, comprising the following steps: Step S100, screening key nodes and related nodes based on historical monitoring data, performing a first analysis and a second analysis on the historical monitoring data of the key nodes and the historical monitoring data of the related nodes to obtain temporal features and spatial features; Step S200, obtaining a first accident severity based on the time characteristics and the spatial characteristics, and obtaining an accident grade based on the first accident severity; Step S300: Perform a third analysis on the historical monitoring data and the accident level to obtain a first impact model, and send a first warning signal according to the first impact model.

[0019] The present invention uses the two-level topology of key nodes and related nodes mined from historical data to pre-deploy AI reasoning models, cache capacity and communication bandwidth to potential accident areas in advance, avoiding indiscriminate redundant calculations across the entire network. The first mapping model from accident severity to accident level is directly embedded in the edge ARM-NPU to achieve local closed-loop decision-making without the need to transmit back to the cloud. The edge nodes regularly transmit new accident samples back to the center for incremental training, and the weights of key nodes are dynamically updated every 24 hours to achieve synchronous resonance between the model and the scenario.

[0020] The historical monitoring data includes historical time, historical natural data and historical road data; The historical time is expressed as a historical date; The historical natural data include visibility and ground friction; The historical road data includes historical road sections, historical traffic flow, historical average speed, historical minimum vehicle spacing and corresponding historical license plate numbers. The historical traffic flow is represented by the traffic flow within the first historical time period, the historical average speed is represented by the average of the average speeds of each vehicle on the road within the first historical time period, the historical minimum vehicle spacing is represented by the minimum value of the distance between the front and rear of any adjacent vehicles in the same lane on the road within the first historical time period, and the historical license plate numbers are represented by the license plate numbers of two or more vehicles belonging to the historical minimum vehicle spacing.

[0021] In specific implementation, by integrating historical time, natural environment and road data, this method has achieved significant improvements in accident prediction accuracy, resource scheduling efficiency and system robustness. Three service areas on the expressway (covering a total of 120km of road sections, including 2 tunnels and 1 viaduct) and 18 months of historical data from January 2022 to June 2024 were used, including 12% of holidays (Spring Festival and National Day) and 88% of weekdays, visibility (15% of foggy days, with a minimum of 50m), ground friction (slippery coefficient μ=0.35 on rainy days and μ=0.25 on snowy days), traffic flow peak (2,400 on Spring Festival vehicles / hour, 600 vehicles / hour on weekdays), historical average vehicle speed (110km / h on sunny days, 85km / h on rainy days, 60km / h on snowy days), and Jetson AGX Orin edge devices (275TOPS computing power) are deployed at key nodes (such as tunnel entrances). Through historical data training, the system only needs to enable full AI detection for the top 20% high-accident risk periods (such as foggy days on holidays), and reduce the frame rate to 15fps (saving 60% computing power) for the rest of the time. The average computing delay is reduced from 1.2 seconds to 0.5 seconds, and the amount of data transmitted back to the edge and cloud is compressed from 8Mbps to 4.4Mbps.

[0022] Before acquiring historical road data, a road segment cutting method is configured, which includes: Obtain the historical road length, set the first length as the segmentation distance, segment the historical road according to the first length to obtain N historical road segments, and number the historical road segments from either end point of the road. The numbers are natural numbers and are distributed from 1 to N.

[0023] In specific implementation, by evenly dividing historical roads into N numbered sections (1 to N) according to a fixed first length L, this method achieves quantifiable improvements in data alignment accuracy, computational efficiency, and anomaly location granularity.

[0024] Methods for screening key nodes and related nodes based on historical monitoring data include: Select any historical road section, obtain the historical minimum vehicle spacing corresponding to the historical road section, perform a first comparison between the historical minimum vehicle spacing and 0, and when the historical minimum vehicle spacing is greater than 0, jump to the next first historical time period of the historical road section and repeat the first comparison process until the historical minimum vehicle spacing is less than or equal to 0, obtain the corresponding historical license plate number, count a first number of historical license plate numbers, traverse the first number of the historical road section in the first historical time period in a first period, and calculate a first sum value of the first number, where the first period is represented as one year; Traverse each historical road section, calculate the first sum value of each historical road section, sort each of the sum values ​​in descending order, select the top three first sum values, obtain the historical road sections corresponding to the top three first sum values, set the historical road sections corresponding to the top three first sum values ​​as key nodes, and set the previous historical road section of the key node as a related node in order of numbering; The key nodes represent the top three road sections with the highest number of traffic accidents in the first historical time period within the first cycle, and the related nodes represent the road sections before the key nodes where attention should be paid to controlling road data.

[0025] In specific implementation, through the screening logic of minimum spacing ≤ 0, license plate counting, annual sum value sorting, the first three sections as key nodes, and their predecessors as related nodes, this method has achieved significant improvements in the accuracy of high-incidence accident identification, preventive deployment efficiency, and system response timeliness. The data of the two-way six-lane highway from 2022 to 2023 (total length of 240 km, divided into 1,200 200 m road sections, numbered 1-1,200). A total of 2,847 traffic accidents were recorded during the year. License plate capture was triggered by "minimum distance ≤ 0", and a total of 5,632 high-risk events (including rear-end collisions, scratches, and sudden stops) were marked. Sorted by annual sum value, the top three sections were **#487, #712, and #954** (sum values ​​of 142, 137, and 129 license plate triggers, respectively). #487 had 138 actual accidents (coverage rate 100%), #712 had 132 actual accidents (coverage rate 100%), and #954 had 127 actual accidents (coverage rate 100%). The false positive rate was 0% (no non-accident section was included in the top 3). Radar speed measurement + LED warning lights were deployed (500 m advance warning). The dynamic speed limit was reduced from 120 km / h to 80 km / h, section #487 (key node): accidents dropped from 21 / quarter to 12 (a decrease of 43%), section #486 (relevant node): accidents dropped from 9 / quarter to 3 (a decrease of 67%), verifying the effectiveness of "preemptive intervention".

[0026] The time feature includes a first change feature and a second change feature, wherein the first change feature of the first quantity of the key nodes in each first historical time period in the first cycle and the second change feature of the first quantity of the key nodes in each first historical time period on the same day are obtained through the first analysis; Performing a first analysis and a second analysis on the historical monitoring data of key nodes and related nodes to obtain temporal features and spatial features; Select any key node and any historical time period, obtain the first number of the key node in the same first historical time period within the first cycle, set the first value as the first number threshold, the first value is expressed as the average value of the first number of the historical time period, compare each first number with the first value, when the first number is greater than or equal to the first value, retain the corresponding date, when the first number is less than the first value, delete the corresponding date, traverse the retained dates corresponding to each historical time period, count the number of times the retained date appears, sort the number of times the retained date appears in descending order, set the first number as the number threshold, compare the number of times the retained date appears with the first number, when the number of times the retained date appears is greater than or equal to the first number, set the retained date as an accident-prone date, and set the accident-prone date as the first change feature; When the number of times a reserved date appears is less than the first number, the reserved date is deleted and the next reserved date is jumped to.

[0027] In the specific implementation, through the closed-loop process of annual and daily double threshold screening, accident-prone date locking, first change feature generation, this method achieves quantifiable benefits in three aspects: mining accident periodicity patterns, calendaring preventive deployment, and time-sharing reuse of system resources. The first historical time period: each day is divided into 96 15-minute segments (00:00~23:45), the first quantity threshold: the average value of the same segment of the year (such as 3.2 incidents / year for the 07:30-07:45 segment), the first number threshold: the top-20% percentile of the number of occurrences of the retained date (such as ≥9 times per year), accident-prone date locking: a total of 36 days will be locked in 2023 (accounting for 9.9% of the whole year), including: the third day of the Spring Festival travel rush (January 20): 7 accidents in the 07:30-07:45 segment (> average 3.2), 11 accidents in the year (> threshold 9 times), National Day return day (October 5th): 9 clip accidents between 5:00 PM and 5:15 PM (>2.8 average), 13 times per year, 35 of 36 days with actual accidents (97.2% accuracy), and 100% hit rate for the top three dates (January 20th, October 5th, and December 31st). The traditional model: additional police officers were deployed throughout holidays (an average of 120 people per day). This method: additional police officers were deployed only on accident-prone dates and time periods (an average of 40 people per day). Incident response time: reduced from an average of 14 minutes to 4.5 minutes (a 68% reduction). Annual computing power utilization: increased from 58% to 89% (saving 1,200 kWh of GPU power per year). Storage costs: the video retention period was compressed from 30 days to 7 days during accident-prone periods and 3 days during other periods, saving 45 TB of storage per year. Obtain the first quantity of the key node in each first historical time period on the same day, calculate the first difference of adjacent first quantities in chronological order, traverse each first difference, select the first historical time period corresponding to the first difference with the largest value, traverse the first historical time period corresponding to each first difference with the largest value in the first cycle, count the number of occurrences of each first historical time period, sort the number of occurrences of each first historical time period in descending order, select the first historical time period corresponding to the largest number of occurrences, record it as the accident-prone time period, and set the accident-prone time period as the second change feature.

[0028] In the specific implementation, through the mechanism of maximum mutation of the difference between adjacent time periods within the day, periodic statistics, and locking of accident-prone time periods, this method realizes the sinking of precise time features from macro dates to micro periods, and obtains quantifiable benefits in three aspects: time period-level early warning deployment, sudden accident interception, and time-sharing release of edge computing power. The difference (Δ) in the number of accidents between each time period and the previous time period is calculated, and the time period with the maximum positive difference (i.e., the time period with the fastest increase in the number of accidents) is extracted. For 365 days a year, a total of 365 "maximum difference time periods" are extracted, and their occurrence frequencies are counted. The most frequent time periods are: 07:45–08:00 (occurred 42 times), 17:30–17 From 12:45 (occurred 38 times) to 12:15–12:30 (occurred 31 times), the system deployed edge AI (Jetson Orin NX) to calculate the "mutation threshold" of Δ>3 / 15 minutes in real time. Once triggered, linkage was immediately implemented: the upstream information board 2 km away switched to "accident warning", the AutoNavi / Baidu Map API pushed the "red congestion" prompt, and the edge node locally broadcast "please slow down". During non-accident-prone periods (accounting for 82.5%), the inference frequency was reduced to 5 fps, and during accident-prone periods (07:45–08:00, 17:30–17:45, etc.), it maintained full speed at 30 fps.

[0029] The third change feature between the first number of related nodes and the first number of key nodes in different first historical time periods on different days in the first cycle is obtained through the second analysis; Obtain any time point in the first cycle, obtain the first number of any key node and the first number of corresponding related nodes, calculate the second difference between the first number of key nodes and the first number of corresponding related nodes, traverse each second difference in any day, calculate the average value of the second difference, traverse the average value of the second difference of each day in the first cycle, calculate the average value of the average value of the second difference, record the average value of the average value of the second difference as the number of associated accidents, and set the number of associated accidents as the third change feature.

[0030] In specific implementation, through the chain calculation of the number of accidents at key nodes, the number of accidents at related nodes, the intraday difference, and the periodic mean, the third change feature quantifies the spatiotemporal coupling intensity between key nodes and related nodes into a single numerical indicator (the number of related accidents).

[0031] Obtain an accident level according to the first accident degree, and obtain a first cutoff value according to historical monitoring data; The accident levels include the first, second, and third levels, where the severity of the accident represented by the first, second, and third levels gradually increases. The method for obtaining the first accident level based on time characteristics and spatial characteristics includes: Counting the number of days of the first change feature as a second quantity, counting the total duration of the second change feature, converting the total duration into days as a third quantity, recording the number of accidents associated with the third associated quantity as a fourth quantity, calculating a second sum of the second quantity, the third quantity, and the fourth quantity, and setting the second sum as the first accident degree; Setting the second value and the third value as a second sum threshold, wherein the second value is less than the third value, and comparing the second sum with the second sum threshold; When the second sum is less than or equal to the second value, the accident level is set to the first level; when the second sum is greater than the second value and less than or equal to the third value, the accident level is set to the second level; when the second sum is greater than the third value, the accident level is set to the third level.

[0032] The construction method of the first impact model includes: Obtain the visibility, ground friction, historical traffic volume, historical average speed corresponding to any key node and the historical average speed of the corresponding related nodes, take the first accident degree as the dependent variable, and visibility, ground friction, historical traffic volume and historical average speed as independent variables, conduct a multiple regression analysis, and obtain the first impact model.

[0033] In the specific implementation, the second quantity (number of days prone to accidents): 36 days per year for key node #487, the third quantity (total duration of accident-prone periods): 57.6 hours per year, equivalent to 2.4 days, the fourth quantity (number of related accidents): 6.3 per year, the second sum value is 44.7, model construction: dependent variable: first accident degree (second sum value) independent variables: visibility (VIS, unit: m), ground friction (μ, dimensionless), historical traffic flow (FLOW, vehicles / hour), historical average speed (SPEED, km / h), speed difference of related nodes (ΔSPEED, km / h), second sum value .

[0034] The method for sending a first early warning signal according to the first impact model includes: Obtaining the current road section on which any vehicle is traveling. If the current road section does not belong to a key node, no first warning signal is sent. If the current road section belongs to a key node, a first warning signal is sent. The first warning signal includes a final accident probability and a warning decision. Obtaining current visibility, current ground friction, current traffic volume, current average speed, and current average speed of relevant nodes to obtain a current first accident severity, obtaining a current accident level based on the current first accident severity, obtaining an accident level corresponding to a key node, subtracting an upper limit of the first accident severity corresponding to the accident level from the first accident severity to obtain a third difference, calculating a ratio of the third difference to an upper limit of the underestimated level, and setting the ratio of the third difference to the upper limit of the underestimated level as the accident probability; Get the current date and the first current time period. If the current date is a time feature or the first current time period, weight the accident probability with 0.1. If the current date is a time feature and the first current time period, weight the accident probability with 0.2 to obtain the final accident probability. The early warning decision includes: setting the expected accident level, obtaining expected parameters based on the first impact model, the expected parameters including expected traffic flow, expected average speed and the expected average speed of the corresponding related nodes, and according to the expected parameters, not restricting or limiting the current traffic flow, not restricting or reducing the current average speed, and not restricting or reducing the current average speed of the related nodes until the expected parameters are met, wherein the traffic flow of the key node or the related node is limited when it is greater than the expected average speed, and is not restricted when the current traffic flow is less than or equal to the expected average speed.

[0035] The present invention uses the two-level topology of key nodes and related nodes mined from historical data to pre-deploy AI reasoning models, cache capacity and communication bandwidth to potential accident areas in advance, avoiding indiscriminate redundant calculations across the entire network. The first mapping model from accident severity to accident level is directly embedded in the edge ARM-NPU to achieve local closed-loop decision-making without the need to transmit back to the cloud. The edge nodes regularly transmit new accident samples back to the center for incremental training, and the weights of key nodes are dynamically updated every 24 hours to achieve synchronous resonance between the model and the scenario.

[0036] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media containing computer-usable program code. The storage medium may be implemented by any type of volatile or non-volatile storage device, or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0037] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A distributed video surveillance integration method based on edge computing, characterized in that: The following steps are involved: Step S100, screening key nodes and related nodes based on historical monitoring data, performing a first analysis and a second analysis on the historical monitoring data of the key nodes and the historical monitoring data of the related nodes to obtain temporal features and spatial features; Step S200, obtaining a first accident severity based on the time characteristics and the spatial characteristics, and obtaining an accident grade based on the first accident severity; Step S300: Perform a third analysis on the historical monitoring data and the accident level to obtain a first impact model, and send a first warning signal according to the first impact model.

2. The distributed video surveillance integration method based on edge computing according to claim 1, characterized in that: The historical monitoring data includes historical time, historical natural data and historical road data; The historical time is expressed as a historical date; The historical natural data include visibility and ground friction; The historical road data includes historical road sections, historical traffic flow, historical average speed, historical minimum vehicle spacing and corresponding historical license plate numbers. The historical traffic flow is represented by the traffic flow within the first historical time period, the historical average speed is represented by the average of the average speeds of each vehicle on the road within the first historical time period, the historical minimum vehicle spacing is represented by the minimum value of the distance between the front and rear of any adjacent vehicles in the same lane on the road within the first historical time period, and the historical license plate numbers are represented by the license plate numbers of two or more vehicles belonging to the historical minimum vehicle spacing.

3. The distributed video surveillance integration method based on edge computing according to claim 2, characterized in that: Before acquiring historical road data, a road segment cutting method is configured, which includes: Obtain the historical road length, set the first length as the segmentation distance, segment the historical road according to the first length to obtain N historical road segments, and number the historical road segments from either end point of the road. The numbers are natural numbers and are distributed from 1 to N.

4. The distributed video surveillance integration method based on edge computing according to claim 1, characterized in that: Methods for screening key nodes and related nodes based on historical monitoring data include: Select any historical road section, obtain the historical minimum vehicle spacing corresponding to the historical road section, perform a first comparison between the historical minimum vehicle spacing and 0, and when the historical minimum vehicle spacing is greater than 0, jump to the next first historical time period of the historical road section and repeat the first comparison process until the historical minimum vehicle spacing is less than or equal to 0, obtain the corresponding historical license plate number, count a first number of historical license plate numbers, traverse the first number of the historical road section in the first historical time period in a first period, and calculate a first sum value of the first number, where the first period is represented as one year; Traverse each historical road section, calculate the first sum value of each historical road section, sort each of the sum values ​​in descending order, select the top three first sum values, obtain the historical road sections corresponding to the top three first sum values, set the historical road sections corresponding to the top three first sum values ​​as key nodes, and set the previous historical road section of the key node as a related node in order of numbering; The key nodes represent the top three road sections with the highest number of traffic accidents in the first historical time period within the first cycle, and the related nodes represent the road sections before the key nodes where attention should be paid to controlling road data.

5. The distributed video surveillance integration method based on edge computing according to claim 1, characterized in that: The time feature includes a first change feature and a second change feature, wherein the first change feature of the first quantity of the key nodes in each first historical time period in the first cycle and the second change feature of the first quantity of the key nodes in each first historical time period on the same day are obtained through the first analysis; Performing a first analysis and a second analysis on the historical monitoring data of key nodes and related nodes to obtain temporal features and spatial features; Select any key node and any historical time period, obtain the first number of the key node in the same first historical time period within the first cycle, set the first value as the first number threshold, the first value is expressed as the average value of the first number of the historical time period, compare each first number with the first value, when the first number is greater than or equal to the first value, retain the corresponding date, when the first number is less than the first value, delete the corresponding date, traverse the retained dates corresponding to each historical time period, count the number of times the retained date appears, sort the number of times the retained date appears in descending order, set the first number as the number threshold, compare the number of times the retained date appears with the first number, when the number of times the retained date appears is greater than or equal to the first number, set the retained date as an accident-prone date, and set the accident-prone date as the first change feature; When the number of times a reserved date appears is less than the first number, the reserved date is deleted and the next reserved date is jumped to.

6. The distributed video surveillance integration method based on edge computing according to claim 5, characterized in that: Obtain the first quantity of the key node in each first historical time period on the same day, calculate the first difference of adjacent first quantities in chronological order, traverse each first difference, select the first historical time period corresponding to the first difference with the largest value, traverse the first historical time period corresponding to each first difference with the largest value in the first cycle, count the number of occurrences of each first historical time period, sort the number of occurrences of each first historical time period in descending order, select the first historical time period corresponding to the largest number of occurrences, record it as the accident-prone time period, and set the accident-prone time period as the second change feature.

7. The distributed video surveillance integration method based on edge computing according to claim 1, characterized in that: The third change feature between the first number of related nodes and the first number of key nodes in different first historical time periods on different days in the first cycle is obtained through the second analysis; Obtain any time point in the first cycle, obtain the first number of any key node and the first number of corresponding related nodes, calculate the second difference between the first number of key nodes and the first number of corresponding related nodes, traverse each second difference in any day, calculate the average value of the second difference, traverse the average value of the second difference of each day in the first cycle, calculate the average value of the average value of the second difference, record the average value of the average value of the second difference as the number of associated accidents, and set the number of associated accidents as the third change feature.

8. The distributed video surveillance integration method based on edge computing according to claim 1, characterized in that: Obtain an accident level according to the first accident degree, and obtain a first cutoff value according to historical monitoring data; The accident levels include the first, second, and third levels, where the severity of the accident represented by the first, second, and third levels gradually increases. The method for obtaining the first accident level based on time characteristics and spatial characteristics includes: Counting the number of days of the first change feature as a second quantity, counting the total duration of the second change feature, converting the total duration into days as a third quantity, recording the number of accidents associated with the third associated quantity as a fourth quantity, calculating a second sum of the second quantity, the third quantity, and the fourth quantity, and setting the second sum as the first accident degree; Setting the second value and the third value as a second sum threshold, wherein the second value is less than the third value, and comparing the second sum with the second sum threshold; When the second sum is less than or equal to the second value, the accident level is set to the first level; when the second sum is greater than the second value and less than or equal to the third value, the accident level is set to the second level; when the second sum is greater than the third value, the accident level is set to the third level.

9. The distributed video surveillance integration method based on edge computing according to claim 1, characterized in that: The construction method of the first impact model includes: Obtain the visibility, ground friction, historical traffic volume, historical average speed corresponding to any key node and the historical average speed of the corresponding related nodes, take the first accident degree as the dependent variable, and visibility, ground friction, historical traffic volume and historical average speed as independent variables, conduct a multiple regression analysis, and obtain the first impact model.

10. The distributed video surveillance integration method based on edge computing according to claim 9, characterized in that: The method for sending a first early warning signal according to the first impact model includes: Obtaining the current road section on which any vehicle is traveling. If the current road section does not belong to a key node, no first warning signal is sent. If the current road section belongs to a key node, a first warning signal is sent. The first warning signal includes a final accident probability and a warning decision. Obtaining current visibility, current ground friction, current traffic volume, current average speed, and current average speed of relevant nodes to obtain a current first accident severity, obtaining a current accident level based on the current first accident severity, obtaining an accident level corresponding to a key node, subtracting an upper limit of the first accident severity corresponding to the accident level from the first accident severity to obtain a third difference, calculating a ratio of the third difference to an upper limit of the underestimated level, and setting the ratio of the third difference to the upper limit of the underestimated level as the accident probability; Get the current date and the first current time period. If the current date is a time feature or the first current time period, weight the accident probability with 0.

1. If the current date is a time feature and the first current time period, weight the accident probability with 0.2 to obtain the final accident probability. The early warning decision includes: setting the expected accident level, obtaining expected parameters based on the first impact model, the expected parameters including expected traffic flow, expected average speed and the expected average speed of the corresponding related nodes, and according to the expected parameters, not restricting or limiting the current traffic flow, not restricting or reducing the current average speed, and not restricting or reducing the current average speed of the related nodes until the expected parameters are met, wherein the traffic flow of the key node or the related node is limited when it is greater than the expected average speed, and is not restricted when the current traffic flow is less than or equal to the expected average speed.

Images