Real-time video analysis system and method based on edge computing

Abstract

The present application relates to the technical field of image processing, in particular to a real-time video analysis system and method based on edge computing, the system comprising a central control service module, an edge computing module, an algorithm adaptation module, a node management optimization module and a network scheduling module.In the present application, by analyzing the video stream frame rate and processing delay variation amplitude, high load periods can be identified in real time and task allocation can be dynamically optimized, ensuring that the frame processing strategy matches the current load state, minimizing system delay, while ensuring the accuracy of video analysis, accurately identifying network influence periods and optimizing task scheduling according to network fluctuations, avoiding inaccurate analysis results or task execution delays due to network problems, and significantly improving the real-time performance and stability of the frequency analysis system by considering load changes and environmental factors, avoiding performance bottlenecks caused by fixed frame intervals or excessive dependence on central servers.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a real-time video analysis system and method based on edge computing. Background Technology

[0002] Image processing technology mainly involves methods and means for acquiring, analyzing, recognizing, and processing static images or dynamic video data. Its core aspects include image acquisition and preprocessing, feature extraction and representation, target detection and recognition, scene understanding and analysis, etc. In practical applications, it is widely used in multiple scenarios such as security monitoring, industrial automation, medical diagnosis, and traffic management. With the development of IoT and AI technologies, this technology field is gradually evolving towards higher real-time performance, stronger processing efficiency, and deeper levels of intelligence. Traditional real-time video analytics systems refer to system architectures that offload some computing tasks to edge nodes closer to the data source to alleviate pressure on the central server and reduce latency. These systems primarily address technical aspects such as high-concurrency processing capabilities, flexible algorithm deployment, and real-time analysis results in video data processing. Traditional solutions accomplish these tasks by centrally processing video streams on a single server, statically deploying algorithms with fixed configurations, and setting fixed frame intervals for image sampling. Some improved solutions attempt to achieve data offloading by introducing distributed computing architectures or by using plug-in structures for algorithm replacement and loading. However, most lack fine-grained management mechanisms for plug-in versions and support for online hot updates. Furthermore, they fail to dynamically adjust the image sampling frequency based on load conditions in frame processing strategies, resulting in dual limitations on processing performance and real-time performance.

[0003] Existing technologies for processing high-concurrency video streams rely on centralized servers for centralized video stream processing. Because the computational tasks cannot be flexibly adjusted according to the actual load, processing latency and real-time performance are affected. Particularly under conditions of significant load fluctuations, the frame processing frequency fails to adjust in a timely manner, making it difficult for the system to maintain stable processing power and low latency under high load. Fixed algorithm configurations and the lack of online hot-update capabilities result in poor algorithm flexibility, making it difficult to cope with complex scene changes and quickly switch algorithms according to different situations, even leading to a decrease in analysis accuracy. Especially under unstable network bandwidth conditions, the scheduling methods of existing technologies fail to consider the actual impact of network fluctuations on task execution, thus affecting the overall system performance and user experience. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing a real-time video analysis system and method based on edge computing.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: A real-time video analysis system based on edge computing includes:

[0006] The central control service module analyzes load fluctuations by comparing the changes in video stream frame rate and processing latency based on edge node operation data, filters out key time periods with high load, extracts critical conditions for task scheduling judgment, and obtains task allocation optimization marker values.

[0007] The edge computing module optimizes the marker value based on the task allocation, matches the video stream data and frame processing strategy for the corresponding time period, compares the frame processing delay point in the marked segment with the current load status, filters the frame processing instructions corresponding to the deviation rate, and obtains the frame processing offset control instruction group.

[0008] The algorithm adaptation module extracts the analysis accuracy value of the corresponding node before and after the adjustment instruction according to the frame processing offset control instruction group, combines the target detection accuracy path under the unit frame with the frame processing trend identification, extracts the change segment to form the response interval, and obtains the node unit adaptation algorithm trend.

[0009] The node management optimization module calls the node unit to adapt to the algorithm trend, collects external environment data of the edge nodes, analyzes the overlap duration between the edge computing optimization algorithm change and the network bandwidth jump period, and outputs the duration of the network impact.

[0010] As a further aspect of the present invention, the task allocation optimization flag value includes frame rate fluctuation threshold, processing latency change range, and load rise rate characteristics; the frame processing offset control instruction group includes frame processing deviation, latency change anomaly value, and load state offset rate; the node unit adaptation algorithm trend includes analysis accuracy change range, target detection accuracy path type, and algorithm response amplitude; and the network impact duration includes bandwidth fluctuation duration, latency jump overlap period, and network disturbance duration.

[0011] As a further aspect of the present invention, the central control service module includes:

[0012] The frame rate fluctuation extraction submodule extracts the video stream time series based on the edge node running data, calculates the frame rate difference between adjacent sampling points, determines the fluctuation period and extracts the peak-to-trough distance, filters out the segments with period differences exceeding the threshold, and obtains the frame rate period fluctuation interval value.

[0013] The latency change calculation submodule calls the frame rate periodic fluctuation interval value, identifies the corresponding latency data, analyzes the absolute latency change amplitude of adjacent time periods, compares it with the set latency amplitude threshold, locates the abrupt change node, and obtains the latency abrupt change amplitude interval value.

[0014] The task status marking submodule identifies the corresponding load and algorithm complexity data based on the latency mutation amplitude range value, extracts the load and algorithm data, calculates the load complexity offset value by combining latency mutation and frame rate fluctuation, sets the offset threshold, marks the time segment exceeding the threshold, and obtains the task allocation optimization marking value.

[0015] As a further aspect of the present invention, the edge computing module includes:

[0016] The frame processing data matching submodule extracts the video stream frame rate and latency data for the corresponding time period based on the task allocation optimization mark value, calculates the latency fluctuation amplitude according to the sampling interval, and aligns the frame rate and latency amplitude according to the same time to obtain the frame latency linkage interval group.

[0017] The frame processing deviation judgment submodule calls the frame delay linkage interval group, extracts the load state sequence, compares the frame processing changes with the state differences, normalizes and compares with the preset offset boundary, calculates the cumulative value of the deviation intensity of the frame processing point, extracts the frame processing position index with the deviation intensity greater than the benchmark judgment value, and establishes the offset intensity index group.

[0018] The frame processing instruction extraction submodule, based on the offset intensity index group, filters the position of the corresponding time point in the task instruction set, extracts the instruction value, sorts it according to the time sequence, removes duplicate instructions, and obtains the frame processing offset control instruction group.

[0019] As a further aspect of the present invention, the algorithm adaptation module includes:

[0020] The accuracy extraction submodule extracts the frame processing accuracy data of the node in the loop before and after adjustment according to the frame processing offset control instruction group, identifies the analysis boundary conditions, and obtains the analysis accuracy difference value.

[0021] The target detection accuracy recognition submodule calls the analysis accuracy difference value to identify the target detection accuracy change trajectory in the unit frame path, judges the target quantity trend corresponding to the accuracy fluctuation, compares the accuracy decrease and target quantity fluctuation in adjacent time periods, calculates the accuracy concentration of the segment in the unit frame path, filters synchronous fluctuation segments, and obtains the target detection accuracy feature segment.

[0022] The algorithm response interval identification submodule analyzes the time and accuracy displacement trend based on the target detection accuracy feature segment, judges the consistency of direction and amplitude change characteristics, filters fluctuation segments and aggregates them to obtain the node unit adaptation algorithm trend.

[0023] As a further aspect of the present invention, the node management optimization module includes:

[0024] The algorithm trend extraction submodule calls the node unit to adapt to the algorithm trend, extracts node task records and periodic algorithm data, identifies the fluctuation amplitude and fluctuation frequency within a single period, and obtains the algorithm change trend value.

[0025] The network leap identification submodule, based on the trend value of the algorithm, selects bandwidth and latency records within the same period according to the external network data of the edge nodes, compares the data change magnitude on a daily basis, filters time nodes with a magnitude greater than the network leap threshold, and obtains the bandwidth and latency leap period.

[0026] The network impact quantification submodule identifies the intersection duration of the algorithm amplitude and the network amplitude based on the bandwidth delay jump period, performs weighted processing, and normalizes the usage duration under the differentiated period by referring to the amplitude frequency, the combined amplitude value and the periodic fluctuation, and outputs the network impact duration.

[0027] As a further aspect of the present invention, the system also includes a network scheduling module:

[0028] Based on the duration of network impact, the network scheduling module filters task scheduling instructions affected by network interference, categorizes load switching time and adjustment trigger frequency, identifies periods of excessive load jumps, and obtains the frequency of load interference impact on video analysis tasks.

[0029] The frequency of video analysis task load interference includes load switching frequency, number of adjustment command triggers, and number of times the jump time period exceeds the limit.

[0030] As a further aspect of the present invention, the network scheduling module includes:

[0031] The instruction filtering submodule filters matching task scheduling instructions based on the duration of network impact, identifies load periods and load states, compares disturbance periods with instruction periods, eliminates low-matching instructions, and obtains a set of instructions affected by network interference.

[0032] The load classification submodule calls the set of instructions affected by network interference, extracts the load switching time and adjustment trigger frequency corresponding to the instructions, arranges them in order of switching time, classifies them by frequency range, statistically analyzes the correspondence between frequency bands and load time periods, and obtains the load adjustment frequency band distribution value.

[0033] The interference identification submodule collects the load jump amplitude and duration within the frequency band based on the load adjustment frequency band distribution value, determines whether the jump amplitude threshold is exceeded, identifies the frequency band interference intensity, sorts the interference intensity of all frequency bands, determines the frequency band and jump time period of the interference intensity, and obtains the frequency of load interference affecting the video analysis task.

[0034] The edge computing-based real-time video analysis method is executed based on the aforementioned edge computing-based real-time video analysis system and includes the following steps:

[0035] S1: Based on edge node running data, analyze frame rate fluctuations and processing latency changes, extract periods of sudden load peak changes and rapid changes in algorithm complexity, identify state switching points that synchronize latency and load changes, mark task levels, and generate task state switching recognition segments.

[0036] S2: Based on the task state switching identification segment, extract frame processing change data and latency change amplitude, compare frame processing abrupt changes with latency fluctuation amplitude, analyze the correspondence between frame processing changes and load response, and obtain the frame processing adjustment response trajectory segment.

[0037] S3: Based on the frame processing adjustment response trajectory segment, extract the change in accuracy value and target detection accuracy gradient before and after the response, analyze the target detection accuracy decrease and target number increase of the node, identify the accuracy decay path and compare it with the original loop data, filter the segments with accuracy decay characteristics, and generate an accuracy adaptation decay path distribution set.

[0038] S4: Based on the accuracy adaptation attenuation path distribution set, identify the external network data within the corresponding time period, analyze the overlapping time periods of bandwidth jump and accuracy attenuation path, extract the associated time periods, and obtain the interference segment of network parameters for frame processing adjustment.

[0039] S5: Based on the network parameter interference segment of the frame processing adjustment, extract the scheduling control strategy record within the time period, analyze the time interval of load switching and the adjustment trigger frequency, filter high-frequency adjustment strategy segments, and obtain the frequency of load interference impact of video analysis task.

[0040] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0041] This invention analyzes the changes in video stream frame rate and processing latency to identify high-load periods in real time and dynamically optimize task allocation. This ensures that the frame processing strategy matches the current load state, thereby adjusting the frame processing frequency to minimize system latency and guarantee the accuracy of video analysis. Furthermore, by combining external environmental data from edge nodes with network bandwidth changes, it accurately identifies periods affected by network fluctuations and optimizes task scheduling accordingly. This allows task load to be adjusted based on actual network conditions, avoiding inaccurate analysis results or task execution delays caused by network issues. By comprehensively considering load changes and environmental factors, the real-time performance and stability of the video analysis system are significantly improved, avoiding performance bottlenecks caused by fixed frame intervals or over-reliance on a central server. Attached Figure Description

[0042] Figure 1 This is a system flowchart of the present invention;

[0043] Figure 2 This is a flowchart of the central control service module of the present invention;

[0044] Figure 3 This is a flowchart of the edge computing module of the present invention;

[0045] Figure 4 The flowchart of the algorithm adaptation module of this invention is shown below;

[0046] Figure 5 This is a flowchart of the node management optimization module of the present invention;

[0047] Figure 6 This is a flowchart of the network scheduling module of the present invention. Detailed Implementation

[0048] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0049] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0050] Please see Figure 1 This invention provides a technical solution: a real-time video analysis system based on edge computing, comprising:

[0051] The central control service module analyzes load fluctuations by comparing the changes in video stream frame rate and processing latency based on edge node operation data, filters out key time periods with high load, extracts critical conditions for task scheduling judgment, and obtains task allocation optimization marker values.

[0052] The edge computing module optimizes the marker value based on task allocation, matches the video stream data and frame processing strategy for the corresponding time period, compares the frame processing delay point in the marked segment with the current load status, filters the frame processing instructions corresponding to the deviation rate, and obtains the frame processing offset control instruction group.

[0053] The algorithm adaptation module extracts the analysis accuracy value of the corresponding node before and after the adjustment instruction based on the frame processing offset control instruction group, and combines the frame processing trend to identify the target detection accuracy path under the unit frame, extracts the change segment to form the response interval, and obtains the node unit adaptation algorithm trend.

[0054] The node management optimization module calls the node unit to adapt to the algorithm trend, collects external environment data of edge nodes, analyzes the overlap time between the edge computing optimization algorithm change and the network bandwidth jump period, and outputs the network impact duration.

[0055] The network scheduling module filters task scheduling instructions affected by network interference based on the duration of network impact, categorizes load switching time and adjustment trigger frequency, identifies periods of excessive load fluctuation, and obtains the frequency of load interference impact on video analysis tasks.

[0056] The task allocation optimization marker values ​​include frame rate fluctuation threshold, processing latency change range, and load rise rate characteristics. The frame processing offset control instruction group includes frame processing deviation, latency change anomaly value, and load state offset rate. The node unit adaptation algorithm trend includes analysis accuracy change range, target detection accuracy path type, and algorithm response amplitude. The network impact duration includes bandwidth fluctuation duration, latency jump overlap period, and network disturbance duration. The video analysis task load interference impact frequency includes load switching frequency, adjustment instruction trigger count, and jump period exceeding limit count.

[0057] Please see Figure 2 The central control service module includes:

[0058] The frame rate fluctuation extraction submodule extracts the video stream time series based on the edge node running data, calculates the frame rate difference between adjacent sampling points, determines the fluctuation period and extracts the peak-to-trough distance, filters out the segments with period differences exceeding the threshold, and obtains the frame rate period fluctuation interval value.

[0059] Based on edge node operation data, the frame rate performance of the video stream is monitored and processed in real time. Frame rate sequences with different timestamps are obtained, such as FPS=[25, 26, 24, 23, 27, 28, 26, 25, 24, 22]. The difference between adjacent frame rates is calculated to obtain [-1, 2, -1, 4, 1, -2, -1, -1, -2]. Then, FFT is used to analyze its periodicity to identify a frame rate fluctuation period of about 2 seconds. The peaks and troughs within the period are extracted and their differences are calculated, such as 28−22=6 frames / second. The frame rate fluctuation threshold is set to 3 frames / second (based on historical data statistics and test verification). The segments exceeding the threshold are selected as segments with severe fluctuations. Finally, the frame rate period fluctuation interval values ​​are obtained as [1s, 3s] and [5s, 7s], which are used to identify performance anomalies.

[0060] The latency change calculation submodule calls the frame rate periodic fluctuation interval value, identifies the corresponding latency data, analyzes the absolute latency change amplitude of adjacent time periods, compares it with the set latency amplitude threshold, locates the abrupt change node, and obtains the latency abrupt change amplitude interval value.

[0061] The frame rate period fluctuation range value is called, such as the frame rate period fluctuation range values ​​[1s, 3s] and [5s, 7s] mentioned above. The corresponding latency data of the range is identified. For example, in the time period from 1s to 3s, the obtained latency data sequence is [50ms, 55ms, 60ms, 65ms, 70ms], and in the time period from 5s to 7s, the obtained latency data sequence is [40ms, 42ms, 45ms, 60ms, 80ms]. The absolute change in latency between adjacent time periods is analyzed. For example, for the time period from 1s to 3s... For the delayed data within a given timeframe, the absolute variation in delay between adjacent sampling points is |55-50|=5ms, |60-55|=5ms, |65-60|=5ms, and |70-65|=5ms. For the delayed data within a timeframe of 5s to 7s, the absolute variation in delay between adjacent sampling points is |42-40|=2ms, |45-42|=3ms, |60-45|=15ms, and |80-60|=20ms. This absolute variation is compared to a set delay amplitude threshold, which is used to identify abrupt changes in delay. For example... The latency amplitude threshold is set to 10ms. This threshold is based on statistics of latency fluctuation range under normal operating conditions. By collecting latency data from edge nodes under stable load for at least 24 hours, the average absolute deviation of latency fluctuation is calculated as the threshold, and appropriate adjustments are made to balance sensitivity and false alarm rate. In practical applications, after several months of operational data accumulation, it has been found that setting the latency amplitude threshold to 10ms can effectively identify more than 90% of latency mutation events while maintaining a low false alarm rate. Comparing the above latency absolute change amplitude sequence with the 10ms threshold, latency changes of 15ms and 20ms exceed the threshold in the time period of 5s to 7s. The mutation nodes exceeding the threshold are located, that is, the specific time points when the latency mutation occurs. For example, in the time period of 5s to 7s, the time points when the latency changes from 45ms to 60ms and from 60ms to 80ms are located as mutation nodes. For example, [5.5s, 6.5s] indicates that the latency mutation occurred in this time period, and the latency mutation amplitude range value [5.5s, 6.5s] is obtained.

[0062] The task status marking submodule identifies the corresponding load and algorithm complexity data based on the latency fluctuation range, extracts the load and algorithm data, and combines latency fluctuations and frame rate fluctuations using the following formula:

[0063] ;

[0064] Calculate the load complexity offset value, set the offset threshold, mark the time intervals that exceed the threshold, and obtain the task allocation optimization mark value;

[0065] in, This represents the load complexity offset value. For the first The rate of load increase at any given moment For the first The algorithm complexity value at time step. For the first The difference in frame rate fluctuation at any given moment. As a delay correction factor, This is the load complexity offset parameter value. Indicates the total number of moments;

[0066] Based on the delayed mutation magnitude range value, such as the delayed mutation magnitude range value obtained above. Identify the load data and algorithm complexity data corresponding to the interval, for example, in arrive Record the rate of load increase at the edge nodes within the specified time period. for The corresponding algorithm complexity value for and frame rate fluctuation difference for Frames per second, load complexity offset parameter for Among them, the rate of increase of load This is calculated by monitoring the changes in parameters such as CPU utilization and memory utilization at edge nodes per unit time. For example, if it takes 1 second for CPU utilization to increase from 50% to 60%, then... The algorithm complexity is 10% / second. The complexity of currently running video analytics algorithms (such as the object detection algorithm YOLOv5 or the image recognition algorithm ResNet) is quantified by evaluating metrics such as computational cost and memory usage. For example, the complexity of the YOLOv5 model can be quantified as follows: After normalization The value is 0.8, representing the frame rate fluctuation difference. This represents the output frame rate fluctuation range and the load complexity offset parameter. This reflects an experimental value representing adjustments to the load or algorithm made to maintain service quality under specific latency spikes. This value is initially set to 0.1 and is dynamically adjusted based on the effectiveness of adjustments in historical latency spike events. For example, if historical adjustments were ineffective, this value will be appropriately increased to give more weight to latency. This is a delay correction factor, set to a value of 0.6. The setting was achieved through experiments simulating real network environments, testing different... The calculated load complexity offset value is compared with the actual performance degradation, and the value selected is chosen so that the calculation result most accurately reflects the actual performance degradation. Value, as proven by experiments, When set to 0.6, the calculated load complexity offset value shows a high degree of consistency with the actual observed performance degradation trend, ensuring the accurate quantification of the formula's impact on latency, and the total number of time steps. This indicates the number of sampling points within a statistical time period, for example, here. Using formula Calculate the load complexity offset value , This represents the load complexity offset value, which quantifies the overall pressure exerted on the system by the current task by comprehensively considering the load escalation rate, algorithm complexity, frame rate fluctuations, and latency correction parameters. (The value in parentheses is...) The term indicates the first The impact of time-based load and algorithm complexity on frame rate fluctuations, among which and The product represents the overall computational burden of the current task. As a divisor, it represents the dilution effect of frame rate fluctuations on the workload. The term represents the effect of delay on the load complexity offset, where It is a delay correction factor used to adjust the delay offset parameter value. The weights in the total offset value enable the formula to more accurately reflect the impact of latency fluctuations on system load. The overall calculation logic of the formula is to combine the computational load of the task itself with the system's sensitivity to latency fluctuations to quantify the offset of resource demand. The larger the value, the greater the system load pressure. The specific calculation process is as follows:

[0067] ;

[0068] Calculated load complexity offset value An offset threshold is set to determine whether the load complexity offset has reached a level requiring task allocation optimization. The offset threshold is set to 0.3, and this threshold setting is based on the analysis of historical stable operating data, through analysis of normal load conditions. The statistical values ​​were used, and the 90th percentile was taken as the threshold. This threshold was then fine-tuned based on the false alarm rate and false negative rate during actual operation. Experiments verified that when the threshold was set to 0.3, it could effectively identify 85% of overload situations while avoiding unnecessary task scheduling overhead. The calculated load complexity offset value was then used. Compared to an offset threshold of 0.3, This indicates that the current load complexity offset exceeds the threshold, marking the time segment exceeding the threshold. The advantage of this formula lies in its ability to comprehensively and accurately quantify the overall load pressure on edge nodes in video analytics tasks by combining load rise rate, algorithm complexity, frame rate fluctuation, and latency correction factor, especially through the introduction of the latency correction factor. and load complexity offset parameter This allows the model to more sensitively capture the system performance degradation caused by latency fluctuations, thereby taking network instability factors into account in task allocation decisions and improving the accuracy of task allocation and the efficiency of system resource utilization.

[0069] Please see Figure 3 The edge computing module includes:

[0070] The frame processing data matching submodule optimizes the marker value based on task allocation, extracts the video stream frame rate and latency data for the corresponding time period, calculates the latency fluctuation amplitude according to the sampling interval, and aligns the frame rate and latency amplitude according to the same time to obtain the frame latency linkage interval group.

[0071] Based on the task allocation optimization marker value, such as the aforementioned task allocation optimization marker value [5.5s, 6.5s], the system extracts the video stream frame rate and latency data for the corresponding time period. For example, in the time period from 5.5s to 6.5s, the system extracts video stream frame rate data of [25, 23, 20, 18] frames / second and corresponding latency data of [70, 85, 100, 120] milliseconds. The system calculates the latency fluctuation amplitude according to the preset sampling interval. For example, if the sampling interval is set to 0.1 seconds, the latency fluctuation amplitude is calculated as the difference in latency between adjacent sampling points, such as |85-70|=15ms, |100-85|=15ms, |120-100|=20ms. s, align the frame rate and latency amplitude with the same time. For example, align the frame rate and latency amplitude data with timestamps at a sampling interval of 0.1 seconds to obtain the frame rate and corresponding latency fluctuation amplitude at each time point. For example, time point 5.6s corresponds to a frame rate of 23 frames / second and a latency fluctuation amplitude of 15ms; time point 5.7s corresponds to a frame rate of 20 frames / second and a latency fluctuation amplitude of 15ms; time point 5.8s corresponds to a frame rate of 18 frames / second and a latency fluctuation amplitude of 20ms, resulting in the frame latency linkage interval group [5.6s, 23fps, 15ms], [5.7s, 20fps, 15ms], [5.8s, 18fps, 20ms].

[0072] The frame processing deviation judgment submodule calls the frame delay linkage interval group, extracts the load state sequence, compares the frame processing changes with the state differences, normalizes the result, and compares it with the preset offset boundary, using the following formula:

[0073] ;

[0074] Calculate the cumulative value of the deviation intensity of the frame processing points, extract the frame processing position index where the deviation intensity is greater than the baseline judgment value, and establish the offset intensity index group;

[0075] in, This represents the cumulative deviation intensity value of the frame processing point. Represents the number of normalized frames. Representing the Frame current state value, Representing the The preset boundary values ​​of the frame. Representing the Frame processing weighting coefficients Representing the Frame state change rate The average value representing the rate of change of frame state;

[0076] Call the frame delay linkage interval group, such as the aforementioned frame delay linkage interval group. The load state sequence is extracted, reflecting the resource consumption of edge nodes when processing video frames. For example, based on real-time monitoring data such as CPU utilization and memory usage, the load state sequence... , , They are respectively (Normalized value between 0 and 1, 1 represents full load). Compare frame processing changes with load state differences. Frame processing changes refer to the correlation between frame rate and latency fluctuations. For example, when the frame rate decreases while latency fluctuations increase, it indicates a decline in frame processing performance. Normalize the frame processing changes with load state differences. For example, normalize the frame rate and latency fluctuations to make them comparable to the load state sequence on the same scale. After normalization, compare with a preset offset boundary. The preset offset boundary defines the normal range of frame processing deviations. For example, the offset boundary is set to 0.15. This offset boundary is based on the relationship between frame processing performance and load state under ideal operating conditions at edge nodes. Through extensive testing under different load intensities, the differences between frame processing and load state are recorded, and the 90th percentile of the difference is used as the offset boundary. In practical applications, through long-term operation and data analysis, we found that setting the offset boundary to 0.15 can effectively identify 92% of frame processing deviations while avoiding misjudgments of normal fluctuations. The formula is used... Calculate the cumulative value of deviation intensity at frame processing points , This represents the cumulative deviation intensity of each frame processing point, quantifying the degree of deviation of each frame processing point relative to the preset normal behavior. The larger the value, the more severe the deviation. Represents the number of normalized frames, indicating the calculation The number of frames considered at the time, in this example, , Representing the The current state value of a frame refers to the frame number processed by the system at the current moment. Normalized load state values ​​at frame time, such as the load state sequence mentioned above. , Representing the The preset boundary values ​​of the frame indicate the system's processing of the first frame. The expected upper limit of the load state at frame rate, for example, a preset boundary value set to 0.7. Representing the Frame processing weighting coefficients are used to adjust the importance of different frames in deviation intensity calculations. For example, when frame processing experiences stuttering or frame drops, the weighting coefficients will be increased accordingly. Values The weighting coefficient The settings are based on priority evaluation of different frame processing scenarios. For critical video streams (such as important areas in security monitoring), their frame processing weight coefficients are set to higher values ​​to ensure that deviations can be identified in a timely manner. Through empirical adjustments and actual testing, it was found that... Setting it between 0.05 and 0.2 effectively reflects the importance of different frames and provides a reasonable balance in deviation calculations. Representing the The rate of state change of a frame refers to the rate of change of state of the first frame. The rate at which the load state of a frame changes relative to the previous frame, for example, calculated from the load state sequence. , The average value representing the rate of change of frame state is calculated for all. The average value, for example ,molecular The denominator represents the difference between the current state and the preset boundary. The deviation strength is then weighted and standardized, where and To ensure the denominator is not zero, and to assign higher deviation sensitivity to high-weight frames and frames with drastic state changes, the overall calculation logic of this formula is to accurately quantify the degree of deviation in frame processing by comparing the current state of a frame with a preset boundary and combining the frame processing weight and the rate of state change. The larger the value, the more severe the deviation. The specific calculation process is as follows:

[0077] for (Time point) ): , , , , ;

[0078] ;

[0079] for (Time point) ): , , , , ;

[0080] ;

[0081] for (Time point) ): , , , , ;

[0082] ;

[0083] ;

[0084] The calculated cumulative value of the deviation intensity of the frame processing point The system extracts the frame processing location indexes where the deviation intensity exceeds the baseline judgment value. The baseline judgment value is used to determine which frame processing points have sufficiently significant deviations that require intervention. For example, the baseline judgment value is set to 0.08. This baseline judgment value is based on statistical analysis of the frame processing deviation intensity under normal operating conditions. It is calculated by collecting a large amount of frame processing data in typical scenarios. The mean and standard deviation were calculated, and the mean was increased by 1.5 times the standard deviation to serve as the baseline judgment value to ensure that most anomalies could be identified. Experimental data showed that when the baseline judgment value was set to 0.08, the accuracy rate of identifying frame processing deviations reached over 90%. The calculated cumulative value of deviation intensity was then used. Compared with the benchmark value of 0.08, This indicates that the deviation intensity exceeds the baseline judgment value. Therefore, an offset intensity index group is established. The index group contains all frame processing positions where the deviation intensity is greater than the baseline judgment value and their corresponding deviation intensity. For example, the index group contains... Frame processing at a given time point, with an offset intensity of 0.1399, yields the offset intensity index group. By quantifying the difference between the current state of a frame and a preset boundary, and introducing frame processing weight coefficients and frame state change rate as weighting factors, the judgment of frame processing deviation is made more refined and accurate. In particular, the design of the denominator effectively avoids the error of dividing by zero and ensures greater sensitivity to deviations in key frames and frames with drastic state changes, thereby improving the ability to identify abnormal frame processing behavior.

[0085] The frame processing instruction extraction submodule filters the position of the corresponding time point in the task instruction set based on the offset intensity index group, extracts the instruction value, sorts it according to the time sequence, removes duplicate instructions, and obtains the frame processing offset control instruction group.

[0086] Based on the offset intensity index group, such as the aforementioned offset intensity index group [5.8s, 0.1399], the position of the corresponding time point in the task instruction set is filtered. The task instruction set contains a series of predefined frame processing optimization instructions. For example, there is instruction A (corresponding to time point 5.8s, instruction content is "reduce video encoding bitrate") and instruction B (corresponding to time point 5.9s, instruction content is "switch to low complexity algorithm") in the task instruction set. The instruction values ​​are extracted. For example, for the time point 5.8s, the value content of instruction A, "reduce video encoding bitrate", is extracted. The extracted instructions are sorted according to the time sequence. For example, if there are multiple time points corresponding to instructions, they are arranged in chronological order. In this example, there is only one time point 5.8s. Duplicate instructions are removed. For example, if there are multiple identical instructions at the same time point, only one is kept. Finally, the frame processing offset control instruction group is obtained, and the instruction group is [5.8s, "reduce video encoding bitrate"].

[0087] Please see Figure 4 The algorithm adaptation module includes:

[0088] The accuracy extraction submodule extracts the frame processing accuracy data of the node in the loop before and after adjustment based on the frame processing offset control instruction group, identifies the analysis boundary conditions, and obtains the analysis accuracy difference value.

[0089] Based on the frame processing offset control instruction group, such as the aforementioned frame processing offset control instruction group [5.8s, "reduce video encoding bitrate"], the frame processing accuracy data of the node is extracted within the loop before and after adjustment. The frame processing accuracy data reflects the accuracy of the video analysis task in target recognition, classification, and other tasks before and after adjustment. For example, before the 5.8s instruction is issued, the frame processing accuracy of the node in the past 5 seconds is recorded as 92%. After the instruction is issued, the frame processing accuracy is recorded as 88% in the next 5 seconds. Identify and analyze boundary conditions. The boundary conditions are used to define the validity and comparability of the accuracy data. For example, it is set that 100 frames of video data are collected before and after adjustment as analysis samples, and the video content type is required to remain consistent to ensure the fairness of the comparison. Finally, the difference value of the analysis accuracy is obtained. For example, if the accuracy drops from 92% to 88%, the difference value of the analysis accuracy is 92%-88%=4%.

[0090] The target detection accuracy recognition submodule calls the analysis accuracy difference value, identifies the target detection accuracy change trajectory in the unit frame path, judges the target quantity trend corresponding to the accuracy fluctuation, compares the accuracy decrease and target quantity fluctuation in adjacent time periods, calculates the accuracy concentration of the segment in the unit frame path, filters synchronous fluctuation segments, and obtains the target detection accuracy feature segment.

[0091] Based on the accuracy difference value (e.g., 4%), the trajectory of target detection accuracy change in the unit frame path is identified, and the fluctuation of detection accuracy and the corresponding target number trend in each frame are tracked. For example, in the 5.8s to 6.0s segment, the accuracy drops from 90% to 85%, and the number of targets decreases from 10 to 8, showing a consistent decrease. By comparing the accuracy decrease and target number fluctuation in adjacent time periods, the accuracy concentration in this segment is calculated. If the concentration is 80%, it reflects the stability of the detection accuracy. Segments where the accuracy and target number fluctuate synchronously are screened, and finally the target detection accuracy feature segment [5.8s, 6.0s] is obtained, indicating that the algorithm performance has decreased significantly in this period.

[0092] The algorithm response interval identification submodule analyzes the time and accuracy displacement trend based on the target detection accuracy feature segment, judges the consistency of direction and amplitude change characteristics, filters fluctuation segments and aggregates them to obtain the node unit adaptation algorithm trend.

[0093] Based on the target detection accuracy characteristic segment analysis, the time and accuracy displacement trends are analyzed. The time displacement trend reflects the time points and duration of accuracy fluctuations, while the accuracy displacement trend shows the trajectory of accuracy changes from normal to abnormal. Taking 5.8s to 6.0s as an example, the accuracy drops from 90% to 85%. The consistency between the direction of accuracy change and the direction of target quantity change is determined, indicating a decrease in accuracy and a reduction in the number of targets, with an amplitude of 5%. Fluctuation segments are selected, referring to time periods in which both accuracy and target quantity undergo abnormal changes simultaneously. Adjacent segments are aggregated into larger response intervals to obtain the node unit adaptation algorithm trend, such as [5.8s, 6.0s, "Accuracy decreases, algorithm adjustment required"]. This indicates that the algorithm performance deteriorates during this time period, requiring adaptation adjustments.

[0094] Please see Figure 5 The node management optimization module includes:

[0095] The algorithm trend extraction submodule calls the node unit to adapt to the algorithm trend, extracts node task records and periodic algorithm data, identifies the fluctuation amplitude and frequency within a single period, and uses the following formula:

[0096] ;

[0097] Obtain the algorithm's trend value;

[0098] in, Values ​​representing the trend of algorithm changes Representing the Within each cycle, the node task records the corresponding fluctuation coefficient. Representing the The original algorithm output values ​​of the nodes in each cycle. This represents the arithmetic mean of the original algorithm output values ​​of the nodes in this cycle. Representing the The maximum variation difference of the node algorithm data sequence within a period. Represents the total number of nodes within the period;

[0099] The node unit is invoked to adapt to algorithm trends, such as the aforementioned node unit adapting to algorithm trends [5.8s, 6.0s, "accuracy decreases, algorithm adjustment is needed"]. Node task records and periodic algorithm data are extracted. Node task records contain performance data of the edge node when performing specific tasks, such as CPU utilization, memory usage, and processing frame count. Periodic algorithm data contains the algorithm's performance indicators at different periods. For example, in the task records, it is identified that the node's CPU utilization was 85%, memory usage was 90%, and processing frame count was 18 frames / second during the period from 5.8s to 6.0s. In the periodic algorithm data, it is identified that the algorithm's fluctuation range within a normal period is 2% (referring to the maximum deviation between the algorithm's result for each run and its average value), with a fluctuation frequency of 1 time / minute. The fluctuation range and frequency within a single period are identified. For example, in... arrive Within this single cycle, the algorithm detected an actual fluctuation amplitude of 5%, with a fluctuation frequency of 2 times per minute. The system uses the formula... Obtain algorithm change trend value ,in, This value represents the algorithm's trend and quantifies the overall fluctuation of the algorithm's performance over a specific period. A larger value indicates greater fluctuation and less stable performance. Represents the total number of nodes within a period, indicating the number of nodes in the calculation. The number of nodes to consider, in this example, (Single node) Representing the The fluctuation coefficient corresponding to the node task record within each period is used to measure the contribution of the node task record to the algorithm's fluctuation. For example... The value is 0.5, which is the volatility coefficient. The settings were derived from correlation analysis between node task record data (such as CPU usage, memory consumption, etc.) and algorithm performance fluctuations. Higher correlation results in a larger coefficient. Through regression analysis and correlation evaluation of a large amount of historical data, it was found that... Setting it to 0.5 can effectively reflect the exacerbating effect of node resource consumption on algorithm fluctuations. Representing the The original algorithm output values ​​of the nodes in the nth cycle, for example, in the nth cycle. In each cycle, the original algorithm output value of the node is (Right now arrive (Frame processing accuracy during the period) This represents the arithmetic mean of the original algorithm output values ​​of the nodes in that period, for example, (Right now arrive (Average expected accuracy during the period) Representing the The maximum variation difference of the node algorithm data sequence within a period refers to the difference between the maximum and minimum values ​​of the algorithm performance within a specific period. For example, in arrive During this period, the algorithm's performance reached a maximum of 90% and a minimum of 85%. The parentheses The term measures the contribution of node task records to the deviation of the algorithm output value from the average value; the squared term... This means that both positive and negative deviations affect the results, and amplifies the impact of large deviations on the results. Item simplified to This directly reflects the maximum fluctuation range of the algorithm's performance. The overall calculation logic of this formula is to measure the deviation of the algorithm's output from its average value by weighting, and combine this with the algorithm's maximum fluctuation range to comprehensively quantify the overall stability of the algorithm within a specific period. The larger the value, the greater the fluctuation of the algorithm and the more unstable its performance. The specific calculation process is as follows:

[0100] for (Single node): , , , , ;

[0101] ;

[0102] Calculated algorithm change trend value The algorithm's trend value is obtained. The advantage of this formula is that it not only considers the deviation of the algorithm's output value from the average value, but also introduces the fluctuation coefficient of the node task record to weigh the importance of this deviation. At the same time, by adding the maximum variation difference of the algorithm itself, the trend value can more comprehensively and sensitively reflect the actual running stability of the algorithm and the coupling effect of node resources, thus providing a more accurate decision-making basis for subsequent network scheduling and task allocation.

[0103] The network leap identification submodule, based on the algorithm's trend value, selects bandwidth and latency records within the same period according to the external network data of the edge nodes, compares the data change magnitude on a daily basis, filters time nodes with magnitudes greater than the network leap threshold, and obtains the bandwidth and latency leap periods.

[0104] Based on the algorithm's trend values ​​and external network data of edge nodes, bandwidth and latency records within the same period are selected, and the daily changes in data are compared. The daily average values ​​of bandwidth and latency are compared with the previous day to calculate the change magnitude. Time nodes with magnitudes greater than preset network jump thresholds are selected. The bandwidth jump threshold is set to 20Mbps, and the latency jump threshold is set to 30ms. The threshold settings are based on historical data statistics to ensure that significant bandwidth and latency jumps can be captured. Experimental tests show that when the bandwidth jump threshold is 20Mbps and the latency jump threshold is 30ms, 95% of abnormal network jump events can be accurately identified. By comparing the bandwidth and latency change magnitudes with the thresholds, the latency jump event at 6.0s is selected, and the final bandwidth and latency jump period is [6.0s, 6.0s].

[0105] The network impact quantification submodule identifies the intersection duration of algorithm amplitude and network amplitude based on the bandwidth delay jump period, performs weighted processing, and normalizes the usage duration under different periods by referring to amplitude frequency, combined amplitude value and periodic fluctuation, and outputs the network impact duration.

[0106] Based on the bandwidth latency surge period, such as the aforementioned bandwidth latency surge period [6.0s, 6.0s], the intersection duration of algorithm fluctuation and network surge is identified. Algorithm fluctuation refers to the period when algorithm performance fluctuates significantly, and network surge refers to the period when network bandwidth or latency changes significantly. The algorithm fluctuation period is identified as 5.8s to 6.0s, and the network surge period is 6.0s. The intersection duration of the two is 0.0s (i.e., the intersection is only at the instant of 6.0s). Weighting is then performed. Weighting is to comprehensively consider the contribution of different factors to the network impact. For example, different weights can be set for algorithm fluctuation frequency, combined fluctuation value, and periodic fluctuation. For example, the weight of fluctuation frequency can be set to 0.3, the weight of combined fluctuation value to 0.4, and the weight of periodic fluctuation to 0.3. The weight setting is based on the analysis of different network and algorithm anomalies in historical data. Through expert experience and regression analysis, the impact of each factor on the final network is determined. To ensure accurate reflection of the actual network impact, the usage duration under different periods is normalized by referring to the frequency of amplitude variation, the combined amplitude value, and periodic fluctuation. The frequency of amplitude variation refers to the number of times the algorithm performance fluctuates within a specific period. For example, the algorithm performance fluctuates twice within the time period of 5.8s to 6.0s. The combined amplitude value refers to the maximum amplitude of the algorithm performance fluctuation. For example, if the algorithm accuracy drops from 90% to 85%, the combined amplitude value is 5%. Periodic fluctuation refers to the range of fluctuation of the algorithm performance throughout the entire period. For example, the periodic fluctuation is 0.0502 (the aforementioned algorithm change trend value). The normalization process is to eliminate the influence of different period lengths on the results and standardize the usage duration under different periods to a uniform scale, ultimately outputting the network impact duration. For example, by comprehensively considering the above factors and performing weighted normalization, the network impact duration is calculated to be 2.5 seconds, resulting in a network impact duration of 2.5s.

[0107] Please see Figure 6 The network scheduling module includes:

[0108] The instruction filtering submodule filters matching task scheduling instructions based on the duration of network impact, identifies load periods and load states, compares the disturbance period with the instruction period, eliminates low-matching instructions, and obtains a set of instructions affected by network interference.

[0109] Based on the duration of network impact, such as the aforementioned 2.5s duration, task scheduling instructions are filtered and matched. These instructions are pre-defined to adjust task allocation on edge nodes. For example, the task scheduling instruction set might include instructions like "reduce video resolution," "pause non-core analysis tasks," and "switch to a backup server." Load periods and load states are identified. Load periods refer to the time when edge nodes are under high load, while load states refer to the current resource utilization of edge nodes. For example, if it's identified that edge nodes are under high load (CPU utilization exceeding 80%) between 5.8s and 6.0s, the system compares the disturbance period with the instruction period. The duration of a network impact refers to the duration of the network effect, while the instruction duration refers to the execution cycle of a task scheduling instruction. For example, if the duration of a network impact is 2.5 seconds, matching instructions are selected. If the instruction duration is close to the duration of the network impact, it is considered a high match, and instructions with low match are discarded. For example, if the execution cycle of an instruction is 10 seconds, which does not match the 2.5-second duration of the network impact, the instruction is discarded. Finally, the set of instructions affected by network interference is obtained. For example, the instruction "reduce video resolution" is selected. The execution cycle of this instruction is 2 seconds, which has a high match with the 2.5-second duration of the network impact, thus obtaining the set of instructions affected by network interference ["reduce video resolution"].

[0110] The load classification submodule calls the set of instructions affected by network interference, extracts the load switching time and adjustment trigger frequency corresponding to the instructions, arranges them in order of switching time, classifies them by frequency range, statistically analyzes the correspondence between frequency bands and load time periods, and obtains the load adjustment frequency band distribution value.

[0111] The system invokes a set of instructions affected by network interference, such as the aforementioned instruction set "reduce video resolution." It extracts the corresponding load switching time and adjustment trigger frequency for each instruction. The load switching time refers to the time required for the load state to change after executing the instruction, and the adjustment trigger frequency refers to the frequency at which the instruction is triggered. For example, the load switching time for the "reduce video resolution" instruction is 1 second, and the adjustment trigger frequency is once every 5 seconds. Instructions are arranged in order of switching time. For instance, if there are multiple instructions, they are sorted by their load switching time from smallest to largest. They are then categorized by frequency range, which groups instructions with similar adjustment frequencies. Instructions that adjust trigger frequency are grouped together. For example, frequency ranges can be set to 0-0.2Hz (low frequency), 0.2-0.5Hz (medium frequency), and 0.5-1.0Hz (high frequency). The relationship between frequency bands and load time periods is statistically analyzed. For example, the adjustment trigger frequency (0.2Hz) of the "reduce video resolution" instruction belongs to the mid-frequency band, and this instruction is executed during periods of high load (e.g., CPU utilization is higher than 80%). Finally, the load adjustment frequency band distribution value is obtained. For example, the load adjustment frequency band distribution value is ["mid-frequency", "high load period"], indicating that mid-frequency instructions mainly play a role during high load periods.

[0112] The interference identification submodule collects the load jump amplitude and duration within the frequency band based on the load adjustment frequency band distribution value, determines whether the jump amplitude threshold is exceeded, identifies the frequency band interference intensity, sorts the interference intensity of all frequency bands, determines the frequency band and jump time period of the interference intensity, and obtains the frequency of load interference impact of the video analysis task.

[0113] Based on the load regulation frequency band distribution values, such as the aforementioned load regulation frequency band distribution values ​​[“mid-frequency”, “high-load period”], the load jump amplitude and duration within the frequency band are collected. The load jump amplitude refers to the instantaneous change in the load level of the edge node from normal to abnormal within a specific frequency band. The duration refers to the duration of this abnormal load state. For example, in the high-load period corresponding to the mid-frequency band, if a load jump amplitude of 20% (CPU utilization jumps from 60% to 80%) and a duration of 10 seconds are collected, it is determined whether the jump amplitude threshold is exceeded. The jump amplitude threshold is used to identify the severity of the load jump. For example, the jump amplitude threshold is set to 15%. This jump amplitude threshold is set based on the statistics of load fluctuations under normal system operation. By collecting a large amount of data from edge nodes under stable load, the average absolute deviation of load fluctuations is calculated, and adjustments are made based on this to balance sensitivity and false alarm rate. In actual deployment, after several months... Based on accumulated operational data, it was found that setting the jump amplitude threshold to 15% can effectively identify 88% of load change events while maintaining a low false alarm rate. Comparing the collected load jump amplitude of 20% with the 15% threshold, 20% > 15%, indicating that the jump amplitude threshold is exceeded. The system identifies the frequency band interference intensity, which is comprehensively evaluated based on the load jump amplitude and duration. For example, a large load jump amplitude and long duration indicate high interference intensity. The system sorts the interference intensity of all frequency bands. For example, if there are multiple frequency bands, they are sorted from high to low interference intensity to finally determine the frequency band and jump time period of the interference intensity. For example, the mid-frequency band is determined to be the frequency band with the highest interference intensity, and the corresponding jump time period is [5.8s, 6.0s]. The frequency of load interference affecting the video analysis task is obtained. For example, the frequency of load interference affecting the video analysis task is 1 time, which occurs during the high load period of the mid-frequency band [5.8s, 6.0s].

[0114] The edge computing-based real-time video analytics method is executed based on the aforementioned edge computing-based real-time video analytics system and includes the following steps:

[0115] S1: Based on edge node running data, analyze frame rate fluctuations and processing latency changes, extract periods of sudden load peak changes and rapid changes in algorithm complexity, identify state switching points that synchronize latency and load changes, mark task levels, and generate task state switching recognition segments.

[0116] S2: Based on the task state switching recognition segment, extract frame processing change data and latency change amplitude, compare frame processing abrupt changes with latency fluctuation amplitude, analyze the correspondence between frame processing changes and load response, and obtain the frame processing adjustment response trajectory segment.

[0117] S3: Based on frame processing, adjust the response trajectory segment, extract the changes in accuracy value and target detection accuracy gradient before and after the response, analyze the target detection accuracy decrease and target number increase of the node, identify the accuracy decay path and compare it with the original loop data, filter the segments with accuracy decay characteristics, and generate an accuracy-adaptive decay path distribution set.

[0118] S4: Based on the accuracy adaptation attenuation path distribution set, identify the external network data in the corresponding time period, analyze the overlapping time period of bandwidth jump and accuracy attenuation path, extract the associated time period, and obtain the interference segment of network parameters for frame processing adjustment.

[0119] S5: Based on the interference segment of network parameters adjusted by frame processing, extract the scheduling control strategy record within the time period, analyze the time interval of load switching and the adjustment trigger frequency, filter high-frequency adjustment strategy segments, and obtain the frequency of load interference impact of video analysis task.

[0120] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A real-time video analysis system based on edge computing, characterized in that, The system includes: The central control service module analyzes load fluctuations by comparing the changes in video stream frame rate and processing latency based on edge node operation data, filters out key time periods with high load, extracts critical conditions for task scheduling judgment, and obtains task allocation optimization marker values. The edge computing module optimizes the marker value based on the task allocation, matches the video stream data and frame processing strategy for the corresponding time period, compares the frame processing delay point in the marked segment with the current load status, filters the frame processing instructions corresponding to the deviation rate, and obtains the frame processing offset control instruction group. The algorithm adaptation module extracts the analysis accuracy value of the corresponding node before and after the adjustment instruction according to the frame processing offset control instruction group, combines the target detection accuracy path under the unit frame with the frame processing trend identification, extracts the change segment to form the response interval, and obtains the node unit adaptation algorithm trend. The node management optimization module calls the node unit to adapt to the algorithm trend, collects external environment data of the edge nodes, analyzes the overlap duration between the edge computing optimization algorithm change and the network bandwidth jump period, and outputs the duration of the network impact.

2. The real-time video analysis system based on edge computing according to claim 1, characterized in that, The task allocation optimization flag values ​​include frame rate fluctuation threshold, processing latency change range, and load rise rate characteristics. The frame processing offset control instruction group includes frame processing deviation, latency change anomaly value, and load state offset rate. The node unit adaptation algorithm trend includes analysis accuracy change range, target detection accuracy path type, and algorithm response amplitude. The network impact duration includes bandwidth fluctuation duration, latency jump overlap period, and network disturbance duration.

3. The real-time video analysis system based on edge computing according to claim 1, characterized in that, The central control service module includes: The frame rate fluctuation extraction submodule extracts the video stream time series based on the edge node running data, calculates the frame rate difference between adjacent sampling points, determines the fluctuation period and extracts the peak-to-trough distance, filters out the segments with period differences exceeding the threshold, and obtains the frame rate period fluctuation interval value. The latency change calculation submodule calls the frame rate periodic fluctuation interval value, identifies the corresponding latency data, analyzes the absolute latency change amplitude of adjacent time periods, compares it with the set latency amplitude threshold, locates the abrupt change node, and obtains the latency abrupt change amplitude interval value. The task status marking submodule identifies the corresponding load and algorithm complexity data based on the latency mutation amplitude range value, extracts the load and algorithm data, calculates the load complexity offset value by combining latency mutation and frame rate fluctuation, sets the offset threshold, marks the time segment exceeding the threshold, and obtains the task allocation optimization marking value.

4. The real-time video analysis system based on edge computing according to claim 3, characterized in that, The edge computing module includes: The frame processing data matching submodule extracts the video stream frame rate and latency data for the corresponding time period based on the task allocation optimization mark value, calculates the latency fluctuation amplitude according to the sampling interval, and aligns the frame rate and latency amplitude according to the same time to obtain the frame latency linkage interval group. The frame processing deviation judgment submodule calls the frame delay linkage interval group, extracts the load state sequence, compares the frame processing changes with the state differences, normalizes and compares with the preset offset boundary, calculates the cumulative value of the deviation intensity of the frame processing point, extracts the frame processing position index with the deviation intensity greater than the benchmark judgment value, and establishes the offset intensity index group. The frame processing instruction extraction submodule, based on the offset intensity index group, filters the position of the corresponding time point in the task instruction set, extracts the instruction value, sorts it according to the time sequence, removes duplicate instructions, and obtains the frame processing offset control instruction group.

5. The real-time video analysis system based on edge computing according to claim 4, characterized in that, The algorithm adaptation module includes: The accuracy extraction submodule extracts the frame processing accuracy data of the node in the loop before and after adjustment according to the frame processing offset control instruction group, identifies the analysis boundary conditions, and obtains the analysis accuracy difference value. The target detection accuracy recognition submodule calls the analysis accuracy difference value to identify the target detection accuracy change trajectory in the unit frame path, judges the target quantity trend corresponding to the accuracy fluctuation, compares the accuracy decrease and target quantity fluctuation in adjacent time periods, calculates the accuracy concentration of the segment in the unit frame path, filters synchronous fluctuation segments, and obtains the target detection accuracy feature segment. The algorithm response interval identification submodule analyzes the time and accuracy displacement trend based on the target detection accuracy feature segment, judges the consistency of direction and amplitude change characteristics, filters fluctuation segments and aggregates them to obtain the node unit adaptation algorithm trend.

6. The real-time video analysis system based on edge computing according to claim 5, characterized in that, The node management optimization module includes: The algorithm trend extraction submodule calls the node unit to adapt to the algorithm trend, extracts node task records and periodic algorithm data, identifies the fluctuation amplitude and fluctuation frequency within a single period, and obtains the algorithm change trend value. The network leap identification submodule, based on the trend value of the algorithm, selects bandwidth and latency records within the same period according to the external network data of the edge nodes, compares the data change magnitude on a daily basis, filters time nodes with a magnitude greater than the network leap threshold, and obtains the bandwidth and latency leap period. The network impact quantification submodule identifies the intersection duration of the algorithm amplitude and the network amplitude based on the bandwidth delay jump period, performs weighted processing, and normalizes the usage duration under the differentiated period by referring to the amplitude frequency, the combined amplitude value and the periodic fluctuation, and outputs the network impact duration.

7. The real-time video analysis system based on edge computing according to claim 1, characterized in that, The system also includes a network scheduling module: Based on the duration of network impact, the network scheduling module filters task scheduling instructions affected by network interference, categorizes load switching time and adjustment trigger frequency, identifies periods of excessive load jumps, and obtains the frequency of load interference impact on video analysis tasks. The frequency of video analysis task load interference includes load switching frequency, number of adjustment command triggers, and number of times the jump time period exceeds the limit.

8. The real-time video analysis system based on edge computing according to claim 7, characterized in that, The network scheduling module includes: The instruction filtering submodule filters matching task scheduling instructions based on the duration of network impact, identifies load periods and load states, compares disturbance periods with instruction periods, eliminates low-matching instructions, and obtains a set of instructions affected by network interference. The load classification submodule calls the set of instructions affected by network interference, extracts the load switching time and adjustment trigger frequency corresponding to the instructions, arranges them in order of switching time, classifies them by frequency range, statistically analyzes the correspondence between frequency bands and load time periods, and obtains the load adjustment frequency band distribution value. The interference identification submodule collects the load jump amplitude and duration within the frequency band based on the load adjustment frequency band distribution value, determines whether the jump amplitude threshold is exceeded, identifies the frequency band interference intensity, sorts the interference intensity of all frequency bands, determines the frequency band and jump time period of the interference intensity, and obtains the frequency of load interference affecting the video analysis task.

9. A real-time video analysis method based on edge computing, characterized in that, The method is used to implement the real-time video analysis system based on edge computing as described in any one of claims 1-8, and includes the following steps: S1: Based on edge node running data, analyze frame rate fluctuations and processing latency changes, extract periods of sudden load peak changes and rapid changes in algorithm complexity, identify state switching points that synchronize latency and load changes, mark task levels, and generate task state switching recognition segments. S2: Based on the task state switching identification segment, extract frame processing change data and latency change amplitude, compare frame processing abrupt changes with latency fluctuation amplitude, analyze the correspondence between frame processing changes and load response, and obtain the frame processing adjustment response trajectory segment. S3: Based on the frame processing adjustment response trajectory segment, extract the change in accuracy value and target detection accuracy gradient before and after the response, analyze the target detection accuracy decrease and target number increase of the node, identify the accuracy decay path and compare it with the original loop data, filter the segments with accuracy decay characteristics, and generate an accuracy adaptation decay path distribution set. S4: Based on the accuracy adaptation attenuation path distribution set, identify the external network data within the corresponding time period, analyze the overlapping time periods of bandwidth jump and accuracy attenuation path, extract the associated time periods, and obtain the interference segment of network parameters for frame processing adjustment. S5: Based on the network parameter interference segment of the frame processing adjustment, extract the scheduling control strategy record within the time period, analyze the time interval of load switching and the adjustment trigger frequency, filter high-frequency adjustment strategy segments, and obtain the frequency of load interference impact of video analysis task.

Images