Intelligent video stream scheduling method and device based on scene adaptive multi-feature fusion, and medium

Abstract

The invention discloses an intelligent video stream scheduling method based on scene adaptive multi-feature fusion, and the method comprises the steps: only extracting the edge change caused by motion through calculating the difference of Canny edge images between continuous frames, and filtering the static background edge interference; and designing an adaptive background modeling algorithm, and dynamically adjusting the learning rate of an MOG2 background subtractor based on the current motion proportion to realize the balance between key motion target protection and scene adaptation. Through mathematical modeling and scene targeted optimization, on the premise of not depending on a deep learning model, accurate perception and intelligent scheduling of video contents are realized, bandwidth occupation is remarkably reduced, meanwhile, real-time transmission of key events is guaranteed, and the method is suitable for large-scale video monitoring scenes.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary technical field of computer vision, edge computing and intelligent video stream scheduling, and specifically relates to a video content perception and scheduling method, device and medium based on scene adaptive multi-feature fusion. Background Technology

[0002] With the deepening of smart city construction in my country, more than 32 million traffic-specific video surveillance devices have been deployed on highways and urban roads. The traditional "camera-central server" direct connection architecture faces severe challenges in actual operation: First, the vast majority of video footage (especially during off-peak hours) consists of routine scenes with no events occurring, yet it is continuously uploaded in high-definition quality, resulting in a huge waste of bandwidth resources; second, during rush hour and other times, hundreds of cameras uploading simultaneously cause severe network congestion, resulting in video glitches and stuttering; finally, a large amount of worthless footage occupies valuable cloud storage and AI computing resources, causing critical event videos to miss the best analysis opportunity due to queuing.

[0003] Existing solutions largely rely on deploying complex deep learning models at the front end for event recognition, but this places extremely high demands on the camera's computing power and power consumption, resulting in high costs and difficulty in large-scale deployment. More importantly, existing traditional computer vision methods have inherent limitations in video stream scheduling scenarios:

[0004] 1. Edge detection misjudgment problem: The traditional Canny operator detects all intensity gradient edges, including a large number of static background edges (such as buildings and road signs). In video surveillance scenarios, these static edges will interfere with the accurate extraction of motion features, resulting in an inflated motion ratio calculation.

[0005] 2. Insufficient adaptability of background modeling: The traditional MOG2 background subtractor uses a fixed learning rate. In critical scenarios such as traffic accidents, stationary vehicles will be absorbed into the background too quickly, resulting in missed detection of "sudden stationary" events. In scenarios with changing lighting, the fixed learning rate cannot adapt quickly, resulting in a large number of false detections.

[0006] 3. Simple stacking of features: Existing methods often simply combine different features, lacking deep fusion strategies for video scheduling scenarios, and cannot effectively distinguish between real motion and environmental interference.

[0007] Therefore, there is an urgent need for a lightweight, scenario-adaptive intelligent scheduling solution that can effectively solve the above-mentioned technical problems without increasing the burden on front-end devices through mathematical modeling and algorithm optimization.

[0008] The following publicly available patents were found as prior art documents:

[0009] I. Chinese Patent CN119182943A discloses a method for accelerating video image scheduling based on a multi-level video networking architecture in airports. The method includes: dividing the monitoring architecture into multi-level layers based on airport layout and security requirements; analyzing and classifying video streams into real-time monitoring streams, playback streams, and low-priority streams, and performing different optimization processes (such as adjusting encoding parameters, compression, and degradation); dynamically adjusting the transmission order and rate based on real-time scheduling algorithms and video stream priorities; setting up a distributed caching mechanism to cache high-frequency or important data to reduce latency; and implementing a dynamic load balancing strategy to allocate video processing resources. This method aims to improve the overall management efficiency and transmission stability of large-scale airport monitoring systems through system architecture optimization, resource management, and network transmission strategies.

[0010] II. US Patent US20220368972A1 discloses a method and apparatus for scheduling compressed video frames in a vehicle-mounted wireless network. The core of this method lies in: allocating time-frequency resources for video frames from multiple video sources based on their compression type (I-frame, P-frame, B-frame); when the total bandwidth requirement of concurrently transmitted frames exceeds a threshold, comparing video stream priorities or frame priorities, and taking action accordingly, such as delaying or canceling the transmission of low-priority frames, or instructing video sources to increase the compression rate of the video streams. The priority of the video streams is determined based on factors such as vehicle location, operating status, external environment, camera location, and resolution. This method focuses on wireless resource allocation at the communication layer, aiming to optimize the transmission efficiency and reliability of multiple vehicle-mounted video streams under limited bandwidth.

[0011] III. Patent CN116824503A discloses a video analysis system and method based on edge computing. The system includes multiple video analysis devices integrating video acquisition, interface matching, and edge computing modules, as well as data scheduling management, a big data center, and a terminal module. The method involves the edge computing module extracting frames from the video stream and using optimized deep learning algorithms (such as the YOLO series) for target detection and recognition. The data scheduling management module dynamically adjusts computing resource strategies based on the combined GPU / CPU load of the edge devices. For example, it uses a high-precision deep convolution mode when the load is low, and switches to a fast, lightweight convolution mode or offloads the computing task to other edge devices for collaborative computing when the load is high. The core of this method is to implement video content semantic analysis based on a deep learning model at the edge and ensure the real-time performance and accuracy of the analysis task through dynamic scheduling of computing resources.

[0012] Upon comparison, the above-mentioned prior art documents differ from this application in the following ways:

[0013] This invention discloses an intelligent video stream scheduling method based on scene-adaptive multi-feature fusion. The core innovation of this method lies in: proposing a motion-aware edge detection algorithm that extracts only edge changes caused by motion by calculating the differences in Canny edge maps between consecutive frames, filtering out static background edge interference; and designing an adaptive background modeling algorithm that dynamically adjusts the learning rate of the Gaussian mixture background subtractor based on the current motion ratio, achieving a balance between protecting key moving targets and scene adaptation. By extracting inter-frame differences, motion edges, and adaptive foregrounds in parallel, and then fusing them in a two-stage process to generate robust comprehensive motion features, the motion ratio of video frames is accurately quantified. The system combines the temporal history of this motion ratio to intelligently identify abnormal patterns such as "sudden stillness" (abrupt change from motion to stillness, such as a traffic accident) and "violent motion." Finally, based on the quantified motion ratio value and the abnormal event state, the video stream is dynamically divided into four upload priorities: low, medium, high, and urgent, and the quality parameters of the uploaded video stream are adjusted accordingly. This invention, through mathematical modeling and algorithm optimization, achieves accurate and lightweight perception and intelligent scheduling of video content without relying on any deep learning model.

[0014] While patent CN119182943A focuses on resource allocation and transmission optimization through a pre-defined multi-level system architecture and static video stream classification rules (such as real-time monitoring streams and playback streams), its decisions rely on external system configuration rather than the video content itself; patent US20220368972A1 focuses on wireless time-frequency resource scheduling based on the grammatical structure of video compression coding (such as I / P / B frames) and vehicle external sensor information, which belongs to communication layer optimization and does not involve understanding the semantic content of images; patent CN116824503A focuses on deploying and scheduling deep learning models such as YOLO at the edge for high-level semantic recognition (such as target classification), the core of which is the allocation of computing resources to ensure the inference performance of the model, which has high requirements for hardware computing power; this invention differs from the above technical paths and proposes a brand-new "lightweight content-aware" scheduling paradigm. The core scientific principle of this invention lies in its innovative motion-aware edge detection and adaptive background modeling algorithms, which robustly extract and quantify the intensity of "motion energy" in a scene directly from low-level visual features (pixel changes, edge changes, foreground segmentation), rather than identifying specific object or event categories. This method bypasses the complex computations, large amounts of labeled data, and model training processes required by deep learning models, relying solely on lightweight traditional computer vision algorithms to achieve motion state perception highly relevant to security monitoring. Compared to semantic recognition based on deep learning, this invention does not answer the question "What is in the picture?" but rather precisely answers the question "How much motion is in the picture and is the motion pattern abnormal?" The latter is the direct and efficient basis for determining transmission priority. Therefore, this invention achieves a leap from "indirect scheduling based on rules or recognition results" to "direct scheduling based on content motion characteristics" without increasing the computing power burden and cost of edge devices, achieving a unique and efficient balance between bandwidth saving and critical event protection. Summary of the Invention

[0015] The purpose of this invention is to overcome the following technical defects in existing video stream scheduling technologies:

[0016] Feature extraction lacks scenario specificity: traditional edge detection contains a lot of static background interference, which affects the accuracy of motion perception; background modeling is not adaptable enough: fixed learning rate cannot balance the needs of key target protection and environmental adaptation; feature fusion is simply stacked: there is a lack of deep fusion strategies for video surveillance scenarios.

[0017] To this end, this invention proposes an intelligent video stream scheduling method, device, and medium based on scene-adaptive multi-feature fusion. Through mathematical modeling and algorithm optimization, it achieves accurate perception and intelligent differentiated transmission of video content.

[0018] To achieve the above objectives, this invention provides an intelligent video stream scheduling method based on scene-adaptive multi-feature fusion, comprising the following core steps: Step 1: Video stream access and state initialization. Multiple video streams are accessed, and an independent state buffer is established and maintained for each video stream, including: the previous frame grayscale image. Canny Edge Historical Images Historical queue of movement proportions Adaptive background subtractor instance, abnormal state timer.

[0019] Step 2: Scene-Adaptive Multi-Feature Extraction. For the current video frame... After preprocessing, the following three feature extraction operations are performed in parallel:

[0020] 2.1 Inter-frame differential motion detection: Calculate the absolute pixel difference between the current frame and the previous frame, and obtain a binary motion mask by thresholding.

[0021]

[0022] in, This is the inter-frame difference threshold, with a typical value of 6.

[0023] 2.2 Motion-Aware Edge Detection: By calculating the differences in Canny edge maps between consecutive frames, only edge changes caused by motion are extracted. Let... and Given the binary edge images extracted by the Canny operator for the current frame and the previous frame, respectively, the motion edge mask is defined as follows:

[0024]

[0025] in, The threshold for edge change detection is 1, with a typical value of 1. This improvement significantly filters out static background edges such as buildings and road signs, enhancing the correlation between edge features and motion events.

[0026] 2.3 Adaptive Background Subtraction: An improved Gaussian mixture background modeling algorithm is adopted, based on the current motion ratio. Dynamically adjust the learning rate The mathematical relationship is as follows:

[0027] ;

[0028] in, , This adaptive mechanism freezes background updates in emergency scenarios such as traffic accidents to prevent stationary vehicles from being absorbed; and accelerates background adaptation in static scenes to reduce false detections of light and shadow.

[0029] Step 3: Two-stage feature fusion. Based on the features extracted in Step 2, a comprehensive motion feature map is generated according to the following rules. :

[0030] First-stage fusion (region enhancement): This stage combines the results of inter-frame difference and background subtraction using a logical "OR" operation.

[0031]

[0032] Second-stage fusion (spatial constraints): Spatial refinement of the initial motion region using motion edge information:

[0033]

[0034] in, Indicates use Structural elements (typical) The morphological dilation operation is applied. This fusion strategy effectively distinguishes between real-world object motion and environmental texture changes.

[0035] Step 4: Motion Proportion Quantization and Anomaly Detection. Calculate the comprehensive motion feature map. motion ratio value :

[0036]

[0037] in, The total number of pixels in the image. This is an indicator function.

[0038] Historical queue based on motion proportions Detect two abnormal patterns:

[0039] Sudden onset of static abnormality: when and Time determination, among which , The historical average proportion of movement;

[0040] Abnormalities in strenuous exercise: When Make a judgment immediately.

[0041] Step 5: Priority Determination and Scheduling. Based on the motion ratio value. In case of abnormal conditions, upload priority is determined based on multi-level thresholds:

[0042] ;

[0043] The upload quality parameters (resolution, frame rate, bit rate) of the video stream are dynamically adjusted based on priority.

[0044] Compared with the prior art, the present invention has the following beneficial effects:

[0045] 1. Improved feature extraction accuracy: Motion-aware edge detection reduces static background edge interference by more than 60%, and motion ratio calculation is more accurate;

[0046] 2. Enhanced protection against critical events: Adaptive background subtraction extends the continuous detection time of sudden static events by 3-5 times, significantly reducing missed detections;

[0047] 3. Optimized scene adaptability: The dynamic learning rate mechanism reduces the false detection rate by 40%-60% while maintaining high sensitivity to real motion.

[0048] 4. Excellent engineering practicality: It is based entirely on traditional computer vision algorithms, has low computational complexity, and can be deployed on low-cost edge devices;

[0049] 5. Significant bandwidth savings: Level 4 priority scheduling can reduce cloud bandwidth usage by 50%-80% while ensuring video quality for critical events.

[0050] 6. This invention is particularly suitable for large-scale video surveillance scenarios such as smart cities, intelligent transportation, and security monitoring. It can achieve intelligent and differentiated transmission of video streams without increasing the burden on front-end equipment. Attached Figure Description

[0051] Figure 1 This is a flowchart of the method of the present invention;

[0052] Figure 2 This is a diagram of the intelligent video stream scheduling system architecture in this invention;

[0053] Figure 3 This is a flowchart of the multi-feature extraction and fusion process in this invention;

[0054] Figure 4 This is a comparison image of motion-aware edge detection in this invention;

[0055] Figure 5 This is the adaptive background subtraction learning rate adjustment diagram in this invention;

[0056] Figure 6 This is a logic diagram for anomaly detection and priority determination in this invention. Detailed Implementation

[0057] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments. Those skilled in the art should understand that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention.

[0058] The present invention provides an intelligent video stream scheduling method based on scene-adaptive multi-feature fusion, the specific implementation steps of which are as follows:

[0059] Step 1: Video frame preprocessing and state initialization

[0060] The system receives image frame data from video files or real-time video streams. The input is a BGR format color image, typically with a resolution of 1920×1080 or 1280×720. The following preprocessing operations are performed first:

[0061] 1.1 Color Space Conversion: Converting a BGR image to a single-channel grayscale image The standard grayscale conversion formula is used:

[0062]

[0063] in, , , These are the pixel intensity values ​​for the red, green, and blue channels, respectively.

[0064] 1.2 Size Downsampling: To reduce computational complexity, based on the downsampling factor... Reduce the size of the grayscale image. Let the original image size be... The size after downsampling is Resampling is performed using a bilinear interpolation algorithm, mathematically expressed as follows:

[0065]

[0066] in, For the target pixel coordinates, For the coordinates of the four nearest neighbor pixels in the original image, These are the bilinear weighting coefficients.

[0067] 1.3 Gaussian Blur Denoising: A 5×5 Gaussian convolution kernel is applied to smooth the downsampled image, eliminating image noise interference. The two-dimensional Gaussian convolution kernel is defined as:

[0068] ;

[0069] in, The standard deviation is used to control the degree of fuzziness. In this embodiment, we take... This corresponds to the default value in OpenCV. The image after Gaussian filtering is as follows:

[0070]

[0071] Step 2: Scene-Adaptive Multi-Feature Extraction

[0072] 2.1 Inter-frame differential motion detection

[0073] Maintain the previous frame grayscale image for each video stream Calculate the current frame. Absolute pixel difference from the previous frame:

[0074]

[0075] Through empirical thresholds Binarization is performed to obtain the inter-frame differential motion mask:

[0076]

[0077] threshold Selection criteria: Based on statistical analysis of typical traffic monitoring scenarios, pixel intensity changes caused by vehicle movement are usually greater than 6, while pixel changes caused by minor changes in illumination and sensor noise are usually less than 6.

[0078] 2.2 Edge Detection for Motion Perception

[0079] Traditional Canny edge detection detects all static background edges, such as buildings and road signs, which are considered interference in video scheduling scenarios. This invention proposes a motion-aware edge detection method:

[0080] First, the traditional Canny operator is applied to the current frame to obtain the original edge image. The gradient calculation of the Canny operator uses the Sobel operator:

[0081]

[0082] The gradient magnitude and direction are:

[0083]

[0084] After nonmaximum suppression and double thresholding (low threshold) High threshold After that, a binary edge image is obtained. The edge pixel value is 255, and the non-edge pixel value is 0.

[0085] Key improvement: Reading the edge image from the previous frame from the state buffer Calculate the difference in edge variation:

[0086]

[0087] because and All are binary images (0 or 255). There are only two possible values:

[0088] 0: The pixel's edge state is the same in both frames (either both are edges or neither is an edge).

[0089] 255: The edge state changes (from nothing to something or from something to nothing).

[0090] By threshold Extracting motion edge masks:

[0091]

[0092] Mathematical meaning: Let For any pixel in the image, if in two consecutive frames, A point changes from a non-edge to an edge ( ) or change from edge to non-edge ( ),but This point is marked as the moving edge. This change is usually caused by the movement of the outline of a moving object.

[0093] This method can filter out about 70% of static background edges, significantly improving the motion relevance of edge features.

[0094] 2.3 Adaptive Background Subtraction

[0095] Traditional MOG2 background subtractors use a fixed learning rate, which cannot adapt to different motion states. This invention designs an adaptive learning rate mechanism:

[0096] First, an improved Gaussian mixture background model is created for each video stream. Let the background model be... Composed of Gaussian distributions (default) ), No. The weights, mean, and variance of the Gaussian distributions are respectively... , , .

[0097] For the current pixel value Calculate its matching degree with each Gaussian distribution:

[0098]

[0099] The key innovation of adaptive learning rate: learning rate It's no longer a fixed value, but rather the current proportion of motion. Functions:

[0100]

[0101] The learning rate is determined by the historical length of the background model. Indirect implementation: Therefore, according to Dynamically set the history length:

[0102]

[0103] The parameter update rule for the Gaussian distribution is:

[0104]

[0105] Mathematical Principle Analysis:

[0106] when (Emergency): Minimal, background model hardly updates, ensuring stationary targets are not absorbed;

[0107] when (Normal exercise): Moderate, slow updates to adapt to changes in lighting;

[0108] when (Still scene): Larger size, faster updates to eliminate false detections of light and shadow.

[0109] The foreground mask is obtained after background subtraction. The foreground pixel value is 255, and the background pixel value is 0.

[0110] Step 3: Two-stage feature fusion

[0111] First-stage fusion (region enhancement): The inter-frame differential mask is then... With MOG foreground mask Perform a logical "OR" operation:

[0112]

[0113] This phase combines the advantages of two algorithms: inter-frame difference is sensitive to fast motion, while MOG2 is robust to continuous foreground.

[0114] Second-stage fusion (spatial constraints): Spatial refinement of the enhancement mask using motion edge information:

[0115] 1. Motion edge mask Perform 3×3 morphological dilation:

[0116]

[0117] in, It represents a 3×3 neighborhood.

[0118] 2. Perform logical AND operations:

[0119]

[0120] 3. Area filtering: Calculation Area of ​​each connected region Filter out areas smaller than Small noise areas of pixels:

[0121]

[0122] in, This indicates that the region was detected by both algorithms simultaneously, and it was preserved even if the area was small.

[0123] Step 4: Motion Proportion Quantification and Anomaly Detection

[0124] 4.1 Calculation of Motion Proportions (Assume the final comprehensive motion characteristic diagram) The size is Calculate the motion ratio :

[0125]

[0126] Among them, indicator function Defined as:

[0127]

[0128] 4.2 Abnormal Mode Detection and Maintenance (Recent) Frame motion scale sequence .

[0129] Sudden Static Anomaly Detection: The following two conditions must be met for a determination to be made:

[0130]

[0131] Detection of abnormal strenuous exercise: Immediate judgment is made when the following conditions are met:

[0132]

[0133] The abnormal state must last at least It took seconds to be confirmed before entering. The abnormal recovery period is 2 seconds, during which the priority is fixed at the emergency level.

[0134] Step 5: Priority Determination and Scheduling Execution

[0135] According to the proportion of exercise In case of abnormal conditions, determine the four-level upload priority:

[0136] low priority ( (and no abnormalities)

[0137] Medium priority ( (and no abnormalities)

[0138] High priority ( (and no abnormalities)

[0139] Emergency Priority ( (or an anomaly detected):

[0140] The scheduling system sends video streams with corresponding quality parameters to the cloud server based on priority. For low and medium priority video streams, a frame extraction strategy (such as uploading 1 frame every 2 frames) can be used to further save bandwidth.

[0141] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A smart video stream scheduling method based on scene-adaptive multi-feature fusion, characterized in that: Includes the following steps: Step 1: Video Stream Access and State Initialization Multiple video streams are accessed, and an independent state buffer is initialized and maintained for each video stream. The state buffer includes: the previous frame grayscale image. Canny Edge Historical Images Historical queue of movement proportions and abnormal state timers; Step 2: Scene-Adaptive Multi-Feature Extraction For the current video frame After preprocessing, the scene adaptive feature extraction module performs the following operations in parallel: 2.1 Inter-frame differential motion detection: Calculate the grayscale image of the current frame Compared to the previous grayscale image absolute pixel difference After threshold After binarization, the inter-frame difference motion mask is obtained. ; 2.2 Edge detection for motion sensing: Calculate the edge image of the current frame. Images of historical edges The absolute difference, after threshold After binarization, the motion edge mask is obtained. ,in: 2.3 Adaptive Background Subtraction: An adaptive Gaussian mixture background modeling algorithm is used, based on the current motion ratio. Dynamically adjust the learning rate of the background model Foreground extraction is performed on the current frame to obtain the background minus foreground mask. ; Step 3: Two-stage feature fusion Based on the scene adaptive multi-feature features extracted in step 2, a comprehensive motion feature map is generated through a two-stage fusion method. : First stage fusion: merging inter-frame differential motion masks Subtracting foreground mask from background Perform a logical OR operation to obtain the enhanced foreground mask. ; Second-stage fusion: will enhance the foreground mask. With motion edge mask Spatial constraint fusion is performed to generate a comprehensive binary motion feature map. ; Step 4: Quantifying the proportion of motion Calculate the comprehensive motion feature map motion ratio value ,in: in The total number of pixels in the image. For indicator functions; Step 5: Maintain the historical sequence of motion ratios. When a motion ratio value is detected... When an abnormal change pattern is observed, it is marked as an abnormal event; the abnormal change pattern includes: Sudden Static Mode: Current And historical average ; Intense exercise mode: Current ; Step 6: Based on the motion ratio value Based on preset multi-level thresholds, the upload priority of the video stream is determined according to the abnormal event status and the priority includes low, medium, high and urgent. Step 7: Based on the upload priority, send video stream data encoded with the corresponding quality of service parameters to the cloud server.

2. The intelligent video stream scheduling method based on scene-adaptive multi-feature fusion according to claim 1, characterized in that: Step 2.2, which describes motion-aware edge detection, specifically includes the following sub-steps: 2.2.1: Apply the Canny edge detection operator to the current frame to obtain the original edge image. ; 2.2.2: Read the edge image of the previous frame from the state buffer. ; 2.2.3: Calculating edge variation differences in images ; 2.2.4: Thresholding is performed on the difference images, where the threshold... The value range is 1-10; 2.2.5: Use the thresholding result as a motion edge mask Output; 2.2.6: Update the edge history image in the state cache, let .

3. The intelligent video stream scheduling method based on scene-adaptive multi-feature fusion according to claim 1, characterized in that: The learning rate for adaptive background subtraction described in step 2.3 Based on the current exercise ratio The following rules will be dynamically adjusted: ； in, ,and , , .

4. The intelligent video stream scheduling method based on scene-adaptive multi-feature fusion according to claim 1, characterized in that: The two-stage fusion method described in step 3 specifically includes: The first phase of fusion uses logical operations: ； The second-stage fusion employs space-constrained computation: ， in This represents a morphological dilation operation on the edge image.

5. The intelligent video stream scheduling method based on scene-adaptive multi-feature fusion according to claim 1, characterized in that: The detection of abnormal change patterns in step 5 employs a sliding time window mechanism, wherein: Maintain the motion scale sequence of the most recent N frames N is a positive integer; Calculate the historical average movement ratio ; When the condition is met and At that time, it was determined to be a sudden abnormal stillness; When the condition is met When the threshold is met, it is determined to be an abnormal, violent movement; where the threshold is met: Typical value: =0.005, =0.02, =0.

1.

6. The intelligent video stream scheduling method based on scene-adaptive multi-feature fusion according to claim 1, characterized in that: The priority determination rule mentioned in step 6 is as follows: when When the priority is low; when At that time, the priority is medium; when When the priority is high; when When an abnormal event is detected, the priority is urgent.

7. An electronic device, characterized in that: The method includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the intelligent video stream scheduling method based on scene-adaptive multi-feature fusion as described in any one of claims 1 to 6.

8. A computer-readable storage medium storing a computer program, characterized in that: The computer program causes the computer to execute the intelligent video stream scheduling method based on scene-adaptive multi-feature fusion as described in any one of claims 1 to 6.

Images