Intelligent Analysis Method and System for Security Video Streams Based on Edge Computing and Cloud Collaboration

By dynamically adjusting the frame extraction density and feature summary compression ratio of the video stream, and combining key events and cloud load status, the problems of bandwidth fluctuation and cache pressure in edge-cloud collaborative video analysis are solved, enabling the complete transmission and analysis of high-value event information.

CN123075221APending Publication Date: 2026-11-10SHANDONG FOCUSED INTELLIGENCE TECHNOLOGY CO LTD
View PDF 1 Cites 0 Cited by

Patent Information

Application Number
CN202611436375.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-16
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

Existing edge-cloud collaborative video analytics methods fail to effectively address dynamic changes in network bandwidth and edge caching status, resulting in poor transmission quality of high-value event information in bandwidth-constrained and highly volatile monitoring scenarios.

Method used

By acquiring network bandwidth jitter and edge cache queue backlog length, the frame sampling density and feature summary compression ratio of the video stream are dynamically adjusted. Combined with key event triggering conditions and cloud load status, adaptive sampling and feature extraction of the video stream are achieved, and the upload strategy is optimized.

Benefits of technology

Ensuring the integrity and accuracy of high-value event information under limited backhaul bandwidth improves the efficiency and reliability of video stream analysis.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123075221A_ABST
    Figure CN123075221A_ABST
Patent Text Reader

Abstract

This disclosure provides an intelligent analysis method for security video streams based on edge computing and cloud collaboration, including: obtaining the current network bandwidth jitter amplitude and edge buffer queue backlog length; determining the frame extraction density and feature summary compression ratio of the video stream after a key event is triggered based on the network bandwidth jitter amplitude and edge buffer queue backlog length; dynamically sampling and extracting features from the video stream according to the frame extraction density and feature summary compression ratio; and uploading the extracted video feature data to the cloud for further intelligent analysis and behavior discrimination. The solution provided by this disclosure addresses how to jointly regulate the frame extraction density and feature summary compression ratio of the video stream after a key event is triggered based on the network bandwidth jitter amplitude and edge buffer queue backlog length, thereby solving the problem of missing high-value event information integrity under limited backhaul bandwidth.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the field of information technology, specifically to a method for intelligent analysis of security video streams based on edge computing and cloud collaboration. Background Technology

[0002] With the advancement of smart city and safe city construction, the number of security surveillance cameras has surged, and the real-time transmission and analysis of massive amounts of video data poses a severe challenge to network bandwidth and cloud computing power. Edge computing and cloud-based collaborative architectures, which perform video preprocessing and feature extraction at edge nodes close to the data source and then upload lightweight feature data to the cloud for in-depth analysis, have become the mainstream technical direction for addressing these bottlenecks. However, most existing edge-cloud collaborative video analysis methods employ fixed frame extraction strategies or single compression ratio settings, lacking the ability to jointly perceive and adaptively adjust dynamic changes in network bandwidth and edge cache status. For example, Chinese patent application CN115103157A discloses a video analysis method based on edge-cloud collaboration, which achieves anomaly detection by comparing the similarity of extracted frames from the surveillance video stream with a set of false alarm images; another existing technology dynamically allocates video data into continuous and non-continuous frame extraction categories based on the computing resource utilization of the edge gateway and cloud server. None of these solutions fully consider the impact of network bandwidth jitter on the quality of critical event video data transmission, making it difficult to guarantee the complete upload and accurate analysis of high-value event information in actual monitoring scenarios with limited bandwidth and significant fluctuations. Summary of the Invention

[0003] In view of this, the present disclosure provides a method for intelligent analysis of security video streams based on edge computing and cloud collaboration, which at least partially solves the problems existing in the prior art.

[0004] Intelligent analysis methods for security video streams based on edge computing and cloud collaboration include: Get the current network bandwidth jitter level and edge buffer queue backlog length; Based on the network bandwidth jitter amplitude and the edge buffer queue backlog length, the video stream frame extraction density and feature summary compression ratio after the key event is triggered are determined. Dynamic sampling and feature extraction are performed on the video stream based on the frame extraction density and feature summary compression ratio. The extracted video feature data is uploaded to the cloud for further intelligent analysis and behavior identification.

[0005] Preferably, the step of determining the frame extraction density and compression ratio based on network bandwidth jitter amplitude and edge buffer queue backlog length further includes: Obtain the current average network bandwidth B_avg and jitter amplitude ΔB; Get the backlog length Q of the edge cache queue; Adjust the frame rate density F=F0 according to the following formula. (ΔB / B_avg)+Q / K, where F0 is the baseline frame sampling density and K is the buffer scaling factor; When ΔB is greater than the preset threshold T_ΔB, the value of F is increased to improve the frame extraction frequency.

[0006] Preferably, the step of determining the frame extraction density and compression ratio based on network bandwidth jitter amplitude and edge buffer queue backlog length further includes: Get the real-time network delay D_delay; Combining the previously calculated frame extraction density F and feature summary compression ratio C; Dynamically adjust C=C0 according to the following formula. (F / F_max) (Q / Q_max), where C0 is the initial compression ratio, F_max is the maximum allowed frame rate, and Q_max is the maximum buffer capacity; When C exceeds the set upper limit C_upper, the value of F is reduced to ensure the integrity of uploaded data.

[0007] Preferably, the step of dynamically sampling the video stream based on the frame extraction density and feature summary compression ratio further includes: Calculate the background feature difference E_diff before the occurrence of the critical event; Determine whether it is in a peak period T_period by combining the current timestamp t; If E_diff > threshold T_diff and T_period is a busy period, then enable a higher precision compression ratio C_high; Otherwise, a low-precision compression ratio C_low is used by default to save bandwidth.

[0008] Preferably, uploading the extracted video feature data to the cloud further includes: Get the current cloud platform load status S_cloud; The upload period P is determined based on the following formula: P=P0 (S_cloud / S_ref), where P0 is the baseline upload period and S_ref is the reference load value; When S_cloud < reference load value, reduce P to shorten the upload interval and speed up the response time; If P is less than the lower limit P_min, then P_min is forcibly maintained to avoid resource abuse.

[0009] Preferably, the scenario segmentation based on key event triggering further includes: Detect key event types and mark them as Event_type; Determine whether to retain the original picture or only the feature information according to Event_type; Evaluate the transmission priority according to the following formula: P_prior=α (Event_type_priority+β ΔB^γ), where α, β, and γ are weight coefficients; High-priority events are uploaded preferentially when bandwidth is insufficient, and the frame skipping density is increased to ensure integrity.

[0010] Preferably, the deploying a multi-level cache structure at the edge node further comprises: Initializing a three-level cache, including a fast cache R_Cache, a medium-speed cache M_Cache, and a slow cache L_Cache); Determine the cache level according to the historical packet loss rate Packet_loss_rate; Make a cache decision using the following formula: if Packet_loss_rate>Threshold_loss, use M_Cache, otherwise use R_Cache; When the L_Cache still has storage space, temporarily store part of non-emergency feature data into the L_Cache.

[0011] Preferably, the adjusting process of the feature summary compression ratio introduces a user permission level, which further comprises: Collect the user's access permission identifier U_level; Set the compression ratio limit C_limit=C0+(U_level λ), where λ is the increment factor corresponding to each permission level; If C<C_limit, a higher-level feature summary compression strategy is forcibly adopted; If C>C_limit, reduce the frame skipping density to ensure the balance between security and integrity.

[0012] Preferably, the optimization measures for network fluctuations in specific scenarios further comprise: Calculate the standard deviation of bandwidth σ_Bandwidth within the past 5 minutes; Determine whether to trigger the dynamic compensation mechanism based on the following formula: if σ_Bandwidth>threshold_B, activate_compensate is true, otherwise it is false; Start an additional feature compression algorithm when compensate is true: C_comp=C_base (σ_Bandwidth / σ_ref); Simultaneously increase the keyframe extraction frequency F_base (σ_Bandwidth / σ_ref) to avoid losing critical information.

[0013] Preferably, the synchronization mechanism based on multi-camera collaboration further includes: Get the timestamp difference T_offset between each camera; A synchronization matrix S_matrix is ​​established to align the video stream content; Image offset is compensated using the following formula: compensation_factor=max(0,(T_offset / T_sync_window)), where T_sync_window is the length of the synchronization window; When the compensation_factor exceeds the set threshold T_comp_max, frame synchronization adjustment is initiated to reduce information errors.

[0014] Preferably, the adaptive strategy based on edge device computing power further includes: Obtain the available computing power A_avail for the edge node; Resource utilization is evaluated using the following formula: Utilization_ratio = A_avail / A_total, where A_total is the total available computing power. The frame rate and compression ratio are automatically adjusted based on Utilization_ratio, i.e., F_new = F_old. (1Utilization_ratio)+F_base Utilization_ratio; To ensure system stability, complex feature extraction operations are reduced under low computing power conditions.

[0015] This disclosure provides an intelligent analysis method for security video streams based on edge computing and cloud collaboration, including: obtaining the current network bandwidth jitter amplitude and edge buffer queue backlog length; determining the frame extraction density and feature summary compression ratio of the video stream after a key event is triggered based on the network bandwidth jitter amplitude and edge buffer queue backlog length; dynamically sampling and extracting features from the video stream according to the frame extraction density and feature summary compression ratio; and uploading the extracted video feature data to the cloud for further intelligent analysis and behavior discrimination. The solution provided by this disclosure addresses how to jointly regulate the frame extraction density and feature summary compression ratio of the video stream after a key event is triggered based on the network bandwidth jitter amplitude and edge buffer queue backlog length, thereby solving the problem of missing high-value event information integrity under limited backhaul bandwidth. Attached Figure Description

[0016] In the accompanying drawings, unless otherwise specified, the same reference numerals throughout the various drawings denote the same or similar parts or elements. These drawings are not necessarily drawn to scale. It should be understood that these drawings depict only some embodiments disclosed in this application and should not be construed as limiting the scope of this application.

[0017] Figure 1 The flowchart shows the intelligent analysis method for security video streams based on edge computing and cloud collaboration of the present invention. Figure 2 This is a block diagram of the security video stream intelligent analysis system based on edge computing and cloud collaboration of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of this disclosure clearer, the embodiments of this disclosure will be further described in detail below with reference to the accompanying drawings. The illustrative implementation methods and descriptions of the embodiments of this disclosure are only used to explain the embodiments of this disclosure and are not intended to limit the embodiments of this disclosure.

[0019] The intelligent analysis method for security video streams based on edge computing and cloud collaboration provided by this invention can be deployed on edge computing nodes (such as embedded AI computing boxes, intelligent network video recorders, or edge servers with GPU acceleration capabilities) of security monitoring systems. These edge nodes communicate with the cloud-based intelligent analysis platform via wired or wireless networks. Figure 1 As shown, the method mainly includes the following four steps.

[0020] Step S101: Obtain the current network bandwidth jitter amplitude and the edge cache queue backlog length.

[0021] In this step, the edge computing node first starts the network status monitoring thread and the cache status monitoring thread to collect the current network bandwidth jitter and the backlog length of the edge cache queue in real time.

[0022] Specifically, edge computing nodes periodically send probe packets to the cloud platform (e.g., a 64-byte ICMP probe packet every 100 milliseconds) and record the round-trip time (RTT) and throughput of each probe to calculate the current network bandwidth. Assuming that within a preset time window T_window (e.g., the past 5 seconds), the edge node collects n bandwidth samples B_1, B_2, ..., B_n, then the current average network bandwidth B_avg = (B_1 + B_2 + ... + B_n) / n. The network bandwidth jitter ΔB is defined as the standard deviation of the bandwidth samples within this time window, i.e., ΔB = sqrt(Σ(B_i - B_avg)). 2The value of ΔB ( / n) is used to quantify the severity of bandwidth fluctuations within that time window. The larger the value of this jitter amplitude ΔB, the more unstable the current network is, and the higher the uncertainty faced by data transmission.

[0023] Meanwhile, edge computing nodes maintain a first-in, first-out (FIFO) buffer queue to temporarily store video frame data and feature data to be uploaded to the cloud. Edge nodes obtain the buffer queue backlog length Q by reading the number of currently occupied bytes or the number of buffered frames in the buffer queue. For example, if the total capacity of the buffer queue is Q_max bytes, and Q_used bytes have been written, then the backlog length Q = Q_used / Q_max, representing the current buffer fullness as a percentage. Edge nodes update the values ​​of B_avg, ΔB, and Q at fixed intervals (e.g., every 200 milliseconds) to ensure the real-time nature of the state information used for subsequent adjustments.

[0024] By sensing the network bandwidth jitter amplitude and the backlog length of the cache queue in real time, accurate and timely status input is provided for the joint control of the subsequent video stream frame extraction density and feature summary compression ratio. This avoids the blindness caused by control based on fixed parameters or empirical values, and enables the system to truly reflect the actual working conditions of the current network and cache.

[0025] Step S102: Based on the network bandwidth jitter amplitude and the edge buffer queue backlog length, determine the video stream frame extraction density and feature summary compression ratio after the key event is triggered.

[0026] In this step, the edge computing node dynamically calculates the video stream frame extraction density F and the feature summary compression ratio C after the key event is triggered, based on the network bandwidth jitter amplitude ΔB and the cache queue backlog length Q obtained in step S101.

[0027] First, edge computing nodes continuously run lightweight object detection models (such as compressed versions based on MobileNetV2-SSD or YOLOv5-Nano) to perform real-time inference on each frame of the video stream to detect whether preset key events (such as intrusion, falls, abnormal running, etc.) have occurred. When a key event is detected, the system immediately switches from regular monitoring mode to event response mode.

[0028] In event-response mode, edge nodes calculate the frame rate density F according to the following formula: F = F0 (ΔB / B_avg) + Q / K. Where F0 is the system's preset baseline frame sampling density (e.g., a default frame sampling rate of 5 frames per second when there are no network fluctuations and the buffer is idle), and K is the buffer scaling factor (which can be pre-calibrated according to the hardware performance of the edge nodes, e.g., K=0.8). The physical meaning of this formula is: when the network bandwidth jitter amplitude ΔB is larger relative to the average bandwidth B_avg, it indicates that the network is more unstable. In this case, the frame sampling density F should be increased to collect more key frame samples within a limited stable transmission window, preventing the loss of critical images due to sudden network interruptions. Simultaneously, when the buffer queue backlog length Q is larger, it indicates that the data to be uploaded has accumulated in large quantities at the edge nodes, facing the risk of overflow. In this case, the frame sampling density F should also be increased to accelerate the consumption and turnover of data in the buffer.

[0029] After determining the frame extraction density F, the edge nodes further calculate the feature summary compression ratio C. The edge nodes obtain the initial compression ratio C0 of the current feature extraction algorithm (e.g., a feature encoder based on a deep convolutional neural network) (e.g., a default compression ratio of 4:1), and dynamically adjust it based on the frame extraction density F and cache backlog: C = C0. (F / F_max) (Q / Q_max). Where F_max is the maximum allowed frame rate supported by the edge node hardware (e.g., limited by the encoder throughput limit, such as 30 frames per second), and Q_max is the maximum capacity of the buffer queue (in frames or bytes). This formula indicates that: when the frame rate density F increases, the amount of data to be uploaded per unit time increases. In this case, the compression ratio C should be increased accordingly to compress the size of a single frame and avoid a surge in uploaded data. When the buffer backlog Q approaches Q_max, the compression ratio should also be increased to accelerate data expulsion.

[0030] To ensure the identifiability of uploaded data, the system also sets an upper limit for the compression ratio C_upper (e.g., 16:1). When the calculated C exceeds C_upper, the edge nodes automatically call back the frame extraction density F (e.g., by proportionally reducing the F value) to avoid loss of analytical value due to severe distortion of feature information caused by excessive compression.

[0031] This step incorporates network bandwidth jitter and cache queue backlog length into a unified control formula, enabling the joint dynamic determination of frame extraction density and feature summary compression ratio. This ensures that the two coordinate and constrain each other, avoiding imbalances such as "high frame extraction rate + high compression ratio" or "low frame extraction rate + low compression ratio" that may result from single-dimensional control. It maximizes the temporal coverage density and single-frame information fidelity of key event video information under limited backhaul bandwidth.

[0032] Step S103: Perform dynamic sampling and feature extraction on the video stream based on the frame sampling density and feature summary compression ratio.

[0033] In this step, the edge computing node performs dynamic sampling and feature extraction operations on the video stream based on the frame sampling density F and feature summary compression ratio C determined in step S102.

[0034] Specifically, edge nodes continuously receive raw video frames from video acquisition modules (such as real-time video streams pushed by IP cameras via the RTSP protocol). The sampling controller within the edge node adaptively extracts frames from the input video stream using a frame extraction density F as the target sampling rate. For example, if the original video stream has a frame rate of 30 frames per second (fps), and the currently calculated frame extraction density F = 10 fps, the sampling controller extracts one frame every three frames as the frame to be processed; if F is dynamically adjusted to 15 fps, then one frame is extracted every two frames. The frame extraction strategy can employ uniform sampling or non-uniform sampling based on the inter-frame differences—increasing the sampling density appropriately during periods of rapid image change (such as when the target moves quickly) and decreasing it appropriately during periods of static image movement, but the overall average sampling rate strictly adheres to the value of F.

[0035] For each frame retained after frame extraction, the edge nodes invoke the feature extraction engine to generate a feature summary. The feature extraction engine first preprocesses the image (including size normalization, color space conversion, etc.), and then extracts high-dimensional feature vectors through a pre-trained feature encoding network (e.g., a backbone network based on ResNet-18 or MobileNetV3, which is lightweighted and deployed to the edge nodes after knowledge distillation). During the extraction process, the compression ratio C controls the quantization precision and dimensionality reduction of the feature vectors: when C is low (e.g., C=2:1), the feature extraction engine retains high-precision floating-point feature data; when C is high (e.g., C=8:1), the engine adopts a more aggressive quantization strategy (e.g., quantizing 32-bit floating-point numbers into 8-bit integers) or prunes the feature dimensions through principal component analysis (PCA) to reduce the size of the feature data per frame.

[0036] The edge nodes also encapsulate the extracted feature data according to a preset data structure. Each feature record includes a timestamp, camera ID, event type label, and corresponding feature vector ontology, which facilitates parsing and association after being received by the cloud.

[0037] Step S104: Upload the extracted video feature data to the cloud for further intelligent analysis and behavior identification.

[0038] In this step, the edge computing node uploads the video feature data encapsulated in step S103 to the cloud-based intelligent analysis platform via wired or wireless network, whereby the cloud performs further in-depth intelligent analysis and behavior discrimination.

[0039] Specifically, the edge node maintains an independent upload thread that continuously monitors the feature data packets awaiting transmission in the upload queue. When a new feature data packet is generated, the upload thread pushes it to the cloud platform's data receiving interface (e.g., a RESTful API based on HTTPS or a message queue based on MQTT) in a first-in, first-out (FIFO) order. During the upload process, the edge node monitors the time and success rate of each upload in real time. If an upload failure is detected (e.g., network timeout or error code returned by the cloud), a retransmission mechanism is automatically triggered, adding the failed data packet back to the end of the queue for a second upload.

[0040] After receiving the feature data, the cloud platform first performs data verification and decoding to reconstruct the timestamp, camera ID, event type label, and feature vector from each feature record. Subsequently, the cloud platform invokes a deep analytics model deployed in the cloud (such as a behavior recognition network based on the Transformer architecture or a multi-object association analysis model based on graph neural networks) to perform temporal modeling and contextual reasoning on the received feature sequences. The granularity of cloud-based analysis is much finer than that of edge nodes—edge nodes, limited by computing power, can typically only perform preliminary object detection and event labeling, while the cloud can fuse data reported by multiple edge nodes to perform cross-camera event correlation, behavior trajectory reconstruction, and abnormal behavior identification in complex scenes.

[0041] After the cloud-based system completes the analysis, it pushes the results (such as "intrusion confirmed, high risk level" or "fall incident, immediate action required") to the security management platform or mobile terminal for security personnel to view and handle. Simultaneously, the cloud also feeds the analysis results back to edge nodes to optimize target detection model parameters or adjust event trigger sensitivity thresholds, forming a closed-loop optimization chain of "collection-processing-upload-analysis-feedback".

[0042] In summary, the present invention achieves dynamic adaptive control of intelligent analysis tasks for security video streams under conditions of network bandwidth fluctuations and changes in cache pressure through the above steps, effectively solving the technical problem of the lack of integrity of high-value event information under limited backhaul bandwidth.

[0043] Next, a preferred embodiment of the present invention is described, which further limits the determination of frame sampling density and compression ratio based on network bandwidth jitter amplitude and edge cache queue backlog length. Specifically, after startup, the edge computing node continuously runs a network status monitoring thread and a cache status monitoring thread. The network status monitoring thread sends a lightweight probe data packet, such as a 64-byte UDP probe packet, to the cloud platform every 100 milliseconds and records the round-trip latency and the amount of data successfully received for each probe, thereby collecting multiple bandwidth samples within a sliding time window (e.g., the past 5 seconds). Let there be n valid bandwidth samples B1, B2, ..., B in this time window. n Then the edge node first calculates the current average network bandwidth B_avg=(B1+B2+…+B n ) / n, and then calculate the network bandwidth jitter amplitude ΔB=sqrt(Σ(B i -B_avg) 2 / n), which is the standard deviation of the bandwidth samples within this time window, is used to quantify the severity of bandwidth fluctuations. Simultaneously, the cache status monitoring thread reads the currently occupied bytes Q_used from the FIFO cache queue maintained in the edge node's memory, and combines this with the total cache queue capacity Q_max to calculate the backlog length Q = Q_used / Q_max, using a normalized value between 0 and 1 to represent the current cache fullness. After obtaining these three key parameters, the edge node uses the formula F = F0 The frame sampling density F is calculated using (ΔB / B_avg) + Q / K, where F0 is the system's preset baseline frame sampling density. For example, in a normal monitoring mode, if the original video stream frame rate is 30 frames / second, then F0 can be set to 5 frames / second, meaning 1 frame is sampled every 6 frames. K is the buffer scaling factor, typically ranging from 0.5 to 1.0. In this embodiment, K = 0.8. The specific value of this factor can be pre-calibrated based on the edge node's CPU processing power, memory bandwidth, and other hardware performance. The physical meaning of this formula is that when the network bandwidth jitter amplitude ΔB is larger relative to the average bandwidth B_avg, it indicates that the network is more unstable. In this case, the frame sampling density F should be increased to collect more key frame samples within a limited stable transmission window, preventing the loss of critical images due to sudden network interruptions. At the same time, when the buffer queue backlog length Q is larger, it indicates that the data to be uploaded has accumulated in large quantities at the edge node and faces the risk of overflow. In this case, the frame sampling density F should also be increased to accelerate the consumption and turnover of data in the buffer. In addition, the system presets a jitter threshold T_ΔB, for example, setting this threshold to 15% of the average bandwidth B_avg. The edge node compares the current ΔB with T_ΔB in each adjustment cycle (e.g., every 500 milliseconds). If ΔB is greater than the threshold, it is determined that the current network is in a state of severe fluctuation. The system will add an additional incremental term based on the calculation result of the above formula, for example, multiplying F by 1.2, to further increase the frame sampling frequency and ensure that enough video frames can be collected under extreme network jitter conditions, thereby ensuring the time coverage integrity of key event information.

[0044] After determining the frame extraction density F, this invention further dynamically adjusts the feature summary compression ratio C, introducing real-time network latency D_delay as another important parameter. Edge nodes record the sending timestamp t_send each time they send a probe data packet, and record the receiving timestamp t_recv when they receive an acknowledgment packet from the cloud. This allows them to calculate the single-transmission network latency D_i = t_recv - t_send. The average of all effective probe latencies within the same time window is then calculated to obtain the real-time network latency D_delay. This parameter directly reflects the current transmission latency of the network link; a higher latency indicates higher network congestion and worse data transmission timeliness. After obtaining the frame extraction density F, the edge node further obtains the system's preset initial compression ratio C0. In this embodiment, C0 is set to 4:1 by default, which means that under normal conditions, the feature data is compressed to one-quarter of the original volume. At the same time, it obtains the maximum allowable frame extraction rate F_max (limited by the edge node's video decoder throughput and neural network inference speed; for example, for devices with deployed NPU acceleration units, F_max can reach 30 frames per second) and the maximum cache capacity Q_max (for example, 2000 frames per second). Subsequently, the edge node uses the formula C=C0. (F / F_max) The dynamic compression ratio C is calculated using (Q / Q_max). This formula indicates that when the frame density F increases, the number of frames to be uploaded per unit time increases. Therefore, the compression ratio C should be increased accordingly to compress the volume of single-frame feature data, preventing a surge in the total uploaded data volume that could lead to bandwidth saturation. Similarly, when the cache backlog Q approaches Q_max, the compression ratio needs to be increased to accelerate data expulsion and alleviate cache pressure. However, excessive compression can cause severe distortion of feature information. Therefore, the system sets an upper limit for the compression ratio C_upper. In this embodiment, based on the fault-tolerance characteristics of the feature extraction algorithm, C_upper is set to 16:1. That is, when the compression ratio exceeds this value, the fidelity of the feature information will drop sharply, potentially leading to a significant decrease in the accuracy of cloud behavior discrimination. After calculating C, the edge node first checks if it is greater than C_upper. If so, a callback mechanism is triggered, and F_new = F (C_upper / C) proportionally reduces the frame sampling density F so that the recalculated compression ratio does not exceed C_upper, thereby ensuring that the feature data uploaded to the cloud still has sufficient discriminative value and avoiding the problem of information unavailability caused by excessive compression.

[0045] To further achieve scene-adaptive compression ratio control, this invention also introduces the background feature difference E_diff before a critical event occurs and the judgment of the peak period T_period. In normal monitoring mode, the edge node continuously runs a lightweight target detection and background modeling module. Specifically, the node maintains a background feature reference library, which stores the baseline background feature vectors when there is no abnormal activity in the current monitoring scene. Every 10 seconds, the node extracts the background region features of the current frame and calculates the difference E_diff with the baseline background features in the reference library. This difference can be quantified using cosine similarity or Euclidean distance between feature vectors. A larger difference indicates a more significant change in the current scene compared to the baseline background, potentially indicating an impending or ongoing critical event. Simultaneously, the edge node maintains a system clock to determine whether the current timestamp t is within the preset peak period T_period. In this embodiment, the peak period is set to 7:00 to 22:00 daily, corresponding to the time period with frequent personnel activity and a high event occurrence rate in security monitoring. Based on the two judgment dimensions mentioned above, the edge node decides the compression ratio value according to the following logic: If E_diff is greater than the preset threshold T_diff (e.g., the normalization threshold of feature difference is 0.6) and the current time is a busy period, it is determined that the current scene is in a high-incidence window of critical events, requiring higher precision feature representation to support fine analysis in the cloud. Therefore, a higher precision compression ratio C_high is enabled. In this embodiment, C_high is 2:1, that is, only light compression is performed to retain higher feature precision; otherwise, that is, if E_diff does not exceed the threshold or the current time is a non-busy period, the system defaults to using a low precision compression ratio C_low to save bandwidth resources. In this embodiment, C_low is 8:1. It should be noted that there is a priority relationship between the values ​​of C_high and C_low and the aforementioned dynamic compression ratio C: when the C_high condition is triggered, the system forces the compression ratio to be set to C_high, overriding the calculation result of C; when the C_high condition is not triggered, the system uses the calculated C value, but with C_low as its upper limit, that is, C=min(C_computed,C_low). This prioritization design ensures that feature accuracy is prioritized during peak periods of critical events, while bandwidth conservation is prioritized during non-critical periods. This maximizes the utilization efficiency of bandwidth resources while ensuring the quality of high-value event analysis.

[0046] After determining the frame extraction density and compression ratio, this invention also dynamically adjusts the upload cycle of feature data based on the cloud load status. A bidirectional communication channel is established between the edge computing nodes and the cloud platform. The cloud platform deploys a load monitoring service that collects real-time data on cloud computing resource usage, including CPU utilization, memory utilization, GPU utilization, and the current number of concurrent requests. Every second, these load status indicators are encapsulated into a status report and pushed to each edge computing node via the downlink channel. Upon receiving the cloud status report, the edge node extracts the key load status parameter S_cloud. In this embodiment, S_cloud is defined as the weighted sum of cloud CPU utilization and memory utilization, i.e., S_cloud = w_CPU. CPU_util+w_mem Mem_util, where w_CPU and w_mem are weighting coefficients, with values ​​of 0.6 and 0.4 respectively, to reflect the greater impact of CPU load on cloud processing capabilities. S_cloud ranges from 0 to 1, with higher values ​​indicating heavier cloud load. Edge nodes are defined according to the formula P=P0. (S_cloud / S_ref) determines the upload cycle P of the feature data, where P0 is the baseline upload cycle, which is set to 500 milliseconds by default in this embodiment. That is, when there is no load pressure, a batch of feature data is sent to the cloud every 500 milliseconds. S_ref is a reference load value, set to 0.5, representing a moderate level of cloud load. The physical meaning of this formula is that when the cloud load S_cloud is higher than the reference value S_ref, it indicates that the cloud processing capacity is approaching saturation. At this time, the upload cycle P should be extended and the data push frequency reduced to avoid data backlog or request timeouts due to the cloud's inability to process data in time. Conversely, when the cloud load is lower than the reference value, the upload cycle P should be shortened to speed up the data push speed, so that the cloud can obtain the latest video feature data for real-time analysis in a timely manner. Meanwhile, the system also sets a minimum upload period P_min, which is set to 100 milliseconds in this embodiment. That is, the minimum upload period is not less than 100 milliseconds, in order to prevent excessively frequent upload requests from consuming too much network connection resources and cloud API call quota. After calculating P, the edge node first determines whether it is less than P_min. If so, it forces P to be set to P_min, thereby achieving a reasonable balance between real-time performance and resource consumption.

[0047] To further ensure priority transmission of high-value events when bandwidth is insufficient, this invention also introduces a transmission priority evaluation mechanism based on key event types. Edge nodes deploy a multi-category event detection model for security scenarios. This model, based on lightweight deep learning architectures such as MobileNetV2-SSD or YOLOv5-Nano, has been fine-tuned for security scenarios and can identify various types of key events, including intrusion detection (person crossing the cordon), fall detection (person falling), abnormal behavior detection (running, fighting, gathering), and vehicle violation detection (illegal parking, driving against traffic). Edge nodes perform inference on each frame of video. When any of the above events is detected, an event tag `Event_type` is generated, and the timestamp and spatial location information of the event are recorded. After determining the Event_type, the edge node queries a pre-defined event type priority mapping table. This table assigns a priority score, Event_type_priority, to each event type. A higher value indicates that the event requires priority in ensuring the integrity of data transmission. For example, a fall event, involving emergency rescue for personal safety, has a priority score of 10; an intrusion event has a priority score of 8; abnormal behavior (such as fighting) has a priority score of 7; and traffic violations have a priority score of 5. The edge node further uses the formula P_prior=α... (Event_type_priority+β The transmission priority P_prior is calculated using ΔB^γ, where α, β, and γ are weighting coefficients, which in this embodiment are α=0.1, β=0.5, and γ=0.8 respectively, and ΔB is the current network bandwidth jitter amplitude. The design idea behind this formula is that transmission priority depends not only on the importance of the event type itself but also on the modulation of the current network fluctuations. When network jitter intensifies, the uncertainty of data transmission increases, requiring a further increase in the transmission priority of high-priority events to ensure they receive priority service within limited transmission resources. At the beginning of each upload cycle, the edge node traverses all feature data packets waiting to be sent in the upload queue, sorts them according to the P_prior value corresponding to each data packet, and places high-priority data packets at the head of the queue for priority upload. Simultaneously, when calculating the frame density according to the aforementioned formula, for high-priority events (i.e., events where P_prior exceeds a preset threshold), the system adds an incremental term to the calculation result of F. Specifically, if P_prior>5, then F=F_prior. (1+0.2 (P_prior-5)) to ensure that high-priority events have a higher sampling density, so that when network bandwidth is insufficient, the most critical security event information can be automatically identified and prioritized to ensure that it is not lost due to bandwidth contention.

[0048] Regarding cache management at edge nodes, this invention further designs a multi-level cache structure and a cache decision mechanism based on historical packet loss rates. Edge nodes initialize three levels of cache in memory: a fast cache (R_Cache), a medium-speed cache (M_Cache), and a slow cache (L_Cache). The fast cache uses high-speed SRAM or high-performance DDR memory areas, offering the lowest read / write latency (microseconds) but with a small capacity (e.g., 50MB), used to store the most urgent and immediately uploadable feature data. The medium-speed cache uses conventional memory areas, with moderate read / write latency (milliseconds) and a medium capacity (e.g., 200MB), used to store feature data with ordinary priority. The slow cache can use slower storage media such as eMMC or SD cards, with higher read / write latency (tens of milliseconds) but the largest capacity (e.g., 2GB), used to store non-urgent or delayable upload feature data. During operation, edge nodes continuously monitor the historical packet loss rate (Packet_loss_rate). Specifically, after each data packet is uploaded, the node determines whether the data packet has arrived successfully based on the confirmation information returned by the cloud. If no confirmation is received within a specified timeout period, it is counted as a packet loss. Within a sliding time window (e.g., the past 60 seconds), the total number of packets sent and the number of packets lost are counted, and the packet loss rate is calculated as Packet_loss_rate = Lost_packets / Total_packets. The edge nodes make cache level decisions based on the packet loss rate: if Packet_loss_rate is greater than the preset threshold Threshold_loss (set to 5% in this embodiment), it indicates that the current network quality is poor and packet loss is severe. In this case, the system will prioritize storing newly generated feature data in the medium-speed cache M_Cache instead of the fast cache R_Cache. This is because even if the data is stored in the fastest cache in a high packet loss environment, successful transmission cannot be guaranteed. Therefore, it is unnecessary to occupy scarce fast cache resources. Storing the data in the medium-speed cache allows for batch uploading when the network improves slightly. Meanwhile, when the slow cache L_Cache still has storage space, the system temporarily stores some non-urgent feature data in L_Cache. Specifically, edge nodes classify data packets in the upload queue according to priority P_prior. Data packets with a priority lower than a certain threshold (e.g., P_prior<3) are considered non-urgent data. When the remaining space of L_Cache is greater than 10%, these non-urgent data are transferred to L_Cache for temporary storage to free up space in the fast cache and medium cache for higher priority data. Thus, the hierarchical classification storage of the three-level cache effectively avoids the overall congestion problem faced by a single cache structure when the network deteriorates.

[0049] Considering that different users in a security monitoring system have different data access permissions, this invention also incorporates user permission levels into the adjustment process of the feature digest compression ratio. During system initialization, the edge node obtains the current user's access permission identifier U_level from the cloud-based permission management service. Permission levels are divided into management level (U_level=2), viewing level (U_level=1), and prohibited access (U_level=0). Management level users have the highest data access permissions, able to obtain the highest precision feature data and even the original video. Viewing level users can only view the compressed feature digest, and prohibited access users have no right to obtain any video data. After calculating the dynamic compression ratio C according to the aforementioned formula, the edge node further sets a compression ratio limit C_limit = C0 + (U_level) based on the user's permission level. λ), where C0 is the initial compression ratio (as mentioned above, 4:1), and λ is the incremental factor corresponding to each level of permission. In this embodiment, λ is set to 2, so for management users, C_limit = 4 + 2 2 = 8:1, for viewer-level users C_limit = 4 + 1 2=6:1, for users with restricted access C_limit=4:1. When determining the final compression ratio, the edge node executes the following judgment logic: If the currently calculated compression ratio C is less than C_limit, meaning the compression level is lower than the upper limit allowed by permissions, then a higher-level feature summary compression strategy is forced. Specifically, the system increases the compression ratio to C_limit, compressing data as much as possible within the permission limits to save bandwidth; if C is greater than C_limit, meaning the compression level exceeds the upper limit allowed by permissions, then the frame extraction density F is reduced to ensure a balance between security and integrity. Specifically, the system follows the formula F_new=F (C_limit / C) Callback frame extraction density. This design ensures that high-privilege users can obtain more precise feature data (lower compression ratio), while low-privilege users can only obtain highly compressed feature summaries, thus achieving a balance between data transmission security and information integrity, and realizing fine-grained data control based on security level.

[0050] To address extreme network fluctuations in specific scenarios, this invention also incorporates a dynamic compensation mechanism. Edge nodes continuously monitor network bandwidth fluctuations over the past 5 minutes, recording the current instantaneous bandwidth value B_instant at fixed sampling intervals (e.g., every 1 second). A sequence containing 300 bandwidth samples is maintained within a 5-minute sliding time window. Then, the standard deviation of the bandwidth within this time window, σ_Bandwidth = sqrt(Σ(B_instant)), is calculated. i -B_mean) 2 / N), where B_mean is the mean of the bandwidth samples within the time window, N=300 is the total number of samples, and σ_Bandwidth reflects the fluctuation range of network bandwidth in the past 5 minutes. The larger the value, the more unstable the network. The edge node compares σ_Bandwidth with the preset bandwidth fluctuation threshold threshold_B. In this embodiment, threshold_B is set to 20% of the average bandwidth B_mean. If σ_Bandwidth > threshold_B, it is determined that the current network is in a state of severe fluctuation, triggering a dynamic compensation mechanism and setting activate_compensate=true; otherwise, it is false. When the compensation flag is true, the edge node starts an additional feature compression algorithm, according to C_comp=C_base. The compensation compression ratio C_comp is calculated using (σ_Bandwidth / σ_ref), where C_base is the currently used baseline compression ratio (i.e., the previously calculated C value), and σ_ref is the reference standard deviation of fluctuation (set to 10% of the average bandwidth in this embodiment). This formula means that the more severe the bandwidth fluctuation, the higher the additional compression ratio, i.e., further compression of feature data to cope with extreme network instability. Simultaneously, the system also increases the keyframe extraction frequency, according to F_key=F_base. (σ_Bandwidth / σ_ref) adjusts the keyframe sampling frequency, where F_base is the current baseline sampling frequency (i.e., the F value calculated above). This means that when network fluctuations intensify, the system not only increases the overall sampling density, but also specifically increases the sampling frequency of key frames (such as frames containing moving targets or event-triggered areas) to ensure that critical information is not missed under extreme network conditions. Thus, the system's adaptability in the face of severe network fluctuations is significantly enhanced through additional compression and sampling dual compensation strategies.

[0051] In multi-camera collaborative monitoring scenarios, this invention also achieves cross-camera video stream alignment based on a time synchronization mechanism. When an edge node connects to multiple cameras, it first performs clock synchronization calibration on each camera using Precision Time Protocol (PTP) or Network Time Protocol (NTP), recording the deviation between the system clock of each camera and the reference clock of the edge node. Let the system clock of camera i be T_i, and the reference clock of the edge node be T_ref, then the timestamp difference T_offset_i = T_i - T_ref. The edge node maintains an independent timestamp difference record for each camera. Based on the timestamp differences of each camera, the edge node establishes a synchronization matrix S_matrix to align the content of multiple video streams. This matrix is ​​an N×N square matrix (N is the number of cameras), and the matrix element S_matrix[i][j] = T_offset_i - T_offset_j represents the timestamp difference between camera i and camera j. Through this matrix system, the frames of any two video streams can be aligned according to the actual acquisition time. During video stream alignment, edge nodes calculate the image offset compensation factor according to compensation_factor=max(0,(T_offset / T_sync_window)), where T_sync_window is the synchronization window length, set to 100 milliseconds in this embodiment, representing the maximum allowed synchronization deviation window. The value of compensation_factor ranges from 0 to 1, with a larger value indicating a more severe time deviation between the two video streams. When compensation_factor exceeds the set threshold T_comp_max (set to 0.8 in this embodiment, corresponding to an 80-millisecond deviation), the system initiates frame synchronization adjustment. Specifically, in the video streams from cameras with large deviations, the two streams are aligned in time by dropping or inserting frames. For example, if camera i is ahead of camera j by T_offset milliseconds, several frames within the corresponding time window are discarded in the video stream of camera i, causing the two streams to realign in time. This provides a reliable data foundation for cross-camera event correlation analysis and multi-view behavior trajectory reconstruction in the cloud, avoiding mis-association of events or trajectory breakage caused by time asynchrony between cameras.

[0052] Furthermore, this invention fully considers the limitations of the edge device's own computing power and designs an adaptive strategy based on computing resources. Edge nodes obtain real-time resource status information such as CPU utilization and memory usage through system call interfaces provided by the operating system (e.g., reading files like / proc / stat and / sys / fs / cgroup / cpuacct in a Linux environment), and calculate the available computing power A_avail. In this embodiment, A_avail is defined as a quantified value of the current idle CPU computing power, measured in floating-point operations per second (FLOPS). This is achieved by reading the CPU's clock frequency, number of cores, and current utilization, according to the formula A_avail = CPU_cores. CPU_freq The CPU utilization is estimated using (1-CPU_util), where CPU_cores is the number of CPU cores, CPU_freq is the current operating frequency (GHz), and CPU_util is the CPU utilization rate (between 0 and 1). Edge nodes are further evaluated for resource utilization using Utilization_ratio = A_avail / A_total, where A_total is the total available computing power of the edge node (i.e., the theoretical peak computing power when the CPU is running at full speed), and Utilization_ratio ranges from 0 to 1, with smaller values ​​indicating more strained computing resources. Based on resource utilization, edge nodes are further evaluated using F_new = F_old. (1-Utilization_ratio)+F_base Utilization_ratio automatically adjusts the frame sampling density and compression ratio, where F_old is the current frame sampling density and F_base is the baseline frame sampling density (same as F0 mentioned above). The meaning of this formula is that when the resource utilization Utilization_ratio is low (i.e., computing power is tight), F_new approaches F_old (maintaining the current frame sampling density), and when the resource utilization is high (i.e., computing power is sufficient), F_new approaches F_base (adopting the baseline frame sampling density). In fact, the effect of this formula is that when computing power is tight, the system tends to maintain a lower frame sampling density to reduce the computational burden, while when computing power is sufficient, the frame sampling density can be increased to collect more frame data. Similar adjustments are also applied to the calculation of the compression ratio. Meanwhile, when the Utilization_ratio is lower than the preset low computing power threshold (e.g., 0.3), the system automatically disables computationally intensive feature extraction operations, such as not enabling fine-grained feature encoding based on attention mechanisms, and instead adopts lighter-weight feature extraction methods (e.g., extracting only traditional features such as HOG or color histograms), to ensure that the system can still maintain basic video processing and uploading functions when computing power is limited, thereby effectively avoiding service crashes caused by overload at edge nodes.

[0053] Corresponding to the above methods, such as Figure 2As shown, the present invention also provides a security video stream intelligent analysis system 200 based on edge computing and cloud collaboration. The system includes an acquisition module 201, a determination module 202, a sampling and feature extraction module 203, and an upload module 204. These four modules are deployed on edge computing nodes and interact with each other through inter-process communication or shared memory to collaboratively complete the intelligent analysis task of the video stream. The acquisition module, serving as the system's data perception layer, is responsible for acquiring the current network bandwidth jitter and edge cache queue backlog length. This module includes a network status monitoring subcomponent and a cache status monitoring subcomponent. The network status monitoring subcomponent runs as an independent thread, sending probe data packets to the cloud platform every 100 milliseconds, recording the round-trip latency and throughput of each probe, and calculating the average network bandwidth B_avg and jitter ΔB within a sliding time window. The cache status monitoring subcomponent calculates the cache queue backlog length Q=Q_used / Q_max by reading the number of currently occupied bytes in the cache queue in the edge node's memory. The acquisition module packages the collected B_avg, ΔB, and Q parameters into a status information structure and publishes it to the determination module via shared memory. The determination module, as the core of the system's decision-making, is responsible for determining the frame extraction density and feature summary compression ratio of the video stream after a critical event is triggered, based on the network bandwidth jitter amplitude and cache queue backlog length provided by the acquisition module. This module first calls a lightweight event detection model (such as a compressed version based on YOLOv5-Nano) deployed on edge nodes to perform real-time inference on each frame in the video stream to detect whether a preset critical event has occurred. When a critical event is detected, the system switches from the regular monitoring mode to the event response mode. In the event response mode, the determination module calculates the frame extraction density F and the feature summary compression ratio C according to the aforementioned formula, and finely adjusts F and C according to various additional conditions. Finally, the values ​​of F and C are written into the control parameter area in shared memory for the sampling and feature extraction module to read. The sampling and feature extraction module, as the system's execution layer, is responsible for dynamically sampling and extracting features from the video stream based on the frame extraction density and feature summary compression ratio output by the determined module. This module includes a dynamic sampling sub-component and a feature extraction sub-component. The dynamic sampling sub-component continuously receives raw video frames from the video acquisition module (such as a real-time video stream pushed by an IP camera via the RTSP protocol) and adaptively extracts frames from the input video stream using the frame extraction density F as the target sampling rate. The feature extraction sub-component calls a feature encoding network (such as a backbone network based on MobileNetV3, which is lightweight and deployed on edge nodes after knowledge distillation) to extract high-dimensional feature vectors for each frame image retained after frame extraction. During the extraction process, the quantization accuracy and dimensionality reduction of the feature vectors are controlled according to the compression ratio C. The extracted feature data is encapsulated according to a preset data structure. Each feature record includes a timestamp, camera ID, event type label, and the corresponding feature vector body. After encapsulation, the data packet is pushed into the upload queue.As the system's output layer, the upload module is responsible for uploading the feature data generated by the sampling and feature extraction modules to the cloud for further intelligent analysis and behavior discrimination. This module maintains an independent upload thread that continuously monitors the feature data packets to be sent in the upload queue. When a new feature data packet is generated, the upload thread pushes the data packet to the cloud platform's data receiving interface (such as a RESTful API based on the HTTPS protocol or a message queue based on the MQTT protocol) according to the upload cycle P that is dynamically adjusted based on the cloud load status. During the upload process, the upload module monitors the time consumption and success rate of each upload in real time. If an upload failure is detected (such as network timeout or error code returned by the cloud), a retransmission mechanism is automatically triggered, and the failed data packet is added back to the end of the queue to wait for a second upload. After receiving the feature data, the cloud platform first performs data verification and decoding to restore the timestamp, camera ID, event type label, and feature vector of each feature record. Then, it calls the deep analysis model deployed in the cloud (such as a behavior recognition network based on the Transformer architecture or a multi-object association analysis model based on graph neural networks) to perform temporal modeling and contextual reasoning on the received feature sequence. The analysis granularity of the cloud is more refined than that of the edge, and it can integrate data reported by multiple edge nodes to perform cross-camera event association, behavior trajectory reconstruction, and abnormal behavior identification in complex scenarios. After the cloud completes the analysis, it pushes the identification results (such as "confirmed intrusion behavior, high risk level" or "fall event, requires immediate action") to the security management platform or mobile terminal for security personnel to view and handle. At the same time, it feeds back the analysis results to the edge nodes to optimize the target detection model parameters on the edge side or adjust the sensitivity threshold of event triggering, thus forming a closed-loop optimization link of "collection-processing-upload-analysis-feedback". Through the clear division and organic coordination of the above four functional modules, the system constructs a complete "perception-decision-execution-upload" architecture. The acquisition module ensures the system's real-time perception of network and cache status, the determination module realizes intelligent decision-making based on multi-dimensional status information, the sampling and feature extraction module completes the efficient simplification and information condensation of video data, and the upload module ensures the reliable transmission of feature data to the cloud. The collaborative work among the four enables the system to automatically complete the dynamic adaptive adjustment of security video stream analysis tasks under complex conditions of network bandwidth fluctuations and cache pressure changes, effectively solving the technical problem of the lack of integrity of high-value event information under limited backhaul bandwidth.

[0054] The exemplary systems and methods of the present invention have been specifically shown and described with reference to the above embodiments, which are merely examples of the best mode for implementing the systems and methods. Those skilled in the art will understand that various changes can be made to the embodiments of the systems and methods described herein without departing from the spirit and scope of the invention as defined in the appended claims when implementing the systems and / or methods.

Claims

1. A security video stream intelligent analysis method based on edge computing and cloud collaboration, characterized in that, include: Get the current network bandwidth jitter level and edge buffer queue backlog length; Based on the network bandwidth jitter amplitude and the edge buffer queue backlog length, the video stream frame extraction density and feature summary compression ratio after the key event is triggered are determined. Dynamic sampling and feature extraction are performed on the video stream based on the frame extraction density and feature summary compression ratio. The extracted video feature data is uploaded to the cloud for further intelligent analysis and behavior identification.

2. The intelligent analysis method for security video streams based on edge computing and cloud collaboration according to claim 1, characterized in that, The method of determining the frame extraction density and compression ratio based on network bandwidth jitter amplitude and edge buffer queue backlog length further includes: Obtain the current average network bandwidth B_avg and jitter amplitude ΔB; Get the backlog length Q of the edge cache queue; Adjust the frame rate density F=F0 according to the following formula. (ΔB / B_avg)+Q / K, where F0 is the baseline frame sampling density and K is the buffer scaling factor; When ΔB is greater than the preset threshold T_ΔB, the value of F is increased to improve the frame extraction frequency.

3. The intelligent analysis method for security video streams based on edge computing and cloud collaboration according to claim 1, characterized in that, The method of determining the frame extraction density and compression ratio based on network bandwidth jitter amplitude and edge buffer queue backlog length further includes: Get the real-time network delay D_delay; Combining the previously calculated frame extraction density F and feature summary compression ratio C; Dynamically adjust C=C0 according to the following formula. (F / F_max) (Q / Q_max), where C0 is the initial compression ratio, F_max is the maximum allowed frame rate, and Q_max is the maximum buffer capacity; When C exceeds the set upper limit C_upper, the value of F is reduced to ensure the integrity of uploaded data.

4. The intelligent analysis method for security video streams based on edge computing and cloud collaboration according to claim 1, characterized in that, The dynamic sampling of the video stream based on the frame extraction density and feature summary compression ratio further includes: Calculate the background feature difference E_diff before the occurrence of the critical event; Determine whether it is in a peak period T_period by combining the current timestamp t; If E_diff > threshold T_diff and T_period is a busy period, then enable a higher precision compression ratio C_high; Otherwise, a low-precision compression ratio C_low is used by default to save bandwidth.

5. The intelligent analysis method for security video streams based on edge computing and cloud collaboration according to claim 1, characterized in that, Uploading the extracted video feature data to the cloud further includes: Get the current cloud platform load status S_cloud; The upload period P is determined based on the following formula: P=P0 (S_cloud / S_ref), where P0 is the baseline upload period and S_ref is the reference load value; When S_cloud < reference load value, reduce P to shorten the upload interval and speed up the response time; If P is less than the lower limit P_min, then P_min is forcibly maintained to avoid resource abuse.

6. The intelligent analysis method for security video streams based on edge computing and cloud collaboration according to claim 1, characterized in that, The scenario segmentation based on key event triggers further includes: Detect key event types and mark them as Event_type; Determine whether the original image or only feature information needs to be retained based on the Event_type; The transmission priority is evaluated using the following formula: P_prior=α (Event_type_priority+β ΔB^γ), where α, β, and γ are weighting coefficients; High-priority events are uploaded first when bandwidth is insufficient, and the frame rate is increased to ensure integrity.

7. The intelligent analysis method for security video streams based on edge computing and cloud collaboration according to claim 1, characterized in that, The deployment of a multi-level caching structure at the edge nodes further includes: Initialize a three-level cache, including a fast cache (R_Cache), a medium cache (M_Cache), and a slow cache (L_Cache). The cache level is determined based on the historical packet loss rate (Packet_loss_rate). The following formula is used for caching decisions: if Packet_loss_rate > Threshold_loss, then use M_Cache; otherwise, use R_Cache. When L_Cache still has storage space, some non-urgent feature data will be temporarily stored in L_Cache.

8. The intelligent analysis method for security video streams based on edge computing and cloud collaboration according to claim 1, characterized in that, The feature summary compression ratio adjustment process, which incorporates user permission levels, further includes: Collect the user's access permission identifier U_level; Set the compression ratio limit C_limit=C0+(U_level) λ), where λ is the increment factor corresponding to each level of permission; If C<C_limit, a higher-level feature summary compression strategy is forcibly adopted; If C>C_limit, the frame sampling density is reduced to ensure the balance between security and integrity.

9. The intelligent analysis method for security video streams based on edge computing and cloud collaboration according to claim 1, characterized in that, The optimization measures for network fluctuations in specific scenarios further include: Calculating the bandwidth standard deviation σ_Bandwidth within the past 5 minutes; Determining whether to trigger the dynamic compensation mechanism based on the following formula: if σ_Bandwidth>threshold_B, activate_compensate is true, otherwise it is false; When compensate is true, initiate an additional feature compression algorithm: C_comp = C_base (σ_Bandwidth / σ_ref); Simultaneously increase the keyframe extraction frequency F_base (σ_Bandwidth / σ_ref) to avoid losing critical information.

10. A security video stream intelligent analysis system based on edge computing and cloud collaboration, characterized in that, comprising: an acquisition module, configured to acquire the current network bandwidth jitter amplitude and the backlog length of the edge cache queue; a determination module, configured to determine the frame sampling density of the video stream and the compression ratio of the feature summary after the key event is triggered, based on the network bandwidth jitter amplitude and the backlog length of the edge cache queue; a sampling and feature extraction module, configured to perform dynamic sampling and feature extraction on the video stream according to the frame sampling density and the compression ratio of the feature summary; an upload module, configured to upload the extracted video feature data to the cloud for further intelligent analysis and behavior discrimination.

Citation Information

Patent Citations

  • Video analysis method and device based on edge cloud collaboration, electronic equipment and medium

    CN115103157A