Cloud computing-based video monitoring big data analysis method

Abstract

This invention discloses a cloud-based video surveillance big data analysis method, belonging to the interdisciplinary field of video surveillance, cloud computing, and big data analysis. The method includes distributed video data acquisition, edge-end collaborative preprocessing, transmission and storage under a cloud computing architecture, cloud-based intelligent analysis, dynamic resource scheduling, multi-terminal feedback, and adaptive iterative optimization of the model. Video data is acquired through heterogeneous cameras, with keyframe extraction, noise reduction, format standardization, and privacy anonymization performed at the edge. The data is then encrypted and fragmented before being transmitted to the cloud. An improved temporal enhancement analysis model performs multi-dimensional analysis, including target detection and tracking. A dynamic scheduling algorithm optimizes resource utilization, and results are synchronously fed back from multiple terminals, iteratively optimizing the model. This invention solves the problems of high latency and insufficient privacy protection in existing methods, reduces cloud pressure, improves resource utilization and model adaptability, ensures data security and reliability, and is adaptable to various large-scale monitoring scenarios, demonstrating strong practicality.

Description

Technical Field

[0001] This invention belongs to the field of interdisciplinary technology of video surveillance, cloud computing and big data analysis. Specifically, it relates to a cloud-based video surveillance big data analysis method, which is applicable to large-scale video surveillance data analysis in various scenarios such as urban security, park management, transportation hubs, and commercial complexes. It can realize efficient processing, intelligent analysis and multi-terminal collaborative feedback of video data, and improve the intelligence level and resource utilization of the monitoring system. Background Technology

[0002] With the rapid development of the security industry, video surveillance systems have been widely used in various scenarios, resulting in an explosive growth in the number of surveillance cameras and a dramatic increase in the amount of video data generated, forming massive video surveillance big data. Most existing video surveillance analysis methods have the following shortcomings: 1. Low data processing efficiency: Traditional methods mostly adopt a centralized processing architecture, in which all video data is transmitted to the cloud for processing. The computation of massive data transmission puts a huge pressure on the cloud, resulting in high analysis latency and failing to meet the needs of real-time monitoring. Especially in scenarios with fast-moving targets and multiple concurrent targets, the false detection and false detection rates remain high. 2. Insufficient privacy protection: Video surveillance data often contains sensitive personal information such as faces and license plates. Most existing methods do not perform targeted desensitization processing on sensitive areas, or the desensitization methods are too simple, which can easily lead to the leakage of personal privacy and does not comply with relevant laws and regulations. 3. Low resource utilization: Cloud computing and storage resources are allocated in a fixed manner and cannot be dynamically adjusted according to the priority of analysis tasks and real-time load. This results in urgent tasks not receiving sufficient resources, while idle resources cannot be effectively recycled, leading to resource waste. 4. Poor model adaptability: Most existing analysis models are fixed parameter models. After training, they cannot adaptively iterate and optimize according to changes in actual application scenarios and historical analysis data. After long-term use, the analysis performance will gradually decline, making it difficult to adapt to the monitoring needs of different scenarios and environments. 5. Inappropriate data transmission and storage: Massive amounts of video data are transmitted directly without lightweight processing, making them susceptible to network bandwidth fluctuations, which can lead to transmission interruptions and data loss. At the same time, the data is not managed in a hierarchical manner, and all data is stored in a uniform way, which increases storage costs and affects the access efficiency of high-frequency data. To address the aforementioned issues, there is an urgent need for a video surveillance big data analysis method that can balance real-time performance, privacy protection, resource utilization, and model adaptability, combining the advantages of cloud computing and edge computing to achieve efficient and intelligent analysis of massive video data. Summary of the Invention

[0003] The purpose of this invention is to overcome the shortcomings of the prior art and provide a cloud computing-based video surveillance big data analysis method. This method solves problems such as high data processing latency, insufficient privacy protection, low resource utilization, and poor model adaptability in existing methods. It enables efficient transmission, storage, and intelligent analysis of video surveillance big data, improves the intelligence level and practicality of the monitoring system, and complies with relevant regulations on personal information protection.

[0004] To address the aforementioned technical problems, this invention provides the following technical solution: a cloud-based video surveillance big data analysis method, specifically comprising the following steps: S1, Distributed Video Data Acquisition A distributed acquisition network composed of multiple heterogeneous surveillance cameras is constructed. These cameras include various types such as high-definition bullet cameras, PTZ cameras, and dome cameras, adaptable to different monitoring scenarios including indoor and outdoor environments, and high-light and low-light conditions. They support video stream acquisition at different resolutions, such as 1080P and 4K. The acquisition network is compatible with multiple mainstream video transmission protocols, including RTSP and HTTP, and can automatically select the optimal transmission protocol based on camera type and network environment. It also features adaptive online / offline capabilities, monitoring camera status in real time through a heartbeat detection mechanism. When a camera is detected to be offline, video data is automatically temporarily stored in the camera's local cache module (cache capacity not less than 128GB). When the camera returns to online status, it is automatically synchronized to the edge computing nodes in chronological order, ensuring no data loss. S2, Edge-end Collaborative Preprocessing In edge computing nodes deployed near surveillance cameras, the video stream data collected in step S1 undergoes collaborative preprocessing to reduce data redundancy and alleviate cloud transmission and computing pressure. The edge computing nodes employ a lightweight hardware architecture and can be deployed in surveillance cabinets or nearby network server rooms. A single edge node can access data from 10-50 surveillance cameras. The preprocessing process specifically includes: (1) Keyframe extraction: The keyframe extraction is achieved by combining the inter-frame difference method with entropy analysis. The absolute value of the inter-frame difference between two adjacent video frames is calculated. The preset inter-frame difference threshold is 5-15 (which can be dynamically adjusted according to the scene). At the same time, the preset effective entropy value range is 3-8. When the absolute value of the inter-frame difference is greater than the preset threshold and the entropy value of the frame is within the preset effective entropy value range, the frame is determined as a keyframe. The keyframe extraction frequency can be dynamically adjusted according to the video scene. The dynamic range is 1-10 frames / second. The extraction frequency is set to 8-10 frames / second for scenes with dense crowds and frequent target movement (such as transportation hubs), and 1-3 frames / second for scenes with sparse crowds and stable scenes (such as park walls). (2) Gaussian denoising: A Gaussian filter with a standard deviation of 0.5-2.0 is used to denoise the extracted key frames, remove random noise and environmental interference generated during video acquisition, and improve the clarity of the key frames. (3) Format standardization: Convert keyframes with different resolutions and encoding formats (such as H.264 and H.265) into 1080P resolution and H.265 encoding format to ensure compatibility and consistency of subsequent cloud analysis. (4) Privacy desensitization processing: By pre-setting the coordinates of sensitive areas (such as areas where faces and license plates often appear in the monitoring screen) or using a lightweight intelligent recognition algorithm, sensitive areas such as faces and license plates in key frames of the video are located; an adaptive dynamic blurring algorithm is used to desensitize the sensitive areas, and the size of the blur kernel is dynamically adjusted according to the size of the sensitive area (the size of the blur kernel ranges from 3×3 to 11×11). At the same time, the positions of sensitive areas in adjacent frames are matched by IoU to ensure the continuity of desensitized areas in adjacent frames and avoid flickering of desensitized areas; the desensitization process complies with the anonymization requirements of the Personal Information Protection Law of the People's Republic of China to ensure that sensitive personal information is not leaked. Through the above preprocessing, lightweight preprocessed data is obtained (the data volume is reduced by 60%-80% compared to the original video stream), laying the foundation for subsequent cloud transmission and analysis. S3, Transmission and Storage under Cloud Computing Architecture The preprocessed data obtained in step S2 is encrypted using the AES-256 encryption algorithm to ensure data security during transmission and prevent data theft or tampering. Combined with an adaptive fragmentation transmission strategy, the data fragment size is dynamically adjusted based on real-time network bandwidth detection results: 128MB for bandwidth greater than 100Mbps; 64MB for bandwidth between 10-100Mbps; and 32MB for bandwidth less than 10Mbps. A retransmission mechanism is also employed; if a fragment transmission failure is detected, the fragment is automatically retransmitted to ensure transmission stability. The encrypted, fragmented, preprocessed data is transmitted to a cloud computing platform. This platform employs a distributed architecture, including compute nodes, storage nodes, and scheduling nodes. The storage nodes utilize a distributed file system (such as HDFS or a similar distributed storage architecture) to fragment the data. A tiered management mechanism for hot and cold data is also implemented. Hot data is defined as data accessed or used for real-time analysis within the last 7 days; this data is stored on hot data nodes (using SSDs with an access latency of ≤10ms). Historical archived data (data not accessed for more than 7 days and not requiring real-time analysis) is stored on cold data nodes (using hard disk drives), achieving a balance between efficient data access and low-cost storage.

[0005] S4, cloud-based video big data intelligent analysis The cloud computing platform invokes an improved time-series augmented analytics model to perform multi-dimensional intelligent analysis on the stored preprocessed data. This analysis includes target detection, target tracking, abnormal behavior identification, and scene traffic statistics, meeting the needs of different monitoring scenarios. (1) Improved temporal enhancement analysis model: Based on the optimization of YOLOv8 model, the core improvements include two parts: First, a temporal attention mechanism is added to the feature extraction network (CSPDarknet) of YOLOv8 to capture the target motion features and correlations of adjacent video key frames, thereby improving the feature extraction capability in fast target motion and occlusion scenarios; Second, a generative adversarial refinement module is introduced in the detection head to optimize the detected target bounding box and segmentation mask, thereby reducing the detection error in target occlusion and complex background scenarios, so that the model inference speed is not less than 30fps and the target detection accuracy is not less than 95%.

[0006] (2) Target detection and tracking: By using an improved temporal enhancement analysis model, targets such as people, vehicles, and objects in video frames are detected, and features such as the size, color, and outline of the targets are extracted. The Kalman filter algorithm is combined with target feature matching to achieve continuous tracking of the targets. Even when the targets are briefly occluded or the scene changes, the tracking continuity can be maintained, and the tracking accuracy is not less than 90%. (3) Abnormal behavior identification: preset rules for various abnormal behaviors such as climbing over, gathering (more than the preset number of people, such as 10 people), leaving objects (staying for more than the preset time, such as 5 minutes), and running fast (speed exceeding the preset threshold, such as 5m / s); extract the target's movement trajectory, speed and behavior characteristics, and perform similarity matching with the preset rules. When the matching degree exceeds 85%, it is judged as abnormal behavior, and the time, location and duration of the abnormal behavior are recorded. (4) Scene traffic statistics: Based on the target detection and tracking results, the number of people and vehicles in the monitored scene per unit time (e.g., per minute) is counted, and traffic statistics reports are generated. Statistical queries can be performed by time period and target type to provide data support for monitoring and management decisions. S5, dynamic scheduling of cloud computing resources The scheduling nodes of the cloud computing platform monitor the CPU utilization, memory usage and network bandwidth of each computing node in real time, with a monitoring frequency of 10 seconds per time, to obtain the real-time load status of each node; at the same time, combined with the dynamic changes in the priority of the analysis tasks in step S4, the analysis tasks are preset into three levels: urgent, normal and low priority. The urgent level corresponds to the real-time detection of abnormal behavior, the normal level corresponds to the routine target statistics task, and the low priority corresponds to the historical data review task. Through an intelligent scheduling algorithm (optimized based on a genetic algorithm), computing and storage resources are dynamically allocated: when the CPU utilization of a computing node exceeds 80% or the memory occupancy exceeds 75%, low-priority tasks on that node are migrated to idle computing nodes; sufficient computing and storage resources are allocated to urgent tasks first, ensuring that the execution latency of urgent anomaly detection tasks does not exceed 500ms; idle computing and storage resources are automatically reclaimed and allocated to high-load nodes or cached high-frequency access data to improve resource utilization, thereby increasing cloud resource utilization to over 80%. S6. Analysis Results Output and Multi-terminal Feedback The analysis results obtained in step S4 (including target detection results, tracking trajectory, abnormal behavior alarm information, and traffic statistics) are formatted to generate a standardized analysis report, which is then pushed to the monitoring terminal and mobile terminal (such as mobile phone or tablet) in real time. (1) Monitoring terminal: Deployed in the monitoring center, it supports multi-screen synchronous display (can display 16-64 monitoring screens and corresponding analysis results at the same time), abnormal event early warning settings (can customize the alarm threshold and alarm mode of abnormal behavior). When a preset abnormal behavior is detected, it triggers an audible and visual alarm to remind the monitoring personnel to handle it in time; it also supports local storage and query of analysis results. (2) Mobile terminal: Receives analysis results through a dedicated APP, supports viewing abnormal segments, previewing real-time monitoring images and remotely controlling camera rotation (such as adjusting the PTZ camera angle and zooming), enabling mobile monitoring personnel to keep abreast of abnormal situations in the monitoring scene even when they are not in the monitoring center. Meanwhile, the analysis results are associated with and stored with the corresponding video keyframes and original video clips, supporting multi-condition retrieval and review by time, scene, target type, and abnormal behavior type, making it convenient for monitoring personnel to trace historical events and investigate potential security risks. S7, Model Adaptive Iterative Optimization The cloud computing platform regularly collects historical analysis data, anomaly detection feedback results, and false positive and false negative samples to achieve adaptive iterative optimization of the improved time-series augmentation analysis model, ensuring continuous improvement in model analysis performance. This includes the following sub-steps: S71. Sample Collection and Labeling: Collect historical analysis data from the past 30 days, and screen out correctly detected samples, falsely detected samples, and missed samples, with falsely detected and missed samples accounting for no less than 20%; manually label the missed samples, labeling the target type, location, and behavioral characteristics, and construct an iterative training sample set, with the ratio of the training sample set to the validation sample set set to 8:2. S72. Incremental Training: The gradient descent algorithm is used to incrementally train the improved time series enhancement analysis model. The number of iterations is set to 100-200, and the learning rate is dynamically adjusted to 0.001-0.0001 (the learning rate gradually decreases as the number of iterations increases). During the training process, only the feature extraction layer and detection head parameters of the model are updated, without retraining the entire model, which reduces the iteration cost and shortens the iteration time (the time for a single iteration does not exceed 2 hours). S73. Model Validation and Replacement: Use a validation sample set to validate the trained model. Validation metrics include detection accuracy, recall, and inference speed. When the model's accuracy and recall both improve by more than 5% and the inference speed is not less than 30fps, replace the original analysis model and update the cloud model parameters. If the improvement criteria are not met, abandon this iteration, retain the original model parameters, and avoid model performance degradation. Furthermore, in step S2, the edge computing node and the cloud computing platform adopt a heartbeat detection mechanism, which performs a communication status check every 30 seconds. The detection content includes network connection status and data transmission rate. When a communication interruption is detected, the edge computing node automatically stores the pre-processed data in the local SSD cache (the cache capacity can be dynamically expanded, up to a maximum of 1TB). After the communication is restored, the data is synchronized to the cloud in sequence according to the transmission priority (urgent and abnormal data is transmitted first) to avoid data loss. One or more technical solutions provided in this application embodiment have at least the following technical effects or advantages compared with the prior art: 1. Improve the real-time performance of data processing and reduce cloud pressure: Adopt an edge-cloud collaborative processing architecture. The edge performs lightweight preprocessing on video data and transmits only the key frames and other effective data to the cloud, reducing the amount of data transmission by 60%-80% and significantly reducing the transmission and computing pressure on the cloud. At the same time, the improved time-series enhancement analysis model improves the analysis speed and ensures that the execution delay of emergency tasks does not exceed 500ms, meeting the needs of real-time monitoring. 2. Enhanced privacy protection and compliance with regulations: Through an adaptive dynamic blurring and desensitization algorithm, sensitive areas such as faces and license plates in the video are targeted for desensitization. Combined with sensitive area matching of adjacent frames, the desensitized areas are prevented from flickering. This ensures both monitoring effectiveness and anonymization of sensitive information, complying with the requirements of the "Personal Information Protection Law of the People's Republic of China" and effectively preventing privacy leaks. 3. Improve resource utilization and reduce costs: Adopt a dynamic scheduling algorithm for cloud computing resources to dynamically allocate resources based on task priority and node load, and reclaim idle resources to increase cloud resource utilization to over 80%; At the same time, build a tiered management mechanism for hot and cold data, with hot data using high-speed storage and cold data using low-cost storage, thereby improving access efficiency while reducing storage costs. 4. Enhance model adaptability and ensure stable analysis performance: Through the model adaptive iterative optimization mechanism, incremental training is adopted, and the model parameters are updated regularly using historical samples, without the need to retrain the entire model, thus reducing iteration costs; at the same time, the model can adapt to changes in different scenarios and environments, continuously improving detection accuracy and recall, and avoiding performance degradation after long-term use. 5. Secure and reliable data transmission and storage: The AES-256 encryption algorithm combined with an adaptive fragmentation transmission strategy ensures the security and stability of data transmission and resists the impact of network fluctuations; the heartbeat detection and local caching mechanism at the edge and cloud avoids data loss caused by camera offline or communication interruption, ensuring data integrity. 6. Multi-terminal collaboration and strong practicality: It supports collaborative feedback between monitoring terminals and mobile terminals, allowing monitoring personnel to view analysis results and receive abnormal alarms in real time, enabling mobile monitoring; it also supports multi-condition retrieval and historical event review, providing data support for monitoring management decisions and adapting to various scenarios such as urban security and park management.

[0007] Other advantages, objectives and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination or study, or may be learned from the practice of the invention. Attached Figure Description

[0008] Figure 1 This is a system end-to-end architecture diagram of the cloud computing-based video surveillance big data analysis method of the present invention; Figure 2 This is a flowchart of the edge collaborative preprocessing process for the cloud-based video surveillance big data analysis method of the present invention. Figure 3 This is a timing diagram of the secure transmission and layered storage of the cloud-based video surveillance big data analysis method of the present invention; Figure 4 This is a timing diagram of the cloud-based intelligent analysis and result output of the video surveillance big data analysis method based on cloud computing in this invention; Figure 5 This is a schematic diagram of the dynamic scheduling of cloud computing resources in the video surveillance big data analysis method based on cloud computing of the present invention. Detailed Implementation

[0009] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0010] It should be noted that the terms "vertical," "horizontal," "up," "down," "left," "right," and similar expressions used in this article are for illustrative purposes only and do not represent the only possible implementation.

[0011] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains; the terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to limit the invention; the term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0012] The following is in conjunction with the appendix Figure 1-5 The present invention will be further described in detail with reference to specific embodiments, so that those skilled in the art can understand and implement it.

[0013] This embodiment provides a cloud-based video surveillance big data analysis method applied to a monitoring scenario in an urban transportation hub (such as a train station). This scenario deploys 40 heterogeneous surveillance cameras (including 20 high-definition bullet cameras, 15 PTZ cameras, and 5 dome cameras), covering areas such as waiting halls, entrances / exits, and parking lots. The system needs to achieve real-time detection and tracking of people and vehicles, identification of abnormal behaviors (such as climbing over railings, crowd gatherings, and leaving objects behind), and traffic statistics. The specific implementation steps are as follows: S1, Distributed Video Data Acquisition A distributed acquisition network consisting of 40 heterogeneous surveillance cameras was constructed. Cameras in the waiting hall and entrances / exits use 4K resolution, while those in the parking lot use 1080P resolution. The acquisition network supports RTSP and HTTP transmission protocols, automatically selecting the appropriate protocol based on the camera type: PTZ cameras use RTSP (for easier real-time control), while bullet and dome cameras use HTTP (to reduce network usage). A heartbeat detection mechanism (every 10 seconds) monitors the camera status in real time. When a PTZ camera at an entrance / exit goes offline due to a network failure, the video data is temporarily stored in the camera's local 128GB cache. After the fault is resolved, the data is automatically synchronized to the edge computing node.

[0014] S2, Edge-end Collaborative Preprocessing Two edge computing nodes (each configured with an Intel Core i7-12700H CPU, 32GB of RAM, and 512GB of SSD cache) were deployed in the railway station monitoring room. Each node connects to data from 20 cameras for collaborative preprocessing. (1) Key frame extraction: The preset inter-frame difference threshold is 10, and the effective entropy value range is 4-7; the key frame extraction frequency of the entrance and exit and waiting hall (where there are dense crowds and frequent target movements) is set to 10 frames / second, and the parking lot (where there are few people) is set to 3 frames / second; key frames are extracted by combining the inter-frame difference method with entropy value analysis, and redundant frames are removed.

[0015] (2) Gaussian denoising: A Gaussian filter with a standard deviation of 1.0 is used to denoise the key frames and remove noise generated by the waiting hall lights and parking lot vehicle lights.

[0016] (3) Format standardization: All keyframes are uniformly converted to 1080P resolution and H.265 encoding format to ensure cloud analysis compatibility.

[0017] (4) Privacy desensitization processing: preset the coordinates of the face sensitive area at the entrance and exit and the waiting hall, and the coordinates of the license plate sensitive area in the parking lot; adopt an adaptive dynamic blur algorithm, with a 7×7 blur kernel for the face area and a 5×5 blur kernel for the license plate area; use IoU to match the sensitive areas of adjacent frames to avoid flickering in the desensitized area and ensure that privacy protection complies with regulatory requirements.

[0018] After preprocessing, the amount of data is reduced by 70% compared to the original video stream, resulting in lightweight preprocessed data.

[0019] S3, Transmission and Storage under Cloud Computing Architecture The preprocessed data is encrypted using the AES-256 encryption algorithm, combined with an adaptive fragmentation transmission strategy: the railway station monitoring network bandwidth is stable at over 100Mbps, so the fragment size is set to 128MB; a retransmission mechanism is used, automatically retransmitting when a fragment transmission failure is detected. The encrypted fragmented data is transmitted to the cloud computing platform (using Alibaba Cloud ECS distributed architecture, including 10 compute nodes, 5 storage nodes, and 1 scheduling node).

[0020] The storage nodes use the HDFS distributed file system to build a tiered management mechanism for hot and cold data: real-time analysis data and frequently accessed abnormal event data within the past 7 days are considered hot data and stored on SSD hot data nodes (access latency ≤ 8ms); historical data older than 7 days is considered cold data and stored on mechanical hard drive cold data nodes to reduce storage costs.

[0021] S4, cloud-based video big data intelligent analysis The improved time-series augmentation analysis model is invoked from the cloud to perform multi-dimensional analysis on the preprocessed data: (1) Improved temporal enhancement analysis model: Based on YOLOv8 optimization, the feature extraction network is equipped with a temporal attention mechanism, and the detection head is introduced with a generative adversarial refinement module. The model inference speed reaches 35fps and the target detection accuracy reaches 96%.

[0022] (2) Target detection and tracking: Detect people in waiting halls and entrances / exits, vehicles and people in parking lots, and extract target features; use Kalman filtering algorithm combined with feature matching to achieve continuous target tracking with a tracking accuracy of 92%, which can effectively deal with crowded and vehicle-obstructed scenarios.

[0023] (3) Abnormal behavior identification: Preset abnormal behavior rules: climbing over guardrails (matching threshold 85%), gathering of people (more than 10 people, stay time ≥ 3 minutes, matching threshold 85%), leaving behind objects (stay time ≥ 5 minutes, matching threshold 88%), running fast (speed ≥ 5m / s, matching threshold 85%). When it is detected that 12 people gather in a certain area of ​​the waiting hall for more than 3 minutes and the matching degree reaches 87%, it is judged as abnormal behavior, and the time, location and duration are recorded.

[0024] (4) Scene traffic statistics: Count the number of people entering and exiting the station and the number of vehicles in the parking lot every minute, generate hourly traffic statistics reports, and push them to the monitoring center to provide data support for passenger flow scheduling.

[0025] S5, dynamic scheduling of cloud computing resources The scheduling node monitors the load status of each computing node every 10 seconds, with preset task priorities: real-time anomaly detection is urgent, traffic statistics are normal, and historical data review is low priority. When the CPU utilization of a computing node reaches 82% and the memory utilization reaches 76%, the historical data review task on that node is migrated to an idle computing node; computing resources are prioritized for anomaly detection tasks, ensuring their execution latency is 450ms, lower than the preset 500ms; idle resources are reclaimed and allocated to high-load nodes, achieving a cloud resource utilization rate of 83%.

[0026] S6. Analysis Results Output and Multi-terminal Feedback After formatting the analysis results, they are pushed to 8 monitoring terminals (supporting simultaneous display of 32 video feeds) in the monitoring center and 10 mobile terminals (dedicated APP) for management personnel. The monitoring terminals trigger audible and visual alarms, displaying images and detailed information of the abnormal behavior area. Management personnel can view abnormal segments through their mobile terminals and remotely control the PTZ cameras in that area to adjust their angles and zoom in for details. Simultaneously, the analysis results are associated with and stored with corresponding video keyframes and original video segments, supporting retrieval and review by time (e.g., day, month, year, hour, 2026), scene (e.g., waiting hall area 1), and anomaly type (e.g., crowd gathering).

[0027] S7, Model Adaptive Iterative Optimization The model is iterated and optimized every 30 days. S71. Sample Collection and Labeling: Collect historical analysis data from the past 30 days, and screen out 10,000 correctly detected samples, 2,000 false positive samples, and 1,500 false negative samples; manually label the false negative samples and construct an iterative training sample set of 13,500 samples (10,800 training samples and 2,700 validation samples).

[0028] S72. Incremental Training: Gradient descent algorithm is used, with 150 iterations and the learning rate gradually decreasing from 0.001 to 0.0001; only the feature extraction layer and detector head parameters of the model are updated, and the time for a single iteration is 1.5 hours.

[0029] S73. Model Validation and Replacement: After validation, the model detection accuracy improved by 6%, the recall rate improved by 7%, and the inference speed remained at 35fps, meeting the improvement standard. The original analysis model was replaced, cloud parameters were updated, and analysis performance was continuously improved.

[0030] Furthermore, the edge computing nodes and the cloud employ a heartbeat detection mechanism (every 30 seconds). When communication between the edge nodes and the cloud is interrupted due to network fluctuations, the edge nodes will temporarily store pre-processed data in the local SSD cache. After communication is restored, the edge nodes will prioritize synchronizing data related to abnormal behavior to ensure that no data is lost.

[0031] In this embodiment, the above method enables efficient and intelligent analysis of big data from video surveillance in urban transportation hubs, solving problems such as poor real-time performance, insufficient privacy protection, and low resource utilization in existing methods. It improves the intelligence level and practicality of the monitoring system and can be widely applied to various large-scale video surveillance scenarios.

[0032] Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Anyone skilled in the art can make various modifications and alterations without departing from the spirit and scope of the present invention. Therefore, the scope of protection of the present invention should be determined by the claims.

Claims

1. A cloud-based video surveillance big data analysis method, characterized in that, Includes the following steps: S1. Distributed video data acquisition: Through an acquisition network composed of multiple heterogeneous surveillance cameras, video stream data of different scenes and resolutions are acquired in real time. The acquisition network supports multiple video transmission protocols such as RTSP and HTTP, and can adapt to the online / offline status of the cameras. When offline, the video data is temporarily stored in the local cache and automatically synchronized when online. S2. Edge-end collaborative preprocessing: In the edge computing nodes deployed near the surveillance camera, the video stream data collected in step S1 is collaboratively preprocessed. The preprocessing includes key frame extraction, Gaussian denoising, format standardization, and privacy desensitization processing based on sensitive area identification to obtain lightweight preprocessed data and reduce cloud transmission and computing pressure. S3. Transmission and storage under cloud computing architecture: The preprocessed data obtained in step S2 is transmitted to the cloud computing platform using the AES-256 encryption algorithm combined with an adaptive fragmentation transmission strategy. The cloud computing platform uses a distributed file system to fragment the data for storage and builds a hot and cold data hierarchical management mechanism. High-frequency access real-time analysis data is stored in hot data nodes, and historical archived data is stored in cold data nodes, so as to achieve efficient data access and low-cost storage. S4. Cloud-based video big data intelligent analysis: The cloud computing platform calls an improved temporal enhancement analysis model to perform multi-dimensional analysis on the stored preprocessed data. The analysis includes target detection, target tracking, abnormal behavior recognition, and scene traffic statistics. The improved temporal enhancement analysis model is optimized based on YOLOv8 and introduces a temporal attention mechanism and a generative adversarial refinement module to improve the accuracy and real-time performance of analysis in target occlusion and fast-moving scenes. S5. Dynamic scheduling of cloud computing resources: The cloud computing platform monitors the CPU utilization, memory usage and network bandwidth of each computing node in real time. Combined with the dynamic changes in the priority of the tasks analyzed in step S4, it dynamically allocates computing and storage resources through intelligent scheduling algorithms, giving priority to the execution of urgent anomaly detection tasks. Idle resources can be dynamically reclaimed to improve resource utilization. S6. Analysis Result Output and Multi-Terminal Feedback: The analysis results obtained in step S4 are formatted and pushed to the monitoring terminal and mobile terminal in real time. When a preset abnormal behavior is detected, an audible and visual alarm is triggered and the abnormal segment is pushed. At the same time, the analysis results are associated with the corresponding video segments and stored, supporting multi-condition retrieval and review by time, scene, and target type. S7. Model Adaptive Iterative Optimization: The cloud computing platform regularly collects historical analysis data, anomaly detection feedback results, and false positives and false negatives. The improved time series enhancement analysis model in step S4 is iteratively optimized using the gradient descent algorithm. Incremental training is used during the iteration process, which eliminates the need to retrain the entire model, reduces iteration costs, and continuously improves analysis performance.

2. The video surveillance big data analysis method based on cloud computing according to claim 1, characterized in that, The key frame extraction in step S2 is achieved by combining the inter-frame difference method with entropy analysis. Specifically, the absolute value of the inter-frame difference between two adjacent video frames is calculated. When the absolute value of the inter-frame difference is greater than a preset threshold and the entropy value of the frame is within a preset effective entropy value range, the frame is determined as a key frame. The key frame extraction frequency can be dynamically adjusted according to the video scene, with a dynamic range of 1-10 frames / second.

3. The video surveillance big data analysis method based on cloud computing according to claim 1, characterized in that, The privacy desensitization process described in step S2 is as follows: sensitive areas such as faces and license plates in the video are located by pre-setting sensitive area coordinates or intelligent recognition algorithms, and the sensitive areas are desensitized by an adaptive dynamic blurring algorithm. The size of the blur kernel is dynamically adjusted according to the size of the sensitive area. At the same time, the positions of sensitive areas in adjacent frames are matched by IoU (Intersection over Union) to avoid flickering of the desensitized areas, which complies with the anonymization requirements of the Personal Information Protection Law of the People's Republic of China.

4. The video surveillance big data analysis method based on cloud computing according to claim 1, characterized in that, The optimization process of the improved temporal enhancement analysis model described in step S4 is as follows: a temporal attention mechanism is added to the feature extraction network of YOLOv8 to capture the target motion features and correlations of adjacent video frames; a generative adversarial refinement module is introduced in the detection head to optimize the detected target bounding box and segmentation mask, reduce the detection error in target occlusion and complex background scenes, and ensure that the model inference speed is not less than 30fps.

5. The video surveillance big data analysis method based on cloud computing according to claim 1, characterized in that, The intelligent scheduling algorithm described in step S5 is as follows: the preset analysis task priority is divided into three levels: emergency level, normal level and low priority. The emergency level corresponds to the real-time detection task of abnormal behavior, the normal level corresponds to the routine target statistics task, and the low priority corresponds to the historical data review task. When the CPU utilization of a computing node exceeds 80% or the memory usage exceeds 75%, low-priority tasks on that node are migrated to idle computing nodes to ensure that the task execution latency does not exceed 500ms.

6. The video surveillance big data analysis method based on cloud computing according to claim 1, characterized in that, In step S2, the edge computing node and the cloud computing platform adopt a heartbeat detection mechanism, which performs a communication status check every 30 seconds. When a communication interruption is detected, the edge computing node automatically stores the preprocessed data in the local SSD cache. The cache capacity can be dynamically expanded. After the communication is restored, the data is synchronized to the cloud in sequence according to the transmission priority to avoid data loss.

7. The video surveillance big data analysis method based on cloud computing according to claim 1, characterized in that, The adaptive iterative optimization of the model described in step S7 specifically includes the following sub-steps: S71. Sample collection and labeling: Collect historical analysis data from the past 30 days, screen out correctly detected samples, falsely detected samples and missed samples, manually label the missed samples, and construct an iterative training sample set; S72. Incremental Training: The improved temporal enhancement analysis model is incrementally trained using the gradient descent algorithm. The number of iterations is set to 100-200, and the learning rate is dynamically adjusted to 0.001-0.0001. Only the feature extraction layer and detection head parameters of the model are updated. S73. Model Validation and Replacement: Use the validation sample set to validate the trained model. When the model's accuracy and recall both improve by more than 5%, replace the original analysis model. If the improvement standard is not met, abandon this iteration and retain the original model parameters.

8. The video surveillance big data analysis method based on cloud computing according to claim 1, characterized in that, The abnormal behavior identification in step S4 specifically involves: pre-setting multiple abnormal behavior rules such as climbing over, gathering, leaving objects, and running fast; extracting the target's movement trajectory, speed, and behavioral characteristics; performing similarity matching with the pre-set rules; and determining abnormal behavior when the matching degree exceeds 85%. At the same time, the time, location, and duration of the abnormal behavior are recorded.

9. The video surveillance big data analysis method based on cloud computing according to claim 1, characterized in that, The adaptive fragmentation transmission strategy described in step S3 is as follows: Based on the real-time network bandwidth detection results, the data fragment size is dynamically adjusted. When the bandwidth is greater than 100Mbps, the fragment size is set to 128MB; when the bandwidth is between 10-100Mbps, the fragment size is set to 64MB; and when the bandwidth is less than 10Mbps, the fragment size is set to 32MB to ensure transmission stability.

10. The video surveillance big data analysis method based on cloud computing according to claim 1, characterized in that, The monitoring terminal mentioned in step S6 supports multi-screen synchronous display and abnormal event warning settings, while the mobile terminal supports analysis result push, abnormal segment viewing and remote control of camera rotation, realizing multi-terminal collaborative monitoring.

Images