Video surveillance data storage device and method
Through deep learning technology, including multi-scale convolution kernels and graph neural networks, combined with deep hash indexes, the storage and retrieval of video surveillance data is optimized, and the processing efficiency of video surveillance data and the utilization rate of storage resources in the existing technology is solved. The problem of this paper realizes efficient video surveillance data storage and retrieval.
Patent Information
- Application Number
- CN202410441302.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-04-12
- Publication Date
- 2024-11-19
- Estimated Expiration
- 2044-04-12
AI Technical Summary
The existing video surveillance data storage technology lacks depth and accuracy, and cannot effectively mine complex spatiotemporal relationships, resulting in low efficiency in processing dynamic information, low utilization rate of storage resources, and low data retrieval efficiency, especially when quickly accessing and analyzing large amounts of video data. Lower.
Deep learning technology is adopted, including data feature extraction module, feature fusion module, spatiotemporal graph construction module, spatiotemporal graph analysis module, index module, decision-making module and storage optimization module, visual features are extracted through multi-scale convolution kernel, dynamic weight fusion, graph 神经网络分析时空图,深度哈希索引优化存储布局,实现高效的视频监控数据存储和检索。
It improves the storage and retrieval efficiency of video surveillance data, captures complex spatio-temporal relationships, maximizes the utilization of storage resources, reduces storage costs, and provides fast and accurate data access and retrieval capabilities.
Smart Images

Figure CN118277607B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of electronic digital data processing, and in particular to a video monitoring data storage device and method. Background Art
[0002] Video surveillance systems play an increasingly important role in security monitoring, traffic management, public safety and other fields. With the rapid development of surveillance technology and the continuous expansion of its application scope, a large amount of video surveillance data has been generated. This data is not only huge in volume, but also contains rich dynamic information and complex spatiotemporal relationships. Effectively storing, managing and retrieving this video data is crucial to improving the effectiveness of surveillance systems, strengthening security precautions and optimizing resource management.
[0003] However, existing video surveillance data storage and processing technologies face many challenges. First, traditional storage methods are often unable to efficiently process and analyze large-scale video data, especially in extracting complex visual features and spatiotemporal relationships. Second, due to the lack of effective indexing and retrieval mechanisms, accessing and analyzing the required video surveillance content is often inefficient. In addition, existing storage strategies are usually inflexible and cannot be dynamically adjusted according to the actual usage patterns of the data, resulting in low utilization of storage resources.
[0004] my country's patent application number: CN202311291041.6, publication date: 2023.12.26, discloses a video surveillance system, method and storage medium, the video surveillance system comprising: a support body and a master control unit and at least one lens group installed on the support body; the at least one lens group comprises lens modules of different types; the master control unit is electrically connected to the at least one lens group, and the parameters of the at least one lens group are controlled to be unified through the control chip of the at least one lens group, and the fused images obtained by the different types of lens groups are spliced, and the spliced images are output through a logical video channel.
[0005] However, the above technologies have at least the following technical problems: the existing technologies lack the depth and accuracy of video data analysis and the mining of complex spatiotemporal relationships, resulting in limited effectiveness when processing dynamic information; the storage strategies are not flexible and efficient enough, resulting in inefficient use of storage resources; the data retrieval efficiency is low and there is a lack of effective indexing mechanisms, especially in application scenarios that require rapid access and analysis of large amounts of video data. Summary of the invention
[0006] The present invention provides a video surveillance data storage device and method, which solves the problems that the prior art lacks depth and accuracy in video data analysis, lacks mining of complex spatiotemporal relationships, resulting in limited effects when processing dynamic information; storage strategies are not flexible and efficient, resulting in inefficient use of storage resources; data retrieval efficiency is low, and there is a lack of effective indexing mechanisms, especially in application scenarios where large amounts of video data need to be quickly accessed and analyzed. An efficient video surveillance data storage and retrieval method is implemented, which optimizes feature extraction, spatiotemporal relationship analysis, and data storage efficiency through deep learning and dynamic storage strategies.
[0007] The present invention provides a video surveillance data storage device and method, which specifically include the following technical solutions:
[0008] A video surveillance data storage device, comprising the following parts:
[0009] Data feature extraction module, feature fusion module, spatiotemporal graph construction module, spatiotemporal graph analysis module, indexing module, decision module, storage optimization module, and execution module;
[0010] Data feature extraction module, feature fusion module, spatiotemporal graph construction module, spatiotemporal graph analysis module, indexing module, decision module, storage optimization module, and execution module;
[0011] The data feature extraction module is used to extract visual elements as visual features from the video frames constituting the video surveillance data; the data feature extraction module is connected to the feature fusion module by means of data transmission;
[0012] The feature fusion module is used to fuse visual features of different scales through dynamic weights to form feature representation; the feature fusion module is connected to the spatiotemporal graph construction module through data transmission;
[0013] The spatiotemporal graph construction module is used to convert the fused visual features into a spatiotemporal graph structure; the spatiotemporal graph construction module is connected to the spatiotemporal graph analysis module by means of data transmission;
[0014] The spatiotemporal graph analysis module processes the spatiotemporal graph through a graph neural network to extract spatiotemporal graph features; the spatiotemporal graph analysis module is connected to the index module through data transmission;
[0015] The index module combines the spatiotemporal graph features with the video surveillance data, generates composite features through a heterogeneous information fusion function, and converts the composite features into an index of the video surveillance data using a deep hash function; the index module is connected to the decision module by means of data transmission;
[0016] The decision module analyzes the data characteristics of the video surveillance data according to the index of the video surveillance data and determines the storage location; the decision module is connected to the storage optimization module by means of data transmission;
[0017] The storage optimization module adjusts the storage layout according to the storage location and index and generates a storage strategy; the storage optimization module is connected to the execution module by means of data transmission;
[0018] The execution module executes the storage strategy according to the instruction of the storage optimization module.
[0019] A video surveillance data storage method comprises the following steps:
[0020] S1. Extract visual features of different scales of video frames, fuse visual features of different scales through dynamic weights, convert the fused visual features into a spatiotemporal graph, and use graph neural network to process the spatiotemporal graph to obtain spatiotemporal graph features;
[0021] S2. Integrate spatiotemporal graph features with video surveillance content to generate composite features, and convert the composite features into indexes of video surveillance data through deep hash functions. Determine the storage location of video surveillance data by analyzing the indexes of video surveillance data, and dynamically adjust the storage layout according to the storage location and index of video surveillance data to generate a storage strategy; further execute the storage strategy.
[0022] Preferably, the S1 further specifically includes:
[0023] Through convolutional neural network technology, convolution kernels of different scales are used to extract visual features of different scales in video frames.
[0024] Preferably, the S1 further comprises:
[0025] A dynamic weight allocation mechanism is introduced; the dynamic weight allocation mechanism allocates dynamic weights to visual features of different scales through saliency evaluation, and fuses visual features extracted from different scales through dynamic weights.
[0026] Preferably, the S1 further comprises:
[0027] The fused visual features are converted into a spatiotemporal graph; in the spatiotemporal graph, each pixel and its neighborhood are regarded as nodes and edges of the spatiotemporal graph.
[0028] Preferably, the S2 further specifically includes:
[0029] Design heterogeneous information fusion functions to fuse spatiotemporal graph features and video surveillance data to generate composite features.
[0030] Preferably, said S2 further comprises:
[0031] The deep hash function uses the hierarchical structure to extract information from the composite features and converts the extracted information into a hash code for retrieval.
[0032] Preferably, said S2 further comprises:
[0033] In the process of deciding the storage location of the video surveillance data, a decision function is introduced to determine the storage location of the video surveillance data.
[0034] Preferably, said S2 further comprises:
[0035] Build a storage policy function to dynamically adjust the storage layout according to the storage location and index of video surveillance data and generate a storage policy.
[0036] The beneficial effects of the technical solution of the present invention are:
[0037] 1. By using multi-scale convolution kernels to extract visual features of video frames and fusing features extracted at multiple scales through dynamic weights, the present invention can capture the content of video frames more comprehensively; comprehensive feature extraction provides a rich information basis for retrieving and analyzing video surveillance data; through spatiotemporal graph construction and graph neural network analysis, the present invention can capture and learn complex spatiotemporal relationships and patterns in video frames, making the storage and retrieval process of video surveillance data more efficient and accurate, especially in processing temporal continuity and spatial relationships;
[0038] 2. Through the decision-making module and the storage optimization module, the present invention can dynamically adjust the storage layout based on the type and access frequency of the video surveillance data to maximize the use of storage resources, which not only improves the storage efficiency but also reduces the storage cost; combined with heterogeneous information fusion and deep hash indexing technology, the present invention provides an efficient indexing mechanism for video surveillance data, so that video surveillance data can be retrieved quickly and accurately, greatly improving the availability and access speed of the data. BRIEF DESCRIPTION OF THE DRAWINGS
[0039] FIG1 is a structural diagram of a video surveillance data storage device provided by an embodiment of the present invention;
[0040] FIG. 2 is a flow chart of a video surveillance data storage method provided by an embodiment of the present invention. DETAILED DESCRIPTION
[0041] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the technical scheme in the embodiment of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiment of the present invention. Obviously, the described embodiment is only a part of the embodiment of the present invention, not all the embodiments. Based on the embodiment of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0042] Unless defined otherwise, 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 belongs.
[0043] In order to better understand the above technical solution, the above technical solution will be described in detail below in conjunction with the accompanying drawings and specific implementation methods.
[0044] Referring to FIG. 1 , a video surveillance data storage device provided by an embodiment of the present invention includes the following parts:
[0045] Data feature extraction module, feature fusion module, spatiotemporal graph construction module, spatiotemporal graph analysis module, indexing module, decision module, storage optimization module, and execution module;
[0046] The data feature extraction module uses convolutional neural network technology to extract visual elements as visual features from video frames constituting video surveillance data; the data feature extraction module is connected to the feature fusion module by means of data transmission;
[0047] The feature fusion module fuses visual features of different scales to form a feature representation; and uses dynamic weights to balance visual features from different scales; the feature fusion module is connected to the spatiotemporal graph construction module by data transmission;
[0048] The spatiotemporal graph construction module converts the fused features into a spatiotemporal graph structure to obtain the temporal continuity and spatial relationship in the video surveillance data; the spatiotemporal graph construction module is connected to the spatiotemporal graph analysis module by means of data transmission;
[0049] The spatiotemporal graph analysis module processes the spatiotemporal graph through a graph neural network to extract spatiotemporal graph features; the spatiotemporal graph analysis module is connected to the index module through data transmission;
[0050] The index module combines the spatiotemporal graph features with the video surveillance data, generates a composite feature through a heterogeneous information fusion function, and converts the composite feature into an index of the video surveillance data using a deep hash function; the index module is connected to the decision module by means of data transmission;
[0051] The decision module analyzes the data characteristics of the video surveillance data according to the index of the video surveillance data and determines the storage location; the decision module is connected to the storage optimization module by means of data transmission;
[0052] The storage optimization module adjusts the storage layout according to the storage location and index and generates a storage strategy; the storage optimization module is connected to the execution module by means of data transmission;
[0053] The execution module executes the storage strategy according to the instruction of the storage optimization module.
[0054] Referring to FIG. 2 , a video surveillance data storage method provided by an embodiment of the present invention includes the following steps:
[0055] S1. Extract visual features of different scales of video frames, fuse visual features of different scales through dynamic weights, convert the fused visual features into a spatiotemporal graph, and use graph neural network to process the spatiotemporal graph to obtain spatiotemporal graph features;
[0056] In order to process and store large amounts of video surveillance data and enable the video surveillance data to be retrieved quickly and accurately, the video surveillance data is deeply analyzed and key visual features are extracted.
[0057] Video surveillance data consists of video frames. The data feature extraction module uses convolutional neural network technology in deep learning to extract visual features of different levels of video frames using convolution kernels of different scales. Visual features of different scales capture various information in the video, from subtle details to overall structure. Each convolution kernel corresponds to a receptive field. The effect of the convolution kernel is not limited to a single pixel, but covers all pixels within its neighborhood. The specific calculation method is as follows:
[0058] ,
[0059] in, Indicates that the kth convolution kernel is at position The convolution output on , It is an activation function, which increases the nonlinearity of the model; is the kth convolution kernel at its local position The weight on Indicated in pixels In the surrounding neighborhood, relative to The value of the pixel whose position is offset by i and j; is a bias term that adjusts the offset of the convolution result. The visual features captured by each convolution kernel reflect the visual information at different scales, thus laying the foundation for capturing complex visual patterns.
[0060] The feature fusion module fuses visual features extracted from multiple scales through dynamic weights, integrating visual information at different levels to form a comprehensive feature representation. During the fusion process, each dynamic weight is dynamically adjusted according to different parts of the video surveillance content, balancing and integrating visual features from different scales to make them more comprehensively represent the content of the video frame. The fusion formula is:
[0061] ,
[0062] in, represents the fused features, It is a dynamic weight that is dynamically adjusted according to the video surveillance content to ensure the effective fusion of visual features of different scales. The dynamic weight allocation mechanism is implemented by analyzing each area in the video frame, evaluating the visual saliency of each area, and assigning dynamic weights accordingly. The saliency evaluation of a feature is based on the distribution of the feature in the local area and the entire video frame. The mathematical expression of saliency evaluation is as follows:
[0063] ,
[0064] ,
[0065] in, is feature k at position The significance score of Represents local significance and evaluates the importance of features in local areas; Represents global saliency, evaluating the distribution of features in the entire video frame; is a balance coefficient; Represents the size radius of the convolution kernel, The purpose of calculating global saliency is to quantify the prominence of each feature relative to the average level of the entire video frame. The global saliency of the feature is obtained by calculating the difference between the feature value of each pixel in the video frame and the average value of the feature in the entire video frame, and dividing the sum of all difference values by the total number of pixels in the video frame.
[0066] Assign dynamic weights to each feature based on the significance score , use the SoftMax function for normalization:
[0067] ,
[0068] The dynamic weight adjustment mechanism ensures the effective integration of visual features at different scales, making the final video surveillance data storage and retrieval process more efficient and accurate.
[0069] Furthermore, the spatiotemporal graph construction module converts the fused features into a spatiotemporal graph structure, which can reveal the temporal continuity and spatial relationship in the video surveillance data. In the spatiotemporal graph structure, each pixel and its neighborhood are regarded as nodes and edges of the spatiotemporal graph, thus forming a complex network that can reflect the spatiotemporal association between pixels. The spatiotemporal graph construction function is expressed as:
[0070] ,
[0071] in, represents the constructed space-time graph, It is a balance parameter used to adjust the importance of the features of the current pixel and its neighboring pixels; is the set of neighborhood pixels, are the pixel coordinates in the neighborhood; It's location Relative to position The weighting coefficient is used to emphasize or weaken the influence of specific neighborhood pixels. is in position The characteristics of the pixels.
[0072] The spatiotemporal graph analysis module uses graph neural networks to process spatiotemporal graphs and extracts high-level features that can reflect the complex spatiotemporal relationships in video frames, namely spatiotemporal graph features. Through the processing of graph neural networks, deep relationships and patterns between nodes in the spatiotemporal graph are learned and extracted, providing rich information for subsequent retrieval and analysis. The calculation formula of the graph neural network is:
[0073] ,
[0074] ,
[0075] in, represents a graph neural network function, is the output of the space-time graph construction function, which is the position The graph node features at time t, It is for location Parameters of graph neural network; It is a parameter that adjusts the importance of its own features and neighboring features; It's location Features of graph nodes at time t.
[0076] S2. Integrate spatiotemporal graph features with video surveillance content to generate composite features, and convert the composite features into indexes of video surveillance data through deep hash functions. Determine the storage location of video surveillance data by analyzing the indexes of video surveillance data, and dynamically adjust the storage layout according to the storage location and index of video surveillance data to generate a storage strategy; further execute the storage strategy.
[0077] The index module combines the spatiotemporal graph features obtained from the graph neural network with the video surveillance data to form an index for data storage and fast retrieval. A heterogeneous information fusion function is designed to fuse the spatiotemporal graph features and the original video surveillance content to generate a composite feature. The calculation formula of the heterogeneous information fusion function is:
[0078] ,
[0079] in, It is the composite feature obtained by the heterogeneous information fusion function; is the weight of the spatiotemporal features at time t; It is the space-time graph feature; is the weight of the original video surveillance data; Indicates that video surveillance data Convert to vector form.
[0080] The index module constructs a deep hash function based on a deep neural network model, and converts the composite features into indexes of video surveillance data through the deep hash function. The deep hash function uses a hierarchical structure to extract information from the composite features and converts the extracted information into a hash code for retrieval; each layer transforms the data of the composite features through a set of weights and biases. The calculation formula of the deep hash function is:
[0081] ,
[0082] in, Represents the index of video surveillance data, displayed in the form of a binary hash code. is the sign function used to convert the result into a binary hash, is the weight of the rth layer; is the weight connecting the sth neuron to the rth layer; is the composite feature input to the sth neuron, is the bias term of the rth layer.
[0083] The decision module identifies the data characteristics of the video surveillance data, including the data type and access frequency, by analyzing the index of the video surveillance data, so as to provide key information for the decision of the storage location. Based on the data characteristics, the optimal storage location is determined for each video surveillance data, and it is decided whether the data should be stored in the fast access area or the long-term storage area. A decision function is used to balance the weight of storage in the fast access area and the long-term storage area, so as to maximize the use of storage resources and meet the fast access requirements. The decision function is:
[0084] ,
[0085] in, represents the decision function, which calculates the proportion of video surveillance data stored in the fast access area according to the data type and access frequency. It is a data type used to distinguish different types of video surveillance data. is the access frequency, which indicates how frequently the data is accessed. , , It is a coefficient obtained through video data analysis and is used to weigh the decision of storage location based on data type and access frequency. The output of the decision function is a value between 0 and 1, indicating the proportion of data stored in the fast access area.
[0086] The storage optimization module dynamically adjusts the storage layout according to the storage location and index, optimizes the overall utilization efficiency of the storage space, and reduces resource waste, which is implemented by the following formula:
[0087] ,
[0088] in, Function represents the storage strategy function, which determines the storage location of each video surveillance data and index Adjust the storage layout and generate the most optimized storage strategy. is a tuning factor used to control the sensitivity of the optimization process.
[0089] The execution module executes the optimized storage strategy to ensure that data is stored according to the optimized storage strategy layout, improve retrieval efficiency and reduce storage costs. The specific formula is:
[0090] ,
[0091] in, The function is responsible for implementing the storage operation, based on the index of the video surveillance data, the storage location determined, and the actual video surveillance data content. To perform storage operations, , The index is used to adjust the degree of influence of indexing and storage location. is a normalization factor that ensures that the results of storage operations are within a reasonable range.
[0092] By implementing the final storage strategy, data is stored in the calculated optimal location, ensuring data storage efficiency and access speed while making full use of storage resources.
[0093] In summary, a video surveillance data storage device and method are completed.
[0094] The present invention is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present invention. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate a device for implementing the functions specified in one or more processes in the flowchart and / or one or more boxes in the block diagram.
[0095] These computer program instructions may also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, whereby the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in one or more processes in the flowchart and / or one or more boxes in the block diagram.
[0096] Although the preferred embodiments of the present invention have been described, those skilled in the art may make other changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.
[0097] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.
Claims
1. A video surveillance data storage method, characterized in that: The following steps are involved: S1. Extract visual features of different scales of video frames in video surveillance data, and fuse visual features of different scales through dynamic weights to obtain fused visual features; convert the fused visual features into a spatiotemporal graph through a spatiotemporal graph construction function, which is expressed as: Among them, Φ(T(x,y),N) represents the constructed spatiotemporal graph; T(x,y) represents the fused visual features; N is the set of neighborhood pixels; α is the balance parameter that adjusts the importance of the features of the current pixel and its neighborhood pixels; (x′,y ′ ) are the pixel coordinates in the neighborhood; α xy,x′y′ is the position (x,y) relative to the position (x ′ ,y ′ ) is the weighting coefficient of the pixel at position (x′, y′); T(x′, y′) is the feature of the pixel at position (x′, y′); Then use the graph neural network to process the spatiotemporal graph to obtain the spatiotemporal graph features. The calculation formula of the graph neural network function is: G t (x,y)=Φ(T(x,y),N) Among them, GNN represents the graph neural network function; GNN (G t ,Θ) represents the obtained spatiotemporal graph features; G t (x, y) is the output of the spatiotemporal graph construction function, which represents the graph node feature at position (x, y) at time t; Θ xy is the graph neural network parameter for position (x, y); η is the parameter that adjusts the importance of its own features and neighborhood features; G t (x′, y′) is the graph node feature at position (x′, y′) at time t; S2. Design a heterogeneous information fusion function to fuse the spatiotemporal graph features with the video surveillance content to generate composite features. The calculation formula of the heterogeneous information fusion function is: Among them, Ψ is the composite feature obtained by the heterogeneous information fusion function; δ t is the weight of the spatiotemporal features at time t; H t is the spatiotemporal graph feature; ω is the weight of the original video surveillance data; Vec(V) represents the conversion of the video surveillance data V into a vector form; Then, the composite features are converted into the index of video surveillance data through the deep hash function, and the storage location of the video surveillance data is determined by analyzing the index of the video surveillance data; a storage strategy function is constructed to dynamically adjust the storage layout according to the storage location and index of the video surveillance data and generate a storage strategy; the storage strategy is further executed; the specific formula of the storage strategy function is: Among them, Optimize(θ,D) represents the storage strategy function; θ and D represent the decision and index of the storage location of each video surveillance data respectively; d is the adjustment factor that controls the sensitivity of the optimization process.
2. A video surveillance data storage method as claimed in claim 1, characterized in that: The S1 specifically includes: Through convolutional neural network technology, convolution kernels of different scales are used to extract visual features of different scales in video frames.
3. A video surveillance data storage method as claimed in claim 2, characterized in that: Said S1 further comprises: A dynamic weight allocation mechanism is introduced; the dynamic weight allocation mechanism allocates dynamic weights to visual features of different scales through saliency evaluation, and fuses visual features extracted from different scales through dynamic weights.
4. A video surveillance data storage method as claimed in claim 3, characterized in that: Said S1 further comprises: The fused visual features are converted into a spatiotemporal graph; in the spatiotemporal graph, each pixel and its neighborhood are regarded as nodes and edges of the spatiotemporal graph.
5. A video surveillance data storage method as claimed in claim 1, characterized in that: The S2 specifically includes: The deep hash function uses the hierarchical structure to extract information from the composite features and converts the extracted information into a hash code for retrieval.
6. A video surveillance data storage method as claimed in claim 1, characterized in that: Said S2 further comprises: In the process of deciding the storage location of the video surveillance data, a decision function is introduced to determine the storage location of the video surveillance data.
7. A video surveillance data storage device, applied to a video surveillance data storage method as claimed in claim 1, characterized in that: Includes the following parts: Data feature extraction module, feature fusion module, spatiotemporal graph construction module, spatiotemporal graph analysis module, indexing module, decision module, storage optimization module, and execution module; The data feature extraction module is used to extract visual elements as visual features from video frames constituting the video surveillance data; The data feature extraction module is connected to the feature fusion module through data transmission; The feature fusion module is used to fuse visual features of different scales through dynamic weights to form feature representation; The feature fusion module is connected to the spatiotemporal graph construction module through data transmission; The spatiotemporal graph construction module is used to convert the fused visual features into a spatiotemporal graph structure; the spatiotemporal graph construction module is connected to the spatiotemporal graph analysis module by means of data transmission; The spatiotemporal graph analysis module processes the spatiotemporal graph through a graph neural network to extract spatiotemporal graph features; The spatiotemporal graph analysis module is connected to the index module through data transmission; The index module combines the spatiotemporal graph features with the video surveillance data, generates composite features through a heterogeneous information fusion function, and converts the composite features into an index of the video surveillance data using a deep hash function; the index module is connected to the decision module by means of data transmission; The decision module analyzes the data characteristics of the video surveillance data according to the index of the video surveillance data and determines the storage location; the decision module is connected to the storage optimization module by means of data transmission; The storage optimization module adjusts the storage layout according to the storage location and index, and generates a storage strategy; The storage optimization module is connected to the execution module by means of data transmission; The execution module executes the storage strategy according to the instruction of the storage optimization module.
