Video monitoring storage system and method based on distributed cloud storage

Abstract

The invention relates to the technical field of cloud storage, and particularly discloses a video monitoring storage system and method based on distributed cloud storage, and the method comprises the steps: collecting a monitoring video stream through a camera, carrying out the time partitioning of the monitoring video stream based on a video retrieval query state of a user, and forming a plurality of video data blocks, a video distributed cloud storage architecture is constructed, the storage position of each video data block in the distributed cloud storage is determined by using a consistent Hash algorithm, and the use frequency of each video data block is monitored in real time, so that the storage priority of each video data block is dynamically adjusted. According to the method, the risk of single-point failure can be effectively avoided, the storage reliability and fault tolerance are improved, the storage priority can be dynamically adjusted according to the use frequency of the video data, and efficient management and access of the data are achieved.

Description

Technical Field

[0001] The present application relates to the field of cloud storage technology, and more specifically, to a video surveillance storage system and method based on distributed cloud storage. Background Art

[0002] With the rapid development of video surveillance technology, the popularity of surveillance cameras has increased year by year, and video surveillance systems have been increasingly used in public safety, traffic management, commercial operations, and other fields. However, with the increase in the number of surveillance cameras and the improvement in video resolution, the storage and management of surveillance video data face huge challenges.

[0003] Video surveillance systems usually need to run 24 hours a day, generating a huge amount of video data, and traditional centralized storage systems often cannot meet the storage needs of large-scale video surveillance data. On the one hand, centralized storage systems may have the risk of single point failure. Once the storage center fails, the entire surveillance system will be paralyzed; on the other hand, as video data continues to increase, the data access efficiency of centralized storage systems will drop significantly, resulting in increased delays in video retrieval and playback.

[0004] Compared with centralized storage systems, distributed cloud storage systems can not only improve data reliability and fault tolerance by distributing video data on multiple nodes, but also achieve parallel processing and access of data, further improving the efficiency of video retrieval and playback. However, traditional distributed cloud storage systems usually adopt a static storage strategy, that is, once video data is assigned to a storage node, its storage location remains fixed. This static storage strategy is not flexible enough when facing frequently changing data access patterns, making it difficult to achieve efficient data management and optimization.

[0005] Therefore, an optimized video surveillance storage system and method based on distributed cloud storage is needed to solve the above technical problems. Summary of the invention

[0006] In order to solve the above technical problems, the present application is proposed. The embodiment of the present application provides a video surveillance storage system and method based on distributed cloud storage, which collects surveillance video streams through cameras, and divides the surveillance video streams into time blocks based on the user's video retrieval query status to form multiple video data blocks, and then constructs a video distributed cloud storage architecture, uses a consistent hashing algorithm to determine the storage location of each video data block in the distributed cloud storage, and monitors the usage frequency of each video data block in real time, so as to dynamically adjust the storage priority of each video data block, which can not only effectively avoid the risk of single point failure and improve the reliability and fault tolerance of storage, but also dynamically adjust the storage priority of video data according to the usage frequency, so as to achieve efficient management and access of data.

[0007] According to one aspect of the present application, a video surveillance storage method based on distributed cloud storage is provided, which includes: Get the surveillance video stream collected by the camera; Dividing the surveillance video stream into a plurality of video data blocks; Determine a target storage node for each of the plurality of video data blocks based on a consistent hashing algorithm; Based on the target storage node of each of the video data blocks, storing the multiple video data blocks in multiple storage nodes of the distributed cloud storage; Monitoring the usage frequency of each video data block among the plurality of video data blocks, and adjusting the storage priority of each video data block based on the usage frequency of each video data block; The step of monitoring the usage frequency of each video data block among the plurality of video data blocks and adjusting the storage priority of each video data block based on the usage frequency of each video data block includes: In response to the usage frequency of the video data block exceeding a first preset threshold, storing the corresponding video data block in a storage node with optimal speed and delay; In response to the usage frequency of the video data block being lower than a second preset threshold, the corresponding video data block is stored in a storage node with the lowest cost in a compressed or archived manner.

[0008] Preferably, the storage node with the best speed and delay is an SSD storage node, and the storage node with the lowest cost is an HDD storage node.

[0009] Preferably, the surveillance video stream is divided into a plurality of video data blocks, including: Obtaining a user query log, wherein the query log records all search requests, wherein the search requests include a time range, a camera number, and a click rate of search results; Performing semantic embedding coding on each retrieval request in the user query log to obtain a set of retrieval request semantic embedding coding vectors; Performing a demand clustering analysis based on unsupervised learning on the set of the retrieval request semantic embedding coding vectors to obtain a video retrieval demand semantic clustering coding vector; Inputting the video retrieval requirement semantic clustering encoding vector into a decoder-based time segmentation recommender to obtain a recommendation value for the time segmentation; Based on the recommended value of the time segmentation, the monitoring video stream is data-separated to obtain the multiple video data blocks.

[0010] Preferably, performing unsupervised learning-based demand clustering analysis on the set of retrieval request semantic embedding coding vectors to obtain video retrieval demand semantic clustering coding vectors includes: Inputting the set of search request semantic embedding coding vectors into the information core coarse-grained aggregation network to obtain a search request coarse-grained semantic aggregation coding vector; Based on the feature difference between each retrieval request semantic embedding coding vector in the set of retrieval request semantic embedding coding vectors and the retrieval request coarse-grained semantic convergent coding vector, the set of retrieval request semantic embedding coding vectors is dynamically compensated and converged to obtain a retrieval request fine-grained semantic compensated and converged coding vector; The retrieval request fine-grained semantic compensation aggregation coding vector and the retrieval request coarse-grained semantic aggregation coding vector are input into a residual unit to obtain the video retrieval requirement semantic clustering coding vector.

[0011] Preferably, based on the feature difference between each retrieval request semantic embedding coding vector in the set of retrieval request semantic embedding coding vectors and the retrieval request coarse-grained semantic converged coding vector, the set of retrieval request semantic embedding coding vectors is dynamically compensated and converged to obtain a retrieval request fine-grained semantic compensated and converged coding vector, including: Calculating the core convergence compensation factor of each retrieval request semantic embedding coding vector in the set of retrieval request semantic embedding coding vectors relative to the retrieval request coarse-grained semantic convergence coding vector to obtain a set of retrieval request semantic core convergence compensation factors; Performing compensation explicit modeling based on a gating function on the set of retrieval request semantic core convergence compensation factors to obtain a set of retrieval request semantic core convergence compensation weight factors; The set of the retrieval request semantic core convergence compensation weight factors, the retrieval request coarse-grained semantic convergence coding vector and the set of the retrieval request semantic embedding coding vector are input into the node fine-grained dynamic compensation convergence network to obtain the retrieval request fine-grained semantic compensation convergence coding vector.

[0012] Preferably, calculating the core convergence compensation factor of each retrieval request semantic embedding coding vector in the set of retrieval request semantic embedding coding vectors relative to the retrieval request coarse-grained semantic convergence coding vector to obtain a set of retrieval request semantic core convergence compensation factors includes: Performing point convolution encoding based on Sigmoid activation function on the retrieval request semantic embedding encoding vector and the retrieval request coarse-grained semantic aggregation encoding vector respectively to obtain a standardized retrieval request semantic embedding encoding vector and a standardized retrieval request coarse-grained semantic aggregation encoding vector; Calculating a position difference vector between the standardized search request semantic embedding coding vector and the standardized search request coarse-grained semantic convergence coding vector, and taking an absolute value of the position difference vector to obtain a search request semantic core convergence difference compensation coding vector; The retrieval request semantic kernel convergence difference compensation encoding vector is input into a compensation feature importance scoring module based on a neural network layer to obtain the retrieval request semantic kernel convergence compensation factor.

[0013] Preferably, the retrieval request semantic core convergence difference compensation encoding vector is input into a compensation feature importance scoring module based on a neural network layer to obtain the retrieval request semantic core convergence compensation factor, including: Multiplying the retrieval request semantic core convergence difference compensation coding vector by a weight parameter matrix, and adding the multiplication result to a bias term to obtain a retrieval request semantic core convergence compensation feature modulation vector; The retrieval request semantic core convergence compensation feature modulation vector is multiplied by the retrieval request semantic compensation feature importance score conversion vector to obtain the retrieval request semantic core convergence compensation factor.

[0014] Preferably, the weight parameter matrix is ​​multiplied by the retrieval request semantic core convergence difference compensation coding vector, and the multiplication result is added to the bias term to obtain the retrieval request semantic core convergence compensation feature modulation vector, including: Calculate the ratio between the Euclidean norm of the search request semantic embedding coding vector and the Euclidean norm of the search request coarse-grained semantic aggregation coding vector, and if the ratio is less than 1, add 1 to the ratio and calculate the logarithm value with base 2 as the bias term; If the ratio is greater than or equal to 1, the ratio is used as the bias term.

[0015] According to another aspect of the present application, a video surveillance storage system based on distributed cloud storage is provided, which includes: A monitoring video stream acquisition module is used to obtain the monitoring video stream collected by the camera; A monitoring video stream segmentation module, used for segmenting the monitoring video stream into a plurality of video data blocks; A storage node allocation module, configured to determine a target storage node for each of the plurality of video data blocks based on a consistent hashing algorithm; A video data block storage module, configured to store the plurality of video data blocks in a plurality of storage nodes of a distributed cloud storage based on a target storage node of each of the video data blocks; The storage priority adjustment module is used to monitor the usage frequency of each video data block among the multiple video data blocks, and adjust the storage priority of each video data block based on the usage frequency of each video data block.

[0016] Preferably, the storage priority adjustment module: in response to the usage frequency of the video data block exceeding a first preset threshold, the corresponding video data block is stored in a storage node with optimal speed and delay; in response to the usage frequency of the video data block being lower than a second preset threshold, the corresponding video data block is stored in a storage node with the lowest cost in a compressed or archived manner.

[0017] This application has at least the following technical effects: Compared with the prior art, the video surveillance storage system and method based on distributed cloud storage provided by the present application collects surveillance video streams through cameras, and time-blocks the surveillance video streams based on the user's video retrieval query status to form multiple video data blocks, thereby constructing a video distributed cloud storage architecture, using a consistent hashing algorithm to determine the storage location of each video data block in the distributed cloud storage, and monitoring the usage frequency of each video data block in real time, so as to dynamically adjust the storage priority of each video data block. This can not only effectively avoid the risk of single point failures and improve the reliability and fault tolerance of storage, but also dynamically adjust the storage priority of the video data according to the usage frequency, thereby realizing efficient management and access of data. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] By describing the embodiments of the present application in more detail in conjunction with the accompanying drawings, the above and other purposes, features and advantages of the present application will become more apparent. The accompanying drawings are used to provide a further understanding of the embodiments of the present application and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the present application and do not constitute a limitation of the present application. In the accompanying drawings, the same reference numerals generally represent the same components or steps.

[0019] Figure 1 The present invention is a flowchart of a video surveillance storage method based on distributed cloud storage according to an embodiment of the present application.

[0020] Figure 2 This is a flowchart of sub-step S2 of the video surveillance storage method based on distributed cloud storage according to an embodiment of the present application.

[0021] Figure 3 Schematic diagram of data flow in sub-step S2 of the video surveillance storage method based on distributed cloud storage according to an embodiment of the present application.

[0022] Figure 4 This is a flowchart of sub-step S23 of the video surveillance storage method based on distributed cloud storage according to an embodiment of the present application.

[0023] Figure 5 This is a flowchart of sub-step S232 of the video surveillance storage method based on distributed cloud storage according to an embodiment of the present application.

[0024] Figure 6 4 is a block diagram of a video surveillance storage system based on distributed cloud storage according to an embodiment of the present application. DETAILED DESCRIPTION

[0025] As shown in this application and claims, unless the context clearly indicates an exception, the words "a", "an", "an" and / or "the" do not refer to the singular and may also include the plural. Generally speaking, the terms "include" and "comprise" only indicate the inclusion of the steps and elements that have been clearly identified, and these steps and elements do not constitute an exclusive list. The method or device may also include other steps or elements.

[0026] Although the present application makes various references to certain modules in the system according to the embodiments of the present application, any number of different modules can be used and run on the user terminal and / or server. The modules are only illustrative, and different aspects of the system and method can use different modules.

[0027] Flowcharts are used in the present application to illustrate the operations performed by the system according to the embodiments of the present application. It should be understood that the preceding or following operations are not necessarily performed accurately in order. On the contrary, various steps may be processed in reverse order or simultaneously as required. Meanwhile, other operations may also be added to these processes, or a certain step or several steps of operations may be removed from these processes.

[0028] Below, the exemplary embodiments according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application, and it should be understood that the present application is not limited to the exemplary embodiments described here.

[0029] It should be noted that in this application, all actions to obtain data are carried out in compliance with the relevant data protection laws and policies of the country where the data is located, and with the authorization given by the owner of the corresponding device.

[0030] Figure 1 FIG. 1 is a flow chart of a video surveillance storage method based on distributed cloud storage according to an embodiment of the present application. Figure 1As shown, the video surveillance storage method based on distributed cloud storage includes the following steps: S1, obtaining a surveillance video stream collected by a camera; S2, dividing the surveillance video stream into multiple video data blocks; S3, determining a target storage node for each of the multiple video data blocks based on a consistent hashing algorithm; S4, storing the multiple video data blocks in multiple storage nodes of distributed cloud storage based on the target storage nodes of each of the video data blocks; S5, monitoring the usage frequency of each of the multiple video data blocks, and adjusting the storage priority of each of the video data blocks based on the usage frequency of each of the video data blocks.

[0031] In the above-mentioned video surveillance storage method based on distributed cloud storage, the step S1 obtains the surveillance video stream collected by the camera. Specifically, as a front-end perception unit, the camera is responsible for capturing images of the surrounding environment. The collected image information is directly related to the quality of subsequent data processing, storage and analysis. Therefore, the selection and configuration of the camera is crucial, and the requirements for cameras in different application scenarios vary significantly. For example, in the field of public safety, high-definition cameras can provide clearer images and help identify suspicious behaviors or people; while in traffic management, panoramic cameras can cover a wider area and effectively monitor road conditions. In addition, with the development of technology, smart cameras have gradually become popular. Such cameras not only have traditional image capture functions, but also have built-in preliminary data analysis capabilities, such as face recognition, behavior analysis, etc. This allows effective data screening at the source, reducing the burden on the back-end system.

[0032] In the above-mentioned video surveillance storage method based on distributed cloud storage, the step S2 divides the surveillance video stream into multiple video data blocks. Specifically, since the complete long video stream is not convenient to store and manage, for example, a large amount of continuous storage space may be required during video storage, which is not conducive to the flexible allocation of storage resources; during video retrieval, if the video content of a specific time period is to be located, it is necessary to traverse the entire long video, which is inefficient. Therefore, the present application can effectively reduce the amount of data stored and retrieved at a single time and improve the flexibility and scalability of data storage by dividing the surveillance video stream into multiple video data blocks. Among them, Figure 2 This is a flowchart of sub-step S2 of the video surveillance storage method based on distributed cloud storage according to an embodiment of the present application. Figure 3 FIG. 2 is a data flow diagram of sub-step S2 of the video surveillance storage method based on distributed cloud storage according to an embodiment of the present application. Figure 2 and Figure 3As shown, the step S2 includes the steps of: S21, obtaining a user query log, wherein the query log records all retrieval requests, wherein the retrieval requests include a time range, a camera number, and a click-through rate of retrieval results; S22, performing semantic embedding coding on each retrieval request in the user query log to obtain a set of retrieval request semantic embedding coding vectors; S23, performing unsupervised learning-based demand clustering analysis on the set of retrieval request semantic embedding coding vectors to obtain a video retrieval demand semantic clustering coding vector; S24, inputting the video retrieval demand semantic clustering coding vector into a decoder-based time block recommender to obtain a recommended value for the time block; S25, based on the recommended value for the time block, performing data separation on the surveillance video stream to obtain the multiple video data blocks.

[0033] Specifically, the step S21 obtains a user query log, and the query log records all retrieval requests, and the retrieval request includes a time range, a camera number, and a click-through rate of the retrieval result. Specifically, the present application takes into account that when the surveillance video stream is processed in blocks, the selection of the time interval will directly affect the efficiency of subsequent storage and retrieval. If the time interval is too long, the amount of information contained in each video data block is too large, and the problem of low storage and retrieval efficiency will still be faced; if the time interval is too short, it will lead to an excessive number of video data blocks, increasing the complexity of storage management and user video retrieval. Therefore, in order to improve the convenience of user video retrieval, the present application further obtains a user query log, and determines the appropriate video data block time interval by analyzing the retrieval request information in the user query log, so that the time period covered by each video data block matches the common time period in the user retrieval request, thereby improving the efficiency and accuracy of user video retrieval.

[0034] Specifically, in order to effectively collect these query logs, it is necessary to consider the system architecture level to ensure seamless integration into the existing platform. Usually, it is necessary to establish a complete log management system, which consists of multiple components, such as log collectors, log storage servers, and log analysis tools. As the front-end part, the log collector is directly connected to the user interface, responsible for capturing each retrieval request initiated by the user and encapsulating the relevant information into a structured data format. This process requires accurate extraction of parameters such as the time range entered by the user and the selected camera number, while also tracking the user's click behavior on the retrieval results to calculate the click-through rate.

[0035] When designing a log collector, it is particularly important to adopt an asynchronous processing mechanism in consideration of stability and reliability in a high-concurrency environment. Through a non-blocking data submission method, system bottlenecks caused by large amounts of log writing can be avoided. In addition, in order to ensure data consistency and integrity, a unique identifier should be generated each time a retrieval request occurs for subsequent log association and analysis. This not only facilitates tracking the complete query path of a single user, but also provides the possibility of in-depth mining of user preferences.

[0036] For the dimension of time range, the log collector needs to have flexible parsing capabilities, and should be able to accurately identify and record both absolute time (such as a specific date and time point) and relative time (such as within the past 24 hours). For camera numbers, since each surveillance camera has a unique identity, it is crucial to clearly record this information in the log. This helps with subsequent data analysis based on different camera locations or types, such as counting surveillance heat in a certain area or evaluating the working efficiency of a specific device.

[0037] Specifically, the step S22 performs semantic embedding coding on each retrieval request in the user query log to obtain a set of semantic embedding coding vectors of the retrieval request. Specifically, since the user's retrieval request record is described in text form, in order to convert it into a computer-processable vector form, the present application further uses semantic embedding coding technology to process each retrieval request in the user query log, so as to map the text information into a high-dimensional semantic space, mine the potential semantic information of each retrieval request, and generate a set of retrieval request semantic embedding coding vectors, so that the semantic correlation between different retrieval requests can be measured based on the vector distance in the semantic space, and the user's video retrieval access pattern can be understood more accurately. In one embodiment of the present application, the Word2Vec semantic embedding coding model is used to implement the semantic embedding coding process for each retrieval request in the user query log.

[0038] Specifically, in step S23, an unsupervised learning-based demand clustering analysis is performed on the set of the retrieval request semantic embedding coding vectors to obtain a video retrieval demand semantic clustering coding vector. Specifically, in order to mine the common features of user video retrieval from a large amount of video retrieval request information to reveal the user's potential video retrieval needs and access patterns, the present application proposes a demand clustering analysis method based on unsupervised learning, which performs coarse-grained aggregation on the set of the retrieval request semantic embedding coding vectors to construct a global feature summary, and combines a dynamic compensation mechanism to perform high-fidelity modeling on the detailed semantic features of each retrieval request to generate a video retrieval demand semantic clustering coding vector with global semantic structure description power and local detail semantic sensitivity. Among them, Figure 4FIG. 1 is a flowchart of sub-step S23 of the video surveillance storage method based on distributed cloud storage according to an embodiment of the present application. Figure 4 As shown, the step S23 includes the steps of: S231, inputting the set of retrieval request semantic embedding coding vectors into the information core coarse-grained aggregation network to obtain a retrieval request coarse-grained semantic aggregation coding vector; S232, based on the feature differences of each retrieval request semantic embedding coding vector in the set of retrieval request semantic embedding coding vectors relative to the retrieval request coarse-grained semantic aggregation coding vector, performing dynamic compensation aggregation coding on the set of retrieval request semantic embedding coding vectors to obtain a retrieval request fine-grained semantic compensation aggregation coding vector; S233, inputting the retrieval request fine-grained semantic compensation aggregation coding vector and the retrieval request coarse-grained semantic aggregation coding vector into the residual unit to obtain the video retrieval demand semantic clustering coding vector.

[0039] More specifically, the step S231 is expressed by the formula: in, represents the set of semantic embedding encoding vectors of the retrieval request, , , and They represent the first, second, and third in the set of semantic embedding encoding vectors of retrieval requests. and vectors, is the number of vectors in the set of semantic embedding encoding vectors of the retrieval request, and Respectively represent the maximum and minimum values, express The median of the characteristic distribution boundary, represents the normalized exponential function, express The attention weight, Represents the coarse-grained semantic aggregation encoding vector of the retrieval request.

[0040] That is, firstly, the set of the retrieval request semantic embedding coding vectors is input into the information core coarse-grained aggregation network, and based on the statistical characteristics of each retrieval request semantic embedding coding vector, its global aggregation weight is calculated to realize the coarse-grained aggregation coding of each retrieval request semantic embedding coding vector, so as to extract the main semantic patterns of the user's retrieval request and obtain a summary expression of the overall semantic characteristics of the set of the retrieval request semantic embedding coding vectors, that is, the retrieval request coarse-grained semantic aggregation coding vector.

[0041] Figure 5FIG. 2 is a flowchart of sub-step S232 of the video surveillance storage method based on distributed cloud storage according to an embodiment of the present application. Figure 5 As shown, the step S232 includes the steps of: S2321, calculating the core convergence compensation factor of each retrieval request semantic embedding coding vector in the set of retrieval request semantic embedding coding vectors relative to the retrieval request coarse-grained semantic convergence coding vector to obtain a set of retrieval request semantic core convergence compensation factors; S2322, performing compensation explicit modeling based on a gating function on the set of retrieval request semantic core convergence compensation factors to obtain a set of retrieval request semantic core convergence compensation weight factors; S2323, inputting the set of retrieval request semantic core convergence compensation weight factors, the retrieval request coarse-grained semantic convergence coding vector and the set of retrieval request semantic embedding coding vectors into a node fine-grained dynamic compensation convergence network to obtain the retrieval request fine-grained semantic compensation convergence coding vector.

[0042] In a specific example of the present application, the step S2321 includes: performing point convolution encoding based on the Sigmoid activation function on the retrieval request semantic embedding coding vector and the retrieval request coarse-grained semantic convergence coding vector to obtain a standardized retrieval request semantic embedding coding vector and a standardized retrieval request coarse-grained semantic convergence coding vector; calculating the positional difference vector between the standardized retrieval request semantic embedding coding vector and the standardized retrieval request coarse-grained semantic convergence coding vector, and taking the absolute value of the positional difference vector to obtain a retrieval request semantic core convergence difference compensation coding vector; inputting the retrieval request semantic core convergence difference compensation coding vector into a compensation feature importance scoring module based on a neural network layer to obtain the retrieval request semantic core convergence compensation factor. More specifically, the retrieval request semantic core convergence difference compensation coding vector is input into a compensation feature importance scoring module based on a neural network layer to obtain the retrieval request semantic core convergence compensation factor, including: using a weight parameter matrix to multiply the retrieval request semantic core convergence difference compensation coding vector, and adding the multiplication result to the bias term to obtain a retrieval request semantic core convergence compensation feature modulation vector; multiplying the retrieval request semantic core convergence compensation feature modulation vector by a retrieval request semantic compensation feature importance scoring conversion vector to obtain the retrieval request semantic core convergence compensation factor.

[0043] The above step S2321 can be expressed as follows: in, represents the compensation factor calculation network, express Activation function, represents a 1×1 convolution operation, and They represent the retrieval request semantic feature weight parameter matrix and the retrieval request coarse-grained semantic aggregation feature weight parameter matrix respectively. represents the normalized retrieval request semantic embedding encoding vector, represents the coarse-grained semantic aggregation encoding vector of the standardized retrieval request, It means to subtract by position point. represents the semantic kernel convergence difference compensation encoding vector of the retrieval request, represents the weight parameter matrix of the difference feature of the semantic kernel convergence of the retrieval request, represents the bias term, represents the retrieval request semantic compensation feature importance score conversion vector, The first in the set of semantic core convergence compensation factors of the retrieval request The semantic kernel convergence compensation factor of the retrieval request.

[0044] Specifically, considering that although the above-mentioned global aggregation method can capture the overall semantic trend of user retrieval requests, the compressive aggregation process of global semantic features may dilute or even completely lose the personalized semantic information of some important retrieval requests. To this end, the present application further introduces the calculation of the kernel convergence compensation factor, which dynamically generates the kernel convergence compensation factor of each retrieval request semantic embedding coding vector by measuring the feature deviation between each retrieval request semantic embedding coding vector and the retrieval request coarse-grained semantic convergence coding vector, and is used to describe the personalized semantic deviation of each retrieval request relative to the overall semantic characteristics, thereby providing strong supplementary support for the subsequent fine-grained compensation modeling of personalized semantic features of retrieval requests.

[0045] In particular, in a preferred example of the present application, a weight parameter matrix is ​​used to multiply the retrieval request semantic core convergence difference compensation coding vector, and the multiplication result is added to the bias term to obtain the retrieval request semantic core convergence compensation feature modulation vector, including: calculating the ratio between the Euclidean norm of the retrieval request semantic embedding coding vector and the Euclidean norm of the retrieval request coarse-grained semantic convergence coding vector, if the ratio is less than 1, then adding 1 to the ratio and calculating the logarithmic value with base 2 as the bias term; if the ratio is greater than or equal to 1, then using the ratio as the bias term, expressed by the formula: in, represents the calculation of the Euclidean norm of a vector, Represents the base 2 logarithmic function.

[0046] Among them, for the deviation compensation between the retrieval request semantic embedding coding vector and the retrieval request coarse-grained semantic convergence coding vector, the performance deviation of the core convergence strategy as a scenario strategy can be measured by quantifying the regret metric based on the information core compression hypothesis in the core convergence decision process, that is, the game counterfactual regret value. Specifically, first, the counterfactual regret value is provided based on the policy action through the norm representation of the vector, that is, the normalization of the vector norm representation of the retrieval request semantic embedding coding vector and the retrieval request coarse-grained semantic convergence coding vector is used as the decision point loss description, and then, according to the possible differences in the vector distribution action game scenarios, the compensation rules of the personalized information of the retrieval request are respectively corrected by the degree of information distribution of the regret value and the relative distribution amplitude of the regret value, so that the personalized information of the retrieval request is considered as the untaken action in the decision, and the biased compensation is performed based on the potential benefits of the information core convergence hypothesis.

[0047] In a specific example of the present application, the step S2322 is expressed by the formula: in, represents the gating threshold, is a natural constant, represents the gating function, The first in the set of semantic core convergence compensation weight factors of the retrieval request The semantic kernel of each retrieval request aggregates the compensation weight factor.

[0048] Specifically, after obtaining the retrieval request semantic core convergence compensation factor, the present application further regulates the effect size of each retrieval request semantic core convergence compensation factor through compensation explicit modeling based on the gating function, and dynamically selects information based on nonlinear constraints to screen the effect strength of each retrieval request semantic core convergence compensation factor, thereby generating a set of retrieval request semantic core convergence compensation weight factors, thereby ensuring that important personalized semantic information is highlighted, while irrelevant or redundant information is suppressed.

[0049] In a specific example of the present application, the step S2323 is expressed by the formula: in, A fine-grained semantic compensation aggregation encoding vector representing the retrieval request.

[0050] Specifically, based on the obtained retrieval request semantic core convergence compensation weight factor, the difference information (i.e., personalized semantic information) between the retrieval request semantic embedding coding vector and the retrieval request coarse-grained semantic convergence coding vector is compensated and aggregated to generate a retrieval request fine-grained semantic compensation convergence coding vector, thereby achieving accurate capture and modeling of the personalized semantic features of each user's retrieval request.

[0051] More specifically, the step S233 is expressed by the formula: in, and represents different weight parameters, Represents the video retrieval requirement semantic clustering encoding vector.

[0052] Specifically, the fine-grained semantic compensation aggregation coding vector of the retrieval request and the coarse-grained semantic aggregation coding vector of the retrieval request are linearly weighted aggregated through the residual unit to combine the global semantic characteristic summary expression and personalized semantic information description of the user's video retrieval needs, and generate a video retrieval requirement semantic clustering coding vector, so as to more accurately describe the user's potential needs and access patterns for video retrieval, and provide key guidance for subsequent video segmentation processing.

[0053] Specifically, in step S24, the video retrieval requirement semantic clustering coding vector is input into a decoder-based time segmentation recommender to obtain a recommended value for the time segmentation. Specifically, the decoder is based on a multi-layer perceptron architecture, and its input layer is used to receive the video retrieval requirement semantic clustering coding vector, and through the nonlinear transformation of the hidden layer, it deeply understands the user's video retrieval access mode, and decodes and generates a recommended value for the time segmentation based on the user's retrieval access characteristics. The recommended value for the time segmentation reflects the appropriate size of the video data block under the current user's video retrieval access mode.

[0054] In a preferred example, considering that each retrieval request semantic embedding coding vector in the set of retrieval request semantic embedding coding vectors respectively represents the retrieval request semantic embedding coding features, a demand clustering analysis based on unsupervised learning is performed on it. Since unsupervised learning does not introduce additional prior information, the video retrieval demand semantic clustering coding vector finally clustered is greatly affected by the noise of local sample points, which causes discontinuities at local sensitive points or edges of its high-dimensional feature manifold, affecting the accuracy of its decoding results, that is, affecting the decoding accuracy of the recommended value of the time block.

[0055] In view of the above technical problems, in the technical solution of the present application, before the video retrieval requirement semantic clustering coding vector is input into a decoder-based time segmentation recommender to obtain a recommendation value for the time segmentation, the video retrieval requirement semantic clustering coding vector is subjected to feature manifold optimization to obtain an optimized video retrieval requirement semantic clustering coding vector, and the process includes the steps of: The phase correlation information between the features of any two positions in the video retrieval requirement semantic clustering coding vector is calculated to obtain the video retrieval requirement semantic clustering coding phase full-granularity mask matrix, which is expressed as: in, and represents the first and The characteristics of the location, represents the first weight hyperparameter, represents the second weight hyperparameter, Representing video retrieval requirements semantic clustering encoding phase full granularity mask matrix The value of the position; Calculating the amplitude correlation information between the features of any two positions in the video retrieval requirement semantic clustering coding vector to obtain the video retrieval requirement semantic clustering coding amplitude full-granularity mask matrix; in, Indicates the video retrieval requirement semantic clustering encoding amplitude full granularity mask matrix The value of the position; Counting the number of effective features in the video retrieval requirement semantic clustering coding vector, and constructing a phase search optimization factor and an amplitude search optimization factor based on the number of effective features; in, , , represents the first , , The characteristics of the location, Indicates a preset threshold, e.g. , can be adjusted according to actual conditions, and this embodiment does not specifically limit it. represents the counting function, represents the number of valid features, represents the phase search optimization factor, represents the amplitude search optimization factor; Based on the video retrieval requirement semantic clustering coding phase full-granularity mask matrix and the video retrieval requirement semantic clustering coding amplitude full-granularity mask matrix, and combined with the phase search optimization factor and the amplitude search optimization factor, the video retrieval requirement semantic clustering coding vector is subjected to dual-stream enhancement based on phase search drive and amplitude search drive to obtain the video retrieval request requirement semantic clustering phase response optimized coding vector and the video retrieval request requirement semantic clustering amplitude response optimized coding vector, which are expressed as: in, represents the semantic clustering encoding vector of the video retrieval requirement, represents the transpose of a vector, represents matrix multiplication, represents vector addition, It means point multiplication by position. Represents the video retrieval requirement semantic clustering encoding phase full granularity mask matrix, Represents the video retrieval requirement semantic clustering encoding amplitude full granularity mask matrix, Represents the video retrieval request requirement semantic clustering phase response optimized coding vector, Represents the video retrieval request requirement semantic clustering magnitude response optimized encoding vector; The optimized video retrieval request semantic clustering phase response encoding vector and the optimized video retrieval request semantic clustering amplitude response encoding vector are fused to obtain an optimized video retrieval request semantic clustering encoding vector, which is expressed as: in, Represents the semantic clustering encoding vector for optimized video retrieval requirements.

[0056] Accordingly, the phase feature spectrum and amplitude feature spectrum for group aggregation quantitative evaluation are constructed through global fine-grained modeling of the phase information and amplitude information of each position of the video retrieval requirement semantic clustering coding vector, and then the amplitude-phase collaborative enhancement framework based on dual retrieval drive is established based on the phase feature spectrum and amplitude feature spectrum. At the same time, in the collaborative enhancement process, the search optimization factor is constructed based on the effective components of the features in the video retrieval requirement semantic clustering coding vector to guide the feature optimization to flow in the predetermined direction. Based on the dynamic configuration index response system of the group aggregation phase-amplitude feature topology in the open probability space, the feedback signal fusion strategy is used to effectively suppress the configuration response redundancy caused by the local overload feature parameters, and finally optimize the accurate probability fitting of the feature space under the premise of ensuring the feature integrity constraint. In this way, the decoding accuracy of the time-block recommendation value obtained by the time-block recommender based on the decoder is improved.

[0057] Specifically, in step S25, based on the recommended value of the time segmentation, the monitoring video stream is divided into data to obtain the multiple video data blocks. That is, based on the recommended value of the time segmentation, the monitoring video stream is divided into blocks to meet the video retrieval requirements of different users, while balancing the integrity and retrieval speed of the video data, optimizing the storage and retrieval efficiency of the video data, and ensuring that the user can quickly obtain the required content.

[0058] In the above-mentioned video surveillance storage method based on distributed cloud storage, the step S3 determines the target storage node of each video data block in the multiple video data blocks based on the consistent hashing algorithm. Specifically, there are multiple storage nodes in the distributed cloud storage environment. In order to achieve a balanced distribution of storage loads among the nodes, avoid excessive loads on some nodes and idleness of some nodes, and ensure that data can be stored and read efficiently, the present application adopts a consistent hashing algorithm to determine the target storage node of each video data block. The consistent hashing algorithm is a distributed hashing algorithm that can achieve uniform distribution and efficient access of data in a dynamically changing distributed system. Specifically, the consistent hashing algorithm abstracts the hash space into a virtual hash ring, and calculates its hash value based on the unique identifier (such as IP address, node number, etc.) of each storage node in the cloud storage environment, and maps it to different positions on the hash ring. Then, for the video data block to be stored, its hash value is also calculated based on its relevant features (such as the number of the data block, timestamp and other combined information), and the storage node closest to the hash value is found clockwise on the hash ring as the target storage node. In this way, when storage nodes increase or decrease, only a small amount of data needs to be migrated to other nodes without causing large-scale data redistribution, thereby ensuring that when nodes change dynamically (such as new nodes, node failures and exits), the storage and reading of data are minimally affected, achieving uniform distribution of data and dynamic balance of storage load.

[0059] In the above-mentioned video surveillance storage method based on distributed cloud storage, the step S4 stores the multiple video data blocks in multiple storage nodes of the distributed cloud storage based on the target storage nodes of each of the video data blocks. That is, the divided video data blocks are distributedly stored according to their corresponding target storage nodes to make full use of the storage resources of multiple nodes in the distributed cloud storage system and realize efficient use of storage resources. At the same time, due to the use of a distributed storage method, even if a storage node fails, other nodes can still continue to provide services, thereby ensuring the continuous and stable operation of the monitoring system. In addition, since multiple storage nodes can respond to data access requests at the same time and realize parallel processing and transmission of data, it can significantly improve data access efficiency and reduce the delay of video retrieval and playback.

[0060] In the above-mentioned video surveillance storage method based on distributed cloud storage, the step S5 monitors the usage frequency of each video data block in the multiple video data blocks, and adjusts the storage priority of each video data block based on the usage frequency of each video data block. Specifically, it is considered that different video data blocks have different access frequencies in actual use. For example, for frequently accessed video data blocks, if they are stored in a storage location with lower performance, the reading speed will be slower, affecting the user experience; and for rarely accessed video data blocks, if they always occupy high-performance storage resources, it will cause resource waste. Therefore, on the basis of distributed cloud storage, the present application further monitors the usage frequency of each video data block, and adjusts the storage priority of each video data block based on the usage frequency of each video data block, so as to achieve efficient utilization and optimal configuration of storage resources. In a specific example of the present application, the step S5 includes: in response to the usage frequency of the video data block exceeding the first preset threshold, the corresponding video data block is stored in the storage node with the best speed and delay; in response to the usage frequency of the video data block being lower than the second preset threshold, the corresponding video data block is stored in the storage node with the lowest cost in a compressed or archived manner. More specifically, the storage node with the best speed and latency is an SSD storage node, and the storage node with the lowest cost is an HDD storage node.

[0061] In summary, the video surveillance storage method based on distributed cloud storage according to the embodiment of the present application is explained, which collects surveillance video streams through cameras, and divides the surveillance video streams into time blocks based on the user's video retrieval query status to form multiple video data blocks, and then constructs a video distributed cloud storage architecture, uses a consistent hashing algorithm to determine the storage location of each video data block in the distributed cloud storage, and monitors the usage frequency of each video data block in real time, so as to dynamically adjust the storage priority of each video data block. In this way, not only can the risk of single point failure be effectively avoided, the reliability and fault tolerance of storage be improved, but also the storage priority of video data can be dynamically adjusted according to its usage frequency, so as to achieve efficient management and access of data.

[0062] Furthermore, a video surveillance storage system based on distributed cloud storage is also provided.

[0063] Figure 6 FIG. 1 is a block diagram of a video surveillance storage system based on distributed cloud storage according to an embodiment of the present application. Figure 6As shown, according to the video surveillance storage system 100 based on distributed cloud storage of the embodiment of the present application, it includes: a surveillance video stream acquisition module 110, which is used to obtain the surveillance video stream collected by the camera; a surveillance video stream block module 120, which is used to divide the surveillance video stream into multiple video data blocks; a storage node allocation module 130, which is used to determine the target storage node of each video data block in the multiple video data blocks based on the consistent hashing algorithm; a video data block storage module 140, which is used to store the multiple video data blocks in multiple storage nodes of distributed cloud storage based on the target storage nodes of each video data block; a storage priority adjustment module 150, which is used to monitor the usage frequency of each video data block in the multiple video data blocks, and adjust the storage priority of each video data block based on the usage frequency of each video data block. Specifically, the storage priority adjustment module: in response to the usage frequency of the video data block exceeding the first preset threshold, the corresponding video data block is stored in the storage node with the best speed and delay; in response to the usage frequency of the video data block being lower than the second preset threshold, the corresponding video data block is stored in the storage node with the lowest cost in a compressed or archived manner.

[0064] The specific operations of each module in the above-mentioned video surveillance storage system based on distributed cloud storage have been referenced above. Figures 1 to 5 The description of the video surveillance storage method based on distributed cloud storage has been introduced in detail, and therefore, its repeated description will be omitted.

[0065] The basic principle of the present invention is described above in conjunction with specific embodiments. However, it should be pointed out that the advantages, strengths, effects, etc. mentioned in the present invention are only examples and not limitations, and it cannot be considered that these advantages, strengths, effects, etc. must be possessed by each embodiment of the present invention. In addition, the specific details of the above embodiments are only for the purpose of illustration and facilitation of understanding, rather than limitation, and the above details do not limit the present invention to being implemented by adopting the above specific details.

[0066] In the above embodiments, the description of each embodiment has its own emphasis. For the parts that are not described or recorded in detail in a certain embodiment, please refer to the relevant description of other embodiments. In the several embodiments provided by the present invention, it should be understood that the disclosed system and method can be implemented in other ways. For example, the system embodiment described above is only schematic. For example, the unit division is only a logical function division, and there may be other division methods in actual implementation. The units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the scheme of this embodiment.

[0067] It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above and that the invention can be implemented in other specific forms without departing from the spirit or essential features of the invention. Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description, and it is intended that all variations falling within the meaning and scope of the equivalent elements of the claims be included in the invention. Any reference to a figure in a claim should not be considered as limiting the claim to which it relates.

[0068] In addition, it is obvious that the word "comprising" does not exclude other units or steps, and the singular does not exclude the plural. The multiple units stated in the system claims can also be implemented by one unit through software or hardware.

[0069] Finally, it should be noted that the above description has been given for the purpose of illustration and description. In addition, the above embodiments are only used to illustrate the technical solution of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, it should be understood by those skilled in the art that the technical solution of the present invention can be modified or replaced by equivalents without departing from the spirit and scope of the technical solution of the present invention.

Claims

1. A video surveillance storage method based on distributed cloud storage, characterized in that: include: Get the surveillance video stream collected by the camera; Dividing the surveillance video stream into a plurality of video data blocks; Determine a target storage node for each of the plurality of video data blocks based on a consistent hashing algorithm; Based on the target storage node of each of the video data blocks, storing the multiple video data blocks in multiple storage nodes of the distributed cloud storage; Monitoring the usage frequency of each video data block among the plurality of video data blocks, and adjusting the storage priority of each video data block based on the usage frequency of each video data block; The step of monitoring the usage frequency of each video data block among the plurality of video data blocks and adjusting the storage priority of each video data block based on the usage frequency of each video data block includes: In response to the usage frequency of the video data block exceeding a first preset threshold, storing the corresponding video data block in a storage node with optimal speed and delay; In response to the usage frequency of the video data block being lower than a second preset threshold, the corresponding video data block is stored in a storage node with the lowest cost in a compressed or archived manner.

2. The video surveillance storage method based on distributed cloud storage according to claim 1 is characterized in that: The storage node with the best speed and delay is the SSD storage node, and the storage node with the lowest cost is the HDD storage node.

3. The video surveillance storage method based on distributed cloud storage according to claim 2 is characterized in that: The monitoring video stream is divided into a plurality of video data blocks, including: Obtaining a user query log, wherein the query log records all search requests, wherein the search requests include a time range, a camera number, and a click rate of search results; Performing semantic embedding coding on each retrieval request in the user query log to obtain a set of retrieval request semantic embedding coding vectors; Performing a demand clustering analysis based on unsupervised learning on the set of the retrieval request semantic embedding coding vectors to obtain a video retrieval demand semantic clustering coding vector; Inputting the video retrieval requirement semantic clustering encoding vector into a decoder-based time segmentation recommender to obtain a recommendation value for the time segmentation; Based on the recommended value of the time segmentation, the monitoring video stream is data-separated to obtain the multiple video data blocks.

4. The video surveillance storage method based on distributed cloud storage according to claim 3 is characterized in that: Performing a demand clustering analysis based on unsupervised learning on the set of the retrieval request semantic embedding coding vectors to obtain a video retrieval demand semantic clustering coding vector, including: Inputting the set of search request semantic embedding coding vectors into the information core coarse-grained aggregation network to obtain a search request coarse-grained semantic aggregation coding vector; Based on the feature difference between each retrieval request semantic embedding coding vector in the set of retrieval request semantic embedding coding vectors and the retrieval request coarse-grained semantic convergent coding vector, the set of retrieval request semantic embedding coding vectors is dynamically compensated and converged to obtain a retrieval request fine-grained semantic compensated and converged coding vector; The retrieval request fine-grained semantic compensation aggregation coding vector and the retrieval request coarse-grained semantic aggregation coding vector are input into a residual unit to obtain the video retrieval requirement semantic clustering coding vector.

5. The video surveillance storage method based on distributed cloud storage according to claim 4 is characterized in that: Based on the feature difference between each retrieval request semantic embedding coding vector in the set of retrieval request semantic embedding coding vectors and the retrieval request coarse-grained semantic converged coding vector, the set of retrieval request semantic embedding coding vectors is dynamically compensated and converged to obtain a retrieval request fine-grained semantic compensated and converged coding vector, including: Calculating the core convergence compensation factor of each retrieval request semantic embedding coding vector in the set of retrieval request semantic embedding coding vectors relative to the retrieval request coarse-grained semantic convergence coding vector to obtain a set of retrieval request semantic core convergence compensation factors; Performing compensation explicit modeling based on a gating function on the set of retrieval request semantic core convergence compensation factors to obtain a set of retrieval request semantic core convergence compensation weight factors; The set of the retrieval request semantic core convergence compensation weight factors, the retrieval request coarse-grained semantic convergence coding vector and the set of the retrieval request semantic embedding coding vector are input into the node fine-grained dynamic compensation convergence network to obtain the retrieval request fine-grained semantic compensation convergence coding vector.

6. The video surveillance storage method based on distributed cloud storage according to claim 5 is characterized in that: Calculating the core convergence compensation factor of each retrieval request semantic embedding coding vector in the set of retrieval request semantic embedding coding vectors relative to the retrieval request coarse-grained semantic convergence coding vector to obtain a set of retrieval request semantic core convergence compensation factors, including: Performing point convolution encoding based on Sigmoid activation function on the retrieval request semantic embedding encoding vector and the retrieval request coarse-grained semantic aggregation encoding vector respectively to obtain a standardized retrieval request semantic embedding encoding vector and a standardized retrieval request coarse-grained semantic aggregation encoding vector; Calculating a position difference vector between the standardized search request semantic embedding coding vector and the standardized search request coarse-grained semantic convergence coding vector, and taking an absolute value of the position difference vector to obtain a search request semantic core convergence difference compensation coding vector; The retrieval request semantic kernel convergence difference compensation encoding vector is input into a compensation feature importance scoring module based on a neural network layer to obtain the retrieval request semantic kernel convergence compensation factor.

7. The video surveillance storage method based on distributed cloud storage according to claim 6 is characterized in that: Inputting the retrieval request semantic core convergence difference compensation encoding vector into a compensation feature importance scoring module based on a neural network layer to obtain the retrieval request semantic core convergence compensation factor, including: Multiplying the retrieval request semantic core convergence difference compensation coding vector by a weight parameter matrix, and adding the multiplication result to a bias term to obtain a retrieval request semantic core convergence compensation feature modulation vector; The retrieval request semantic core convergence compensation feature modulation vector is multiplied by the retrieval request semantic compensation feature importance score conversion vector to obtain the retrieval request semantic core convergence compensation factor.

8. The video surveillance storage method based on distributed cloud storage according to claim 7 is characterized in that: The method of multiplying the retrieval request semantic core convergence difference compensation coding vector by a weight parameter matrix and adding the multiplication result to a bias term to obtain a retrieval request semantic core convergence compensation feature modulation vector includes: Calculate the ratio between the Euclidean norm of the search request semantic embedding coding vector and the Euclidean norm of the search request coarse-grained semantic aggregation coding vector, and if the ratio is less than 1, add 1 to the ratio and calculate the logarithm value with base 2 as the bias term; If the ratio is greater than or equal to 1, the ratio is used as the bias term.

9. A video surveillance storage system based on distributed cloud storage, characterized in that: include: A monitoring video stream acquisition module is used to obtain the monitoring video stream collected by the camera; A monitoring video stream segmentation module, used for segmenting the monitoring video stream into a plurality of video data blocks; A storage node allocation module, configured to determine a target storage node for each of the plurality of video data blocks based on a consistent hashing algorithm; A video data block storage module, configured to store the plurality of video data blocks in a plurality of storage nodes of a distributed cloud storage based on a target storage node of each of the video data blocks; The storage priority adjustment module is used to monitor the usage frequency of each video data block among the multiple video data blocks, and adjust the storage priority of each video data block based on the usage frequency of each video data block.

10. The video surveillance storage system based on distributed cloud storage according to claim 9, characterized in that: The storage priority adjustment module: In response to the usage frequency of the video data block exceeding a first preset threshold, storing the corresponding video data block in a storage node with optimal speed and delay; In response to the usage frequency of the video data block being lower than a second preset threshold, the corresponding video data block is stored in a storage node with the lowest cost in a compressed or archived manner.

Images