A video vector fusion analysis method and system based on deep learning

Abstract

The present invention discloses a video vector fusion analysis method and system based on deep learning, which relates to the technical field of video vector fusion, and includes collecting real-time and historical video and image data and performing preprocessing; performing object detection on the preprocessed data and extracting feature vectors of the objects; constructing an adaptive depth fusion algorithm to fuse the feature vectors, classifying and labeling the fused feature vectors; performing object association based on the classified feature vectors, creating tasks for tracking and generating object trajectories. By creating an algorithm library to select a feature extraction algorithm and constructing an adaptive depth fusion algorithm, classifying and labeling the fused vectors and performing intelligent object tracking, the present invention realizes efficient, accurate and second-level response in massive data, effectively improves the intelligence level of the video system, not only can effectively improve the fusion efficiency of feature vectors, but also can automatically adjust the fusion strategy according to different application scenarios, significantly improving the video analysis efficiency and accuracy.

Description

Technical Field

[0001] The present invention relates to the technical field of video vector fusion, and particularly to a video vector fusion analysis method and system based on deep learning. Background Art

[0002] With the rapid development of artificial intelligence technology, the application of deep learning in the field of video analysis has become increasingly widespread. Through in-depth understanding of video content, video analysis technology supports functions such as object detection, behavior recognition, and event prediction in various scenarios, and is widely used in multiple fields such as urban security, intelligent transportation, and personal authentication. Especially in video surveillance systems, through in-depth analysis of real-time and historical video data, effective object tracking and behavior analysis can be carried out, thereby improving the intelligence level and application efficiency of the surveillance system. However, existing technologies face challenges in processing large-scale video data, such as low data processing efficiency, inaccurate feature extraction, and difficult object association. The data types provided by intelligent front-ends are mainly face and license plate data, and human bodies and non-motor vehicles cannot be recognized, and the data source types are relatively single. This is mainly because traditional video analysis methods lack effective adaptive mechanisms in feature extraction and vector fusion, making it difficult to adapt to the analysis requirements of different video contents and complex scenarios, thus affecting the accuracy and practicality of analysis results. Summary of the Invention

[0003] In view of the problems existing in the above-mentioned existing video vector fusion analysis methods and systems based on deep learning, the present invention is proposed.

[0004] Therefore, the problem to be solved by the present invention is that traditional video analysis methods lack effective adaptive mechanisms in feature extraction and vector fusion, making it difficult to adapt to the analysis requirements of different video contents and complex scenarios, thus affecting the accuracy and practicality of analysis results.

[0005] To solve the above technical problems, the present invention provides the following technical solution: A video vector fusion analysis method based on deep learning, which includes collecting real-time and historical video and image data and performing preprocessing; performing object detection on the preprocessed data and extracting feature vectors of the objects; constructing an adaptive depth fusion algorithm to fuse the feature vectors, classifying and labeling the fused feature vectors; performing object association based on the classified feature vectors, creating tasks for tracking and generating object trajectories.

[0006] As a preferred solution of the video vector fusion analysis method based on deep learning according to the present invention, wherein: the step of collecting real-time and historical video and image data and performing preprocessing includes the following steps,

[0007] Collect real-time video streams and historical video data from surveillance cameras and historical archives;

[0008] Screen the collected video data and remove video clips with poor quality;

[0009] The filtered video data is decomposed into individual frames at a fixed frequency, and each image frame is adjusted to a uniform size. The image is denoised, scaled, normalized, and histogram equalized.

[0010] As a preferred solution of the video vector fusion analysis method based on deep learning of the present invention, the target detection and target feature vector extraction of the preprocessed data include the following steps:

[0011] Create an algorithm library to store feature extraction algorithms, allowing administrators to modify the algorithms through a visual interface and select feature extraction algorithms based on real-time video stream processing requirements;

[0012] The pre-trained deep learning object detection model YOLOv5 is selected. The pre-processed image is input into the object detection model to detect faces, human bodies, motor vehicles, and non-motor vehicles in the image, and the bounding box and confidence score of each object are output.

[0013] We selected ResNet-50 from the pre-trained deep convolutional neural network as the feature extraction model, removed the last fully connected layer of ResNet-50, and retained the convolutional and pooling layers.

[0014] Extract the feature vector of each detected target using the following formula:

[0015]

[0016] Where F is the final extracted feature vector, E k represents the kth feature in the original feature vector E extracted from the pre-trained deep learning model, α k is the weight of the kth feature, M is the number of targets, E k,j′ and α k,j′ Represent the kth feature of the j′th target and its corresponding weight respectively;

[0017] The extracted feature vectors are normalized and bound to the timestamp and position information of the source video frame and stored in a vector database.

[0018] As a preferred solution of the video vector fusion analysis method based on deep learning of the present invention, wherein: the fusion feature vector of the adaptive deep fusion algorithm is expressed as:

[0019]

[0020] Where V fis the integrated comprehensive feature vector, T is the time length of the observation window, λ is the time decay coefficient, β i is the weight of the i-th feature vector, γ j is the weight of the j-th feature in the feature vector, n is the total number of input feature vectors, F ij represents the j-th element of the i-th feature vector F i where m is the dimension of the feature vector F i ∈ is a positive number, and t is the time variable.

[0021] As a preferred solution of the video vector fusion analysis method based on deep learning according to the present invention, wherein: the classification and labeling of the fused feature vector includes the following steps

[0022] Select SVM as the classification model, and divide the preprocessed data into a training set and a test set;

[0023] Use the training data set with labels to train the SVM model, and optimize the SVM model parameters through cross-validation and grid search;

[0024] After each training cycle, use the test set to evaluate the model performance;

[0025] If there is no obvious change in the verification difference after the training cycle, stop the model training to obtain the trained SVM model;

[0026] Input the fused feature vector into the SVM model for classification, and assign the corresponding class label to each detected target feature vector according to the classification result and store it in the target database.

[0027] As a preferred solution of the video vector fusion analysis method based on deep learning according to the present invention, wherein: the target association based on the classified feature vector includes the following steps

[0028] Initialize the target association matrix to record the association degree between different targets. The association degree calculation formula is

[0029]

[0030] In the formula, R AB is the association degree between target A and target B, V A is the feature vector of target A, V B is the feature vector of target B, δ(A,B) is the distance between target A and target B in the physical space, δ max is the maximum space distance, τ A and τ B are the timestamps of target A and target B, τ maxis the maximum time difference, and ω1, ω2, and ω3 are weight coefficients;

[0031] Update the target association matrix according to the calculated correlation degree.

[0032] As a preferred solution of the video vector fusion analysis method based on deep learning according to the present invention, wherein: the creation of tasks for tracking and generating target trajectories includes the following steps,

[0033] Retrieve matching targets in the target database according to the query conditions input by the user, determine other targets highly associated with the query target through the target association matrix, and create a tracking task for each relevant target;

[0034] Use the Deep SORT tracking algorithm to track the positions and movement trajectories of targets in consecutive video frames. During the tracking process, dynamically update the states of the targets and the information of the tracking tasks;

[0035] Construct the movement trajectories of each target in the video sequence according to the target position information recorded by the tracking algorithm;

[0036] Display the target trajectories and tracking results through the user interface.

[0037] Another object of the present invention is to provide a video vector fusion analysis system based on deep learning, which includes,

[0038] A collection module for collecting real-time and historical video and image data for preliminary processing;

[0039] A vector extraction module for identifying and locating targets in the preprocessed video and image data, and extracting the feature vectors of the identified targets;

[0040] A vector fusion module for dynamically fusing the extracted feature vectors and classifying and labeling the fused feature vectors;

[0041] A target association module for calculating the target association degree through the classified target feature vectors in the target database and updating the target association matrix;

[0042] A target tracking module for retrieving matching targets according to the query conditions given by the user and tracking the targets to generate target trajectories.

[0043] A computer device includes: a memory and a processor; the memory stores a computer program, and when the processor executes the computer program, the steps of the video vector fusion analysis method based on deep learning are implemented.

[0044] A computer-readable storage medium stores a computer program thereon, and when the computer program is executed by a processor, the steps of a video vector fusion analysis method based on deep learning are implemented.

[0045] The beneficial effects of the present invention are as follows: By creating an algorithm library to select feature extraction algorithms and constructing an adaptive depth fusion algorithm, classifying and labeling the fusion vectors and performing intelligent target tracking, the present invention achieves efficient, accurate and second-level response in massive data, effectively improves the intelligent level of the video system, not only can effectively improve the fusion efficiency of feature vectors, but also can automatically adjust the fusion strategy according to different application scenarios, significantly improving the video analysis efficiency and accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative work.

[0047] Figure 1 It is a schematic flowchart of a video vector fusion analysis method based on deep learning.

[0048] Figure 2 It is a schematic diagram of the implementation of a video vector fusion analysis method based on deep learning.

[0049] Figure 3 It is a schematic diagram of the structure of a video vector fusion analysis system based on deep learning. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0050] In order to make the above objects, features and advantages of the present invention more obvious and understandable, the following will describe the specific embodiments of the present invention in detail with reference to the drawings of the specification.

[0051] In the following description, many specific details are set forth in order to fully understand the present invention. However, the present invention can also be implemented in other ways different from those described herein. Those skilled in the art can make similar generalizations without departing from the connotation of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed below.

[0052] Secondly, the so-called "one embodiment" or "embodiment" herein refers to a specific feature, structure or characteristic that can be included in at least one implementation manner of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it an embodiment that is separate or selectively mutually exclusive with other embodiments.

[0053] Embodiment 1

[0054] Reference Figure 1 and Figure 2 , which is the first embodiment of the present invention. This embodiment provides a method for video vector fusion analysis based on deep learning. The method for video vector fusion analysis based on deep learning includes the following steps:

[0055] S1. Collect real-time and historical video and image data and perform preprocessing;

[0056] Specifically, collecting real-time and historical video and image data and performing preprocessing includes the following steps.

[0057] Collect real-time video streams and historical video data from surveillance cameras and historical archives;

[0058] Screen the collected video data and eliminate video clips with poor quality;

[0059] Decompose the screened video data into individual frames at a fixed frequency, adjust each image frame to a unified size, and perform denoising, scaling, normalization processing, and histogram equalization processing on the images.

[0060] By collecting real-time video streams and historical video data from surveillance cameras and historical archives, the extensiveness and diversity of the data are ensured, providing a rich information source for subsequent analysis. Further screening and preprocessing operations, such as eliminating video clips with poor quality, adjusting image size, denoising, scaling, normalization, and histogram equalization processing, not only improve the quality of the data but also ensure the consistency and accuracy of the data in subsequent processing, optimize the input data, reduce the computational burden of the system, and improve the efficiency and accuracy of object detection and feature extraction in subsequent steps.

[0061] S2. Perform object detection on the preprocessed data and extract the feature vectors of the objects;

[0062] Specifically, performing object detection on the preprocessed data and extracting the feature vectors of the objects includes the following steps.

[0063] Create an algorithm library to store feature extraction algorithms, allowing administrators to modify the algorithms through a visualization interface and select feature extraction algorithms according to the real-time video stream processing requirements;

[0064] Select the pre-trained deep learning object detection model YOLOv5, input the preprocessed images into the object detection model, detect the face, body, motor vehicle, and non-motor vehicle objects in the images, and output the bounding box and confidence score of each object;

[0065] Select ResNet-50 in the pre-trained deep convolutional neural network as the feature extraction model, remove the last fully connected layer of ResNet-50, and retain the convolutional layer and pooling layer;

[0066] Extract the feature vector of each detected target, and the formula is as follows:

[0067]

[0068] In the formula, F is the finally extracted feature vector, and E k represents the k-th feature in the original feature vector E extracted from the pre-trained deep learning model, and α k is the weight of the k-th feature, which is used to adjust the contribution degree of different features in the final feature vector. M is the number of targets, and E k,j′ and α k,j′ respectively represent the k-th feature of the j'-th target and its corresponding weight;

[0069] Normalize the extracted feature vector and bind the feature vector with the timestamp and position information of the source video frame and store it in the vector database.

[0070] By using the pre-trained deep learning object detection model YOLOv5 and the feature extraction model ResNet-50, efficient and accurate object detection and feature vector extraction are achieved. Through the capabilities of the deep learning model, key objects such as faces, human bodies, motor vehicles, and non-motor vehicles can be detected from the video, and the feature vectors of these objects can be accurately extracted, allowing for the rapid identification and extraction of key information in the video, laying a solid foundation for more in-depth video content analysis. The administrator can adjust the feature extraction algorithm through the visualization interface, increasing the flexibility and adaptability of the system.

[0071] S3. Construct an adaptive deep fusion algorithm to fuse the feature vectors and classify and label the fused feature vectors;

[0072] Specifically, the construction of the adaptive deep fusion algorithm to fuse the feature vectors is expressed as

[0073]

[0074] In the formula, V f is the fused comprehensive feature vector, T is the time length of the observation window, λ is the time decay coefficient, which controls the decay rate of the feature weight over time, β i is the weight of the i-th feature vector, γ j is the weight of the j-th feature in the feature vector, n is the total number of input feature vectors, and F ij represents the j-th element of the i-th feature vector F i of mi The dimension, ∈ is a positive number, and t is a time variable.

[0075] Through the constructed adaptive depth fusion algorithm, it is able to effectively fuse the feature vectors detected from different targets, creating a comprehensive feature representation. This fusion not only considers the mutual relationship between features but also introduces temporal dynamics through the temporal decay coefficient, enabling the fused feature vectors to reflect the changes in video content over time, enhancing the expressive power of the feature vectors, making them more capable of accurately reflecting the key information and dynamic changes in the video, and providing a high-quality data basis for subsequent classification and target association.

[0076] Furthermore, classifying and labeling the fused feature vectors includes the following steps.

[0077] Select SVM as the classification model and divide the preprocessed data into a training set and a test set.

[0078] Train the SVM model using the labeled training dataset and optimize the SVM model parameters through cross-validation and grid search.

[0079] After each training cycle, evaluate the model performance using the test set.

[0080] If there is no obvious change in the validation difference after the training cycle, stop the model training to obtain the trained SVM model.

[0081] Input the fused feature vectors into the SVM model for classification, and assign corresponding class labels to each detected target feature vector according to the classification results and store them in the target database.

[0082] Using the SVM classification model to classify and label the fused feature vectors realizes the efficient parsing and understanding of video content. Through precise classification, each detected target is assigned a clear class label, which not only enhances the accuracy and interpretability of video analysis but also provides a reliable basis for subsequent target tracking and behavior analysis. The SVM model parameters optimized by cross-validation and grid search ensure the efficiency and accuracy of the classification process, further improving the overall performance of the system.

[0083] S4. Perform target association based on the classified feature vectors, create tasks for tracking, and generate target trajectories.

[0084] Specifically, performing target association based on the classified feature vectors includes the following steps.

[0085] Initialize the target association matrix to record the association degree between different targets. The formula for calculating the association degree is

[0086]

[0087] In the formula, R AB is the correlation degree between target A and target B, representing the overall similarity and relevance of the two targets. V A is the feature vector of target A, and V B is the feature vector of target B. δ(A,B) is the distance between target A and target B in physical space, and δ max is the maximum space distance, τ A and τ B are the timestamps of target A and target B, and τ max is the maximum time difference. ω1, ω2, and ω3 are weight coefficients;

[0088] Update the target association matrix according to the calculated correlation degree.

[0089] By initializing the target association matrix and calculating the correlation degree between different targets, a quantitative description of the relationship between targets in the video is achieved. The calculation of the correlation degree comprehensively considers the feature similarity, spatial distance, and time difference between targets, thereby providing a method for comprehensively evaluating target relevance. It can not only reveal the potential connections between targets but also provide an important basis for subsequent target tracking and behavior analysis, improving the accuracy of target association analysis. By comprehensively considering multiple factors, the possibility of misjudgment and missed judgment is reduced, and the adaptability of the system to complex scenarios is enhanced, enabling more accurate identification and association of targets in crowded or occluded situations.

[0090] Furthermore, creating tasks for tracking and generating target trajectories includes the following steps:

[0091] Retrieve matching targets in the target database according to the query conditions input by the user, determine other targets highly associated with the query target through the target association matrix, and create a tracking task for each relevant target;

[0092] Use the Deep SORT tracking algorithm to track the positions and movement trajectories of targets in consecutive video frames. During the tracking process, dynamically update the states of the targets and the information of the tracking tasks;

[0093] Construct the movement trajectories of each target in the video sequence according to the target position information recorded by the tracking algorithm;

[0094] Display the target trajectories and tracking results through the user interface.

[0095] The Deep SORT algorithm is an algorithm for multi-object tracking that combines deep learning techniques to improve the accuracy and robustness of tracking. By retrieving objects based on user query conditions, determining relevant objects using an object association matrix and creating tracking tasks, and using the Deep SORT algorithm to track object positions and generate trajectories, it achieves effective tracking and behavior analysis of specific objects in a video. By dynamically updating object states and tracking information, the system can adapt to changes in objects in the video and maintain the continuity of tracking. Finally, by displaying object trajectories and tracking results through a user interface, it provides an intuitive way to observe and analyze object behavior, improving the flexibility and pertinence of object tracking. By targeting specific objects based on user query conditions, it reduces interference from irrelevant objects.

[0096] Embodiment 2

[0097] Refer to Figure 3 , which is the second embodiment of the present invention. This embodiment is different from the previous one and provides a video vector fusion analysis system based on deep learning, which includes

[0098] A collection module for collecting real-time and historical video and image data for preliminary processing;

[0099] A vector extraction module for identifying and positioning objects in the preprocessed video and image data and extracting feature vectors of the identified objects;

[0100] A vector fusion module for dynamically fusing the extracted feature vectors and classifying and labeling the fused feature vectors;

[0101] An object association module for calculating object association degrees and updating an object association matrix through the classified object feature vectors in an object database;

[0102] An object tracking module for retrieving matching objects according to query conditions given by the user and tracking the objects to generate object trajectories.

[0103] When the above-mentioned functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present invention. The aforementioned storage medium includes: various media such as USB flash drives, mobile hard disks, read-only memories (ROM, Read-Only Memory), random access memories (RAM, Random Access Memory), magnetic disks, or optical discs that can store program codes.

[0104] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a definite sequence list of executable instructions for implementing logical functions, and can be specifically implemented in any computer-readable medium for use by an instruction execution system, apparatus, or device (such as a computer-based system, a system including a processor, or other systems that can fetch and execute instructions from the instruction execution system, apparatus, or device), or in combination with these instruction execution systems, apparatus, or devices. For the purposes of this specification, a "computer-readable medium" can be any device that can contain, store, communicate, propagate, or transport a program for use by or in combination with an instruction execution system, apparatus, or device.

[0105] More specific examples (non-exhaustive list) of computer-readable media include the following: an electrical connection part (electronic device) having one or more wirings, a portable computer disk cartridge (magnetic device), a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber device, and a portable compact disc read-only memory (CDROM). Additionally, a computer-readable medium can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other media, then editing, interpreting, or otherwise processing it as appropriate, and then storing it in a computer memory.

[0106] It should be understood that various parts of the present invention can be implemented by hardware, software, firmware or a combination thereof. In the above embodiments, multiple steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, any one of the following technologies well known in the art or a combination thereof can be used: discrete logic circuits having logic gate circuits for implementing logic functions on data signals, application specific integrated circuits having appropriate combinational logic gate circuits, programmable gate arrays (PGAs), field programmable gate arrays (FPGAs), etc.

Claims

1. A video vector fusion analysis method based on deep learning, characterized by: include, Collect real-time and historical video and image data and perform pre-processing; Perform target detection on the preprocessed data and extract the target's feature vector; Construct an adaptive deep fusion algorithm to fuse feature vectors, and classify and label the fused feature vectors; Target association is performed based on the classified feature vectors, and tasks are created for tracking and generating target trajectories. The adaptive deep fusion algorithm fusion feature vector is expressed as: Where V f is the integrated feature vector after fusion, T is the time length of the observation window, λ is the time attenuation coefficient, β i is the weight of the i-th eigenvector, γ j is the weight of the jth feature in the feature vector, n is the total number of input feature vectors, F ij represents the i-th eigenvector F i The jth feature of F, m is the feature vector F i The dimension of ,∈ is a positive number, and t is a time variable.

2. The video vector fusion analysis method based on deep learning according to claim 1, characterized in that: The collection of real-time and historical video and image data and pre-processing includes the following steps: Collect real-time video streams and historical video data from surveillance cameras and historical archives; Screen the collected video data and remove video clips with poor quality; The filtered video data is decomposed into individual frames at a fixed frequency, and each image frame is adjusted to a uniform size. The image is denoised, scaled, normalized, and histogram equalized.

3. The video vector fusion analysis method based on deep learning according to claim 2, characterized in that: The target detection and target feature vector extraction of the pre-processed data include the following steps: Create an algorithm library to store feature extraction algorithms, allowing administrators to modify the algorithms through a visual interface and select feature extraction algorithms based on real-time video stream processing requirements; The pre-trained deep learning object detection model YOLOv5 is selected. The pre-processed image is input into the object detection model to detect faces, human bodies, motor vehicles, and non-motor vehicles in the image, and the bounding box and confidence score of each object are output. We selected ResNet-50 from the pre-trained deep convolutional neural network as the feature extraction model, removed the last fully connected layer of ResNet-50, and retained the convolutional and pooling layers. Extract the feature vector of each detected target using the following formula: Where F is the final extracted feature vector, E k represents the kth feature in the original feature vector E extracted from the pre-trained deep learning model, α k is the weight of the kth feature, M is the number of targets, and E k,j′ and α k,j′ Represent the kth feature of the j′th target and its corresponding weight respectively; The extracted feature vectors are normalized and bound to the timestamp and position information of the source video frame and stored in a vector database.

4. The video vector fusion analysis method based on deep learning according to claim 3, characterized in that: The classification and labeling of the fused feature vectors includes the following steps: Select SVM as the classification model and divide the preprocessed data into training set and test set; Train the SVM model using a labeled training dataset and optimize the SVM model parameters through cross-validation and grid search; After each training cycle, the test set is used to evaluate the model performance; If the verification difference does not change significantly after the training cycle, the model training is stopped to obtain the trained SVM model; The fused feature vector is input into the SVM model for classification. According to the classification results, each detected target feature vector is assigned a corresponding category label and stored in the target database.

5. The video vector fusion analysis method based on deep learning according to claim 4, characterized in that: The target association based on the classified feature vector includes the following steps: Initialize the target association matrix and record the association between different targets. The association calculation formula is: Where R AB is the correlation between target A and target B, V A is the feature vector of target A, V B is the feature vector of target B, δ(A,B) is the distance between target A and target B in physical space, δ max is the maximum spatial distance, τ A and τ B are the timestamps of target A and target B, τ max is the maximum time difference, ω1, ω2, ω3 are weight coefficients; Update the target association matrix based on the calculated association.

6. The video vector fusion analysis method based on deep learning according to claim 5, characterized in that: The creation of a task to track and generate a target trajectory includes the following steps: Retrieve matching targets from the target database based on the query conditions entered by the user, determine other targets that are highly correlated with the query target through the target association matrix, and create a tracking task for each related target; Use the Deep SORT tracking algorithm to track the target's position and motion trajectory in consecutive video frames. During the tracking process, dynamically update the target's status and tracking task information; Based on the target position information recorded by the tracking algorithm, the motion trajectory of each target in the video sequence is constructed; Display target trajectory and tracking results through the user interface.

7. A system based on the deep learning-based video vector fusion analysis method according to any one of claims 1 to 6, characterized in that: include, Collection module, used to collect real-time and historical video and image data for preliminary processing; A vector extraction module is used to identify and locate targets in preprocessed video and image data and extract feature vectors of identified targets; Vector fusion module, used to dynamically fuse the extracted feature vectors and classify and label the fused feature vectors; A target association module, used to calculate the target association degree and update the target association matrix through the target feature vectors classified in the target database; The target tracking module is used to retrieve matching targets based on the query conditions given by the user and track the targets to generate target trajectories.

8. A computer device comprising: memory and processor; The memory stores a computer program, characterized in that when the processor executes the computer program, the steps of the deep learning-based video vector fusion analysis method according to any one of claims 1 to 6 are implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the deep learning-based video vector fusion analysis method according to any one of claims 1 to 6 are implemented.

Images