Retrieval method and device fusing attribute features and vector embedding and storage medium

By integrating attribute features and vector embedding, the problems of false detection, missed detection and artificial fatigue in video retrieval are solved, and rapid and accurate retrieval in complex scenarios is achieved, and the ability to identify human targets is improved.

CN118916515APending Publication Date: 2024-11-08SHANGHAI WONDERTEK SOFTWARE CORP LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410971650.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-19
Publication Date
2024-11-08

AI Technical Summary

Technical Problem

Existing video retrieval technology has problems such as false detection, missed detection and manual fatigue when processing large-scale monitoring data. Especially in complex scenarios such as lighting changes, occlusion, motion blur, etc., the search accuracy is affected and cross-mode There are problems such as modal mismatch and semantic gap in state matching.

Method used

Using the search method of fusion attribute features and vector embedding, images are extracted by monitoring video streams, face and body feature vectors are extracted using AI graph model and deep learning model, and structured attribute features are obtained in combination with the logistic Stie regression model and stored in In the vector library, similarity calculation and sorting are performed to achieve fast and accurate retrieval.

Benefits of technology

Improve the accuracy of matching face and human features, better capture the diversity and complexity of human objects, enhance the recognition ability of different scenes and environments, and ensure that the videos are quickly and accurately retrieved under multiple cameras. Personnel information.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118916515A_ABST
    Figure CN118916515A_ABST
Patent Text Reader

Abstract

The invention relates to the technical field of video retrieval, in particular to an attribute feature and vector embedding fused retrieval method, which comprises the following steps of: S1, regularly extracting pictures of a video stream, coding the pictures and caching the pictures into a message queue; s2, the AI picture cutting model returns a coded picture ID and a coordinate point position; s3, obtaining a first face feature vector, a first human body feature vector, a first text vector and a structured attribute feature from the message queue, and further obtaining a second face feature vector; s4, storing the second human face feature vector, the first human body feature vector and the first text vector into a vector library; s5, storing the structured attribute features into a search engine; s6, extracting a retrieval vector according to the retrieval information, and obtaining a retrieval result according to a similarity calculation result; and S7, dotting the retrieval result, merging the continuous click areas, and performing video playback by the user according to the marked click areas. According to the method, efficient, stable and intelligent video retrieval in a massive camera environment is realized.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of video retrieval technology, and in particular to a retrieval method, device and storage medium integrating attribute features with vector embedding. Background Art

[0002] With the continuous development and popularization of technology, cameras have become an indispensable part of our lives. From public places to private areas, the number of cameras has increased exponentially, which has provided unprecedented convenience for monitoring and security. However, this large-scale monitoring network has also brought a severe challenge: the processing and retrieval of monitoring data has become extremely difficult.

[0003] Since surveillance data may have different formats, resolutions, viewing angles, and environmental conditions, this poses a greater challenge to data processing and matching. For example, if you want to find a person's video playback from a video, the traditional data processing method may use manual playback, which requires manual search frame by frame or segment by segment, which requires a lot of time and human resources. Especially in large-scale video surveillance systems, it may take a long time to find the target person. In addition, humans are prone to fatigue during long video playback, and it is difficult to concentrate. They may miss the appearance of the target person or important details, which may lead to missing important information or misidentifying the target person. At the same time, people's subjective judgment may be affected by factors such as emotions and fatigue, resulting in instability in the search results.

[0004] Although there are some video retrieval methods based on similarity metrics, there are still problems with false detection and missed detection. Especially in complex scenes, such as lighting changes, occlusion, motion blur, etc., the accuracy of retrieval may be affected. In some cases, video retrieval requires cross-modal matching, that is, matching text queries with video content. This may face problems such as modality mismatch and semantic gap, resulting in poor retrieval results. Summary of the invention

[0005] The purpose of the present invention is to solve the shortcomings of the prior art and provide a retrieval method integrating attribute features and vector embedding, comprising the following steps:

[0006] S1: monitor a video stream, periodically extract pictures of the video stream, encode the pictures to obtain encoded pictures, and cache the encoded pictures in a message queue;

[0007] S2: The AI ​​cropping model returns the coded image ID and coordinate point position. If the coordinate point position has a value, the coded image is saved to the disk;

[0008] S3: The deep learning model extracts the information on the coded image from the message queue to obtain a first facial feature vector, a first human feature vector and a first text vector, and normalizes the first facial feature vector to obtain a second facial feature vector. The logistic regression model obtains structured attribute features according to the coded image.

[0009] S4: the deep learning model detects the first coordinate point position among the coordinate point positions, and associates the second face feature vector, the first body feature vector and the first text vector according to the picture ID and stores them in a vector library;

[0010] S5: the logistic regression model detects the second coordinate point position of the coordinate point position, associates the structured attribute feature, the second face feature vector and the first body feature vector according to the picture ID to implement storing the structured attribute feature in the search engine;

[0011] S6: The deep learning model extracts a search vector according to the search information, performs similarity calculation between the search vector and the data in the vector library, and sorts the search results according to the calculation results of the similarity to obtain the search results;

[0012] S7: Mark the search results, merge the continuous clicked areas, and the user can review the video according to the marked clicked areas to play the previous and next clips of the corresponding time point.

[0013] Preferably, in step S3, the structured attribute features include:

[0014] The structured attribute features include features such as clothing style, color, and gender.

[0015] Preferably, in step S2, normalizing the first facial feature vector to obtain a second facial feature vector further comprises:

[0016] The calculation process of the normalization process is as follows:

[0017] v=[v1,v2,…,v i ],

[0018] Where v is the eigenvector, v i represents the i-th element in the feature vector;

[0019] Calculate the L2 norm based on the feature vector v,

[0020]

[0021] where v a represents the ath element in the feature vector;

[0022] Each element in the feature vector v is multiplied by a normalization factor to obtain a normalized second face feature vector, wherein the calculation formulas of the normalization factor normalizationFactor and the second face feature vector normalizationVector are as follows:

[0023] normalizationFactor=1 / ||v||2,

[0024] The calculation formula of the second face feature vector is as follows:

[0025] normalizationVector=[v1 / ||v||2, v2 / ||v||2,…,v n / ||v||2].

[0026] Preferably, in step S4, the deep learning model extracts a search vector according to the search information, further comprising:

[0027] The deep learning model extracts search vectors through image search or cross-modal text search as follows:

[0028] The image retrieval is that the deep learning model automatically extracts the first human feature vector and the third face feature vector according to the video crop selected by the user or the uploaded image file;

[0029] The cross-modal text retrieval is that the deep learning model extracts a second text vector based on the text vector features input by the user.

[0030] Preferably, in step S6, the search vector is similarly calculated with the data in the vector library, and the search results are sorted according to the similarity calculation results, further comprising:

[0031] When the search vector is a third facial feature vector, the third facial feature vector is similarly calculated with the data in the vector library, and matched with the fourth facial feature vector and the second human body feature vector in the vector library, and similarity calculation is performed. The system performs an intersection-and-union (IOU) calculation association based on the human body coordinates in the query result and the human body coordinates stored in the search engine library;

[0032] When the search vector is the second text vector, the second text vector is similarly calculated and sorted with the first text vector in the vector library, and a search result is obtained by searching according to the second attribute feature selected by the user and in combination with the fifth human body attribute vector.

[0033] Preferably, similarity calculation is performed between the third feature vector of the face and the data in the vector library, and the search results are sorted according to the calculation results of the similarity, further comprising:

[0034] The cosine similarity calculation formula of the third face feature vector and the fourth face feature vector is as follows:

[0035] cosine_similarity(A,B)=(A·B) / (||A||*||B||),

[0036] Where A·B represents the dot product of vector A and vector B, and ||A|| and ||B|| represent the L2 norm of vector A and vector B respectively.

[0037] The fourth face feature vector is subjected to deduplication processing using a human body deduplication model;

[0038] The second human body feature is subjected to cosine similarity calculation with the first human body feature in the vector library, sorted according to the calculation result, and a matching result is returned;

[0039] Querying the corresponding second facial feature vector in the vector library according to the matching result, performing cosine similarity calculation with the fifth facial feature vector selected by the user, and further sorting the results to obtain the query result;

[0040] The third human body coordinates in the query result are associated with the second human body coordinates stored in the search engine library by intersection and union calculation, and are compared with the picture in combination with the structured attribute features in the search engine library.

[0041] Preferably, performing a search based on the second attribute feature selected by the user in combination with the fifth human body attribute vector to obtain a search result further includes:

[0042] The second attribute feature is fused with the second human feature in the search engine library to obtain a fused feature;

[0043] The fusion features are matched and sorted using an intersection sorting method to obtain a retrieval result.

[0044] Preferably, the step S7 further comprises:

[0045] The search results are displayed in the form of thumbnails and sorted by dots, and video image information is displayed. The user can review the video by clicking on the marked dotted area and play the previous and next clips of the corresponding time point. The video image information includes information such as video location, video time, human clothing style, human clothing color, and gender.

[0046] The present invention also provides a computer device, including a memory and a processor, wherein the memory stores computer-readable instructions, and when the computer-readable instructions are executed by the processor, the processor executes the steps of the cross-modal retrieval method for integrating human attribute features with vector embedding as described in any one of the above.

[0047] The present invention also provides a storage medium storing computer-readable instructions, which, when executed by one or more processors, enables the one or more processors to perform the steps of the cross-modal retrieval method for integrating human attribute features and vector embedding as described in any one of the above.

[0048] Compared with the prior art, the present invention has the following beneficial effects:

[0049] 1) By correlating the features of the face and the body, the system can more accurately identify and match the target, thereby improving the overall matching accuracy;

[0050] 2) The present invention can better capture the diversity and complexity of human targets through integrated learning of human structured attributes and vector embedding, improve the ability to recognize human targets in different scenes and environments, and perform secondary retrieval and filtering according to attributes such as human clothing color, style and gender to improve matching accuracy.

[0051] 3) The present invention introduces face and body fusion technology and structured attribute features, combined with a deep learning model, can quickly and accurately retrieve the information of people in the video under multiple cameras, thereby efficiently and stably generating video playback of the corresponding time period. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Various other advantages and benefits will become apparent to those of ordinary skill in the art by reading the following detailed description of the preferred embodiment.The drawings are only for the purpose of illustrating the preferred embodiments and are not to be construed as limiting the invention.

[0053] Figure 1 It is a flow chart of the retrieval method of the present invention integrating attribute features and vector embedding;

[0054] Figure 2 A schematic diagram of recording the coded picture ID and coordinate point positions of the present invention;

[0055] Figure 3 is a schematic diagram of human body structured data of the present invention;

[0056] Figure 4 Another flow chart of the retrieval method of the present invention that integrates attribute features and vector embedding;

[0057] Figure 5The present invention is a flowchart of user retrieval in the retrieval method integrating attribute features and vector embedding. DETAILED DESCRIPTION

[0058] In order to make the purpose, technical scheme and advantages of the present invention clearer, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. Obviously, the described embodiments are part of the embodiments of the present application, rather than all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in the field without making creative work are within the scope of protection of this application.

[0059] Those skilled in the art will appreciate that, unless otherwise stated, the singular forms "a", "an", and "the" used herein may also include plural forms. It should be further understood that the term "comprising" used in the specification of the present invention refers to the presence of the features, integers, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or groups thereof.

[0060] Example 1

[0061] See also Figure 1 , Figure 2 , Figure 3 and Figure 4 As shown, the retrieval method for integrating attribute features and vector embedding provided in this embodiment includes the following steps:

[0062] S1: monitor the video stream, regularly extract pictures of the video stream, encode the pictures to obtain encoded pictures, and cache the encoded pictures to a message queue. Specifically, in this embodiment, Base64 encoding is performed on each picture;

[0063] S2: The AI ​​cropping model returns the coded image ID and the coordinate point position. If the coordinate point position has a value, the coded image is saved to the disk. Specifically, in this embodiment, Figure 2 As shown, the picture saved on the disk is used for subsequent interface retrieval and display, and the coded picture ID and coordinate point position are recorded;

[0064] S3: The deep learning model extracts information on the coded image from the message queue to obtain a first face feature vector, a first body feature vector and a first text vector, and normalizes the first face feature vector to obtain a second face feature vector. The logistic regression model obtains structured attribute features according to the coded image. Specifically, in this embodiment, a convolutional neural network collects a large amount of image data and performs preprocessing operations such as cropping, size normalization, brightness and contrast adjustment. The trained deep learning model extracts the first face feature vector, the first body feature vector, etc. from the preprocessed coded image.

[0065] S4: The deep learning model detects the first coordinate point position in the coordinate point position, associates the second face feature vector, the first body feature vector and the first text vector according to the picture ID and stores them in the vector library, so as to facilitate subsequent retrieval and analysis. Specifically, in this embodiment, the body structured data such as Figure 3 The shown is stored in the vector library;

[0066] S5: The logistic regression model detects the second coordinate point position of the coordinate point position, associates the structured attribute features, the second face feature vector and the first body feature vector according to the picture ID to store the structured attribute features in the search engine, and realizes the fusion of body attribute features. This fusion method can improve the understanding and analysis ability of subsequent retrieval of body attributes, so that the feature information of the body can be more accurately identified and extracted;

[0067] S6: The deep learning model extracts a search vector based on the search information, calculates the similarity between the search vector and the data in the vector library, and sorts the search results based on the similarity calculation results;

[0068] S7: Mark the search results and merge the continuous clicked areas. The user can review the video according to the marked clicked areas and play the previous and next clips of the corresponding time point.

[0069] Preferably, the AI ​​cropping model returns the encoded image ID and coordinate point position by calling the AI ​​interface, including the following steps:

[0070] S21: providing a list of multiple encoded image URLs, where the encoded image URLs point to the encoded images that need to be processed;

[0071] S22: specify the access address of the AI ​​service, which is used to send a request to the AI ​​cropping model for image processing;

[0072] S23: After activating the AI ​​photo service, you can view the Appld and Token on the AI ​​photo page for identity authentication and authorization;

[0073] S24: used to track the processing status of the request and provide a unique identifier of the request, wherein the request status indicates the processing result of the request, such as success, failure, etc., and the request status specific information provides detailed information about the request processing, such as an error code or a success message;

[0074] S25 returns the inspection result, wherein the inspection result includes detailed information on the image processing result, such as the encoded image ID and coordinate point location.

[0075] Preferably, in step S3, the structured attribute features include:

[0076] Structured attribute features include clothing style, color, and gender.

[0077] Preferably, in step S3, normalizing the first facial feature vector to obtain the second facial feature vector further comprises:

[0078] The calculation process of normalization is as follows:

[0079] v=[v1,v2,…,v i ],

[0080] Where v is the eigenvector, v i represents the i-th element in the feature vector;

[0081] Calculate the L2 norm based on the feature vector v,

[0082]

[0083] where v a represents the ath element in the feature vector;

[0084] Multiply each element in the feature vector v by the normalization factor to obtain the normalized second face feature vector, where the calculation formulas of the normalization factor normalizationFactor and the second face feature vector normalizationVector are as follows:

[0085] normalizationFactor=1 / ||v||2,

[0086] The calculation formula of the second face feature vector is as follows:

[0087] normalizationVector=[v1 / ||v||2, v2 / ||v||2,…,v n / ||v||2].

[0088] Preferably, in step S4, the deep learning model extracts a retrieval vector according to the retrieval information, further comprising:

[0089] The deep learning model extracts the search vector through image search or cross-modal text search as follows:

[0090] Image retrieval is a deep learning model that automatically extracts the first human feature vector and the third face feature vector based on the video crop selected by the user or the uploaded image file;

[0091] Cross-modal text retrieval is a deep learning model that extracts a second text vector based on the text vector features input by the user.

[0092] Preferably, in step S6, similarity calculation is performed between the search vector and the data in the vector library, and the search results are sorted according to the similarity calculation results to obtain the search results, further comprising:

[0093] When the search vector is the third facial feature vector, the third facial feature vector is calculated similarly with the data in the vector library, and the fourth facial feature vector and the second human feature vector in the vector library are matched to perform similarity calculation. The system will perform intersection-over-union (IOU) calculation and association based on the human body coordinates in the query result and the human body coordinates stored in the search engine library;

[0094] When the search vector is the second text vector, the second text vector is similarly calculated and sorted with the first text vector in the vector library, and the search result is obtained by searching according to the second attribute feature selected by the user and in combination with the fifth human body attribute vector.

[0095] Preferably, similarity calculation is performed between the third feature vector of the face and the data in the vector library, and the search results are sorted according to the calculation results of the similarity, further comprising:

[0096] The cosine similarity calculation formula of the third face feature vector and the fourth face feature vector is as follows:

[0097] cosine_similarity(A,B)=(A·B) / (||A||*||B||),

[0098] Among them, AB represents the dot product of vector A and vector B, ||A|| and ||B|| represent the L2 norm of vector A and vector B respectively.

[0099] The fourth facial feature vector is processed by removing duplicates using a human body deduplication model;

[0100] The second human feature is subjected to cosine similarity calculation with the first human feature in the vector library, sorted according to the calculation results, and the matching result is returned;

[0101] Associating the corresponding second facial feature vector in the query vector library according to the matching result, and performing cosine similarity calculation with the fifth facial feature vector selected by the user, and further sorting the results to obtain the query result;

[0102] The third human coordinates in the query result are associated with the second human coordinates stored in the search engine library by intersection and union calculation, and are compared with the picture in combination with the structured attribute features in the search engine library.

[0103] Preferably, the fourth face feature vector is processed by removing duplicates using a human body deduplication model; including:

[0104] S31: The face deduplication model calculates the similarity between feature vectors in the vector library based on the fourth face feature vector. Commonly used similarity measurement methods include cosine similarity, Euclidean distance, and Hamming distance.

[0105] S32: According to the result of similarity calculation, a similarity threshold is set. If the similarity of two objects exceeds the threshold, they are considered to be duplicates.

[0106] S33: For object pairs with a similarity exceeding a threshold, one of the objects is identified as a duplicate object, and is processed accordingly, such as being deleted from a database or undergoing other processing.

[0107] Different strategies can be used for optimization according to the actual application scenario. A grid filter can be used to project the point cloud into a grid, and then retain a point in each grid as a representative, or a hash table data structure can be used to quickly find and delete duplicate points.

[0108] See also Figure 5 As shown, according to the search information, that is, the user's captured picture is compared with the face library in the vector library, the picture category is determined according to the comparison result, and the search is performed in the vector library and the search engine according to the judgment result of the picture category.

[0109] When the judgment result of the image category is non-stranger, the standard face library in the face library is searched and the similarity is calculated. If the calculation result does not match the face data, the human body vector data is directly retrieved from the vector library. If the face data is matched, clustering and deduplication are performed to obtain human body attributes. According to the human body attributes, the human body vector data is retrieved from the vector library. After deduplication processing is performed on the obtained human body vector data, the confidence of the image data carrying the time period is adjusted to 0.99 and sent to the Web end;

[0110] If the judgment result of the image category is a stranger with a face, the face library of the vector library (human body film library) is searched and the similarity is calculated. If the calculation result does not match the face data, the human body vector data is directly retrieved from the vector library. If the face data is matched, clustering and deduplication are performed to obtain human body attributes. According to the human body attributes, the human body vector data is retrieved from the vector library. After deduplication processing is performed on the obtained human body vector data, the confidence of the image data carrying the time period is adjusted to 0.99 and sent to the Web end;

[0111] If the judgment result of the image category is a stranger without a face, the human body vector data is retrieved from the vector library. After deduplication processing is performed on the obtained human body vector data, the confidence of the image data carrying the time period is adjusted to 0.99 and sent to the Web end.

[0112] Preferably, performing a search based on the second attribute feature selected by the user in combination with the fifth human body attribute vector to obtain a search result further includes:

[0113] The second attribute feature is fused with the second human feature in the search engine library to obtain a fused feature;

[0114] The intersection sorting method is used to match and sort the fused features to obtain the retrieval results.

[0115] Preferably, step S7 further comprises:

[0116] The search results are displayed in the form of thumbnails and sorted by dots, and the video and picture information is displayed. Users can review the video by clicking on the marked dotted area and play the clips before and after the corresponding time point. The video and picture information includes information such as the video location, video time, body clothing style, body clothing color, and gender.

[0117] Example 2

[0118] In some embodiments of the present application, a computer device is also provided, including a memory and a processor, wherein the memory stores computer-readable instructions, and when the computer-readable instructions are executed by the processor, the processor executes the steps of the retrieval method for fusing attribute features and vector embedding as described in any one of Example 1.

[0119] The present invention also provides a storage medium storing computer-readable instructions, which, when executed by one or more processors, enables the one or more processors to execute the steps of the retrieval method for fusing attribute features and vector embedding as described in any one of Example 1.

[0120] It can be understood that, for the aforementioned cross-modal retrieval method integrating human attribute features and vector embedding, if they are all implemented in the form of software function modules 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 is essentially or part of the contribution to the prior art or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions for a computer device (which can be a personal computer server, or a network device, etc.) to perform all or part of the steps of the methods of each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk and other media that can store program code.

[0121] Computer readable storage media may include data signals propagated in baseband or as part of a carrier wave, wherein readable program codes are carried. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. The readable storage medium may also be any readable medium other than a readable storage medium, which may send, propagate, or transmit a program for use by or in conjunction with an instruction execution system, device, or device. The program codes contained on the readable storage medium may be transmitted using any suitable medium, including but not limited to wireless, wired, optical cable, RF, etc., or any suitable combination thereof.

[0122] The above is only a preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technicians in this technical field, some improvements and modifications without departing from the principle of the present invention should also be regarded as the protection scope of the present invention.

Claims

1. A retrieval method integrating attribute features and vector embedding, characterized in that: The following steps are involved: S1: monitor a video stream, periodically extract pictures of the video stream, encode the pictures to obtain encoded pictures, and cache the encoded pictures in a message queue; S2: The AI ​​cropping model returns the coded image ID and coordinate point position. If the coordinate point position has a value, the coded image is saved to the disk; S3: The deep learning model extracts the information on the coded image from the message queue to obtain a first facial feature vector, a first human feature vector and a first text vector, and normalizes the first facial feature vector to obtain a second facial feature vector. The logistic regression model obtains structured attribute features according to the coded image. S4: the deep learning model detects the first coordinate point position among the coordinate point positions, and associates the second face feature vector, the first body feature vector and the first text vector according to the picture ID and stores them in a vector library; S5: the logistic regression model detects the second coordinate point position of the coordinate point position, associates the structured attribute feature, the second face feature vector and the first body feature vector according to the picture ID to implement storing the structured attribute feature in the search engine; S6: The deep learning model extracts a search vector according to the search information, performs similarity calculation between the search vector and the data in the vector library, and sorts the search results according to the calculation results of the similarity to obtain the search results; S7: Mark the search results, merge the continuous clicked areas, and the user can review the video according to the marked clicked areas to play the previous and next clips of the corresponding time point.

2. The retrieval method of integrating attribute features and vector embedding according to claim 1 is characterized in that: In step S3, the structured attribute features include: The structured attribute features include features such as clothing style, color, and gender.

3. The retrieval method of integrating attribute features and vector embedding according to claim 2 is characterized in that: In step S3, normalizing the first facial feature vector to obtain a second facial feature vector further includes: The calculation process of the normalization process is as follows: v=[v1,v2,…,v i ], Where v is the eigenvector, v i represents the i-th element in the feature vector; Calculate the L2 norm based on the feature vector v, where v a represents the ath element in the feature vector; Each element in the feature vector v is multiplied by a normalization factor to obtain a normalized second face feature vector, wherein the calculation formula of the normalization factor normalizationFactor and the second face feature vector Z is as follows: normalizationFactor=1 / ||v||2, The calculation formula of the second face feature vector is as follows: normalizationVector=[v1 / ||v||2,v2 / ||v||2,…,v n / ||v||2]。 4. The retrieval method of integrating attribute features and vector embedding according to claim 3 is characterized in that: In step S4, the deep learning model extracts a search vector according to the search information, further comprising: The deep learning model extracts search vectors through image search or cross-modal text search as follows: The image retrieval is that the deep learning model automatically extracts the first human feature vector and the third face feature vector according to the video crop selected by the user or the uploaded image file; The cross-modal text retrieval is that the deep learning model extracts a second text vector based on the text vector features input by the user.

5. The retrieval method of integrating attribute features and vector embedding according to claim 4 is characterized in that: In step S6, the search vector is similarly calculated with the data in the vector library, and the search results are sorted according to the similarity calculation results to obtain search results, further comprising: When the search vector is a third facial feature vector, the third facial feature vector is similarly calculated with the data in the vector library, and matched with the fourth facial feature vector and the second human body feature vector in the vector library, and similarity calculation is performed. The system performs an intersection-and-union (IOU) calculation association based on the human body coordinates in the query result and the human body coordinates stored in the search engine library; When the search vector is the second text vector, the second text vector is similarly calculated and sorted with the first text vector in the vector library, and a search result is obtained by searching according to the second attribute feature selected by the user and in combination with the fifth human body attribute vector.

6. The retrieval method of integrating attribute features and vector embedding according to claim 5 is characterized in that: The third feature vector of the face is similarly calculated with the data in the vector library, and the search results are sorted according to the similarity calculation results to obtain the search results, further comprising: The cosine similarity calculation formula of the third face feature vector and the fourth face feature vector is as follows: cosine_similarity(A,B)=(A·B) / (||A||*||B||), Among them, A·B represents the dot product of vector A and vector B, ||A|| and ||B|| represent the L2 norm of vector A and vector B respectively, The fourth face feature vector is subjected to deduplication processing using a human body deduplication model; The second human body feature is subjected to cosine similarity calculation with the first human body feature in the vector library, sorted according to the calculation result, and a matching result is returned; Querying the corresponding second facial feature vector in the vector library according to the matching result, performing cosine similarity calculation with the fifth facial feature vector selected by the user, and further sorting the results to obtain the query result; The third human body coordinates in the query result are associated with the second human body coordinates stored in the search engine library by intersection and union calculation, and are compared with the picture in combination with the structured attribute features in the search engine library.

7. The retrieval method of integrating attribute features and vector embedding according to claim 6 is characterized in that: The method further comprises: performing a search based on the second attribute feature selected by the user and combining the fifth human body attribute vector to obtain a search result. The second attribute feature is fused with the second human feature in the search engine library to obtain a fused feature; The fusion features are matched and sorted using an intersection sorting method to obtain a retrieval result.

8. The retrieval method of integrating attribute features and vector embedding according to claim 7 is characterized in that: The step S7 further comprises: The search results are displayed in the form of thumbnails and sorted by dots, and video image information is displayed. The user can review the video by clicking on the marked dotted area and play the previous and next clips of the corresponding time point. The video image information includes information such as video location, video time, human clothing style, human clothing color, and gender.

9. A computer device, characterized in that: It includes a memory and a processor, wherein the memory stores computer-readable instructions, and when the computer-readable instructions are executed by the processor, the processor executes the steps of the retrieval method for fusing attribute features and vector embedding as described in any one of claims 1 to 8.

10. A storage medium storing computer-readable instructions, characterized in that: When the computer-readable instructions are executed by one or more processors, the one or more processors execute the steps of the retrieval method for fusing attribute features with vector embedding as described in any one of claims 1 to 8.