A method and system for retrieving human posture database images

Abstract

This invention relates to a method and system for retrieving human pose database images. The method includes: segmenting a human image into three types of human regions: human regions, non-human regions, and blurred regions; extracting features and locating key points; classifying key points into multiple levels based on key point confidence; retrieving a human pose database based on key points to determine a pose sample pool; and determining a target pose sample from the pose sample pool. This invention constructs a key point representation by utilizing the approximate location distribution of key points, enabling bidirectional squeeze retrieval, reducing retrieval interference from complex image backgrounds, and improving pose retrieval efficiency.

Description

[Technical Field]

[0001] This invention belongs to the field of posture retrieval technology based on artificial intelligence, and particularly relates to a method and system for retrieving human posture database images. [Background Technology]

[0002] Pose recognition technology is a computer vision-based method for analyzing human posture. It analyzes human posture, extracts key points and skeletal information, and models and recognizes the posture. With the development of deep learning technology, pose recognition technology has been widely applied and researched. Compared with other biometric recognition technologies, pose recognition has advantages such as being non-contact, non-invasive, easy to perceive, and difficult for targets to hide or disguise. Furthermore, pose analysis can easily distinguish different behavioral patterns of an individual, such as whether they are walking, running, or carrying heavy objects. Pose recognition is an important research direction in the field of computer vision, aiming to detect and recognize human postures in images or videos. Human posture information can be used to further perform tasks such as action recognition, human-computer interaction, and abnormal behavior detection. In practical applications, pose recognition technology is widely used in intelligent monitoring, human-computer interaction, virtual reality, and motion analysis. In the field of motion analysis, pose recognition can be used to monitor and evaluate athletes' movements in real time, improving the effectiveness of sports training. In the field of virtual reality, pose recognition technology allows users to perform natural body movements in a virtual environment, improving the immersion and realism of virtual reality. Posture recognition is particularly suitable for access control systems, security monitoring, human-machine interaction, and medical diagnosis. In the security field, posture recognition technology has wide applications and economic value, effectively monitoring and preventing criminal activity.

[0003] However, on the one hand, with the diversification of application scenarios, the complexity of the human body's environment greatly increases the complexity of the image background, especially in the field of security monitoring, where factors such as human occlusion further increase the complexity of the background in the image, significantly reducing the success rate of effective retrieval. On the other hand, there is a close relationship between artificial intelligence models and pose recognition. Artificial intelligence models, especially deep neural network models, are widely used in various steps of pose recognition. Artificial intelligence models for specific scenarios need to be trained specifically for specific vertical applications, and dedicated artificial intelligence models need to be set up to integrate the analysis results of artificial intelligence models used in various processing steps such as image analysis and image segmentation. The superposition of artificial intelligence models leads to an exponential increase in complexity; moreover, the accuracy of training is affected by various environmental factors, which will greatly increase the computational complexity. Based on the above problems, this invention constructs a key point representation method by constructing the approximate location distribution of key points, performs different optimization indicators, and adopts a two-way squeeze retrieval with a combination of precision and leniency, reducing the retrieval interference caused by complex image backgrounds and improving pose retrieval efficiency. [Summary of the Invention]

[0004] To address the aforementioned problems in the prior art, this invention proposes a method and system for retrieving human posture database images, the method comprising:

[0005] Step S1: Segment the image into three types: human body region, non-human body region, and blurred region; where: blurred region is the region between human body region and non-human body region;

[0006] Step S2: Extract human features and locate key points;

[0007] Step S3: Divide keypoints into multiple levels based on keypoint confidence;

[0008] Step S4: Search the human pose database based on key points to determine the pose sample pool; determine the target pose sample from the pose sample pool using a two-way squeeze method based on the positional relationship between key points;

[0009] Step S4 specifically includes the following steps:

[0010] Step S41: Using the specific location as the center point, divide the segmentation result containing the human body region and the blurred region into N sub-regions; determine the key points falling into each sub-region, and construct the current N-gram identifier (cls). n ), cls n =(p nm ), where: p nm For keypoint type or keypoint location in image / segmentation result; element cls n It is M nMeta set, containing M located in the nth subregion n One key point; m is cls n The element numbers in the array; m = 1 to M n M n It is cls n The number of elements;

[0011] Step S42: Based on the current N-gram identifier, retrieve the human pose database using the key point type and its distribution sub-region to obtain the pose sample pool; specifically: put the pose samples in the human pose database whose number of elements of the N-gram identifier is equal to the number of elements in the current N-gram identifier, and whose key point types are the same in each element of the identifier, into the pose sample pool.

[0012] Step S43: Determine the target pose sample from the pose sample pool; specifically: select the highest-level keypoints to form the initial value of the left keypoint set; set the initial value of the right keypoint set to be equal to the keypoint set composed of all keypoints; search the pose sample pool from the left keypoint set and the right keypoint set respectively, and perform bidirectional squeezing until the search result meets the convergence condition and the target pose sample is obtained; use the pose of the target pose sample as the search result.

[0013] Step S43 specifically includes the following steps:

[0014] Step S431: Determine the sub-region into which each keypoint in the left keypoint set falls. The keypoints falling in sub-region n constitute the left keypoint set cls of the nth keypoint set. nl ; Construct its identifier (cls) nl );

[0015] Step S432: Based on the set of left-side key points cls for each nl-th key point nl The positional relationship between each element E1 and any two other left keypoint sets containing elements E2 and E3 is used to select pose samples from the pose sample pool that meet the conditions, and these samples are placed into the left candidate pose sample set; the condition is that there are elements in the pose sample that have the same positional relationship as E1, E2, and E3.

[0016] Step S433: Based on all keypoint types in the right keypoint set and their positions in the image, retrieve the pose sample pool. If there is a pose sample that matches all keypoint types and their positions, add it to the right candidate pose sample set.

[0017] Step S434: If there is an intersection between the pose samples in the left candidate pose sample set and the right candidate pose sample set, then the pose samples in the intersection are taken as the target pose samples.

[0018] Step S435: Update the left keypoint set and the right keypoint set; if neither the left keypoint set nor the right keypoint set can be updated, or the left keypoint set and the right keypoint set are the same, proceed to the next step; otherwise, return to step S431.

[0019] Step S44: Use the pose of the target pose sample as the retrieval result.

[0020] Furthermore, an artificial intelligence model is used to perform three-class classification-based segmentation of the images.

[0021] Furthermore, a semantic segmentation-based deep neural network is used to recover the category of each pixel from the abstract features of the convolutional layers.

[0022] Furthermore, the human body images are derived from video surveillance, professional motion capture equipment, or still images.

[0023] Furthermore, the human body image is preprocessed before performing step S1.

[0024] Furthermore, morphological operations are performed on the segmentation results.

[0025] Furthermore, the morphological operations include removing small objects and / or filling holes.

[0026] A human posture database image retrieval server includes a processor coupled to a memory, the memory storing program instructions, and the human posture database image retrieval method is implemented when the program instructions stored in the memory are executed by the processor.

[0027] A human posture database image retrieval system is provided, the system being used to implement the human posture database image retrieval method.

[0028] A computer-readable storage medium includes a program that, when run on a computer, causes the computer to perform the human pose database image retrieval method.

[0029] The beneficial effects of this invention include:

[0030] (1) By constructing a key point representation method through the approximate location distribution of key points, different optimization indicators are applied, and a two-way squeeze retrieval with precise and relaxed matching is performed, which reduces the retrieval interference caused by complex image backgrounds and improves the efficiency, accuracy and robustness of pose retrieval.

[0031] (2) The influence of complex environmental factors is used in the process of determining the confidence of key points, providing a quantitative reliability index for the location of key points. Based on the classification of the confidence of key points and the spatial relationship between key points, a two-way squeeze search strategy is used to start from the set of key points with high confidence and gradually approach the set of key points with low confidence until the target pose sample is found. This can improve search efficiency and accurately retrieve the sample that best matches the query pose from the pose sample pool.

[0032] Furthermore, by considering the interior angle between key points, this method can robustly match pose samples and obtain accurate retrieval results even when there are certain errors in the key point positions.

[0033] (3) The construction of N-gram identifiers provides a flexible way to organize and retrieve key points, which facilitates the construction of a sample pool containing multiple poses, providing rich sample resources for subsequent accurate retrieval. The value of N can also be adjusted according to different application requirements to obtain better retrieval performance, thereby reducing manual intervention and improving the automation level of pose retrieval. [Attached Image Description]

[0034] The accompanying drawings, which are provided to further illustrate the invention and form part of this application, are not intended to unduly limit the invention. In the drawings:

[0035] Figure 1 This is a schematic diagram of the human posture database image retrieval method provided by the present invention.

Detailed Implementation Methods

[0036] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. The illustrative embodiments and descriptions are merely for explaining the present invention and are not intended to limit the scope of the invention.

[0037] This invention proposes a method and system for retrieving human posture database images, as shown in the appendix. Figure 1 As shown, the method includes the following steps:

[0038] Step S1: Segment the image into three types: human body region, non-human body region, and blurred region; where: blurred region is the human body image region between the human body region and the non-human body region;

[0039] Preferred method: Use an artificial intelligence model to perform three-class classification segmentation on the image. The output of the artificial intelligence model is the probability that each pixel belongs to the human body. Pixels with a probability greater than the high confidence threshold are human body pixels, and pixels with a probability less than the low confidence threshold are non-human body pixels. The human body segmentation result is obtained when the probability is within the confidence threshold.

[0040] Preferred method: Employ a deep neural network based on semantic segmentation to recover the category of each pixel from the abstract features of the convolutional layers;

[0041] Preferred approach: Post-processing the segmentation results, such as morphological operations, removal of small objects, and filling of holes, to improve the segmentation results;

[0042] Preferably, the human body image is sourced from video surveillance, professional motion capture equipment, or still images; alternatively, the image is precisely labeled to distinguish between human body areas, non-human body areas, and blurred areas. This typically requires manual operation or the use of semi-automated tools to improve efficiency.

[0043] Preferred method: Preprocessing human body images, such as resizing, normalizing brightness and contrast, to reduce the influence of factors such as lighting and scale;

[0044] Step S2: Extract human features and locate key points;

[0045] The human body feature extraction specifically involves extracting human body features from an image; these human body features include edge features and texture features.

[0046] Preferred method: Use commonly used deep learning network structures for human feature extraction, such as convolutional neural networks and recurrent neural networks;

[0047] The key point localization process specifically involves: using a key point determination model to locate key points based on human body features; for example: using a key point determination model based on linear regression and support vector machine algorithms to predict the positions of various key points on the human body based on extracted human body features; or using a key point determination model based on traditional image analysis methods to locate key points.

[0048] Preferably, the key points include multiple types located in various parts of the human body; when there are multiple key points of the same type, the key point with the highest confidence is selected and then classified into levels.

[0049] Replaceable: When there are multiple keypoints of the same type, select the keypoints with the highest confidence and then classify them into levels.

[0050] Preferred locations include: the center of the head, the nose or eyes, the position where the head and body connect at the back of the neck, the top of the two shoulders, the elbows of the two arms, the position of the two wrists, the hands, the sides of the hips, the position of the two knees, the position of the ankles, the position of the bottom or heels of the feet, the position of the two eyes, the top or center of the ears, the rotation point of the shoulders, the rotation point of the hips, the rotation point of the elbows, the rotation point of the wrists, the rotation point of the knees, and / or the rotation point of the ankles.

[0051] Step S3: Determine the confidence level of keypoints and classify keypoints into multiple levels based on the confidence level; specifically: classify keypoints into multiple levels based on the confidence level range into which the model output falls for each keypoint; the higher the confidence level, the higher the level.

[0052] Preferred method: Divide the complete confidence interval so that the confidence of key points falls evenly into the divided confidence interval;

[0053] The alternative is that the confidence interval is divided into three levels: high, medium, and low. Based on the confidence score of each keypoint, it is assigned to the corresponding level: keypoints above the high confidence threshold are classified as high level; keypoints below the medium confidence threshold but above the low confidence threshold are classified as medium level; and keypoints below the low confidence threshold are classified as low level.

[0054] Alternative: The confidence level of the key point is determined based on the type and area of ​​the human features involved in the key point localization process. The key point is divided into multiple levels according to the confidence level interval into which the confidence level falls. Specifically, the confidence level fa of the key point is determined using the following formula (1). Where: α is the confidence coefficient, and cb, ncb, and uncb are the areas of the human region, fuzzy region, and non-human region involved in the key point localization process, respectively.

[0055] fa=(∑cb×1+ncb×α+uncb×1) / (cb+ncb+uncb) (1)

[0056] Preferred value: α = 0.98; By adjusting the confidence coefficient α, the model can adapt to different application scenarios and human segmentation accuracy, thus improving the model's adaptability;

[0057] Alternative: Set α to the accuracy of human segmentation;

[0058] Preferred method: When the key points are in three-dimensional coordinates, projection is performed to convert the three-dimensional key points into two-dimensional key points;

[0059] Step S4: Search the human pose database based on key points to determine the pose sample pool; determine the target pose sample from the pose sample pool using a two-way squeeze method based on the positional relationship between key points;

[0060] Step S4 specifically includes the following steps:

[0061] Step S41: Using the specific location as the center point, divide the segmentation result containing the human body region and the blurred region into N sub-regions; determine the key points falling into each sub-region, and construct the current N-gram identifier (cls). n ), clsn =(p nm ), where: p nm This refers to the keypoint type or the location of the keypoint in the human body image; element cls n It is M n Meta set, containing M located in the nth subregion n One key point; m is cls n The element numbers in the array; m = 1 to M n M n It is cls n The number of elements; that is, the number of key points NM in each sub-region is the same or different; for example: 4-tuple identifier ((p1, p5)(p3)(p4)

[0062] (p2)), located in the first sub-region are left eye p1 and right eye center point p5; ((p1, p5)(p3)(p4)(p2));

[0063] Preferably, the specific location is the center location;

[0064] Preferably, N is equal to 4 to 16; rays are uniformly constructed from the center position to divide the human body region into N sub-regions;

[0065] Preferably, the human body region division method is a preset division method, for example: when N=4, a two-dimensional coordinate axis is constructed based on the center position to divide the human body region into 4 sub-regions;

[0066] Step S42: Based on the current N-gram identifier, retrieve the human pose database using key point type to obtain a pose sample pool; specifically: put pose samples in the human pose database whose number of elements in the N-gram identifier is equal to the number of elements in the current N-gram identifier, and whose key point types are the same in each element of the identifier, into the pose sample pool.

[0067] Preferably: the N-ary identifier is set as an index;

[0068] Alternative: Use the entire human pose database as a pose sample pool;

[0069] Step S43: Determine the target pose sample from the pose sample pool; specifically: select the highest-level keypoints to form the initial value of the left keypoint set; set the initial value of the right keypoint set to be equal to the keypoint set composed of all keypoints; search the pose sample pool from the left keypoint set and the right keypoint set respectively, and perform bidirectional squeezing until the search result meets the convergence condition and the target pose sample is obtained; use the pose of the target pose sample as the search result.

[0070] Step S43 specifically includes the following steps:

[0071] Step S431: Determine the sub-region into which each keypoint in the left keypoint set falls. The keypoints falling in sub-region n constitute the left keypoint set cls of the nth keypoint set. nl ; Construct its identifier (cls) nl );

[0072] Step S432: Based on the set of left-side key points cls for each nl-th key point nl The positional relationship between each element and any two other elements in the left keypoint set is used to select pose samples that meet the conditions from the pose sample pool and put them into the left candidate pose sample set.

[0073] Preferably, the condition is that there exists an element in the pose sample that has the same positional relationship as E1, E2, and E3;

[0074] Preferably, the condition is that the pose sample contains an element that has the same key point type as E1 and the same positional relationship as E1, E2, and E3;

[0075] Step S4321: Select an unprocessed pose sample from the pose sample pool as the current pose sample; and set the left keypoint set cls. nl The initial value of number nl is 1;

[0076] Preferred: If there are no unprocessed pose samples, then end step S432;

[0077] Step S4322: From the set of key points cls on the left side of the nth rank. nl Select an unprocessed keypoint p from the middle nlc If a keypoint p can be selected from any two other sets of left-side keypoints, nl1 and nl2... nl1 p nl2 ; so that there exists a pose sample (cls) in the pose sample pool that satisfies condition (2). ns If condition (3) is satisfied, then the set of left-side key points cls of the nth lth node is added and updated. nl The proportion or frequency of key points satisfying the conditions; where: θ(a, b, c) is the interior angle between the line formed by key points a and b and the line formed by key points b and c.

[0078]

[0079] Step S4323: Determine if there are still unprocessed key points. If yes, return to step S4322; otherwise, if the proportion or number of times the condition is met is greater than the preset proportion or preset number of times, proceed to the next step; otherwise, continue to execute step S4321 to judge other posture samples in the posture sample pool.

[0080] Step S4324: If nl is greater than N, proceed to the next step; otherwise, set nl = nl + 1 and return to step S4322.

[0081] Step S4325: Add the current pose sample to the left-hand candidate pose sample set;

[0082] Alternatively, step S432 specifically involves: if there exists a pose sample (cls) in the pose sample pool that satisfies conditions (4) and (5) nb ), so that condition (6) is satisfied; then the posture sample is placed in the left candidate posture sample set; where: θ(a, b, c) is the inner angle between the line formed by key points a and b and the line formed by key points b and c.

[0083] in:

[0084] in:

[0085]

[0086] Alternatively: Step S432 specifically involves: if there exists a pose sample and its identifier (cls) in the pose sample pool that satisfies conditions (4) and (5). nb If condition (7) is satisfied, then the posture sample is placed in the left-hand candidate posture sample set.

[0087] Where: θl1, θl2∈θ(pnb1, pnb2, pnb3)orθ(pnb1, pnb3, pnb2)orθ(pnb2, pnb3, pnb1); θs1, θs2∈(pnl1, pnl2, pnl3)orθ(pnl1, pnl3, pnl2)orθ(pnl2, pnl3, pnl1) (7);

[0088] Preferred: In formulas (3) and (7), ≈ means that the angles are approximately equal, and a certain deviation threshold can be used to limit the conventional stem;

[0089] Step S433: Based on all keypoint types and their positions in the image in the right keypoint set, retrieve the pose sample pool. If there is a pose sample that matches all keypoint types and their positions, put it into the right candidate pose sample set; that is, find pose samples with the same keypoint types and their position distribution.

[0090] Preferred method: Obtain a pose sample from the pose sample pool, fix the position of a key point in the right key point set and the corresponding key point in the pose sample of the same type in the image; scale and rotate the other key points as a whole to determine whether the positions of the other key points in the image overlap or approximately overlap, and if so, determine that the pose sample is a match.

[0091] Step S434: If there is an intersection between the pose samples in the left candidate pose sample set and the right candidate pose sample set, then the pose samples in the intersection are taken as the target pose samples.

[0092] Step S435: Update the left keypoint set and the right keypoint set; if neither the left keypoint set nor the right keypoint set can be updated, or the left keypoint set and the right keypoint set are the same, proceed to the next step; otherwise, return to step S431.

[0093] The update of the left key point set specifically involves: selecting the highest-level key point from the key points that do not belong to the current left key point set and adding it to the left key point set;

[0094] The update of the right-side key point set specifically involves deleting the lowest-level key points from the right-side key point set. Therefore, the search criteria based on the right-side key point set are continuously relaxed, while the search criteria based on the left-side key point set are continuously tightened. Furthermore, the two optimizations are performed from different dimensions.

[0095] Replaceable: Delete the lowest-ranking keypoint from the keypoint set on the right;

[0096] Step S44: Use the pose of the target pose sample as the search result; if the target pose sample does not exist, the search fails.

[0097] Based on the same inventive concept, the present invention also provides a human posture database image retrieval system, the system being used to implement the above-mentioned human posture database image retrieval method;

[0098] Based on the same inventive concept, the present invention also provides a human posture database image retrieval server, the server being used to implement the above-mentioned human posture database image retrieval method;

[0099] Based on the same inventive concept, the present invention also provides a human posture database image retrieval device.

[0100] The device is used to implement the above-mentioned human posture database image retrieval method;

[0101] Based on the same inventive concept, the present invention also provides a human posture database image retrieval platform.

[0102] The platform is used to implement the above-mentioned human posture database image retrieval method;

[0103] A computer program (also referred to as a program, software, software application, script, or code) can be written in any form of programming language, including assembly or interpreted languages, declarative or procedural languages, and can be deployed in any form, including as a standalone program or as a module, component, subroutine, object, or other unit suitable for use in a computing environment. A computer program may, but does not necessarily, correspond to a file in a file system. A program can be stored as part of a file that holds other programs or data (e.g., one or more scripts stored in a markup language document), in a single file dedicated to said program, or in multiple co-located files (e.g., a file storing one or more modules, subroutines, or code portions). A computer program can be deployed to execute on a single computer or on multiple computers located at a single site or distributed across multiple sites and interconnected by a communications network.

[0104] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0105] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0106] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1The function specified in one or more boxes.

[0107] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0108] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for retrieving images from a human posture database, characterized in that, The method includes: Step S1: Segment the image into three types: human body region, non-human body region, and blurred region; where: blurred region is the region between human body region and non-human body region; Step S2: Extract human features and locate key points; Step S3: Divide keypoints into multiple levels based on keypoint confidence; Step S4: Search the human pose database based on key points to determine the pose sample pool; determine the target pose sample from the pose sample pool using a two-way squeeze method based on the positional relationship between key points; Step S4 specifically includes the following steps: Step S41: Using a specific location as the center point, divide the segmentation result containing the human body region and the blurred region into N sub-regions; determine the key points falling into each sub-region, and construct the current N-gram identifier cls. n cls n =(p nm ), where: p nm For keypoint type or keypoint location in the image; element cls n It is M n Meta set, containing M located in the nth subregion n One key point; m is cls n The element numbers in the array; m = 1 to M n M n It is cls n The number of elements; where: the specific position is the center position; Step S42: Based on the current N-gram identifier, retrieve the human pose database using the key point type and its distribution sub-region to obtain the pose sample pool; specifically: put the pose samples in the human pose database whose number of elements of the N-gram identifier is equal to the number of elements in the current N-gram identifier, and whose key point types are the same in each element of the identifier, into the pose sample pool. Step S43: Determine the target pose sample from the pose sample pool; specifically: select the highest-level keypoints to form the initial value of the left keypoint set; set the initial value of the right keypoint set to be equal to the keypoint set composed of all keypoints; search the pose sample pool from the left keypoint set and the right keypoint set respectively, and perform bidirectional squeezing until the search result meets the convergence condition and the target pose sample is obtained; use the pose of the target pose sample as the search result. Step S43 specifically includes the following steps: Step S431: Determine the sub-region into which each keypoint in the left keypoint set falls. The keypoints falling in sub-region n constitute the left keypoint set cls of the nth keypoint set. nl ; Construct its identifier (cls) nl ); Step S432: Based on the set of left-side key points cls for each nl-th key point nl The positional relationship between each element E1 and any two other left keypoint sets containing elements E2 and E3 is used to select pose samples from the pose sample pool that meet the conditions, and these samples are placed into the left candidate pose sample set; the condition is that there are elements in the pose sample that have the same positional relationship as E1, E2, and E3. Step S433: Based on all key point types and their positions in the image in the right key point set, retrieve the pose sample pool. If there is a pose sample that matches all key point types and their positions, put it into the right candidate pose sample set. Step S434: If there is an intersection between the pose samples in the left candidate pose sample set and the right candidate pose sample set, then the pose samples in the intersection are taken as the target pose samples. Step S435: Update the left keypoint set and the right keypoint set; if neither the left keypoint set nor the right keypoint set can be updated, or the left keypoint set and the right keypoint set are the same, proceed to the next step; otherwise, return to step S431. Step S44: Use the pose of the target pose sample as the retrieval result.

2. The human posture database image retrieval method according to claim 1, characterized in that, Use an artificial intelligence model to perform three-class classification-based image segmentation.

3. The human posture database image retrieval method according to claim 2, characterized in that, A semantic segmentation-based deep neural network is used to recover the category of each pixel from the abstract features of the convolutional layers.

4. The human posture database image retrieval method according to claim 3, characterized in that, Before performing step S1, the human body image is preprocessed.

5. The human posture database image retrieval method according to claim 4, characterized in that, Perform morphological operations on the segmentation results.

6. The human posture database image retrieval method according to claim 5, characterized in that, The morphological operations include removing small objects and / or filling holes.

7. A human posture database image retrieval server, characterized in that, The method includes a processor coupled to a memory, the memory storing program instructions, which, when executed by the processor, implement the human posture database image retrieval method according to any one of claims 1-4.

8. A human posture database image retrieval system, characterized in that, The system is used to implement the human posture database image retrieval method according to any one of claims 1-4.

9. A computer-readable storage medium, characterized in that, Includes a program that, when run on a computer, causes the computer to perform the human posture database image retrieval method according to any one of claims 1-4.

Images