Behavior Recognition Model Training, Behavior Recognition Method, Device, and Computer Storage Medium

Abstract

The present application provides a method, device and computer storage medium for training a behavior recognition model and behavior recognition, including obtaining respective normalized features, respective symmetric features, and respective centralized features corresponding to each of the bone nodes identified from training samples, obtaining a plurality of training feature groups of the training samples according to the respective normalized features, respective symmetric features, and respective centralized features corresponding to each of the bone nodes, training a plurality of recognition sub-models according to the plurality of training feature groups of the training samples and the behavior labels of the training samples, and obtaining a trained behavior recognition model based on the trained recognition sub-models. Accordingly, the present application can effectively improve the training effect of the behavior recognition model by expanding the input of training features of the training samples.

Description

Technical Field

[0001] The embodiments of the present application relate to the technical field of image recognition, and in particular, to a method, device, and computer storage medium for training a behavior recognition model and behavior recognition. Background Art

[0002] With the continuous improvement of the demand for the application of artificial intelligence technology in various fields, computer vision, as an important branch of artificial intelligence technology, has also been greatly developed and applied.

[0003] In recent years, video understanding has gradually become an important research direction in computer vision. Among them, human behavior recognition, as an important field in video understanding, has wide applications in scenarios such as human-computer interaction and intelligent video surveillance.

[0004] However, at present, the input of the behavior recognition method based on human body bone points only includes bone point coordinates, and the features are relatively sparse, resulting in less information available for relevant algorithm models to learn and utilize, and leading to unsatisfactory performance of the final algorithm. Summary of the Invention

[0005] In view of the above problems, the present application provides a method, device, and computer storage medium for training a behavior recognition model and behavior recognition, which can improve the model training effect and enhance the behavior recognition accuracy of the model.

[0006] The first aspect of the present application provides a method for training a behavior recognition model, including: obtaining, according to each bone node identified from a training sample, each normalized feature, each symmetric feature, and each centralized feature corresponding to each bone node; obtaining a plurality of training feature groups of the training sample according to each normalized feature, each symmetric feature, and each centralized feature corresponding to each bone node; training a plurality of recognition sub-models according to each training feature group of the training sample and the behavior label of the training sample; and obtaining a trained behavior recognition model based on the trained plurality of recognition sub-models.

[0007] The second aspect of the present application provides a behavior recognition method, including: identifying each bone node of a target object in an image to be recognized, and obtaining each normalized feature, each symmetric feature, and each centralized feature corresponding to each bone node; using a behavior recognition model to perform behavior prediction according to each normalized feature, each symmetric feature, and each centralized feature corresponding to each bone node, and obtaining a behavior prediction result of the target object; wherein, the behavior recognition model is trained by using the behavior recognition model training method described in the first aspect above.

[0008] In a third aspect of the present application, a computer storage medium is provided, in which instructions for executing the steps in the behavior recognition model training method described in the first aspect above, or instructions for executing the steps in the behavior recognition method described in the second aspect above are stored.

[0009] In a fourth aspect of the present application, a behavior recognition model training device is provided, including: a bone feature acquisition module, configured to acquire respective normalized features, respective symmetric features, and respective centralized features corresponding to each of the bone nodes according to each of the bone nodes identified from the training samples; a training feature generation module, configured to acquire a plurality of training feature groups of the training samples according to the respective normalized features, the respective symmetric features, and the respective centralized features corresponding to each of the bone nodes; and a training module, configured to train a plurality of recognition sub-models according to the plurality of training feature groups of the training samples and the behavior labels of the training samples, and obtain a trained behavior recognition model based on the trained plurality of recognition sub-models.

[0010] In a fifth aspect of the present application, a behavior recognition device is provided, including: an acquisition module, configured to identify each of the bone nodes of a target object in an image to be recognized, and acquire respective normalized features, respective symmetric features, and respective centralized features corresponding to each of the bone nodes; and an identification module, configured to perform behavior prediction by using a behavior recognition model according to the respective normalized features, the respective symmetric features, and the respective centralized features corresponding to each of the bone nodes, and obtain a behavior prediction result of the target object; wherein the behavior recognition model is trained by using the behavior recognition model training device described in the fourth aspect above.

[0011] In summary, the behavior recognition model training and behavior recognition solutions provided by the embodiments of the present application can effectively expand the training feature input of the training samples by generating the normalized features, symmetric features, and centralized features of each of the bone nodes in the training samples, thereby improving the training effect of the behavior recognition model and the accuracy of the model recognition result. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the drawings described below are only some embodiments recorded in the embodiments of the present application, and those of ordinary skill in the art can also obtain other drawings based on these drawings.

[0013] Figure 1 It is a schematic flowchart of the behavior recognition model training method according to an exemplary embodiment of the present application.

[0014] Figure 2 It is a schematic flowchart of the behavior recognition model training method according to another exemplary embodiment of the present application.

[0015] Figure 3 It is a schematic flowchart of a method for training a behavior recognition model according to another exemplary embodiment of the present application.

[0016] Figure 4 It is a schematic flowchart of a behavior recognition method according to an exemplary embodiment of the present application.

[0017] Figure 5 It is a schematic architecture diagram of a device for training a behavior recognition model according to an exemplary embodiment of the present application.

[0018] Figure 6 It is a schematic architecture diagram of a behavior recognition device according to an exemplary embodiment of the present application. Detailed implementation manners

[0019] In order to enable those skilled in the art to better understand the technical solutions in the embodiments of the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the embodiments of the present application, all other embodiments obtained by those of ordinary skill in the art shall fall within the protection scope of the embodiments of the present application.

[0020] With the increasing maturity of human skeleton point detection algorithms, behavior recognition based on human skeleton points is an important research direction in the field of human behavior recognition. The advantages of the behavior recognition method based on human skeleton points are as follows: the skeleton information represents human behaviors more naturally and compactly; the input is only the skeleton information of the human body, which can effectively avoid the interference of background information on human behavior recognition; compared with other behavior recognition algorithms, the behavior recognition method based on human skeleton points is more efficient and the model is more lightweight.

[0021] However, the input of the current behavior recognition method based on human skeleton points only contains skeleton point coordinates, and the features are relatively sparse, resulting in less information available for relevant algorithm models to learn and utilize, and also resulting in unsatisfactory performance of the final algorithm.

[0022] In view of this, the present application proposes a behavior recognition model training solution, which can improve various problems in the above-mentioned prior art. The following will describe the specific embodiments of the present application in detail with reference to the accompanying drawings.

[0023] Figure 1 Shows the processing flow of the behavior recognition model training method according to an exemplary embodiment of the present application. As shown in the figure, this embodiment mainly includes the following steps:

[0024] Step S102: Obtain the normalized features, symmetrized features, and centered features corresponding to each bone node identified from the training samples.

[0025] Optionally, each bone node may be a human bone node, including but not limited to: left shoulder node, right shoulder node, left elbow node, right elbow node, left wrist node, right wrist node, left hip node, right hip node, left knee node, right knee node, left ankle node, right ankle node, top of head node, neck node.

[0026] Optionally, each bone node may also be a bone node of various animals, and the present application does not limit this.

[0027] Step S104: Obtain multiple training feature groups of the training samples according to the normalized features, symmetrized features, and centered features corresponding to each bone node.

[0028] In this embodiment, any combination of at least one of the normalized features, symmetrized features, and centered features of the bone nodes can be performed to generate multiple training feature groups of the training samples.

[0029] For example, based on the normalized features of each bone node, a normalized feature group of the training sample can be generated; based on the symmetrized features of each bone node, a symmetrized feature group of the training sample can be generated; based on the centered features of each bone node, a centered feature group of the training sample can be generated.

[0030] For another example, any combination of at least two of the normalized features, symmetrized features, and centered features of each bone node (for example, the fusion of normalized features and symmetrized features, the fusion of normalized features and centered features, the fusion of centered features and symmetrized features, the fusion of normalized features, symmetrized features, and centered features) can also be used to generate one or more groups of fused feature groups of the training samples.

[0031] Step S106: Train multiple recognition sub-models according to the training feature groups of the training samples and the behavior labels of the training samples.

[0032] In this embodiment, multiple recognition sub-models can be constructed, and each recognition sub-model corresponds one-to-one with each training feature group. Using each training feature group as the input of each recognition sub-model and the behavior label of the training sample as the output of each recognition sub-model to train each recognition sub-model.

[0033] Step S108: Obtain a trained behavior recognition model based on the trained recognition sub-models.

[0034] In summary, the method for training the behavior recognition model according to this embodiment can effectively expand the training features by generating the normalized features, symmetric features, and centralized features corresponding to each bone node, thereby improving the problem of sparse input features in the behavior recognition method based on bone nodes in the prior art, fully mining the training features of existing training samples, and effectively improving the training effect of the model.

[0035] Figure 2 The processing flow of the method for training the behavior recognition model according to another exemplary embodiment of the present application is shown. This embodiment mainly shows the specific implementation of the above step S102. As shown in the figure, this embodiment mainly includes the following steps:

[0036] Step S202, identify the target object in the training sample, and obtain the target detection box of the target object and the bone node coordinates corresponding to each bone node of the target object.

[0037] Optionally, the target object may include a human object.

[0038] In this embodiment, the identified target object is a single human object, but it is not limited thereto, and it may also be a single animal object.

[0039] Optionally, the bone node coordinates of the bone node may include two-dimensional coordinates, and this bone node coordinate is the coordinate relative to the training sample (the whole picture).

[0040] Optionally, sequence annotation may be performed on each identified bone node.

[0041] For example, assuming that the identified target object is a single human object, the bone nodes of the human body can be sequentially labeled as: 0 - left shoulder node, 1 - right shoulder node, 2 - left elbow node, 3 - right elbow node, 4 - left wrist node, 5 - right wrist node, 6 - left hip node, 7 - right hip node, 8 - left knee node, 9 - right knee node, 10 - left ankle node, 11 - right ankle node, 12 - top of the head node, 13 - neck node.

[0042] Step S204, according to the target detection box, perform normalization processing on each bone node coordinate to obtain the normalized features corresponding to each bone node.

[0043] Optionally, according to the height and width of the target detection box, normalization processing may be performed on each bone node coordinate to obtain the normalized coordinates corresponding to each bone node.

[0044] In this embodiment, the normalized coordinate of the bone node can be expressed as:

[0045]

[0046] Among them, is the normalized coordinate of the bone node, (x i , y i ) is the coordinate of the bone node, W is the width of the target detection box, and H is the height of the target detection box.

[0047] Step S206: Obtain the symmetric features corresponding to each bone node according to the central axis determined by each bone node and the normalized features corresponding to each bone node.

[0048] Specifically, in many human behaviors, such as running and walking, the movements of each bone node have certain symmetry or antisymmetry. According to this action principle, this embodiment further expands the symmetry features of each bone node.

[0049] Optionally, in the state of normal human standing, the central axis can be determined according to the connection line formed by the vertex node and the neck node in the bone nodes, and based on the central axis, the symmetric first bone node and second bone node can be determined.

[0050] For example, the left shoulder node and the right shoulder node, the left elbow node and the right elbow node, the left hip node and the right hip node, the left knee node and the right knee node, etc. are all symmetric based on the central axis.

[0051] Optionally, the symmetric coordinates of the first bone node and the second bone node can be determined according to the normalized coordinates of the first bone node and the second bone node with a symmetric relationship.

[0052] In this embodiment, the symmetric coordinate of the first bone node can be expressed as: The symmetric coordinate of the second bone node can be expressed as:

[0053] Wherein, is the normalized coordinate of the first bone node, is the normalized coordinate of the second bone node.

[0054] In this embodiment, the symmetric coordinate of the bone node located on the central axis can be determined as (0, 0), that is, the symmetric point of the bone node located on the central axis is itself.

[0055] Step S208: Obtain the centered features corresponding to each bone node according to the center point determined by each bone node and the normalized features corresponding to each bone node.

[0056] Optionally, the center point may be determined according to the vertex nodes in the bone nodes, and the centralized coordinates corresponding to each bone node may be determined according to the normalized coordinates of the center point and the normalized coordinates corresponding to each bone node.

[0057] In this embodiment, the centralized coordinate of the bone node may be expressed as:

[0058]

[0059] Wherein, represents the centralized coordinate of the bone node, represents the normalized coordinate of the center point, represents the normalized coordinate of the bone node.

[0060] Figure 3 FIG. shows the processing flow of another exemplary embodiment of the present application. This embodiment is a specific implementation of the above steps S104 and S106. As shown in the figure, this embodiment mainly includes the following steps:

[0061] Step S302, for each of the bone nodes, arbitrarily combine at least one of the normalized feature, symmetrized feature, and centralized feature of the bone node to obtain multiple training features for each bone node.

[0062] In this embodiment, for each bone node, by arbitrarily combining the normalized coordinate, symmetrized coordinate, and centralized coordinate of the bone node, up to 6 training features of the bone node can be obtained.

[0063] Step S304, integrate the training features corresponding to each bone node with the same combination content to obtain multiple training feature groups of the training sample.

[0064] For example, a normalized feature group of the training sample may be generated based on the normalized coordinates of each bone node, a symmetrized feature group of the training sample may be generated based on the symmetrized coordinates of each bone node, and a centralized feature group of the training sample may be generated based on the centralized coordinates of each bone node.

[0065] For another example, a first fused feature group of the training sample may also be generated based on the first fused features of each bone node (for example, the first fused feature may be composed of the normalized coordinate and the symmetrized coordinate), a second fused feature group of the training sample may be generated based on the second fused features of each bone node (for example, the second fused feature may be composed of the normalized coordinate and the centralized coordinate), and so on.

[0066] In this embodiment, up to 6 training feature groups of the training sample can be obtained.

[0067] Step S306: Based on multiple training feature groups of the training samples, construct respective recognition sub-models corresponding to the training feature groups, and use each training feature group and the behavior label of the training sample to train each recognition sub-model.

[0068] In one embodiment, if the training feature group is a normalized feature group, a normalized feature sub-model can be constructed based on the normalized feature group, and the normalized feature group is used as the input and the behavior label is used as the output to train the normalized feature sub-model.

[0069] In one embodiment, if the training feature group is a symmetrized feature group, a symmetrized feature sub-model can be constructed based on the symmetrized feature group, and the symmetrized feature group is used as the input and the behavior label is used as the output to train the symmetrized feature sub-model.

[0070] In one embodiment, if the training feature group is a centralized feature group, a centralized feature sub-model can be constructed based on the centralized feature group, and the centralized feature group is used as the input and the behavior label is used as the output to train the centralized feature sub-model.

[0071] In one embodiment, if the training feature group is a fused feature group composed of at least two of the centralized features, symmetrized features, and centered features corresponding to each bone node (for example, the first fused feature group composed of normalized coordinates and symmetrized coordinates), a fused feature sub-model (for example, the first fused feature sub-model) can be constructed based on the fused feature group (for example, the first fused feature group), and the fused feature group (for example, the first fused feature group) is used as the input and the behavior label is used as the output to train the fused feature sub-model (for example, the first fused feature sub-model).

[0072] In this embodiment, each recognition sub-model (for example, the normalized feature sub-model, the symmetrized feature sub-model, the centralized feature sub-model, the fused feature sub-model) may include a spatio-temporal graph convolutional model or an improved model based on the spatio-temporal graph convolutional model.

[0073] Optionally, the improved model based on the spatio-temporal graph convolutional model includes, but is not limited to: the action-structure graph convolutional model, the adaptive graph convolutional model).

[0074] In summary, this embodiment uses training feature inputs of different dimensions to train different recognition sub-models, so as to improve the prediction accuracy of the behavior recognition model.

[0075] Figure 4 The processing flow of the behavior recognition method according to an exemplary embodiment of the present application is shown. As shown in the figure, this embodiment mainly includes the following steps:

[0076] Step S402: Identify each bone node of the target object in the image to be recognized, and obtain the corresponding normalized features, symmetric features, and centralized features of each bone node.

[0077] In this embodiment, the target object in the training sample can be recognized, the target detection box of the target object and the bone node coordinates corresponding to each bone node of the target object are obtained, and based on the target detection box, normalization processing is performed on each bone node coordinate to obtain the corresponding normalized features of each bone node. Then, based on the corresponding normalized features of each bone node, the corresponding symmetric features and centralized features of each bone node are calculated.

[0078] Regarding the specific acquisition methods of the above-mentioned normalized features, symmetric features, and centralized features, reference can be made to the description of the above Figure 2 illustrated embodiment, which will not be elaborated here.

[0079] Step S404: Use the behavior recognition model to perform behavior prediction based on the corresponding normalized features, symmetric features, and centralized features of each bone node, and obtain the behavior prediction result of the target object.

[0080] In this embodiment, the behavior recognition model is trained by using the behavior recognition model training method described in the above embodiments.

[0081] Optionally, each recognition sub-model in the behavior recognition model can be used to perform behavior prediction on the image to be recognized respectively, and obtain the corresponding behavior prediction labels of each recognition sub-model.

[0082] Specifically, the normalized feature sub-model can be used to perform behavior prediction based on the corresponding normalized features of each bone node to obtain the normalized behavior prediction label of the normalized feature sub-model; the symmetric feature sub-model can be used to perform behavior prediction based on the corresponding symmetric features of each bone node to obtain the symmetric behavior prediction label of the symmetric feature sub-model; the centralized feature sub-model can be used to perform behavior prediction based on the corresponding centralized features of each bone node to obtain the centralized behavior prediction label of the centralized feature sub-model; the fused feature sub-model can be used to perform behavior prediction based on the corresponding fused features of each bone node to obtain the fused behavior prediction label of the fused feature sub-model.

[0083] Optionally, a voting method can be used to fuse the corresponding behavior prediction labels of each recognition sub-model to obtain the behavior prediction result of the image to be recognized.

[0084] In summary, this embodiment uses the behavior recognition model trained by the above-mentioned behavior recognition model training methods to perform behavior prediction on the target object in the image to be recognized, which can improve the accuracy of the behavior prediction result.

[0085] Figure 5 The figure shows a schematic architecture diagram of a behavior recognition model training device according to an exemplary embodiment of the present application. As shown in the figure, the recognition model training device 500 of this embodiment mainly includes:

[0086] A bone feature acquisition module 502, configured to obtain respective normalized features, respective symmetric features, and respective centered features corresponding to each of the bone nodes according to each bone node identified from the training samples.

[0087] A training feature generation module 504, configured to obtain multiple training feature groups of the training samples according to the respective normalized features, the respective symmetric features, and the respective centered features corresponding to each of the bone nodes.

[0088] A training module 506, configured to train multiple recognition sub-models according to the multiple training feature groups of the training samples and the behavior labels of the training samples, and obtain a trained behavior recognition model based on the trained multiple recognition sub-models.

[0089] Optionally, the bone feature acquisition module 502 is further configured to: identify a target object in the training sample, and obtain a target detection box of the target object and respective bone node coordinates corresponding to each of the bone nodes of the target object; perform normalization processing on each of the bone node coordinates according to the target detection box to obtain respective normalized features corresponding to each of the bone nodes; obtain respective symmetric features corresponding to each of the bone nodes according to a central axis determined by each of the bone nodes and the respective normalized features corresponding to each of the bone nodes; and obtain respective centered features corresponding to each of the bone nodes according to a center point determined by each of the bone nodes and the respective normalized features corresponding to each of the bone nodes.

[0090] Optionally, the bone feature acquisition module 502 is further configured to: perform normalization processing on each of the bone node coordinates according to the height and width of the target detection box to obtain respective normalized coordinates of the bone nodes; the normalized coordinates of the bone nodes can be expressed as:

[0091] wherein, the is the normalized coordinate of the bone node, the (x i , y i ) is the bone node coordinate, the W is the width of the target detection box, and the H is the height of the target detection box.

[0092] Optionally, the bone feature acquisition module 502 is further configured to: determine the central axis according to the connection line formed by the vertex node and the neck node in the bone nodes; determine symmetric first and second bone nodes based on the central axis; determine the symmetric coordinates of the first and second bone nodes according to the respective normalized coordinates of the first and second bone nodes; the symmetric coordinates of the first bone node can be expressed as: The symmetric coordinates of the second bone node can be expressed as: Wherein, the is the normalized coordinate of the first bone node, and the is the normalized coordinate of the second bone node.

[0093] Wherein, the is the normalized coordinate of the first bone node, and the is the normalized coordinate of the second bone node.

[0094] Optionally, the bone feature acquisition module 502 is further configured to: determine the symmetric coordinates of the bone nodes located on the central axis as (0, 0).

[0095] Optionally, the bone feature acquisition module 502 is further configured to: determine the center point according to the vertex node in the bone nodes; determine the centralized coordinates corresponding to each of the bone nodes according to the normalized coordinate of the center point and the respective normalized coordinates corresponding to each of the bone nodes; the centralized coordinates of the bone node can be expressed as:

[0096]

[0097] Wherein, the represents the centralized coordinate of the bone node, the represents the normalized coordinate of the center point, and the represents the normalized coordinate of the bone node.

[0098] Optionally, the bone feature acquisition module 502 is further configured to: for each of the bone nodes, arbitrarily combine at least one of the normalized features, symmetric features, and centralized features of the bone nodes to obtain multiple training features for each of the bone nodes; integrate the respective training features corresponding to the bone nodes with the same combination content to obtain the multiple training feature groups of the training samples.

[0099] Optionally, the training module 506 is further configured to: if the training feature group is a normalized feature group composed of the normalized features corresponding to the respective bone nodes, construct a normalized feature sub-model based on the normalized feature group, and use the normalized feature group as the input and the behavior label as the output to train the normalized feature sub-model; if the training feature group is a symmetric feature group composed of the symmetric features corresponding to the respective bone nodes, construct a symmetric feature sub-model based on the symmetric feature group, and use the symmetric feature group as the input and the behavior label as the output to train the symmetric feature sub-model; if the training feature group is a centralized feature group composed of the centralized features corresponding to the respective bone nodes, construct a centralized feature sub-model based on the centralized feature group, and use the centralized feature group as the input and the behavior label as the output to train the centralized feature sub-model; if the training feature group is a fused feature group composed of at least two of the centralized features, symmetric features, and centralized features corresponding to the respective bone nodes, construct a fused feature sub-model based on the fused feature group, and use the fused feature group as the input and the behavior label as the output to train the fused feature sub-model.

[0100] Figure 6 The architecture diagram of the behavior recognition device according to an exemplary embodiment of the present application is shown. As shown in the figure, the behavior recognition device 600 of this embodiment includes:

[0101] An acquisition module 602, configured to identify each bone node of the target object in the image to be recognized, and acquire the respective normalized features, symmetric features, and centralized features corresponding to each bone node.

[0102] An identification module 604, configured to perform behavior prediction according to the respective normalized features, symmetric features, and centralized features corresponding to each bone node by using a behavior recognition model, and obtain a behavior prediction result of the target object; wherein, the behavior recognition model is trained by using the behavior recognition model training device described in the above embodiment.

[0103] Optionally, the identification module 604 is further configured to: perform behavior prediction on the image to be recognized respectively by using each identification sub-model in the behavior recognition model, and obtain the respective behavior prediction labels corresponding to each identification sub-model; and fuse the respective behavior prediction labels corresponding to each identification sub-model by using a voting method to obtain the behavior prediction result of the image to be recognized.

[0104] In addition, the exemplary embodiments of the present application further provide a computer storage medium, in which instructions for executing the steps in the behavior recognition model training method described in each embodiment, or instructions for executing the steps in the behavior recognition method described in each of the above embodiments are stored.

[0105] In summary, for the behavior recognition model solutions provided in the embodiments of the present application, by simply calculating the normalized features, symmetric features, and centralized features of each bone node, the training features of existing training samples can be fully exploited.

[0106] In addition, the present application solves the problem of sparse input features in the existing behavior recognition method based on human body bone points. By expanding the input features for model training, the accuracy of model recognition can be effectively improved.

[0107] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the embodiments of the present application, and are not intended to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for training a behavior recognition model, characterized in that, Including: Obtain respective normalized features, respective symmetric features, and respective centered features corresponding to each of the bone nodes identified from the training samples; Obtain multiple training feature groups of the training samples according to the respective normalized features, respective symmetric features, and respective centered features corresponding to each of the bone nodes; the multiple training feature groups include a normalized feature group, a symmetric feature group, a centered feature group, and a fused feature group; Train multiple recognition sub-models according to the respective training feature groups of the training samples and the behavior labels of the training samples; the multiple recognition sub-models include a normalized feature sub-model, a symmetric feature sub-model, a centered feature sub-model, and a fused feature sub-model; Obtain a trained behavior recognition model based on the trained multiple recognition sub-models; Wherein, the obtaining the multiple training feature groups of the training samples according to the respective normalized features, respective symmetric features, and respective centered features corresponding to each of the bone nodes includes: Perform arbitrary combinations on at least one of the respective normalized features, respective symmetric features, and respective centered features corresponding to each of the bone nodes to generate multiple training feature groups of the training samples.

2. The method for training a behavior recognition model according to claim 1, wherein The obtaining respective normalized features, respective symmetric features, and respective centered features corresponding to each of the bone nodes according to the bone nodes identified from the training samples includes: Identify a target object in the training sample, and obtain a target detection box of the target object and respective bone node coordinates corresponding to each of the bone nodes of the target object; Perform normalization processing on each of the bone node coordinates according to the target detection box to obtain respective normalized features corresponding to each of the bone nodes; Obtain respective symmetric features corresponding to each of the bone nodes according to a central axis determined by each of the bone nodes and the respective normalized features corresponding to each of the bone nodes; Obtain respective centered features corresponding to each of the bone nodes according to a center point determined by each of the bone nodes and the respective normalized features corresponding to each of the bone nodes.

3. The method for training the behavior recognition model according to claim 2, wherein The performing normalization processing on each of the bone node coordinates according to the target detection box to obtain respective normalized features corresponding to each of the bone nodes includes: Perform normalization processing on each of the bone node coordinates according to the height and width of the target detection box to obtain respective normalized coordinates corresponding to each of the bone nodes; The normalized coordinate of the bone node can be expressed as: Among them, the is the normalized coordinate of the bone node, the (x i , y i ) is the coordinate of the bone node, the W is the width of the target detection box, and the H is the height of the target detection box.

4. The method for training an action recognition model according to claim 3, wherein The obtaining respective symmetric features corresponding to each of the bone nodes according to a central axis determined by each of the bone nodes and the respective normalized features corresponding to each of the bone nodes includes: Determine the central axis according to a connection line formed by a vertex node and a neck node among the bone nodes; [[ID=…]]Based on the central axis, determine a first symmetric bone node and a second symmetric bone node; Determine respective symmetric coordinates of the first bone node and the second bone node according to the respective normalized coordinates of the first bone node and the second bone node; The symmetric coordinate of the first bone node can be expressed as: The symmetric coordinates of the second bone node can be expressed as: Among them, the is the normalized coordinate of the first bone node, and the is the normalized coordinate of the second bone node.

5. The method for training the behavior recognition model according to claim 4, wherein The method further includes: Determining the symmetric coordinates of the bone node located on the central axis as (0, 0).

6. The method for training the behavior recognition model according to claim 3, wherein The obtaining of the centralized features corresponding to each of the bone nodes according to the center points determined by each of the bone nodes and the normalized features corresponding to each of the bone nodes includes: Determining the center point according to the vertex node in the bone node; Determining the centralized coordinates corresponding to each of the bone nodes according to the normalized coordinates of the center point and the normalized coordinates corresponding to each of the bone nodes; The centralized coordinates of the bone node can be expressed as: Among them, the represents the centralized coordinates of the bone node, and the represents the normalized coordinates of the center point, and the represents the normalized coordinates of the bone node.

7. The method for training a behavior recognition model according to claim 1, wherein The obtaining of multiple training feature groups of the training sample according to the normalized features, symmetric features, and centralized features corresponding to each of the bone nodes includes: For each of the bone nodes, arbitrarily combining at least one of the normalized feature, symmetric feature, and centralized feature of the bone node to obtain multiple training features for each of the bone nodes; Integrating the training features corresponding to each of the bone nodes with the same combination content to obtain the multiple training feature groups of the training sample.

8. The method for training a behavior recognition model according to claim 1, wherein The training of multiple recognition sub-models according to the multiple training feature groups of the training sample and the behavior labels of the training sample includes: If the training feature group is a normalized feature group composed of the normalized features corresponding to each of the bone nodes, based on the normalized feature group, constructing a normalized feature sub-model, taking the normalized feature group as the input, and taking the behavior label as the output to train the normalized feature sub-model; If the training feature group is a symmetric feature group composed of the symmetric features corresponding to each of the bone nodes, based on the symmetric feature group, constructing a symmetric feature sub-model, taking the symmetric feature group as the input, and taking the behavior label as the output to train the symmetric feature sub-model; If the training feature group is a centralized feature group composed of the centralized features corresponding to each of the bone nodes, based on the centralized feature group, constructing a centralized feature sub-model, taking the centralized feature group as the input, and taking the behavior label as the output to train the centralized feature sub-model; If the training feature group is a fused feature group composed of at least two of the centralized features, symmetric features, and centralized features corresponding to each of the bone nodes, based on the fused feature group, constructing a fused feature sub-model, taking the fused feature group as the input, and taking the behavior label as the output to train the fused feature sub-model.

9. A behavior recognition method, including: Identifying each bone node of the target object in the image to be recognized, and obtaining the normalized features, symmetric features, and centralized features corresponding to each of the bone nodes; Using a behavior recognition model to perform behavior prediction according to the normalized features, symmetric features, and centralized features corresponding to each of the bone nodes, and obtaining a behavior prediction result of the target object. Among them, the behavior recognition model is trained by using the behavior recognition model training method described in any one of claims 1 to 8, and the behavior recognition model includes a normalization feature sub-model, a symmetrization feature sub-model, a centralization feature sub-model, and a fusion feature sub-model; The behavior prediction performed by using the behavior recognition model according to the respective normalization features, respective symmetrization features, and respective centralization features corresponding to each of the bone nodes includes: Using each recognition sub-model in the behavior recognition model, performing behavior prediction on the image to be recognized respectively, and obtaining respective behavior prediction labels corresponding to each of the recognition sub-models; using a voting method to fuse the respective behavior prediction labels corresponding to each of the recognition sub-models, and obtaining the behavior prediction result of the image to be recognized.

10. A computer storage medium, characterized in that, The computer storage medium stores instructions for executing each step in the behavior recognition model training method described in any one of claims 1 to 8, or instructions for executing each step in the behavior recognition method described in claim 9 above.

11. A behavior recognition model training device, comprising: A bone feature acquisition module, configured to acquire respective normalization features, respective symmetrization features, and respective centralization features corresponding to each of the bone nodes according to each bone node identified from the training samples; A training feature generation module, configured to obtain multiple training feature groups of the training samples according to the respective normalization features, respective symmetrization features, and respective centralization features corresponding to each of the bone nodes; The multiple training feature groups include a normalization feature group, a symmetrization feature group, a centralization feature group, and a fusion feature group; A training module, configured to train multiple recognition sub-models according to the multiple training feature groups of the training samples and the behavior labels of the training samples, and obtain a trained behavior recognition model based on the trained multiple recognition sub-models; the multiple recognition sub-models include a normalization feature sub-model, a symmetrization feature sub-model, a centralization feature sub-model, and a fusion feature sub-model; Among them, the training feature generation module is configured to perform any combination on at least one of the respective normalization features, respective symmetrization features, and respective centralization features corresponding to each of the bone nodes, and generate multiple training feature groups of the training samples.

12. A behavior recognition device, comprising: An acquisition module, configured to identify each bone node of the target object in the image to be recognized, and acquire respective normalization features, respective symmetrization features, and respective centralization features corresponding to each of the bone nodes; A recognition module, configured to perform behavior prediction by using a behavior recognition model according to the respective normalization features, respective symmetrization features, and respective centralization features corresponding to each of the bone nodes, and obtain the behavior prediction result of the target object; among them, the behavior recognition model is trained by using the behavior recognition model training device described in claim 11, and the behavior recognition model includes a normalization feature sub-model, a symmetrization feature sub-model, a centralization feature sub-model, and a fusion feature sub-model; The recognition module is further configured to use each recognition sub-model in the behavior recognition model to perform behavior prediction on the image to be recognized respectively, so as to obtain each behavior prediction label corresponding to each recognition sub-model; and fuse each behavior prediction label corresponding to each recognition sub-model by means of voting to obtain the behavior prediction result of the image to be recognized.

Images