Body movement recognition method, device, computer equipment and storage medium

Abstract

The present application discloses a method, device, computer equipment and storage medium for body movement recognition, which belongs to the field of image recognition. The method includes: obtaining a target time-series picture set, which is a set of pictures of a group of body movements with a time series; inputting the target time-series picture set into a first network in sequence, obtaining a target tensor output by the first network, the first network is used to process the target time-series picture into a feature matrix, the target tensor is used to characterize the feature matrix set corresponding to the target time-series picture set, inputting the target tensor into a second network, obtaining a body movement recognition result, and extracting at least one sub-tensor from the target tensor and inputting it into the second network to obtain the probability of the action category corresponding to each sub-tensor, and obtaining the body movement recognition result according to the weighted summation. The technical solution of the present application can solve the problem of low accuracy due to inconsistent continuity of actions when performing body movement recognition.

Description

Technical Field

[0001] The present invention relates to the field of image recognition, and in particular to a body movement recognition method, device, computer equipment and storage medium. Background Art

[0002] With the research and advancement of artificial intelligence technology, artificial intelligence technology has been studied and applied in many fields, such as smart devices. Motion recognition technology enables smart devices to understand the actions of interactive objects and make corresponding responses. For example, users trigger a function of a smart device through body movements. Body movements are the key technology to realize intelligent interaction.

[0003] The existing technology usually adopts an analysis method based on the long short-term memory network (LSTM). However, this method controls the duration of the time series image by the memory module in the LSTM, and it is usually difficult to guide the network to carry out targeted training. Therefore, the existing technology has low accuracy when facing inconsistent duration of limb movements during limb recognition. Summary of the invention

[0004] Based on this, the embodiments of the present application provide a method, apparatus, computer device and storage medium for body motion recognition, which can solve the problem of low accuracy due to inconsistent continuity of motion when performing body motion recognition.

[0005] In a first aspect, a method for identifying body movements is provided, the method comprising:

[0006] Acquire a target time-series picture set, where the target time-series picture set is a set of pictures of a group of body movements with a time sequence;

[0007] Inputting the target time series picture set into the first network in sequence, obtaining a target tensor output by the first network, wherein the first network is used to process the target time series picture into a feature matrix, and the target tensor is used to represent a feature matrix set corresponding to the target time series picture set;

[0008] The target tensor is input into the second network to obtain a body movement recognition result.

[0009] In one embodiment, the method further comprises:

[0010] Extracting at least one sub-tensor from the target tensor;

[0011] Inputting the at least one sub-tensor into the second network respectively to obtain the action category probability corresponding to each sub-tensor in the at least one sub-tensor;

[0012] The action category probabilities corresponding to the sub-tensors are weighted and summed to obtain the limb action recognition result.

[0013] In one of the embodiments, the sub-tensor is a set of feature matrices corresponding to consecutive frame images in a target temporal image set.

[0014] In one of the embodiments, the sub-tensor includes at least a feature matrix corresponding to an end frame picture in a target temporal picture set.

[0015] In one embodiment, the target time series picture set is sequentially input into a first network, and a target tensor output by the first network is obtained, including:

[0016] Inputting the target time series picture set into the first network in sequence to obtain a plurality of feature matrices with time series;

[0017] The multiple feature matrices are accumulated in the time dimension to obtain a target tensor output by the first network.

[0018] In one embodiment, acquiring a target time-series picture set includes:

[0019] Acquire a target video, wherein the target video includes at least one set of body movements;

[0020] The target video is frame extracted to obtain at least one target time-series picture set.

[0021] In a second aspect, a body movement recognition device is provided, the body movement recognition device comprising:

[0022] An acquisition module is used to acquire a target time-series picture set, where the target time-series picture set is a set of pictures of a group of body movements with a time sequence;

[0023] A first network module, configured to sequentially input the target time series picture set into a first network to obtain a target tensor output by the first network;

[0024] The second network module is used to input the target tensor into the second network to obtain a body movement recognition result.

[0025] In one embodiment, the second network module is further used to: extract at least one sub-tensor from the target tensor; input the at least one sub-tensor into the second network respectively to obtain the action category probability corresponding to each sub-tensor in the at least one sub-tensor; and perform weighted summation of the action category probabilities corresponding to each sub-tensor to obtain a limb movement recognition result.

[0026] In a third aspect, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the body movement recognition method described in any one of the first aspects is implemented.

[0027] In a fourth aspect, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the body movement recognition method described in any one of the first aspects is implemented.

[0028] The beneficial effects brought by the technical solution provided by the embodiment of the present application include at least:

[0029] In the technical solution provided by the embodiment of the present application, the target time-series picture set is first obtained, and then the target time-series picture set is sequentially input into the first network, the target time-series picture set is processed into a corresponding feature matrix set, i.e., a target tensor, and finally the obtained target tensor is input into the second network to obtain a limb action recognition result. Compared with the existing method of directly inputting the time-series picture set into the LSTM network to obtain the action recognition result, the technical solution provided by the embodiment of the present application can solve the problem of low accuracy due to inconsistent action continuity when performing limb recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 A schematic diagram of an implementation environment provided for an embodiment of the present application;

[0031] Figure 2 A flowchart of a body movement recognition method provided in an embodiment of the present application;

[0032] Figure 3 A flowchart of a body movement recognition method provided in an embodiment of the present application;

[0033] Figure 4 A flowchart of a body movement recognition method provided in an embodiment of the present application;

[0034] Figure 5 A block diagram of a body movement recognition device provided in an embodiment of the present application;

[0035] Figure 6 A schematic diagram of using an embodiment of the present application to perform body movement recognition;

[0036] Figure 7 A schematic diagram of a computer device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0037] In order to make the objectives, technical solutions and advantages of the present application clearer, the implementation methods of the present application will be further described in detail below with reference to the accompanying drawings.

[0038] With the research and advancement of artificial intelligence technology, artificial intelligence technology has been studied and applied in many fields, such as smart devices. Motion recognition technology enables smart devices to understand the actions of interactive objects and make corresponding responses. For example, users trigger a function of a smart device through body movements. Body movements are the key technology to realize intelligent interaction.

[0039] The existing technology usually adopts an analysis method based on the long short-term memory network (LSTM). However, this method controls the duration of the time series image by the memory module in the LSTM, and it is usually difficult to guide the network to carry out targeted training. Therefore, when the existing technology is used for limb recognition, when the duration of limb movements is inconsistent, it has low accuracy.

[0040] In view of this, the embodiments of the present application provide a method, apparatus, computer device and storage medium for body movement recognition, which can improve the recognition accuracy of body movements in a time-series picture set.

[0041] Below, the implementation environment involved in the body movement recognition method provided in the embodiment of the present application will be described.

[0042] Figure 1 The present invention provides a schematic diagram of the implementation environment involved in the body movement recognition method. Figure 1 As shown, the implementation environment may include a terminal 101 and a server 102, wherein the terminal 101 and the server 102 may communicate with each other via a wired or wireless manner.

[0043] The terminal 101 can take a picture of the interactive object to obtain a set of pictures with body movements, and the terminal 101 can send the set of pictures with body movements to the server 201. The server 200 outputs the corresponding body movement recognition result to the terminal 101 by executing the body movement recognition method provided in the embodiment of the present application.

[0044] The terminal 101 may be a smart phone, a tablet computer, a desktop computer, a wearable device, an MP3 player (Moving Picture Experts Group Audio Layer III), an MP4 player (Moving Picture Experts Group Audio Layer IV), an e-book reader, etc. The server 102 may be a single server or a server cluster consisting of multiple servers.

[0045] Please refer to Figure 2, which shows a flow chart of a body movement recognition method provided by an embodiment of the present application, and the body movement recognition method can be applied to the server 102 in the implementation environment described above. Figure 2 As shown, the body movement recognition method may include the following steps:

[0046] Step 201, obtaining a target time-series picture set.

[0047] The target time-series picture set may be a set of pictures of a group of body movements with a time sequence.

[0048] For example, the body movement may be the right arm of the interactive object moving to the left, and the duration of the movement of the right arm of the interactive object moving to the left may be 0.5s. Then, n picture frames may be included in this 0.5s, that is, these n picture frames together constitute the movement of the right arm of the interactive object moving to the left. These n picture frames can be recorded as T -n ,T -n+1 ,T -n+2 ...T0, where T -n is the initial frame of n picture frames, indicating the initial picture of the right arm of the interactive object moving to the left, T0 is the end frame of n picture frames, indicating the end picture of the right arm of the interactive object moving to the left, T -n+1 In the time dimension, the distance T -n The most recent picture frame, these n picture frames can be recorded as arranged according to the time dimension, and the remaining ones are T -n+2 …T0, these n picture frames have a time sequence.

[0049] In an optional embodiment of the present application, acquiring a target time-series picture set may first acquire a target video, wherein the target video includes at least one set of body movements.

[0050] When the target video includes a set of body movements, the target video may be directly frame extracted to obtain a set of pictures of the body movements in a time series.

[0051] When the target video includes multiple sets of body movements, the target video can be used for pattern recognition using existing technology to identify and segment the body movements of different interactive objects, and then frames of different interactive objects are extracted to obtain a set of pictures of the body movements of different interactive objects in a time series.

[0052] Step 202: input the target time series picture set into the first network in sequence to obtain the target tensor output by the first network.

[0053] The first network may be a network that processes an image into a feature matrix, the feature matrix is ​​used to represent key points in the image, and the tensor may be a collection of feature matrices.

[0054] In an embodiment of the present application, the obtained target time series picture set is sequentially input into the first network in time series, wherein the first network may be a 2D convolutional network.

[0055] For example, when the target time series image set is T -n ,T -n+1 ,T -n+2 ...T0, T -n ,T -n+1 ,T -n+2 …T0 is input to the first network, and T -n ,T -n+1 ,T -n+2 …T0’s feature matrix, the obtained n feature matrices are then superimposed in time series to obtain the target tensor.

[0056] Step 203, input the target tensor into the second network to obtain the body movement recognition result.

[0057] In the embodiment of the present application, the second network may be a network that obtains a recognition result by performing a network operation on the tensor, wherein the second network may be a temporal convolutional network (TCN) or a multi-level temporal convolutional network (Multi-Level TCN).

[0058] The obtained target tensor is input into the second network, and the action with the largest probability value is the classification result, that is, the body movement recognition result.

[0059] In the technical solution provided by the embodiment of the present application, the target time-series picture set is first obtained, and then the target time-series picture set is sequentially input into the first network, the target time-series picture set is processed into a corresponding feature matrix set, i.e., a target tensor, and finally the obtained target tensor is input into the second network to obtain a limb action recognition result. Compared with the existing method of directly inputting the time-series picture set into the LSTM network to obtain the action recognition result, the technical solution provided by the embodiment of the present application can solve the problem of low accuracy due to inconsistent action continuity when performing limb recognition.

[0060] In an optional embodiment of the present application, Figure 3 As shown, the technical process after obtaining the target tensor output by the first network includes the following steps:

[0061] Step 301: extract at least one sub-tensor from the target tensor.

[0062] Among them, a sub-tensor refers to a low-dimensional tensor extracted from a high-dimensional tensor. For example, in the present application, the sub-tensor can be an n-dimensional target tensor obtained by superimposing n feature matrices in a time series, and a tensor with a dimension less than n is extracted. In the embodiment of the present application, the feature matrix sets between the sub-tensors may have intersections.

[0063] In an optional embodiment of the present application, the sub-tensor is a set of feature matrices corresponding to consecutive frame images in the target time-series image set. For example, the sub-tensor can be T -n+2 …T0, or it may be a set of feature matrices of T4…T0, that is, the sub-tensor is a set of feature matrices corresponding to consecutive frame images.

[0064] In an optional embodiment of the present application, the sub-tensor at least includes the feature matrix corresponding to the end frame picture in the target time series picture set, that is, each extracted sub-tensor must include the feature matrix corresponding to the picture at time T0.

[0065] Step 302: input at least one sub-tensor into the second network respectively to obtain the action category probability corresponding to each sub-tensor in the at least one sub-tensor.

[0066] In an embodiment of the present application, the obtained sub-tensors are respectively input into the second network to obtain the action category probability corresponding to the number of sub-tensors.

[0067] Step 303: Perform weighted summation on the action category probabilities corresponding to the sub-tensors to obtain the body action recognition result.

[0068] The action category probabilities corresponding to the number of sub-tensors are weighted summed up, and the action with the largest probability value is the classification result, that is, the body movement recognition result.

[0069] Compared with the existing temporal convolutional network for processing time-related inputs, which is aimed at 2D tensors (for example, the paper An Empirical Evaluation of Generic Convolutional and Recurrent Networks for Sequence Modeling), the embodiment of the present application can effectively process 3D tensors by extracting at least one sub-tensor from the target tensor, inputting the sub-tensor into the second network for calculation and performing weighted summation based on the obtained probability values, so that the obtained body movement recognition results are more accurate and efficient than the existing technology.

[0070] In an optional embodiment of the present application, Figure 4 As shown, the target time series picture set is sequentially input into the first network, and the technical process of obtaining the target tensor output by the first network includes the following steps:

[0071] Step 401, input the target time series picture set into the first network in sequence to obtain a plurality of feature matrices with time series.

[0072] Step 402: Multiple feature matrices are accumulated in the time dimension to obtain a target tensor output by the first network.

[0073] For example, when the target time series image set is T -n ,T -n+1 ,T -n+2 ...T0, T -n ,T -n+1 ,T -n+2 …T0 is input to the first network, and T -n ,T -n+1 ,T -n+2 …T0’s feature matrix, the obtained n feature matrices are then superimposed in time series to obtain the target tensor.

[0074] Please refer to Figure 5 , which shows a block diagram of a body movement recognition device 500 provided in an embodiment of the present application, and the body movement recognition device 500 can be configured in a server. Figure 5 As shown, the body motion recognition device 500 may include: an acquisition module 501 , a first network module 502 , and a second network module 503 .

[0075] The acquisition module 501 is used to acquire a target time-series picture set, where the target time-series picture set is a set of pictures of a group of body movements in a time sequence.

[0076] The first network module 502 is used to input the target time series picture set into the first network in sequence to obtain the target tensor output by the first network.

[0077] The second network module 503 is used to input the target tensor into the second network to obtain the body movement recognition result.

[0078] In an optional embodiment of the present application, the second network module 503 is also used to extract at least one sub-tensor from the target tensor; input the at least one sub-tensor into the second network respectively to obtain the action category probability corresponding to each sub-tensor in the at least one sub-tensor; and perform weighted summation on the action category probabilities corresponding to each sub-tensor to obtain the limb movement recognition result.

[0079] In an optional embodiment of the present application, the sub-tensor is a set of feature matrices corresponding to consecutive frame images in the target time-series image set.

[0080] In an optional embodiment of the present application, the sub-tensor includes at least a feature matrix corresponding to an end frame picture in a target time-series picture set.

[0081] In an optional embodiment of the present application, the first network module 502 is specifically used to input the target time series picture set into the first network in sequence to obtain multiple feature matrices with time series; and accumulate the multiple feature matrices in the time dimension to obtain the target tensor output by the first network.

[0082] In an optional embodiment of the present application, the acquisition module 501 is specifically used to acquire a target video, where the target video includes at least one set of body movements; and extract frames from the target video to obtain at least one target time-series picture set.

[0083] For the specific definition of the body movement recognition device, please refer to the definition of the body movement recognition method above, which will not be repeated here. Each module in the above-mentioned body movement recognition device can be implemented in whole or in part by software, hardware and a combination thereof. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory in the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.

[0084] Please refer to Figure 6 In an optional embodiment of the present application, the first network is a key point extraction network, and the second network is a Multi-LevelTCN.

[0085] First, the server obtains T -(n+3) ,T -(n+2) ,T -(n+1) …T0 has a total of n+4 target time-series images, and these n+4 target time-series images are input into the key point extraction network (2D convolutional network) and all the features obtained are superimposed in the time dimension and input into the Multi-LevelTCN network.

[0086] In the Multi-Level TCN network, the accumulated tensor is a 3D tensor with n+4 channels. Multiple subsets are taken from the tensor as the input of the corresponding TCN module. Here, the subsets are divided into three categories, and a subset is taken every 10 frames, that is, T0 to T -(n+3) is a 3D tensor with n+4 channels. This tensor is input into a TCN with n+4 receiving channels to obtain a set of probability values ​​for each action category. Then T0 to T -(n-7) is a 3D tensor with n-6 channels, which is input into a TCN with n-6 receiving channels, and a set of probability values ​​for each action category is obtained; finally, T0 to T -(n-17)A 3D tensor with n-16 channels is input into a TCN with n-16 receiving channels, and a set of probability values ​​for each action category is obtained. Here, the number of channels of each TCN module differs by 10, that is, the time difference is controlled to about 10 frames, which is about half a second for a video stream with a fps of 25. This will neither make the time difference too large to cause the model to not cover all possible situations, nor make the time difference too small to cause repeated calculations. The frame difference is a hyperparameter that can be modified according to different action tasks.

[0087] Finally, the three groups of action category probabilities output by Multi-LevelTCN are weighted and summed, and the action with the largest value (argmax operation) is the classification result.

[0088] In one embodiment, a computer device is provided. The computer device may be a server, and its internal structure diagram may be as follows: Figure 7 As shown. The computer device includes a processor, a memory and a network interface connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store body movement recognition data. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, a body movement recognition method is implemented.

[0089] Those skilled in the art will understand that Figure 7 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0090] In one embodiment, a computer device is provided, including a memory and a processor, wherein a computer program is stored in the memory, and when the processor executes the computer program, the following steps are implemented:

[0091] A target time series picture set is obtained, which is a set of pictures of a group of body movements with a time series; the target time series picture set is sequentially input into the first network, and a target tensor output by the first network is obtained, the first network is used to process the target time series picture into a feature matrix, and the target tensor is used to characterize the feature matrix set corresponding to the target time series picture set; the target tensor is input into the second network to obtain a body movement recognition result.

[0092] In one embodiment of the present application, when the processor executes the computer program, it also implements the following steps: dividing the target tensor into at least one sub-tensor; inputting at least one sub-tensor into the second network respectively to obtain the action category probability corresponding to each sub-tensor in the at least one sub-tensor; and performing weighted summation on the action category probabilities corresponding to each sub-tensor to obtain the limb movement recognition result.

[0093] In one embodiment of the present application, when the processor executes the computer program, it also implements the following steps: inputting the target time series picture set into the first network in sequence to obtain multiple feature matrices with time series; accumulating the multiple feature matrices in the time dimension to obtain the target tensor output by the first network.

[0094] In one embodiment of the present application, when the processor executes the computer program, it also implements the following steps: acquiring a target video, the target video including at least one set of body movements; extracting frames from the target video to obtain at least one target time-series picture set.

[0095] In one embodiment of the present application, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the following steps are implemented:

[0096] A target time series picture set is obtained, which is a set of pictures of a group of body movements with a time series; the target time series picture set is sequentially input into the first network, and a target tensor output by the first network is obtained, the first network is used to process the target time series picture into a feature matrix, and the target tensor is used to characterize the feature matrix set corresponding to the target time series picture set; the target tensor is input into the second network to obtain a body movement recognition result.

[0097] In one embodiment of the present application, when the computer program is executed by a processor, the following steps are also implemented: dividing the target tensor into at least one sub-tensor; inputting at least one sub-tensor into the second network respectively to obtain the action category probability corresponding to each sub-tensor in the at least one sub-tensor; and performing weighted summation on the action category probabilities corresponding to each sub-tensor to obtain the limb movement recognition result.

[0098] In one embodiment of the present application, when the computer program is executed by the processor, the following steps are also implemented: inputting the target time series picture set into the first network in sequence to obtain multiple feature matrices with time series; accumulating the multiple feature matrices in the time dimension to obtain the target tensor output by the first network.

[0099] In one embodiment of the present application, when the computer program is executed by a processor, the following steps are also implemented: obtaining a target video, the target video includes at least one set of body movements; extracting frames from the target video to obtain at least one target time-series picture set.

[0100] The computer-readable storage medium provided in this embodiment has similar implementation principles and technical effects to those of the above method embodiments, and will not be described in detail here.

[0101] Those of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing related hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in M ​​forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (SyMchliMk) DRAM (SLDRAM), memory bus (RaMbus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0102] The technical features of the above-described embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above-described embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0103] The above-described embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be construed as limiting the scope of the patent application. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent application shall be subject to the attached claims.

Claims

1. A body movement recognition method, characterized in that: The method comprises: Acquire a target time-series picture set, where the target time-series picture set is a set of pictures of a group of body movements with a time sequence; Inputting the target time series picture set into the first network in sequence, obtaining a target tensor output by the first network, wherein the first network is used to process the target time series picture into a feature matrix, and the target tensor is used to characterize a feature matrix set corresponding to the target time series picture set; wherein the first network is a 2D convolutional network; Inputting the target tensor into the second network to obtain a body movement recognition result; After obtaining the target tensor output by the first network, the method further includes: extracting at least one sub-tensor from the target tensor; inputting the at least one sub-tensor into the second network respectively to obtain the action category probability corresponding to each sub-tensor in the at least one sub-tensor; performing weighted summation on the action category probabilities corresponding to the sub-tensors to obtain a limb action recognition result; The sub-tensor is a set of feature matrices corresponding to consecutive frame images in the target time series image set; The sub-tensor at least includes a feature matrix corresponding to an end frame picture in a target time series picture set; Inputting the target time series picture set into the first network in sequence to obtain the target tensor output by the first network, including: inputting the target time series picture set into the first network in sequence to obtain a plurality of feature matrices with a time series; accumulating the plurality of feature matrices in the time dimension to obtain the target tensor output by the first network; The method of acquiring a target time-series picture set includes: acquiring a target video, wherein the target video includes at least one set of body movements; and extracting frames from the target video to obtain at least one target time-series picture set.

2. A body movement recognition device, characterized in that: The device comprises: An acquisition module is used to acquire a target time-series picture set, where the target time-series picture set is a set of pictures of a group of body movements with a time sequence; A first network module, used to sequentially input the target time series picture set into the first network to obtain a target tensor output by the first network; wherein the first network is a 2D convolutional network; A second network module, used for inputting the target tensor into a second network to obtain a body movement recognition result; The second network module is further used to: extract at least one sub-tensor from the target tensor; input the at least one sub-tensor into the second network respectively to obtain the action category probability corresponding to each sub-tensor in the at least one sub-tensor; perform weighted summation on the action category probabilities corresponding to the sub-tensors to obtain the limb action recognition result; The sub-tensor is a set of feature matrices corresponding to consecutive frame images in the target time series image set; The sub-tensor at least includes a feature matrix corresponding to an end frame picture in a target time series picture set; Inputting the target time series picture set into the first network in sequence to obtain the target tensor output by the first network, including: inputting the target time series picture set into the first network in sequence to obtain a plurality of feature matrices with a time series; accumulating the plurality of feature matrices in the time dimension to obtain the target tensor output by the first network; The method of acquiring a target time-series picture set includes: acquiring a target video, wherein the target video includes at least one set of body movements; and extracting frames from the target video to obtain at least one target time-series picture set.

3. A computer device, characterized in that: It includes a memory and a processor, the memory stores a computer program, and when the computer program is executed by the processor, the body movement recognition method as claimed in claim 1 is implemented.

4. A computer-readable storage medium, characterized in that: A computer program is stored thereon, and when the computer program is executed by a processor, the body movement recognition method as claimed in claim 1 is implemented.

Images