CN110148212B

CN110148212B - Action sequence generation method and device, electronic equipment and storage medium

Info

Publication number: CN110148212B
Application number: CN201910412248.1A
Authority: CN
Inventors: 颜思捷; 李治中; 熊元骏; 林达华
Original assignee: Beijing Sensetime Technology Development Co Ltd
Current assignee: Beijing Sensetime Technology Development Co Ltd
Priority date: 2019-05-17
Filing date: 2019-05-17
Publication date: 2023-01-31
Anticipated expiration: 2039-05-17
Also published as: CN110148212A

Abstract

The present disclosure relates to an action sequence generation method and apparatus, an electronic device, and a storage medium, wherein the method includes: acquiring a first noise vector sequence generated by a plurality of random processes; and processing the first noise vector sequence by using a convolutional neural network to generate a target human skeleton action sequence. Embodiments of the present disclosure may generate human skeletal motion sequences at multiple time steps.

Description

Action sequence generation method and device, electronic equipment and storage medium

Technical Field

The present disclosure relates to the field of computer technologies, and in particular, to a method and an apparatus for generating an action sequence, an electronic device, and a storage medium.

Background

Skeletal motion generation is an important issue for computer vision and human behavioral understanding. Skeletal motion generation may have important applications in many fields, such as 3D model motion control, video generation, augmented Reality (AR), and so on.

However, the generation of skeletal actions is somewhat difficult. On the one hand, the skeletal action of the human body is influenced by many factors, and it is necessary to keep the center of gravity stable while being restricted by muscular strength and flexibility of the human body. On the other hand, the conventional bone motion generation technology is difficult to generate a rich variety of motions and a long motion sequence, and is easy to repeatedly generate the same motions.

Disclosure of Invention

The present disclosure provides an action sequence generation technical solution.

According to an aspect of the present disclosure, there is provided an action sequence generation method including:

acquiring a first noise vector sequence generated by a plurality of random processes;

and processing the first noise vector sequence by using a convolutional neural network to generate a target human skeleton action sequence.

In one or more optional embodiments, before the obtaining the first noise vector sequence generated by the plurality of random processes, the method further includes:

acquiring a known human skeleton action sequence;

generating a second noise vector sequence based on the known human skeleton motion sequence;

generating the first sequence of noise vectors by a plurality of random processes based on the second sequence of noise vectors.

In one or more optional embodiments, the obtaining a first sequence of noise vectors generated by a plurality of random processes comprises:

generating a noise vector by each of a plurality of random processes;

and combining the plurality of noise vectors generated in the plurality of random processes to obtain a first noise vector sequence.

In one or more optional embodiments, the processing the first noise vector sequence with a convolutional neural network to generate a target human skeletal action sequence includes:

processing the first noise vector sequence by using a convolutional neural network to obtain information of a plurality of human skeleton nodes, wherein the information of the plurality of human skeleton nodes comprises position information of the plurality of human skeleton nodes, spatial correlation information of the plurality of human skeleton nodes and displacement information of the plurality of human skeleton nodes in each time step in a plurality of time steps;

generating the target human bone action sequence based on the information of the plurality of human bone nodes.

In one or more optional embodiments, the convolutional neural network is a graph convolutional neural network, and the processing the first noise vector sequence by using the convolutional neural network to generate a target human bone action sequence includes:

generating an initial space-time map based on the noise vector sequence;

and processing the initial space-time diagram by using a graph convolution neural network to generate a target human skeleton action sequence.

In one or more alternative embodiments, the initial space-time map includes a human bone node.

In one or more optional embodiments, the processing the initial space-time map by using a graph convolutional neural network to generate a target human bone action sequence includes:

and performing at least one stage of convolution processing on the initial space-time map by using a graph convolution neural network to generate a target human skeleton action sequence, wherein each stage of processing in the at least one stage of convolution processing comprises upsampling processing and/or convolution processing.

In one or more optional embodiments, the time-space diagram generated by performing at least one stage of convolution processing on the initial time-space diagram includes a plurality of human bone nodes, and different human bone nodes correspond to different weights in the at least one stage of convolution processing.

In one or more optional embodiments, the upsampling process comprises:

the method comprises the steps of conducting upsampling processing on an input space-time graph in at least one dimension of time and space to obtain the space-time graph output by the upsampling processing, wherein the number of human skeleton nodes included in the output space-time graph is larger than or equal to the number of human skeleton nodes included in the input space-time graph, and the number of time steps corresponding to the output space-time graph is larger than or equal to the number of time steps corresponding to the input space-time graph.

In one or more optional embodiments, the method further comprises:

determining target skeleton node information corresponding to each stage of processing in the at least one stage of convolution processing, wherein the target skeleton node information comprises the number of a plurality of target skeleton nodes and the spatial incidence relation of the plurality of target skeleton nodes;

the method for performing at least one stage of convolution processing on the initial space-time diagram by using the diagram convolution neural network to generate the target human skeleton action sequence comprises the following steps:

and performing each stage of processing based on the target skeleton node information corresponding to each stage of processing in the at least one stage of convolution processing to obtain the target human skeleton action sequence.

In one or more optional embodiments, the determining the target bone node information corresponding to each stage of the at least one stage of convolution processing includes:

carrying out at least one-stage downsampling processing on a preset bone structure comprising a plurality of human body bone nodes to obtain a bone structure comprising one human body bone node, wherein the stage number of the at least one-stage downsampling processing is the same as the stage number of the at least one-stage convolution processing;

and determining the target bone node information of the up-sampling treatment contained in the at least one-stage convolution treatment according to the human bone node information obtained by each-stage down-sampling treatment in the at least one-stage down-sampling treatment.

In one or more optional embodiments, the method further comprises:

determining a target time step corresponding to the target human skeleton action sequence;

determining a length of a noise vector generated by each of the plurality of random processes based on the target time step.

In one or more optional embodiments, the stochastic process comprises a gaussian process.

According to an aspect of the present disclosure, there is provided an action sequence generating apparatus including:

an obtaining module, configured to obtain a first noise vector sequence generated through a plurality of random processes;

and the generating module is used for processing the first noise vector sequence by utilizing a convolutional neural network to generate a target human skeleton action sequence.

In one or more optional embodiments, the apparatus further comprises:

the noise sequence generation module is used for acquiring a known human skeleton action sequence; generating a second noise vector sequence based on the known human skeleton motion sequence; generating the first sequence of noise vectors by a plurality of random processes based on the second sequence of noise vectors.

In one or more optional embodiments, the obtaining module is specifically configured to,

generating a noise vector by each of a plurality of random processes;

In one or more optional embodiments, the generating module is configured to process the first noise vector sequence by using a convolutional neural network to obtain information of a plurality of human bone nodes, where the information of the plurality of human bone nodes includes position information of the plurality of human bone nodes, spatial correlation information of the plurality of human bone nodes, and displacement information of the plurality of human bone nodes at each of a plurality of time steps;

In one or more optional embodiments, the convolutional neural network is a graph convolutional neural network, and the generating module is configured to,

generating an initial space-time map based on the noise vector sequence;

and processing the initial space-time map by using a graph convolutional neural network to generate a target human skeleton action sequence.

In one or more alternative embodiments, the initial space-time map includes a human skeletal node.

In one or more optional embodiments, the generating module is configured to perform at least one stage of convolution processing on the initial space-time map by using a map convolution neural network to generate a target human bone action sequence, where each stage of the at least one stage of convolution processing includes an upsampling processing and/or a convolution processing.

In one or more optional embodiments, the generating module includes:

the up-sampling processing sub-module is configured to perform up-sampling processing on the input space-time graph in at least one dimension of time and space to obtain a space-time graph output by the up-sampling processing, where the number of human skeleton nodes included in the output space-time graph is greater than or equal to the number of human skeleton nodes included in the input space-time graph, and the number of time steps corresponding to the output space-time graph is greater than or equal to the number of time steps corresponding to the input space-time graph.

In one or more optional embodiments, the apparatus further comprises:

a first determining module, configured to determine target bone node information corresponding to each stage of processing in the at least one stage of convolution processing, where the target bone node information includes a number of multiple target bone nodes and an association relationship of the multiple target bone nodes in space;

and the generating module is used for performing each stage of processing based on the target skeleton node information corresponding to each stage of processing in the at least one stage of convolution processing to obtain the target human skeleton action sequence.

In one or more optional embodiments, the first determining module is configured to,

In one or more optional embodiments, the apparatus further comprises:

the second determination module is used for determining a target time step corresponding to the target human skeleton action sequence; determining a length of a noise vector generated by each of the plurality of stochastic processes based on the target time step.

In one or more alternative embodiments, the stochastic process comprises a gaussian process.

According to an aspect of the present disclosure, there is provided an electronic device including:

a processor;

a memory for storing processor-executable instructions;

wherein the processor is configured to: the above-described action sequence generation method is executed.

According to an aspect of the present disclosure, there is provided a computer-readable storage medium having stored thereon computer program instructions which, when executed by a processor, implement the above-described action sequence generation method.

In the embodiment of the present disclosure, a first noise vector sequence generated by a plurality of random processes may be acquired, and the first noise vector sequence is processed by using a convolutional neural network to generate a target human skeleton motion sequence. The first noise vector sequence generated by the random process is variable, so that the target human skeleton action sequence generated based on the first noise vector sequence is variable, and different requirements of a user on the length of the human skeleton action sequence are met. Meanwhile, the target human skeleton action sequence is generated by utilizing the convolutional neural network, so that the quality of the obtained human skeleton action sequence is improved.

It is to be understood that both the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the disclosure.

Other features and aspects of the present disclosure will become apparent from the following detailed description of exemplary embodiments, which proceeds with reference to the accompanying drawings.

Drawings

The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present disclosure and, together with the description, serve to explain the principles of the disclosure.

Fig. 1 shows a flow chart of an action sequence generation method according to an embodiment of the present disclosure.

FIG. 2 illustrates an exemplary schematic of noise vectors generated by different stochastic processes in an embodiment of the disclosure.

Fig. 3 illustrates a flow diagram for obtaining a first sequence of noise vectors in some embodiments according to the present disclosure.

Fig. 4 shows an example of a space-time diagram corresponding to a target human skeletal action sequence according to an embodiment of the present disclosure.

Fig. 5 shows a flowchart of an example of generating a target human skeletal action sequence according to an embodiment of the present disclosure.

Fig. 6 shows a schematic diagram of an example of a downsampling process according to an embodiment of the present disclosure.

Fig. 7 shows a flowchart of an example of generating a target human skeletal action sequence according to an embodiment of the present disclosure.

Fig. 8 shows a flowchart of an example of generating a target human skeletal action sequence according to an embodiment of the present disclosure.

Fig. 9 shows a flowchart of an example of generating a target human skeletal action sequence according to an embodiment of the present disclosure.

Fig. 10 shows a block diagram of an action sequence generating device according to an embodiment of the present disclosure.

Fig. 11 shows a block diagram of an example of an electronic device according to an embodiment of the disclosure.

Detailed Description

Various exemplary embodiments, features and aspects of the present disclosure will be described in detail below with reference to the accompanying drawings. In the drawings, like reference numbers can indicate functionally identical or similar elements. While the various aspects of the embodiments are presented in drawings, the drawings are not necessarily drawn to scale unless specifically indicated.

The word "exemplary" is used exclusively herein to mean "serving as an example, embodiment, or illustration. Any embodiment described herein as "exemplary" is not necessarily to be construed as preferred or advantageous over other embodiments.

The term "and/or" herein is merely an association describing an associated object, meaning that three relationships may exist, e.g., a and/or B, may mean: a exists alone, A and B exist simultaneously, and B exists alone. Additionally, the term "at least one" herein means any one of a variety or any combination of at least two of a variety, for example, including at least one of A, B, C, and may mean including any one or more elements selected from the group consisting of A, B and C.

Furthermore, in the following detailed description, numerous specific details are set forth in order to provide a better understanding of the present disclosure. It will be understood by those skilled in the art that the present disclosure may be practiced without some of these specific details. In some instances, methods, means, elements and circuits that are well known to those skilled in the art have not been described in detail so as not to obscure the present disclosure.

According to the action sequence generation scheme provided by the embodiment of the disclosure, a first noise vector sequence generated through a plurality of random processes can be obtained, and then the first noise vector sequence can be processed by using a convolutional neural network to generate a target human skeleton action sequence representing human skeleton actions at a plurality of time steps.

In some embodiments, the time step corresponding to the target human bone action sequence depends on the network structure of the convolutional neural network and the parameter of the first noise vector sequence, so that under the condition that the network structure of the convolutional neural network remains unchanged, the target human bone action sequences corresponding to different time steps can be obtained by adjusting the parameter of the first noise vector sequence, thereby meeting the requirements of the user on the bone action sequences at different time steps. For example, the noise vector generated by the random process may be of any length, so that the target human bone motion sequence generated by processing the first noise vector sequence with the convolutional neural network may also have any time step, but the embodiment of the present disclosure does not limit this.

The technical solution provided by the embodiment of the present disclosure may be applied to control of a virtual character, automatic generation of a video, expansion of a skeletal motion data set, and the like, and the embodiment of the present disclosure is not limited thereto.

Fig. 1 shows a flow chart of an action sequence generation method according to an embodiment of the present disclosure. The action sequence generating method may be performed by a terminal device, a server, or other type of electronic device, where the terminal device may be a User Equipment (UE), a mobile device, a user terminal, a cellular phone, a cordless phone, a Personal Digital Assistant (PDA), a handheld device, a computing device, a vehicle-mounted device, a wearable device, or the like. In some possible implementations, the sequence of actions generation method may be implemented by a processor invoking computer readable instructions stored in a memory. The following describes an operation sequence generation method according to an embodiment of the present disclosure, taking an electronic device as an execution subject.

S11, a first noise vector sequence generated through a plurality of random processes is obtained.

In the disclosed embodiment, the obtained first noise vector sequence may come from a plurality of random processes. A random process is understood to be a process in which random variables form observations at different times. The observed values of the random variables in the random process may be noise data, and the noise data generated by the random variables of the plurality of random processes for an arbitrary length of time may form a first noise vector sequence. Since the random process has a temporal characteristic, the first noise vector sequence also has a temporal characteristic. In acquiring the first noise vector sequence, the plurality of random processes may be sampled at the same time interval, for example, the plurality of random processes may be sampled at 1s intervals within 15s with the same time axis as a reference, noise data sampled by each random process may form one noise vector, and a plurality of noise vectors obtained by the plurality of random processes may form the first noise vector sequence. The first noise vector sequence may be of any length in time, corresponding to any duration. The random process here may be any random process, for example, a second moment random process, a stationary random process, a normal random process, and the like. For example, the first sequence of noise vectors may be generated using a plurality of gaussian processes with mean at zero and covariance at radial basis functions.

Here, the random variable of each random process may be subject to a corresponding random distribution, and the random distribution may be understood as a probability law of the value of the random variable, for example, a bernoulli distribution, a normal distribution, a uniform distribution, and the like. The random distribution has no time characteristics, so that the noise vector sequence acquired according to the random distribution has no time length, and a human skeleton action sequence at any time step cannot be generated.

In some embodiments, the obtained first noise vector sequence may be generated locally, or may be obtained by other devices, and a specific obtaining manner is not limited herein.

Here, each random process generates noise having a corresponding variation characteristic, and different random processes generate noise having different variation characteristics. The variation characteristic may be understood as a characteristic of the noise that varies with time, for example, a time scale, a frequency, an amplitude, etc. of the noise. The variation characteristics of the noise vectors generated by the different random processes are explained below with reference to fig. 2.

FIG. 2 illustrates an exemplary schematic of noise vectors generated by different stochastic processes in some embodiments according to the disclosure. As shown in fig. 2, each random process may correspond to a curve that changes in magnitude over time. Sigma _c Can represent the time scale, σ _c The larger the noise is, the more gradual a curve formed by the noise is, and the higher the correlation between the noises is; sigma _c The smaller the noise, the steeper the curve formed by the noise, and the lower the correlation between the noises. So that it can be based on σ _c To distinguish different random processes, each sigma _c The corresponding curve can correspond to a noise vector by obtaining different sigma _c The noise data represented by the corresponding curves may generate a plurality of noise vectors, e.g., t may be obtained ₁ To t ₂ Is different from the time length of _c Noise vectors of the same length are generated for the corresponding noise data, and a first noise vector sequence is formed by a plurality of noise vectors.

Fig. 3 shows a flow chart for obtaining a first sequence of noise vectors according to an embodiment of the disclosure.

S111, generating a noise vector through each random process of a plurality of random processes;

and S112, merging the plurality of noise vectors generated in the plurality of random processes to obtain a first noise vector sequence.

Here, in the case where the first noise vector sequence is generated locally, when the first noise vector sequence generated by the plurality of random processes is obtained, the noise vector generated by each of the plurality of random processes may be obtained, and a time length corresponding to each of the noise vectors may be the same, so that the plurality of noise vectors generated by the plurality of random processes may be combined to form the first noise vector sequence. The plurality of random processes may be any random processes, or may also be random processes that are specified in advance, and the embodiment of the present disclosure does not limit specific random processes. In this way, the first noise vector sequence with any length can be generated through a plurality of random processes, so that the target human bone motion sequence with any time step can be generated by using the noise vector sequence with any length, and the human bone motion sequence without changing the time step can be generated by using single noise.

Here, the first noise vector sequence may be represented as (C) ₀ ，1，T ₀ ) Wherein, T ₀ May represent the length of the first noise vector sequence, i.e. the number of noise vectors comprised by the first noise vector sequence, C ₀ May represent the dimension of each noise vector, and a 1 may represent that the first noise vector sequence is 1 in the spatial dimension. The length of the first noise vector sequence may be less than the length of the generated target human bone motion sequence. If the length of the target human skeleton motion sequence is represented as T, the length T of the acquired first noise vector sequence ₀ ＜T。

And S12, processing the first noise vector sequence by using a convolutional neural network to generate a target human skeleton action sequence.

In the embodiment of the present disclosure, the convolutional neural network obtained by training may be utilized to process the first noise vector sequence generated by the multiple random processes, so as to obtain a target human bone motion sequence representing human bone motion. The noise vector sequence may be directly used as an input of the convolutional neural network, or the first noise vector sequence may be preprocessed, and the preprocessed first noise vector sequence is used as an input of the convolutional neural network. The target human skeletal motion sequence can be used as the output of the convolutional neural network. The preprocessing can comprise deconvolution operation, space-time diagram building and other preprocessing, and the preprocessed first noise vector sequence is used as the input of the convolutional neural network, so that the processing effect of the convolutional neural network can be improved, and the high-quality target human skeleton action sequence can be obtained.

In some embodiments, the first noise vector sequence is processed by a convolutional neural network to generate a target human skeleton action sequence, and the first noise vector sequence may be processed by the convolutional neural network to obtain information of a plurality of human skeleton nodes; generating the target human bone action sequence based on the information of the plurality of human bone nodes. The information of the plurality of human bone nodes comprises position information of the plurality of human bone nodes, spatial association information of the plurality of human bone nodes and displacement information of the plurality of human bone nodes at each of a plurality of time steps.

Here, after the first noise vector sequence is processed by the convolutional neural network, information of a plurality of human bone nodes can be obtained. The plurality of human bone nodes may constitute a human bone, and the position information of the plurality of human bone nodes in the information of the plurality of human bone nodes may indicate a position of each human bone node in the human bone. For example, in the case of mapping a human skeleton into an image, the position information may be an image position of each human skeleton node in the image, and for example, in the case of mapping a human skeleton into a certain coordinate system, the position information may be position coordinates of each human skeleton node in the coordinate system. The spatial correlation information of the plurality of human skeleton nodes may be a connection relationship between at least two human skeleton nodes, for example, a human skeleton node a is connected with a human skeleton node B and a human skeleton node C, and the human skeleton node B is connected with the human skeleton node a. The displacement information of the plurality of human skeleton nodes at each of the plurality of time steps can indicate the position change of each human skeleton node at each of the time steps, so that the position information of each human skeleton node at each time can be determined according to the displacement information at each of the time steps, for example, the position information of the human skeleton node A at the first time step is indicated by a space point 1 to a space point 2, which indicates that the position information of the skeleton node A at the first time step is changed from the space point 1 to the space point 2. And generating a target human skeleton action sequence according to the information of the plurality of human skeleton nodes.

Fig. 4 shows an example of a space-time diagram corresponding to a target human skeletal action sequence according to an embodiment of the present disclosure. As shown in fig. 4, the target human bone motion sequence corresponds to a plurality of time steps, and the number of the time steps may be T, which may be understood as the length of the target human bone motion sequence, i.e., the time length of the target human bone motion sequence. The space-time diagram can represent the target human skeleton action sequence in space and time, in the case of representing the space-time diagram by a coordinate system, x and y can represent the coordinates of human skeleton nodes in space, and t can represent the length of the target human skeleton action sequence in time. The human skeleton comprises a plurality of human skeleton nodes, the number of the human skeleton nodes can be V, and each human skeleton node is connected with at least one other human skeleton node. The position information of the human bone node may represent the dimension of the motion vector associated with each vertex, and the position information and the displacement information of each human bone node may be represented as C. The space-time diagram corresponding to the target human skeleton motion sequence can be represented by (C, V, T).

In some embodiments, the convolutional neural network may be a graph convolutional neural network, and the first noise vector sequence may be processed using the graph convolutional neural network to generate the target human bone motion sequence. In some embodiments, the convolutional neural network may also be a convolutional neural network in various forms such as a visual geometry group network, a residual neural network, and the like, and the first noise vector sequence may be converted into a processing format matched with the convolutional neural network and then input into the corresponding convolutional neural network. For example, the first sequence of noise vectors may be converted into a processing format such as image, text, graphics, and so forth. The following describes generation of a target human skeleton motion sequence by using a convolutional neural network, taking a convolutional neural network as an example.

S121, generating an initial space-time diagram based on the first noise vector sequence;

and S122, processing the initial space-time map by using a graph convolution neural network to generate a target human skeleton action sequence.

In some embodiments, when the graph convolution neural network is used to process the noise vector sequence, the noise vector sequence may be first converted into an initial space-time diagram, and then the initial space-time diagram is processed by the graph convolution neural network to obtain a target human skeleton action sequence. The initial space-time diagram converted by the noise vector sequence can include a human skeleton node. In some embodiments, the initial space-time diagram of the first noise vector sequence conversion may include a plurality of human bone nodes, each human bone node may correspond to a human bone at a time, that is, each human bone at a time includes a human bone node, and each human bone node may correspond to a noise vector in the noise vector sequence.

In this way, when the initial space-time map is processed by using the graph convolution neural network to generate the target human body bone action sequence, the initial space-time map can be subjected to at least one-stage convolution processing by using the graph convolution neural network to generate the target human body bone action sequence. Wherein each stage of the at least one stage of convolution processing comprises an upsampling processing and/or a convolution processing.

Here, the graph convolution neural network may include an input layer, an intermediate layer, and an output layer. The input layer may input the first noise vector sequence to generate an initial space-time diagram, and in some embodiments, may also directly input the first noise vector sequence, or input other preprocessed first noise vector data, for example, input the first noise vector data that has undergone deconvolution processing. The input layer may output a target human skeletal sequence. The intermediate layer can be at least one stage, and each stage of intermediate layer can correspond to one stage of convolution processing. Each of the at least one stage of convolution processing includes up-sampling processing and/or convolution processing, i.e., each intermediate layer may perform only up-sampling processing, only convolution processing, or both up-sampling processing and convolution processing. And performing upsampling processing and convolution processing on the initial space-time diagram by using a graph convolution neural network to generate a target human skeleton action sequence.

In some examples, the graph convolution neural network may obtain information of a plurality of human bone nodes obtained by at least one convolution processing on the initial space-time graph after performing the at least one convolution processing on the initial space-time graph, and then may generate the space-time graph generated by performing the at least one convolution processing on the initial space-time graph based on the information of the plurality of human bone nodes obtained by the at least one convolution processing. And in a plurality of human skeleton nodes included in the time-space diagram generated by performing at least one stage of convolution processing on the initial time-space diagram, different human skeleton nodes correspond to different weights in the at least one stage of convolution processing. That is to say, different weights can be set for each network node in the graph convolutional neural network, and weight sharing is not performed between the network nodes, so that a target human skeleton action sequence with high quality can be obtained. In some examples, the time-space diagram generated by performing at least one stage of convolution processing on the initial time-space diagram includes a plurality of human bone nodes, and the plurality of human bone nodes may correspond to the same weight in the at least one stage of convolution processing. The plurality of human skeleton nodes can be all human skeleton nodes or partial nodes in all human skeleton nodes, so that the plurality of network nodes in the graph convolution neural network can share the weight, and the training process of the graph convolution neural network is accelerated.

In some examples, the upsampling includes upsampling the input space-time map in at least one of time and space dimensions to obtain an output space-time map of the upsampling. The number of human skeleton nodes included in the output space-time diagram is greater than or equal to the number of human skeleton nodes included in the input space-time diagram, and the number of time steps corresponding to the output space-time diagram is greater than or equal to the number of time steps corresponding to the input space-time diagram. That is, it is to be understood that the upsampling process may be performed only in the spatial dimension, only in the temporal dimension, or both in the spatial and temporal dimensions.

Here, the upsampling process may expand the generated target human bone motion sequence both temporally and spatially compared to the noise vector sequence. Each stage of upsampling processing may be to upsample the space-time map input to the stage in at least one of time and space dimensions to obtain a space-time map for the stage of upsampling processing. Under the condition that the up-sampling processing of the stage is performed on the time dimension, the number of time steps corresponding to the obtained space-time diagram is larger than that of the space-time diagram input by the up-sampling processing of the stage; in the case that the up-sampling processing at this stage is performed in the spatial dimension, the number of human skeleton nodes included in the obtained space-time graph is greater than the number of human skeleton nodes included in the space-time graph input by the up-sampling processing at this stage. The space-time diagram input by each stage of up-sampling processing is assumed to comprise the number of human skeleton nodes as V _k The space-time graph output by the up-sampling processing at the stage comprises the number V of human skeleton nodes _k+1 Then V is _k+1 ≥V _k . Suppose that a certain stage of upsampling processing is upsampling processing in a time dimension, and the number of time steps corresponding to a space-time diagram input by the stage of upsampling processing is T _k The number of time steps corresponding to the space-time diagram output by the up-sampling processing is T _k+1 Then T is _k+1 ＝nT _k Wherein k is a natural number, and n is greater than 0.

For example, assuming that the first noise vector sequence may correspond to a human bone node, the length of the first noise vector sequence may be 1s, after the upsampling process of the graph convolution neural network, the generated target human bone motion sequence may correspond to 46 human bone nodes, and the length of the target human bone motion sequence may be 32s. In this way, a sequence of actions can be generated that is increased in both time and space relative to the first sequence of noise vectors.

In some examples, preset target bone node information may be utilized when the initial space-time graph is subjected to at least one stage of convolution processing by using the graph convolution neural network to generate the target human bone action sequence. Therefore, target skeleton node information corresponding to each stage of processing in the at least one stage of convolution processing can be determined, and each stage of processing is carried out based on the target skeleton node information corresponding to each stage of processing in the at least one stage of convolution processing to obtain the target human body skeleton action sequence. Wherein the target bone node information includes a number of a plurality of target bone nodes and an association relationship of the plurality of target bone nodes in space.

Here, the target bone node information corresponding to each stage of processing may be set in advance, or the target bone node information may be determined according to a user input, for example, a user may set the number of target bone nodes, a bone architecture, and the like according to his or her preference. In some columns, the target bone node information may include the number of the plurality of target bone nodes and the spatial association relationship between the plurality of target bone nodes, so that the graph convolution neural network may perform each stage of processing based on the target bone node information, where the number of human bone nodes included in the space-time graph obtained by each stage of processing is equal to the number of target bone nodes corresponding to the stage of processing, and the spatial association information of the human bone nodes in the space-time graph obtained by each stage of processing is consistent with the spatial association relationship between the plurality of target bone nodes corresponding to the stage of processing. In some examples, the target bone node information may further include information of a region where each target bone node is located, so that an approximate position where a human bone node is located in each space-time graph may be determined according to the information of the region where the target bone node is located, so that each human bone node in the space-time graph obtained through each level of upsampling may be located within a certain region range, and it is ensured that each human bone node in the obtained space-time graph is located in a correct bone structure region.

The number of human skeleton nodes in the initial space-time graph can be increased by utilizing the graph convolution neural network to perform spatial up-sampling processing on the initial space-time graph, and a plurality of human skeleton nodes have corresponding spatial incidence relations, so that a large number of human skeleton nodes obtained by sampling on a small number of human skeleton nodes have certain difficulty, and the up-sampling of a large number of human skeleton nodes obtained by a small number of human skeleton nodes can be realized through the determined target skeleton node information.

In this example, when determining the target bone node information corresponding to each stage of processing in the at least one stage of convolution processing, at least one stage of downsampling processing may be performed on a preset bone structure including a plurality of human bone nodes to obtain a bone structure including one human bone node, and the target bone node information of the upsampling processing included in the at least one stage of convolution processing is determined according to the human bone node information obtained by each stage of downsampling processing in the at least one stage of downsampling processing. Wherein the number of levels of the at least one stage of downsampling processing is the same as the number of levels of the at least one stage of convolution processing. Here, the target bone node information corresponding to each stage of processing may be determined by performing at least one stage of downsampling processing through a preset bone structure, and the downsampling processing will be described below with reference to the accompanying drawings.

Fig. 6 shows a schematic diagram of a downsampling process according to an embodiment of the present disclosure. As shown in FIG. 6, G _k It can be said that the down-sampling process is performed in the spatial dimension, k can represent the number of stages of the down-sampling process, and k is a natural number. G ₀ The corresponding skeleton structure may represent a preset skeleton structure, and the preset skeleton structure may correspond to a space-time diagram of the target human skeleton sequence, that is, it may be understood that the number of human skeleton nodes in the space-time diagram of the target human skeleton sequence is the same as the number of target human skeleton nodes in the preset skeleton structure, and the spatial association information of the plurality of human skeleton nodes in the space-time diagram of the target human skeleton sequence is consistent with the spatial association information of the plurality of target human skeleton nodes in the preset skeleton structure. G ₁ The corresponding bone structure can represent the bone structure obtained by carrying out first-level down-sampling processing on a preset bone structure, the bone result corresponds to the last-but-second-level up-sampling processing of the graph convolution neural network, and so on, G _k The processed bone structure, which corresponds to the initial space-time map, can be downsampled at the last stage. By performing downsampling processing on the preset bone structure, target bone node information corresponding to each stage of processing in at least one stage of convolution processing can be determined.

In some embodiments, the length of the acquired first noise vector sequence may also be determined according to a target time step of a target human bone motion sequence that needs to be generated. Before the first noise vector sequence is obtained, a target time step corresponding to the target human skeleton action sequence can be determined; determining a length of a noise vector generated by each of the plurality of random processes based on the target time step. The target time step is related to the length of the first noise vector sequence, so that the lengths of a plurality of noise vectors included in the first noise vector sequence can be determined through the target time step corresponding to the target human skeleton action sequence, and thus, the target human skeleton action sequence of the target time step can be obtained through the noise vector sequence formed by the noise vectors with the determined lengths, so that the noise vectors with a certain length can be obtained according to actual requirements, and the requirements of users on the target human skeleton action sequences with different lengths are met.

Fig. 7 shows a flowchart of an example of generating a target human skeletal action sequence according to an embodiment of the present disclosure. As shown in fig. 7, a first noise vector sequence obtained by sampling is generated by a plurality of random processes, and the number of time steps of the first noise vector sequence may be T ₀ After the upsampling processing of the graph neural network, the target human skeleton action sequence may be obtained from the first noise vector sequence, and the number of time steps of the target human skeleton action sequence may be T _k ，T _k ＝T ₀ ·n ^k (ii) a Where n is greater than 0,k is the number of stages employed in proceeding in the time dimension. Correspondingly, the number of the human skeleton nodes corresponding to the first noise vector sequence is V ₀ The number of human skeleton nodes corresponding to the target human skeleton action sequence is V, and V is greater than V ₀ . Each person in the expanded space-time diagram can represent the human skeletal action corresponding to each moment.

By the action sequence generation method, the target human skeleton action sequence can be generated by utilizing the graph convolution neural network and the space-time graph corresponding to the first noise vector sequence generated by a plurality of random processes, the target human skeleton action sequence can be of any length, and the target time step of the target human skeleton action sequence can also be of any length, so that the requirement of a user on the action sequence of any length can be met.

In some embodiments, the incomplete action sequence may be complemented or predicted by using the action sequence generation scheme provided in the embodiments of the present disclosure.

S21, acquiring a known human skeleton action sequence;

s22, generating a second noise vector sequence based on the known human skeleton motion sequence;

and S23, generating the first noise vector sequence through a plurality of random processes based on the second noise vector sequence.

Here, it is known that the human skeleton motion sequence may be a defective discontinuous human skeleton motion sequence, and the defective human skeleton motion sequence needs to be completed. Alternatively, the known human bone motion sequence may be a continuous human bone motion sequence, and the subsequent human bone motion sequence needs to be predicted for the continuous human bone motion sequence. When the known human body skeleton motion sequence is supplemented or predicted, the known human body skeleton motion sequence can be converted into a second noise vector sequence, and then the second noise vector sequence can be supplemented or predicted by utilizing a plurality of random processes to obtain a first noise vector sequence generated by the plurality of random processes. When generating the second noise vector sequence based on the known human skeleton motion sequence, a pre-trained inverse neural network may be utilized, which may convert a human skeleton motion sequence of any length into a noise vector sequence. The inverse neural network may be designed in a manner similar to the structure of the convolutional neural network used to generate the target human skeletal action sequence, and the network structure at each level may be the inverse of the network structure of the convolutional neural network.

In some examples, the inverse neural network may be utilized to take the known human bone motion sequence as a second noise vector sequence, and the known human bone motion sequence is a defective human bone motion sequence due to the need to observe or complement the known human bone motion sequence. So that the second noise vector sequence generated from the known human skeletal motion sequence is also a missing noise vector sequence. The noise vector sequence may be derived from a plurality of random processes having a corresponding random distribution, and accordingly the noise vector sequence is in accordance with the corresponding random distribution of the plurality of random processes, so that this feature can be used to complement or predict a second noise vector sequence, and the first noise vector sequence associated with the second noise vector sequence generated by the plurality of random processes can be understood as a missing noise vector sequence in the complete noise vector sequence.

Fig. 9 shows a flowchart of an example of generating a target human skeletal action sequence according to an embodiment of the present disclosure. As shown in fig. 9, a continuous sequence of known human bone actions may be used to predict a human bone action sequence following the known human bone action sequence. For example, the known human bone motion sequence may be converted into a second noise vector sequence, then a plurality of random processes are sampled according to the second noise vector sequence to obtain a first noise vector sequence after the second noise vector sequence, and then the first noise vector sequence is used to generate a target human bone motion sequence after the known human bone motion sequence. And the human body skeleton action sequence missing in the middle of the plurality of discontinuous known human body skeleton action sequences can be complemented by utilizing the plurality of discontinuous known human body skeleton action sequences. For example, a plurality of known human body skeleton motion sequences may be converted into a plurality of second noise vector sequences, then a plurality of random processes are sampled according to the plurality of second noise vector sequences to obtain a missing first noise vector sequence among the plurality of second noise vector sequences, and then the missing human body skeleton motion sequence among the plurality of known human body skeleton motion sequences is generated by using the first noise vector sequence. Where H may represent a process of converting a human bone motion sequence into a second noise vector sequence, and G may represent a process of converting the second noise vector sequence into a first human bone motion sequence. In this way, known human skeletal motion sequences can be complemented or predicted.

The above action sequence generation process may utilize a convolutional neural network. When the convolutional neural network is trained, a target human body skeleton action sequence obtained by a noise vector sequence can be input into a discriminator, the discriminator can identify whether the target human body skeleton action sequence is an action sequence representing human body skeleton actions by utilizing predetermined skeleton action characteristics, and obtain an identification result, the identification result can judge whether the action sequence represents human body skeleton actions, for example, when the identification result of the discriminator is 1, the target human body skeleton action sequence can represent human body skeleton actions, and when the identification result of the discriminator is 0, the target human body skeleton action sequence can not represent human body skeleton actions. And then, the recognition result of the discriminator can be fed back to the convolutional neural network, so that the convolutional neural network is adjusted according to the recognition result fed back continuously by the discriminator, namely, the weight used in the process of generating the action sequence by the convolutional neural network is adjusted until the proportion of the recognition result obtained by the discriminator is true is greater than a preset threshold value, and the trained convolutional neural network can be obtained.

In some embodiments, the convolutional neural network may be a graph convolutional neural network. For example, a graph convolution neural network may include an input layer, an intermediate layer, and an output layer. Wherein the input layer may input an initial space-time map derived based on the first sequence of noise vectors. The intermediate layer may correspond to at least one convolution process as described above. The processing at level 1 of the middle layer can comprise activation processing, normalization processing and convolution processing, and the initial space-time diagram can be expanded in the spatial dimension; the processing at the 2 nd stage may include activation processing, normalization processing, convolution processing, and upsampling processing in the time dimension, so that the space-time map output by the processing at the 1 st stage may be increased in the time dimension, and may be kept unchanged in the space dimension, resulting in a space-time map with a larger number of corresponding time steps than the number of time steps of the initial space-time map. The 3 rd stage processing can comprise activation processing, normalization processing, convolution processing and upsampling processing in two dimensions of time and space, so that the space-time diagram obtained by the 2 nd stage processing can be increased in both the time dimension and the space dimension. The 4-stage processing may include activation processing, normalization processing, convolution processing, and upsampling processing in the time dimension, and the spatial dimension is kept unchanged, so that the space-time diagram obtained by the 3-stage processing may be further increased in the time dimension. The 5-stage processing can comprise activation processing, normalization processing, convolution processing and upsampling processing in two dimensions of time and space, so that the space-time diagram obtained by the 4-stage processing can be added in both the time dimension and the space dimension to obtain a space-time diagram of the final target human body bone action sequence. Each time dimension up-sampling process can expand the processed space-time map by n times, such as 1.5 times and 2 times, in terms of the number of time steps, where n is greater than 0; each time the up-sampling process of the spatial dimension can increase the number of human bone nodes included in the space-time diagram, the number can be increased from 1 human bone node in the initial space-time diagram to 46 final human bone nodes. The output layer can perform activation processing and convolution processing on the space-time diagram output by the middle layer (namely, the space-time diagram output by the 5 th stage processing), and output a target human skeleton action sequence. Here, the activation process may use an activation function such as a linear unit function with leakage correction, a hyperbolic tangent function, or the like.

The action sequence generation scheme provided by the embodiment of the disclosure can generate a target human skeleton action sequence at any time step, can complement or predict a partially-missing human skeleton action sequence, and can ensure that the generated target human skeleton action sequence has higher quality and diversity by performing quality evaluation on the human skeleton action sequence. The following describes a process of quality evaluation of a target human bone motion sequence.

In some embodiments, the evaluation index may be used to evaluate the generation quality of the motion sequence, obtain an evaluation score, and then determine the generation quality of the target human skeleton motion sequence according to the evaluation score.

After the convolutional neural network generates the target human body skeleton action sequence, the evaluation index can be used for evaluating the generation quality of the target human body skeleton action sequence, so that the convolutional neural network can be adjusted according to the evaluation result, and the convolutional neural network can generate the target human body skeleton action sequence with high quality and diversity. The evaluation index may include an Initiation Score (IS), a Fraschet Initiation Distance (FID). Wherein, the IS can be used for evaluating the diversity of the target human skeleton motion sequence, and the FID can be used for evaluating the quality of the target human skeleton motion sequence. When the generation quality of the target human body skeleton action sequence IS evaluated, the evaluation scores of the IS and the FID can be respectively calculated, and then the evaluation scores of the two evaluation indexes can be combined to determine the generation quality of the action sequence, wherein the generation quality can be understood as the diversity of the target human body skeleton action sequence and the quality of the skeleton framework in the space-time diagram converted from the target human body skeleton action sequence. In this way, the high quality and diversity of target human skeletal action sequences generated by the convolutional neural network can be ensured.

Here, the FID and the IS may each include three evaluation scores, i.e., a short segment evaluation score, an integrated evaluation score, and a time evaluation score. The short-segment evaluation score can represent the quality of the convolutional neural network for the generation of a shorter human skeleton action sequence, the integrated evaluation score can represent the generation quality of a longer human skeleton action sequence and the diversity of a plurality of generated human skeleton action sequences, and the time evaluation score can represent the diversity of human skeleton action changes in a longer human skeleton action sequence. The shorter human body skeleton motion sequence may be a part of the longer human body skeleton motion sequence, or may also be a human body skeleton motion sequence with a length smaller than a preset length. Here, the FID and IS can be used to evaluate the target human skeletal motion sequence generated by the above convolutional neural network trained by the two data sets respectively.

Table 1 shows structural examples of the above convolutional neural network according to an embodiment of the present disclosure corresponding to FID and IS, in which CSGN may represent the above convolutional neural network. The convolutional neural network has three evaluation scores corresponding to FID of 6.03, 5.86 and 15.40 on the first data set and three evaluation scores corresponding to FID of 15.39, 12.71 and 22.90 on the second data set; on the first data set, the three assessment scores corresponding to the IS are 22.60, 23.49 and 34.74 respectively, on the second data set, the three assessment scores corresponding to the IS are 20.71, 20.32 and 18.78 respectively, wherein the lower the assessment score corresponding to the FID, the better the assessment score corresponding to the IS, and the higher the assessment score corresponding to the IS. The evaluation scores for the human skeleton motion sequences generated by other networks are listed in table 1, and it can be seen that the evaluation scores corresponding to the convolutional neural network are generally superior to those of other networks, such as recurrent neural networks (ERD and acLSTM), and antagonistic neural networks (HP-GAN and Two-Stage).

TABLE 1

According to the action sequence generation scheme provided by the embodiment of the disclosure, the target human skeleton action sequence with any length can be generated by the noise vector sequence, and the target human skeleton action sequence has any time step, so that the requirement of a user on the target human skeleton action sequence with any length can be met, and the partially-missing human skeleton action sequence can be supplemented or predicted. In addition, a convolutional neural network can be utilized in the process of generating the action sequence, so that the quality of the generated target human skeleton action sequence is improved, and the diversification of the target human skeleton action sequence is realized. Therefore, the target human skeleton action sequence generated by the embodiment of the disclosure can be applied to the aspect of driving a human three-dimensional model, so that a static human skeleton picture moves to obtain a human skeleton action video. In addition, the motion sequence generation scheme provided by the embodiment of the disclosure can be used for complementing or predicting the shielded and missing human skeleton motion to form a complete human skeleton motion.

It is understood that the above-mentioned method embodiments of the present disclosure can be combined with each other to form a combined embodiment without departing from the logic of the principle, which is limited by the space, and the detailed description of the present disclosure is omitted.

In addition, the present disclosure also provides an apparatus, an electronic device, a computer-readable storage medium, and a program, which can be used to implement any one of the action sequence generating methods provided by the present disclosure, and the descriptions and corresponding descriptions of the corresponding technical solutions and the corresponding descriptions in the method section are omitted for brevity.

It will be understood by those skilled in the art that in the method of the present invention, the order of writing the steps does not imply a strict order of execution and any limitations on the implementation, and the specific order of execution of the steps should be determined by their function and possible inherent logic.

Fig. 10 shows a block diagram of an action sequence generating apparatus according to an embodiment of the present disclosure, which includes, as shown in fig. 10:

an obtaining module 101, configured to obtain a first noise vector sequence generated through a plurality of random processes;

and the generating module 102 is configured to process the first noise vector sequence by using a convolutional neural network to generate a target human skeleton motion sequence.

In one or more optional embodiments, the apparatus further comprises:

In one or more optional embodiments, the obtaining module 101 is specifically configured to,

generating a noise vector by each of a plurality of random processes;

In one or more optional embodiments, the generating module 102 is configured to process the first noise vector sequence by using a convolutional neural network to obtain information of a plurality of human bone nodes, where the information of the plurality of human bone nodes includes location information of the plurality of human bone nodes, spatial correlation information of the plurality of human bone nodes, and displacement information of the plurality of human bone nodes at each of a plurality of time steps;

In one or more optional embodiments, the convolutional neural network is a graph convolutional neural network, and the generating module 102 is configured to,

generating an initial space-time map based on the noise vector sequence;

In one or more optional embodiments, the generating module 102 is configured to perform at least one stage of convolution processing on the initial space-time map by using a graph convolution neural network to generate a target human bone action sequence, where each stage of the at least one stage of convolution processing includes an upsampling processing and/or a convolution processing.

In one or more optional embodiments, the generating module 102 includes:

In one or more optional embodiments, the apparatus further comprises:

a first determining module, configured to determine target bone node information corresponding to each stage of processing in the at least one stage of convolution processing, where the target bone node information includes a number of a plurality of target bone nodes and an association relationship of the plurality of target bone nodes in space;

carrying out at least one stage of downsampling processing on a preset bone structure comprising a plurality of human body bone nodes to obtain a bone structure comprising one human body bone node, wherein the stage number of the downsampling processing of the at least one stage is the same as the stage number of the convolution processing of the at least one stage;

In one or more optional embodiments, the apparatus further comprises:

the second determination module is used for determining a target time step corresponding to the target human skeleton action sequence; determining a length of a noise vector generated by each of the plurality of random processes based on the target time step.

In some embodiments, functions of or modules included in the apparatus provided in the embodiments of the present disclosure may be used to execute the method described in the above method embodiments, and for specific implementation, reference may be made to the description of the above method embodiments, and for brevity, details are not described here again

An embodiment of the present disclosure further provides an electronic device, including: a processor; a memory for storing processor-executable instructions; wherein the processor is configured as the above method.

The electronic device may be provided as a terminal, server, or other form of device.

Fig. 11 is a block diagram illustrating an electronic device 1900 in accordance with an example embodiment. For example, the electronic device 1900 may be provided as a server. Referring to fig. 11, electronic device 1900 includes a processing component 1922 further including one or more processors and memory resources, represented by memory 1932, for storing instructions, e.g., applications, executable by processing component 1922. The application programs stored in memory 1932 may include one or more modules that each correspond to a set of instructions. Further, the processing component 1922 is configured to execute instructions to perform the above-described method.

The electronic device 1900 may also include a power component 1926 configured to perform power management of the electronic device 1900, a wired or wireless network interface 1950 configured to connect the electronic device 1900 to a network, and an input/output (I/O) interface 1958. The electronic device 1900 may operate based on an operating system stored in memory 1932, such as Windows Server, mac OS XTM, unixTM, linuxTM, freeBSDTM, or the like.

In an exemplary embodiment, a non-transitory computer readable storage medium, such as the memory 1932, is also provided that includes computer program instructions executable by the processing component 1922 of the electronic device 1900 to perform the above-described methods.

The present disclosure may be systems, methods, and/or computer program products. The computer program product may include a computer-readable storage medium having computer-readable program instructions embodied thereon for causing a processor to implement various aspects of the present disclosure.

The computer readable storage medium may be a tangible device that can hold and store the instructions for use by the instruction execution device. The computer readable storage medium may be, for example, but not limited to, an electronic memory device, a magnetic memory device, an optical memory device, an electromagnetic memory device, a semiconductor memory device, or any suitable combination of the foregoing. More specific examples (a non-exhaustive list) of the computer readable storage medium would include the following: a portable computer diskette, a hard disk, a Random Access Memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), a Static Random Access Memory (SRAM), a portable compact disc read-only memory (CD-ROM), a Digital Versatile Disc (DVD), a memory stick, a floppy disk, a mechanical coding device, such as punch cards or in-groove projection structures having instructions stored thereon, and any suitable combination of the foregoing. Computer-readable storage media as used herein is not to be construed as transitory signals per se, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through a waveguide or other transmission medium (e.g., optical pulses through a fiber optic cable), or electrical signals transmitted through electrical wires.

The computer-readable program instructions described herein may be downloaded from a computer-readable storage medium to a respective computing/processing device, or to an external computer or external storage device via a network, such as the internet, a local area network, a wide area network, and/or a wireless network. The network may include copper transmission cables, fiber optic transmission, wireless transmission, routers, firewalls, switches, gateway computers and/or edge servers. The network adapter card or network interface in each computing/processing device receives computer-readable program instructions from the network and forwards the computer-readable program instructions for storage in a computer-readable storage medium in the respective computing/processing device.

The computer program instructions for carrying out operations of the present disclosure may be assembler instructions, instruction Set Architecture (ISA) instructions, machine-related instructions, microcode, firmware instructions, state setting data, or source or object code written in any combination of one or more programming languages, including an object oriented programming language such as Smalltalk, C + + or the like and conventional procedural programming languages, such as the "C" programming language or similar programming languages. The computer readable program instructions may execute entirely on the user's computer, partly on the user's computer, as a stand-alone software package, partly on the user's computer and partly on a remote computer or entirely on the remote computer or server. In the case of a remote computer, the remote computer may be connected to the user's computer through any type of network, including a Local Area Network (LAN) or a Wide Area Network (WAN), or the connection may be made to an external computer (for example, through the Internet using an Internet service provider). In some embodiments, aspects of the disclosure are implemented by personalizing an electronic circuit, such as a programmable logic circuit, a Field Programmable Gate Array (FPGA), or a Programmable Logic Array (PLA), with state information of computer-readable program instructions, which can execute the computer-readable program instructions.

Various aspects of the present disclosure are described herein with reference to flowchart illustrations and/or block diagrams of methods, apparatus (systems) and computer program products according to embodiments of the disclosure. 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-readable program instructions.

These computer-readable program instructions may be provided to a processor of a general purpose computer, special purpose computer, 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, create means for implementing the functions/acts specified in the flowchart and/or block diagram block or blocks. These computer-readable program instructions may also be stored in a computer-readable storage medium that can direct a computer, programmable data processing apparatus, and/or other devices to function in a particular manner, such that the computer-readable medium storing the instructions comprises an article of manufacture including instructions which implement the function/act specified in the flowchart and/or block diagram block or blocks.

The computer readable program instructions may also be loaded onto a computer, other programmable data processing apparatus, or other devices to cause a series of operational steps to be performed on the computer, other programmable apparatus or other devices to produce a computer implemented process such that the instructions which execute on the computer, other programmable apparatus or other devices implement the functions/acts specified in the flowchart and/or block diagram block or blocks.

The flowchart and block diagrams in the figures illustrate the architecture, functionality, and operation of possible implementations of systems, methods and computer program products according to various embodiments of the present disclosure. In this regard, each block in the flowchart or block diagrams may represent a module, segment, or portion of instructions, which comprises one or more executable instructions for implementing the specified logical function(s). In some alternative implementations, the functions noted in the block may occur out of the order noted in the figures. For example, two blocks shown in succession may, in fact, be executed substantially concurrently, or the blocks may sometimes be executed in the reverse order, depending upon the functionality involved. It will also be noted that each block of the block diagrams and/or flowchart illustration, and combinations of blocks in the block diagrams and/or flowchart illustration, can be implemented by special purpose hardware-based systems which perform the specified functions or acts, or combinations of special purpose hardware and computer instructions.

Having described embodiments of the present disclosure, the foregoing description is intended to be exemplary, not exhaustive, and not limited to the disclosed embodiments. Many modifications and variations will be apparent to those of ordinary skill in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen in order to best explain the principles of the embodiments, the practical application, or technical improvements to the technology in the marketplace, or to enable others of ordinary skill in the art to understand the embodiments disclosed herein.

Claims

1. An action sequence generation method, comprising:

processing the first noise vector sequence by using a convolutional neural network to generate a target human skeleton action sequence;

wherein the processing the first noise vector sequence with a convolutional neural network to generate a target human skeleton motion sequence comprises:

processing the first noise vector sequence by using a convolutional neural network to obtain information of a plurality of human skeleton nodes;

2. The method of claim 1, wherein prior to obtaining the first sequence of noise vectors generated by the plurality of random processes, further comprising:

acquiring a known human skeleton action sequence;

3. The method of claim 1 or 2, wherein the obtaining a first sequence of noise vectors generated by a plurality of random processes comprises:

generating a noise vector by each of a plurality of random processes;

4. The method according to any one of claims 1 to 3, wherein the information of the plurality of human bone nodes comprises position information of the plurality of human bone nodes, spatial correlation information of the plurality of human bone nodes, and displacement information of the plurality of human bone nodes at each of a plurality of time steps.

5. The method of any one of claims 1 to 4, wherein the convolutional neural network is a graph convolutional neural network, and the processing the first noise vector sequence by using the convolutional neural network to generate a target human bone action sequence comprises:

generating an initial space-time diagram based on the first noise vector sequence, wherein the initial space-time diagram is used for representing position information and displacement information of human skeleton nodes in space and the length of a target human skeleton action sequence in time;

6. The method of claim 5 wherein said initial space-time map includes a human skeletal node.

7. The method according to claim 5 or 6, wherein the processing the initial space-time map by using a graph convolutional neural network to generate a target human skeletal action sequence comprises:

and performing at least one stage of convolution processing on the initial space-time diagram by using a diagram convolution neural network to generate a target human body skeleton action sequence, wherein each stage of processing in the at least one stage of convolution processing comprises upsampling processing and/or convolution processing.

8. The method according to claim 7, wherein the time-space diagram generated by performing at least one stage of convolution processing on the initial space-time diagram comprises a plurality of human bone nodes, and wherein different human bone nodes correspond to different weights in the at least one stage of convolution processing.

9. The method according to claim 7 or 8, wherein the upsampling process comprises:

performing upsampling processing on an input space-time diagram in at least one dimension of time and space to obtain a space-time diagram output by the upsampling processing, wherein the number of human skeleton nodes included in the output space-time diagram is greater than or equal to the number of human skeleton nodes included in the input space-time diagram, the number of time steps corresponding to the output space-time diagram is greater than or equal to the number of time steps corresponding to the input space-time diagram, and the input space-time diagram includes one of: the initial space-time map is generated by the up-sampling processing and/or convolution processing of the previous stage of the current up-sampling processing.

10. The method according to any one of claims 7 to 9, further comprising:

the method for performing at least one stage of convolution processing on the initial space-time map by using the graph convolution neural network to generate a target human skeleton action sequence comprises the following steps:

11. The method of claim 10, wherein said determining the target bone node information corresponding to each of said at least one level of convolution processing comprises:

12. The method according to any one of claims 1 to 11, further comprising:

determining a length of a noise vector generated by each of the plurality of stochastic processes based on the target time step.

13. The method according to any one of claims 1 to 12, wherein the stochastic process comprises a gaussian process.

14. An action sequence generating apparatus, comprising:

the generating module is used for processing the first noise vector sequence by utilizing a convolutional neural network to generate a target human skeleton action sequence;

the generation module is specifically configured to:

15. The apparatus of claim 14, further comprising:

16. The device according to claim 14 or 15, characterized in that the acquisition module, in particular for,

generating a noise vector by each of a plurality of random processes;

17. The apparatus according to any one of claims 14 to 16, wherein the information of the plurality of human bone nodes comprises position information of the plurality of human bone nodes, spatial correlation information of the plurality of human bone nodes, and displacement information of the plurality of human bone nodes at each of a plurality of time steps.

18. The apparatus according to any one of claims 14 to 17, wherein the convolutional neural network is a graph convolutional neural network, and the generating means is configured to,

19. The apparatus of claim 18 wherein said initial space-time map comprises a human skeletal node.

20. The apparatus of claim 18 or 19,

the generating module is used for performing at least one stage of convolution processing on the initial space-time map by using a map convolution neural network to generate a target human skeleton action sequence, wherein each stage of processing in the at least one stage of convolution processing comprises upsampling processing and/or convolution processing.

21. The apparatus according to claim 20, wherein the time-space diagram generated by performing at least one stage of convolution on the initial space-time diagram comprises a plurality of human bone nodes, and wherein different human bone nodes correspond to different weights in the at least one stage of convolution.

22. The apparatus of claim 20 or 21, wherein the generating module comprises:

an upsampling processing sub-module, configured to perform upsampling processing on an input space-time map in at least one dimension of time and space to obtain a space-time map output by the upsampling processing, where the number of human skeleton nodes included in the output space-time map is greater than or equal to the number of human skeleton nodes included in the input space-time map, the number of time steps corresponding to the output space-time map is greater than or equal to the number of time steps corresponding to the input space-time map, and the input space-time map includes one of: an initial space-time diagram, a space-time diagram generated by an upsampling process and/or a convolution process at a stage previous to the upsampling process.

23. The apparatus of any one of claims 20 to 22, further comprising:

24. The apparatus of claim 23, wherein the first determining module is configured to,

25. The apparatus of any one of claims 14 to 24, further comprising:

26. The apparatus of any one of claims 14 to 25, wherein the stochastic process comprises a gaussian process.

27. An electronic device, comprising:

a processor;

a memory for storing processor-executable instructions;

wherein the processor is configured to invoke the memory-stored instructions to perform the method of any one of claims 1 to 13.

28. A computer readable storage medium having computer program instructions stored thereon, which when executed by a processor implement the method of any one of claims 1 to 13.