KR20210076659A - Method and Apparatus for Action Recognition Using Sequential Feature Data - Google Patents

Abstract

Disclosed are an action recognition method using sequential feature data and an apparatus therefor. An action recognition learning method performed by a computing device including one or more processors and a memory for storing one or more programs executed by the processors according to an embodiment of the present invention includes: a natural language acquisition step of acquiring a natural language vector; a natural language processing step of generating natural language feature data including at least one feature value by inputting the natural language vector; a generation processing step of generating source feature data of a source image and target feature data for classification based on the natural language feature data; and a discrimination processing step of processing each classification of a sequence and a segment based on at least one of the source feature data, the natural language feature data, and the target feature data and performing the action recognition of an object. The present invention can recognize unlearned action through real images.

Description

Method and Apparatus for Action Recognition Using Sequential Feature Data

The present invention relates to a method for recognizing an action in an image using sequential sequence data and an apparatus therefor.

The content described in this section merely provides background information on the embodiments of the present invention and does not constitute the prior art.

In the case of moving picture data, in the case of moving image data, even if multiple feature vectors are extracted from one moving image, behavior is recognized by using an average vector for multiple extracted vectors in order to process it in a manner similar to an image.

In other words, in spite of the fact that the conventional zero-shot action recognition technology uses sequential data containing time series information, in order to apply a method similar to the zero-shot image classification study, the sequential extracted through a deep neural network Feature vectors were converted into averaged feature vectors and used for behavior recognition. However, by ignoring the time series included in the sequential feature vector, this method leads to an erroneous judgment result when there is a similar behavior in the intermediate process. For example, as shown in FIG. 1 , when behavior recognition is performed on a video for running behavior and jumping behavior, the time series of the feature vector 10 extracted from the images for each of running behavior and jumping behavior is ignored. If you do, a problem arises in which the running action and the jumping action are mistaken for the same action. That is, as shown in FIG. 1 , as the flow of the time series is lost, the characteristic data 10 cannot be accurately divided and generated, and an intermediate step may be incorrectly classified as a similar action.

The present invention provides a behavior recognition method using sequential feature data capable of recognizing unlearned behavior through an actual image by generating behavioral characteristic data for an image viewed for the first time based on a natural language vector and performing learning, and an apparatus therefor. Its main purpose is to provide

According to one aspect of the present invention, a behavior recognition learning method performed by a computing device comprising one or more processors and a memory for storing one or more programs executed by the processor for achieving the above object, obtaining a natural language vector natural language acquisition step; a natural language processing step of generating natural language feature data including at least one feature value by inputting a natural language vector; a generation processing step of generating source feature data of a source image and target feature data for classification based on the natural language feature data; and a discrimination processing step of processing classification for each of a sequence and a segment based on at least one of the source feature data, the natural language feature data, and the target feature data so that the behavior recognition of the object is performed. can do.

In addition, according to another aspect of the present invention, a behavior recognition apparatus for achieving the above object, one or more processors; and a memory storing one or more programs executed by the processor, wherein when the programs are executed by the one or more processors, a natural language acquisition step of acquiring, in the one or more processors, a natural language vector; a natural language processing step of generating natural language feature data including at least one feature value by inputting a natural language vector; a generation processing step of generating source feature data of a source image and target feature data for classification based on the natural language feature data; and a discrimination processing step of processing classification for each of a sequence and a segment based on at least one of the source feature data, the natural language feature data, and the target feature data so that the behavior recognition of the object is performed. actions can be performed.

In addition, according to another aspect of the present invention, a behavior recognition method performed by a computing device comprising one or more processors for achieving the above object and a memory for storing one or more programs executed by the processor is a source that has never been seen It receives an image, determines a behavior by applying a first learning result of learning sequence feature data and a second learning result of learning segment feature data to the source feature data of the source image data, and outputs the determined behavior recognition result can do.

As described above, the present invention has the effect of performing behavior recognition by generating sequential data on behavior based on a natural language vector.

In addition, the present invention has the effect of improving behavior recognition performance by generating behavioral feature data based on a natural language vector to recognize a behavior (new behavior) that has not been seen during learning.

1 is a view for explaining the problems of the prior art and the schematic operation characteristics of the present invention.
2 is a block diagram schematically illustrating a behavior recognition apparatus according to an embodiment of the present invention.
3 is a block diagram schematically illustrating an operation configuration for learning of a processor according to an embodiment of the present invention.
4 is a flowchart illustrating a learning method for behavior recognition according to an embodiment of the present invention.
5 is a block diagram schematically illustrating an operation configuration for behavior recognition of a processor according to an embodiment of the present invention.
6 is a flowchart illustrating a behavior recognition method according to an embodiment of the present invention.
7 is an exemplary diagram for explaining a learning operation of the behavior recognition apparatus according to an embodiment of the present invention.
8 is an exemplary diagram for explaining an operation of generating feature data by processing an input image according to an embodiment of the present invention.
9 is an exemplary diagram for explaining an operation of generating feature data by processing a natural language vector according to an embodiment of the present invention.
10 is a diagram illustrating an operation configuration of an encoder according to an embodiment of the present invention.
11 is a diagram illustrating an operation configuration of a discriminator according to an embodiment of the present invention.

Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings. In describing the present invention, if it is determined that a detailed description of a related known configuration or function may obscure the gist of the present invention, the detailed description thereof will be omitted. In addition, preferred embodiments of the present invention will be described below, but the technical spirit of the present invention is not limited thereto and may be variously implemented by those skilled in the art without being limited thereto. Hereinafter, a method for recognizing a behavior using sequential feature data proposed by the present invention and an apparatus therefor will be described in detail with reference to the drawings.

As shown in FIG. 1 , the present invention generates sequential feature data 20 rather than the average of feature vectors in order not to lose time series information of the original video, and through this, an apparatus for improving the performance of recognizing an action seen for the first time and methods are proposed.

2 is a block diagram schematically illustrating a behavior recognition apparatus according to an embodiment of the present invention.

The behavior recognition apparatus 100 according to the present embodiment includes an input unit 110 , an output unit 120 , a processor 130 , a memory 140 , and a database 150 . The behavior recognition apparatus 100 of FIG. 2 is according to an embodiment, and not all blocks shown in FIG. 2 are essential components, and in another embodiment, some blocks included in the behavior recognition apparatus 100 are added or changed. Or it can be deleted. Meanwhile, the behavior recognition apparatus 100 may be implemented as a computing device, and each component included in the behavior recognition apparatus 100 may be implemented as a separate software device, or as a separate hardware device combined with software. can

The behavior recognition apparatus 100 receives a natural language vector as an input, and generates target characteristic data through a generator by inputting natural language characteristic data generated by giving sequential information to the natural language vector as an input, and a source through at least two discriminators interworking with the generator Source feature data, natural language feature data, target feature data, etc. of an image (original video) are classified and processed to recognize a behavior in an image viewed for the first time.

The input unit 110 means a means for inputting or obtaining a signal or data for performing a behavior recognition operation in the behavior recognition apparatus 100 . The input unit 110 may interwork with the processor 130 to input various types of signals or data, or may obtain signals or data through interworking with an external device and transmit the signals or data to the processor 130 . Here, the input unit 110 may be implemented as a module for inputting a source image (original video), a natural language vector, a random variable, and the like, but is not limited thereto.

The output unit 120 may output various information such as a sequence learning result based on the feature data, a segment learning result based on the feature data, and a behavior recognition result in conjunction with the processor 130 . The output unit 120 may output various information through a display (not shown) provided in the behavior recognition apparatus 100 , but is not limited thereto, and may perform output in various forms.

The processor 130 performs a function of executing at least one instruction or program included in the memory 140 .

The processor 130 according to this embodiment performs machine learning based on the natural language vector and the source image obtained from the input unit 110 or the database 150, and the first image that is not previously learned based on the machine learning result Perform actions to recognize actions for

The processor 130 receives a source image and performs pre-processing based on the source image to generate source feature data. In addition, the processor 130 receives a natural language vector as an input, provides sequential information to the natural language vector to generate natural language feature data, and generates target feature data by receiving the natural language feature data as an input.

In addition, the processor 130 processes classification for each of a sequence and a segment based on at least one of source feature data, natural language feature data, and target feature data to perform behavior recognition of an object. Here, in order to perform behavior recognition in consideration of sequential features of the image, the processor 130 processes the classification of the sequence using the source feature data and the target feature data to generate a first learning result. In addition, in order to perform behavior recognition in consideration of the characteristics of the preset unit of the image, the processor 130 processes the classification of the segment using the object combination data combining the source characteristic data, the natural language characteristic data, and the target characteristic data. generate a second learning result. The processor 130 performs behavior recognition of an image that has not been input for the first time during learning based on the first learning result and the second learning result generated by processing the classification for each sequence and segment.

Detailed operations of the processor 130 according to the present embodiment will be described with reference to FIGS. 3 to 6 .

The memory 140 includes at least one instruction or program executable by the processor 130 . The memory 140 includes an operation of generating source feature data, an operation of generating natural language characteristic data, an operation of generating target characteristic data, an operation of generating target combined data, an operation of processing classification for a sequence, and classification of segments It may include an instruction or a program for an operation of processing and the like. In addition, the memory 140 may include a command or program for an operation to apply a learning result, an operation to perform behavior recognition, and the like.

The database 150 refers to a general data structure implemented in the storage space (hard disk or memory) of a computer system using a database management program (DBMS), and performs data search (extraction), deletion, editing, addition, etc. Relational database management system (RDBMS) such as Oracle, Infomix, Sybase, DB2, Gemston, Orion ), an object-oriented database management system (OODBMS) such as O2, and an XML Native Database such as Excelon, Tamino, Sekaiju, etc. It can be implemented according to the requirements, and has appropriate fields or elements to achieve its function.

The database 150 according to the present embodiment may store data related to behavior recognition and provide pre-stored data related to behavior recognition.

The data stored in the database 150 include source images, feature data (eg, source feature data, natural language feature data, target feature data, target combination data, etc.), learning results (eg, first learning results, second learning results, behaviors, etc.) recognition learning results, etc.), behavior recognition results, and the like. The database 140 is described as being implemented in the behavior recognition apparatus 100, but is not necessarily limited thereto, and may be implemented as a separate data storage device.

3 is a block diagram schematically illustrating an operation configuration for learning of a processor according to an embodiment of the present invention.

The processor 130 included in the behavior recognition apparatus 100 according to the present embodiment performs an operation of recognizing a behavior in an image viewed for the first time based on machine learning. Here, the machine learning is preferably learning using a generative adversarial network (GAN), but is not necessarily limited thereto.

The processor 130 included in the behavior recognition apparatus 100 receives a source image, performs pre-processing based on the source image to generate source feature data, receives input, a natural language vector, and sequential information is provided to the natural language vector. to generate natural language feature data, and a model that generates target feature data by inputting natural language feature data as an input, based on at least one of source feature data, natural language feature data, and target feature data. It enables an action to recognize a behavior that has never been seen based on a model that processes classification, etc., and can be mounted on any device that performs behavior recognition or can be linked with software that performs behavior recognition.

The processor 130 according to the present embodiment includes an image acquiring unit 310, a preprocessing unit 320, an image feature value processing unit 322, a natural language vector acquiring unit 330, an encoder 340, a first feature value processing unit ( 342 , a generator 350 , a second feature value processing unit 352 , and a discriminator 360 may be included. The processor 130 of FIG. 3 is according to an embodiment, and not all blocks shown in FIG. 3 are essential components, and in other embodiments, some blocks included in the processor 130 may be added, changed, or deleted. have. On the other hand, each component included in the processor 130 may be implemented as a separate software device, or may be implemented as a separate hardware device combined with software.

The image acquisition unit 310 performs an operation of acquiring a source image. Here, the source image refers to a video clip of a source video, and the video clip may be composed of a plurality of image segments. Here, the image segment includes a plurality of motion vector image frames. A difference image may be additionally included between motion vector image frames, and the difference image refers to an image generated through a difference between two adjacent motion vector image frames.

The preprocessor 320 generates source feature data for the source image by receiving the source image as an input. The source feature data generated by the preprocessor 320 includes feature values for each of a plurality of segment units.

The preprocessor 320 may generate source feature data by performing pre-training for learning a convolutional neural network (CNN) on the source image. Here, since the technology for pre-training is a commonly known technology, a detailed description thereof will be omitted.

The image feature value processor 322 transmits the source feature data output from the preprocessor 320 to the discriminator 360 . The image feature value processing unit 322 transmits the source feature data to the first discriminator 372 and the second discriminator 374 , respectively.

Meanwhile, the image feature value processor 322 may be omitted when the preprocessor 320 directly transmits the source feature data to the discriminator 360 , or may be implemented in a form included in the preprocessor 320 .

The natural language vector acquisition unit 330 acquires a natural language vector corresponding to a preset condition. Here, the natural language vector refers to a vector generated based on natural language for a predetermined action without including time-series information.

The encoder 340 receives a natural language vector as an input and generates natural language feature data including at least one feature value.

The encoder 340 adds sequential information to a natural language vector, expands it into a plurality of vectors, and generates each of at least one feature value corresponding to each of the plurality of vectors.

The encoder 340 generates each of at least one feature value having a distribution on a normal distribution of the natural language vector by using at least one of the mean, standard deviation, and noise of the natural language vector.

The encoder 340 expands a natural language vector into a plurality of vectors based on a recurrent neural network (RNN), and each of the plurality of vectors including sequential information may be generated based on a vector generated at a previous time. .

The first feature value processing unit 342 transmits the natural language feature data output from the encoder 340 to the generator 350 . The first feature value processing unit 342 may additionally combine a random variable (a random variable with respect to latent noise) to the natural language feature data and transmit it to the generator 350 .

Meanwhile, the first feature value processing unit 342 may be omitted when the natural language feature data is directly transmitted from the encoder 340 to the generator 350 , or may be implemented in a form included in the encoder 340 .

The generator 350 generates source feature data of a source image and target feature data for classification based on the natural language feature data.

The generator 350 generates target feature data for a fake image based on the natural language feature data and a pre-generated random variable. Here, it is preferable that the generator 350 generates the target characteristic data through convolutional neural network (CNN) learning, but is not limited thereto.

The generator 350 generates target feature data including at least one feature value. Here, the generator 350 generates the target feature data in the same number of segments as the natural language feature data. Here, the segment unit may be divided into respective feature values included in the target feature data.

The second feature value processing unit 352 transmits the target feature data output from the generator 350 to the discriminator 360 . The second feature value processing unit 352 transmits the target feature data to the first discriminator 372 and the second discriminator 374 , respectively.

Meanwhile, when the target feature data is directly transmitted from the generator 350 to the discriminator 360 , the second feature value processing unit 352 may be omitted or implemented in a form included in the generator 350 .

The discriminator 360 processes classification for each of a sequence and a segment based on at least one of source feature data, natural language feature data, target feature data, and the like, so that behavior recognition of an object is performed. The discriminator 360 according to the present embodiment includes a first discriminator 372 and a second discriminator 374 .

The first discriminator 372 performs an operation of processing classification for a sequence using the target feature data and the source feature data. The first discriminator 372 may receive the target characteristic data and the source characteristic data, and determine whether the target characteristic data is authentic or not.

Specifically, the first discriminator 372 compares source feature data combining a plurality of source feature values including sequential information with target feature data combining a plurality of target feature values including sequential information to determine the authenticity of the target feature data. Outputs the first learning result of learning whether or not.

The first discriminator 372 transmits feedback information to the generator 350 that generates target feature data based on the first learning result, and compares the source feature data with the target feature data so that the target feature data corresponds to a true signal. It is possible to learn whether the target feature data is authentic or not. Here, it is preferable that the first discriminator 372 performs learning based on a generative adversarial network (GAN) in order to classify the target feature data to correspond to the true signal in conjunction with the generator 350. It is not necessarily limited to this.

The second discriminator 374 performs an operation of processing classification for a segment using the target combination data and source feature data that are combined with the natural language feature data and the target feature data. The second discriminator 374 may receive the target combination data and the source feature data, and determine whether the target combination data is authentic or not.

Specifically, the second discriminator 374 outputs a second learning result obtained by comparing the segment unit of the source feature data and the segment unit of the target combined data to learn whether the object combined data is authentic. Here, the second discriminator 374 compares the segment unit data of the source feature data and the segment unit target combined data in which the feature value of the natural language feature data and the feature value of the target feature data are combined to process the classification of the segment. can

The second discriminator 374 transmits feedback information to the generator 350 that generates the target feature data based on the second learning result, and compares the source feature data with the target combined data so that the target combined data corresponds to a true signal. It is possible to learn the authenticity of the target binding data by iterating until Here, it is preferable that the second discriminator 374 performs learning based on a generative adversarial network (GAN) in order to classify the target binding data to correspond to the true signal in conjunction with the generator 350. It is not necessarily limited to this.

4 is a flowchart illustrating a learning method for behavior recognition according to an embodiment of the present invention.

The behavior recognition apparatus 100 checks whether a source image is input (S410).

When the source image is input in step S410, the behavior recognition apparatus 100 acquires the source image (S420). The behavior recognition apparatus 100 preprocesses the source image to generate a plurality of image feature values, and generates source feature data including the plurality of image feature values ( S430 ).

Meanwhile, when a natural language vector is input instead of a source image in step S410, the behavior recognition apparatus 100 acquires a natural language vector (S440).

The behavior recognition apparatus 100 generates natural language feature data including at least one feature value (a first feature value) by inputting a natural language vector ( S450 ).

In addition, the behavior recognition apparatus 100 includes, as an input, a feature value (a first feature value) included in the natural language feature data, and at least one feature value (second feature value) for classification with the source feature data of the source image. Target feature data is generated (S460).

The behavior recognition apparatus 100 generates a first learning result through classification of a sequence (a first discrimination process) using the target feature data and the source feature data (S470). Specifically, the behavior recognition apparatus 100 compares source feature data combining a plurality of source feature values including sequential information with target feature data combining a plurality of target feature values including sequential information to determine whether the target feature data is authentic or not. Outputs the first learning result learned.

In addition, the behavior recognition apparatus 100 performs a second learning result through processing (second discrimination processing) classification for a segment using the target combination data and the source characteristic data in which the natural language characteristic data and the target characteristic data are combined. generated (S480). Specifically, the behavior recognition device 100

By comparing the segment-unit data of the source feature data and the segment-unit object-combined data in which the feature values of the natural language feature data and the feature values of the target feature data are combined, the second learning result of learning whether the object-combined data is authentic or not is output.

Although it is described that each step is sequentially executed in FIG. 4 , the present invention is not limited thereto. In other words, since it may be applicable to changing and executing the steps described in FIG. 4 or executing one or more steps in parallel, FIG. 4 is not limited to a time-series order.

The behavior recognition learning method according to the present embodiment described in FIG. 4 may be implemented as an application (or program) and recorded in a recording medium readable by a terminal device (or computer). The recording medium in which the application (or program) for implementing the behavior recognition learning method according to the present embodiment is recorded and the terminal device (or computer) can read is any type of recording device in which data that can be read by the computing system is stored. or media.

5 is a block diagram schematically illustrating an operation configuration for behavior recognition of a processor according to an embodiment of the present invention.

The processor 130 included in the behavior recognition apparatus 100 according to the present embodiment includes an input image acquisition unit 510 , a neural network processing unit 520 , a learning result application unit 530 , an image determination unit 540 , and a result output. part 550 . The processor 130 of FIG. 5 is according to an embodiment, and not all blocks shown in FIG. 5 are essential components, and in other embodiments, some blocks included in the processor 130 may be added, changed, or deleted. have. On the other hand, each component included in the processor 130 may be implemented as a separate software device, or may be implemented as a separate hardware device combined with software.

The input image acquisition unit 510 acquires an unseen source image for behavior recognition. Here, the unseen source image means an image that is not input during learning for behavior recognition.

The neural network processing unit 520 generates source feature data by inputting the acquired source image. The neural network processing unit 520 may generate source feature data by performing preprocessing based on convolutional neural network (CNN) learning. Here, the source feature data may include a plurality of image feature values.

The learning result application unit 530 applies the first learning result of learning the sequence characteristic data and the second learning result of learning the segment characteristic data to the source characteristic data of the source image data, and the image determining unit 540 applies the applied learning It recognizes the behavior of the source video based on the result.

The result output unit 550 outputs a behavior recognition result based on the recognized behavior.

6 is a flowchart illustrating a behavior recognition method according to an embodiment of the present invention.

The behavior recognition apparatus 100 acquires an unseen source image for behavior recognition ( S610 ). Here, the unseen source image means an image that is not input during learning for behavior recognition.

The behavior recognition apparatus 100 receives the acquired source image as an input and performs neural network learning-based preprocessing to extract image feature values to generate source feature data (S620). The behavior recognition apparatus 100 may generate source feature data by performing preprocessing based on convolutional neural network (CNN) learning.

The behavior recognition apparatus 100 compares the feature values by applying the pre-learned learning result ( S630 ). Specifically, the behavior recognition apparatus 100 applies the first learning result of learning the sequence feature data and the second learning result of learning the segment feature data to the source feature data of the source image data, and compares the feature values.

The behavior recognition apparatus 100 determines the behavior of the source image (input image) based on the applied learning result (S640), and outputs the behavior recognition result based on the recognized behavior (S650).

Although it is described that each step is sequentially executed in FIG. 6 , the present invention is not limited thereto. In other words, since it may be applicable to changing and executing the steps described in FIG. 6 or executing one or more steps in parallel, FIG. 6 is not limited to a time-series order.

The behavior recognition method according to the present embodiment described in FIG. 6 may be implemented as an application (or program) and recorded in a terminal device (or computer) readable recording medium. A recording medium in which an application (or program) for implementing the behavior recognition method according to the present embodiment is recorded and a terminal device (or computer) readable recording medium is any type of recording device in which data that can be read by a computing system is stored or includes media.

7 is an exemplary diagram for explaining a learning operation of the behavior recognition apparatus according to an embodiment of the present invention.

When the video data cannot be used for the learning process or there is no video data for learning, the conventional behavior recognition device averages the features extracted from the video, and the image is inherited through ZSIC (Zero-shot Image Classification). It performs an action that recognizes the action through the method. However, since the conventional behavior recognition method recognizes the behavior by ignoring time-series sequential information of the video, a recognition error may occur for the entire behavior included in the video.

In order to solve this conventional problem, the behavior recognition apparatus 100 according to the present embodiment is capable of synthesizing a series of motions for a class that has not been seen, not a single sample, through a sequence generative model in consideration of sequential information, and , transforms the perception of first-time behavior into a fully supervised learning method.

The behavior recognition apparatus 100 according to the present embodiment may include an attribute encoder 340 , a generator 350 , a discriminator 360 , and the like to generate a sequence for recognizing a first-time behavior. Specifically, the attribute encoder 340 may convert a natural language vector into a plurality of vectors to provide sequential information to generate a sequence. In addition, when synthesizing the generated sequence, the sequence generative model of the behavior recognition apparatus 100 samples not only the segment of the behavior but also the entire sequence of behavior as an actual distribution through the sequence discriminator. Here, the behavior recognition apparatus 100 may be implemented as a Sequence Feature Generative Adversarial Network (SFGAN) based on sequential feature data.

The behavior recognition apparatus 100 includes a generative model for generating a feature sequence of a behavior, and this model generates a sequence under a single condition, and the generated sequence must be realistic.

The behavior recognition apparatus 100 includes an attribute encoder 340 based on a recursive neural network to search a semantic embedding space containing temporal information and develop a sequence cue in a condition. In addition, the behavior recognition apparatus 100 includes a sequence discriminator for applying a penalty to a generator ignoring the sequence of behavior. In addition, the behavior recognition apparatus 100 may generate a characteristic of the first seen behavior corresponding to a time condition of zero-shot learning (ZSL).

The behavior recognition apparatus 100 according to the present invention generates an unseen behavior in order to convert the existing semi-supervised learning into a fully-supervised learning.

The behavior recognition apparatus 100 applies a sequence-based generative adversarial network (GAN) model that generates sequences for behavioral features, unlike the conventional zero-shot behavior recognition method that generates averaged features, , including an attribute encoder 340 , a generator 350 , a discriminator 360 , and the like for processing sequential data.

와 같이 표현될 수 있다. 여기서, x_v는

의 RGB 시각적 특징이고, x_f는

의 광학 흐름 특징이고, y는 Y^s의 클레스 라벨(Class label)을 나타내며, c(y)는 클레스의 의미를 의미론적으로 나타낸 클레스 y의 자연어 임베딩을 의미한다. The data set for the class seen in this embodiment can be defined as Ds, and the data set Ds seen is

can be expressed as where x _v is

is the RGB visual feature of , and x _f is

is the optical flow characteristic of , y denotes ^{the class label of Y s} , and c(y) denotes the natural language embedding of class y semantically indicating the meaning of the class.

와 같이 표현될 수 있다.Similarly, a data set for a class not seen in this embodiment may be defined as Du, and Du is separated ^{from Y s .} The data set Du, which is not seen in this example, is

can be expressed as

및

를 만족하도록 설정된다. In the behavior recognition apparatus 100 according to the present embodiment, based on a constraint for recognizing an action seen for the first time (ZSAR: Zero-shot Action Recognition), there are two data sets (Ds) and unseen (Du) The containment relationship between the data sets is

and

is set to satisfy

A sequence of actions can be expressed as a feature vector length of N, where N means the temporal length of the sequence. The seen data can be accessed at the learning stage for behavior recognition, but the RGB characteristics and flow characteristics of the unseen data can only be accessed at the test stage.

Hereinafter, a generative hostile learning operation (GAN for Zero-shot Action Recognition) for recognizing a first-time behavior used in the behavior recognition apparatus 100 according to the present embodiment will be described.

A generative adversarial network (GAN) applied to the behavior recognition device 100 is a sample from an actual distribution through a minimum maximization algorithm between a generator 350 and a discriminator 360. aims to create Here, the generator 350 generates a fake sample to deceive the discriminator 360 , while the discriminator 360 tries to distinguish the real sample from the fake sample.

In addition, the behavior recognition apparatus 100 according to the present embodiment adjusts the Wasserstein distance as an objective function with a gradient penalty for learning stability of the generative adversarial neural network. In order to generate a sample from a class that has not been seen in the behavior recognition apparatus 100 , a generative model is generated based on a conditional Wasserstein GAN (WGAN).

The objective function used in the behavior recognition apparatus 100 may be defined as in Equation (1).

는 생성자(350)의 출력을 의미하고,

는 x와

의 보간을 의미하며, 마지막 항은 페널티를 주어 그라디언트의 폭발(Gradient Exploding)하는 것을 방지하는 정규화항이며, γ는 항의 매개 변수를 의미한다. where P _r and P _g mean the actual distribution and the generated distribution,

means the output of the constructor 350,

is x and

means the interpolation of , and the last term is a normalization term that prevents gradient exploding by giving a penalty, and γ refers to the parameter of the term.

Hereinafter, an operation of generating an action sequence that has not been seen in the action recognition apparatus 100 according to the present embodiment (Generating Unseen Action Sequence) will be described.

Generating a video for behavioral recognition is a more difficult operation than generating a single frame. Since video is more complex with the time axis, the gaps between each segment must be connected naturally when the generated segments are assembled to complete a sequence of motions.

Accordingly, the behavior recognition apparatus 100 according to the present embodiment generates a video feature sequence of an unseen class based on two conditions. The first condition is to generate a sequence from a single condition, and the second condition is to create a sequence by combining a plurality of features to ensure the fidelity of the sequence. Here, the single condition preferably means one natural language vector, but is not limited thereto.

When only a single condition is provided for generating a feature sequence of a video that the behavior recognition apparatus 100 has not seen, the generator 350 may synthesize the sequence using two methods. Here, the two methods may be a one-to-many mapping that simply creates an entire function from a single condition, and a one-to-one mapping that expands to a plurality of conditions of an expected length before generating a given condition. Here, referring to the operation of generating an actual video, when performing one-to-many mapping, it is difficult to generate a feature sequence of an unseen video due to insufficient conditions and network capacity. Accordingly, the behavior recognition apparatus 100 of the present invention includes an attribute encoder 340 capable of developing time information under a single condition through a recurrent neural network (RNN). That is, the semantic embedding space including time information is searched through the attribute encoder 340 .

Next, the behavior recognition apparatus 100 must ensure the fidelity of the feature sequence of the generated unseen video. In the behavior recognition apparatus 100 , a single condition is expanded to a plurality of conditions, a segment is generated in the expanded condition, and a behavior sequence is generated by collecting the generated segments. The flow of the entire generated action sequence must be connected as naturally as the actual action sequence.

Therefore, generative models must simultaneously explore the distribution of segments and sequences in visual space. To this end, the discriminator 360 of the behavior recognition apparatus 100 of the present invention includes a sequence discriminator 372 for discriminating an actual behavior sequence from a fake sequence.

Hereinafter, the operation of the action feature-based generative adversarial networks applied in the behavior recognition apparatus 100 according to the present embodiment will be described.

7 shows a detailed structure of a sequence feature generative adversarial network (SFGAN) based on a behavior sequence feature applied to the behavior recognition apparatus 100 according to the present invention. Referring to FIG. 7 , the behavior recognition apparatus 100 recognizes an unseen behavior through a generative adversarial neural network based on a behavior sequence feature composed of an encoder 340 , a generator 350 , and a discriminator 372 , 374 , and the like. learn to do

Hereinafter, the attribute encoder 340 included in the behavior recognition apparatus 100 according to the present embodiment will be described.

The encoder 340 encodes the input single condition and outputs an output value c(y). Here, the encoder 340 uses a recursive neural network to solve a time stream of a single input condition. For example, the encoder 340 may receive a natural language vector as a single condition, encode the natural language vector, and output natural language feature data.

In addition, the encoder 340 may be configured as a Gated Recurrent Unit (GRU) cell, and the GRU operation of the encoder 340 may be defined as in Equation (2).

= c(y)이고 k 는 0 < k < N, k ∈ N을 만족시킵니다.here

= c(y) and k satisfies 0 < k < N, k ∈ N.

The encoder 340 generates a discontinuity in the latent space by the operation of solving from a single condition to a plurality of conditions. Accordingly, the encoder 340 further uses Conditioning Augmentation.

에서 표본으로 다시 매개 변수화되며, 여기서 μ는 평균을 의미하고, Σ는 공분산 행렬을 의미한다. Each condition extended to a number of conditions is a Gaussian distribution

is parameterized back to the sample in , where μ stands for the mean and Σ stands for the covariance matrix.

The encoder 340 uses KL-divergence (Kullback-Leibler divergence) as a normalization term to prevent excessively adjusting the semantic space and enhancing smoothness.

에서 매개 변수화된 조건

는 생성자(350)으로 전달되어 생성자(350)의 입력 조건의 역할을 한다. Accordingly, as shown in FIG. 10 , the encoder 340 is

parameterized condition in

is passed to the constructor 350 and serves as an input condition of the constructor 350 .

는 행동 사이의 관계정보를 포함해야 한다. 이에, 본 발명의 인코더(340)는 삼중항 손실함수를 사용하며, 삼중항 손실 함수는 GRU에 의해 처리된 조건을 원래 조건과 유사하게 처리하고 다른 행동의 조건과는 다르게 처리한다. 인코더(340)에 삼중항 손실함수에서 사용되는 목적 함수 및 정규화 용어는 수학식 3 및 4와 같이 정의될 수 있다. Also, in order to create a feature never seen in the constructor 350, the condition

should contain information about the relationship between the behaviors. Accordingly, the encoder 340 of the present invention uses a triplet loss function, and the triplet loss function treats the condition processed by the GRU similarly to the original condition and differently from the conditions of other actions. The objective function and normalization term used in the triplet loss function in the encoder 340 may be defined as in Equations 3 and 4.

,

각각은 앵커(anchor), 파지티브 샘플 및 네거티브 샘플이다. m 은 삼중항 손실의 마진이며, 코사인 유사성을 삼중항 손실 거리 측정법으로 사용한다.

는 동일한 클립의 피처에서 샘플링되고 네거티브는 다른 동작의 클립에서 샘플링된다.Here, d ⁺ means the distance of the positive pair, d ^- means the distance of the negative pair, c(y),

,

Each is an anchor, a positive sample and a negative sample. m is the margin of triplet loss, using cosine similarity as a measure of triplet loss distance.

is sampled from features in the same clip and negatives are sampled from clips in different motions.

Hereinafter, the generator 350 included in the behavior recognition apparatus 100 according to the present embodiment will be described.

The behavior recognition apparatus 100 according to the present embodiment recognizes a behavior through a complete supervised learning method, and it is preferable to use an optical flow characteristic in this method.

The behavior recognition apparatus 100 includes a generator 350 for behavior recognition that has not been seen before, and the generator 350 generates a combined feature in which an RGB feature and a flow feature are combined.

The generator 350 generates a combined feature in which the RGB feature and the flow feature are combined with the ^{parameterized condition a t and the latent noise vector z as inputs.}

As the flow features are extracted from the original RGB vision, the generator 350 is constructed with fully connected layers to model the relationship between the RGB features and the flow features. The operation of the generator 350 may be defined as in Equation 5.

where z is a random variable for latent noise, and n is the nth embedded parameterized condition.

Hereinafter, the discriminator 360 included in the behavior recognition apparatus 100 according to the present embodiment will be described.

The discriminator 360 included in the behavior recognition apparatus 100 determines a difference between the distribution of the feature generated by the generator 350 and the actual distribution, and provides feedback to the generator 350 .

As shown in FIG. 11 , the discriminator 360 according to the present embodiment may include a segment discriminator 372 for determining a segment and a sequence discriminator 374 for discriminating a sequence.

Each of the segment discriminator 372 and the sequence discriminator 374 may consist of a plurality of fully connected layers for distinguishing real features and real sequences from fakes.

The segment discriminator 372 simultaneously processes the feature and the condition, and the sequence discriminator 374 processes only the feature.

Since the behavior recognition apparatus 100 according to the present embodiment generates an unseen behavior sequence, a class bias may occur due to excessive conditioning during training. Accordingly, the behavior recognition apparatus 100 should be configured as a discriminator 360 including a sequence discriminator 374 .

Hereinafter, an objective function used in the behavior recognition apparatus 100 according to the present embodiment will be described.

The objective function for model learning of the behavior recognition apparatus 100 according to the present embodiment is based on a conditional Wasserstain GAN. The sequence discriminator 374 included in the behavior recognition device 100 is unconditionally designed to generate an unseen behavior sequence, and the sequence discriminator 374 uses a general Wasserstain distance. Also, a gradient penalty for the sequence discriminator 374 is defined as in Equation 6.

Here, R _uncond denotes unconditional normalization for the sequence discriminator 374 , and R _cond denotes conditional normalization for the segment discriminator 372 . Therefore, the loss function for the generative model can be defined as Equation (7).

은 생성자(350)에서 생성된 대상 특징 데이터를 나타내며,

이고,

이다. R_uncond 및 R_cond는 각각 Dseq 및 Dseg에 대한 정규화 용어를 의미한다. Here, 0 ≤ n < N, a ⁿ represents an encoded condition (feature data), and x is a sample of the actual feature data. Also,

represents the target feature data generated by the generator 350,

ego,

to be. R _uncond and R _cond refer to normalization terms for Dseq and Dseg, respectively.

As a result, the overall objective function of the parameterized end-to-end model used in the behavior recognition apparatus 100 may be defined by Equation (8).

Hereinafter, an operation for recognizing a behavior that has not been seen in the behavior recognition apparatus 100 according to the present embodiment will be described.

을 생성한다. The behavior recognition device 100 learns the seen data set Ds through a generative adversarial neural network, and then learns the behavior characteristic from the unseen class condition.

create

The behavior recognition apparatus 100 processes a problem for recognizing a behavior seen for the first time in a fully supervised learning method for behavior recognition, and uses a multi-layer perceptron classier for evaluation. Here, the classifier is optimized by minimizing the negative log-likelihood loss, and may be defined as Equation (9).

또는 Du를 의미한다. 또한, 분류를 위한 예측 함수는 수학 식 10과 같이 정의될 수 있다. Here, θ is the weight of the fully connected layer in the classifier, and F is GZSL or ZSL when

or Du. Also, the prediction function for classification may be defined as in Equation 10.

이며, GZSL에서 y ∈ Y^s ∪ Y^u, ZSL에서 y ∈ Y^u를 의미한다. where the softmax function is

And, means for ^{y ∈ Y s ∪ Y u,} y ∈ Y u in ZSL in GZSL.

8 is an exemplary diagram for explaining an operation of generating feature data by processing an input image according to an embodiment of the present invention.

Referring to FIG. 8 , the behavior recognition apparatus 100 acquires a source image. Here, the source image means a video clip 810 , and the video clip 810 may include five image segments 811 , 812 , 813 , 814 , and 815 . Here, the video clip 810 may be a clip including an action for basketball, baseball, exit, and the like. Each of the image segments 811 , 812 , 813 , 814 , and 815 may include 32 motion vector image frames. A difference image 821 may be additionally included between the motion vector image frames 820 , and the difference image 821 means an image generated through a difference between two adjacent motion vector image frames 820 .

Referring to FIG. 8 , the behavior recognition apparatus 100 may generate source feature data X by performing pre-training for learning a convolutional neural network (CNN). Here, the source feature data includes feature values 831, 832, 833, 834, and 835 for each segment unit, and each feature value may be a matrix feature value having a size of 1×1024.

9 is an exemplary diagram for explaining an operation of generating feature data by processing a natural language vector according to an embodiment of the present invention.

The encoder 340 receives the natural language vector 910 as an input and generates natural language feature data including at least one feature value. Here, the natural language vector refers to a vector generated based on natural language for a predetermined action without including time-series information.

The encoder 340 adds sequential information to the natural language vector 910 to expand it into a plurality of vectors, and generates at least one feature value 921 , 922 , 923 , 924 , 925 corresponding to each of the plurality of vectors, respectively. .

The encoder 340 generates at least one feature value 921 , 922 , 923 , 924 , 925 having a distribution on a normal distribution of the natural language vector using at least one of the mean, standard deviation, and noise of the natural language vector, respectively. .

The first feature value processing unit 342 transmits the natural language feature data including at least one feature value 921 , 922 , 923 , 924 , and 925 output from the encoder 340 to the generator 350 . . The first feature value processing unit 342 may additionally combine a random variable (a random variable with respect to latent noise) to the natural language feature data and transmit it to the generator 350 .

Also, the first feature value processing unit 342 transmits natural language feature data including at least one feature value 921 , 922 , 923 , 924 , and 925 output from the encoder 340 to the second discriminator 374 . do.

Meanwhile, the first feature value processing unit 342 may be omitted when the natural language feature data is directly transmitted from the encoder 340 to the generator 350 , or may be implemented in a form included in the encoder 340 .

The generator 350 generates source feature data of a source image and target feature data for classification based on the natural language feature data.

The generator 350 generates target feature data for a fake image based on the natural language feature data and a pre-generated random variable. Here, it is preferable that the generator 350 generates the target characteristic data through convolutional neural network (CNN) learning, but is not limited thereto.

The generator 350 generates target feature data including at least one feature value 931 , 932 , 933 , 934 , and 935 . Here, the generator 350 generates the target feature data in the same number of segments as the natural language feature data. Here, the segment unit may be divided into respective feature values included in the target feature data.

The second feature value processing unit 352 transmits the target feature data output from the generator 350 to the discriminator 360 . The second feature value processing unit 352 transmits the target feature data to the first discriminator 372 and the second discriminator 374 , respectively. Meanwhile, when the target feature data is directly transmitted from the generator 350 to the discriminator 360 , the second feature value processing unit 352 may be omitted or implemented in a form included in the generator 350 .

10 is a diagram illustrating an operation configuration of an encoder according to an embodiment of the present invention.

The encoder 340 provides sequential information to the natural language vector to perform expansion into a plurality of vectors. Here, the natural language vector of a single condition may be extended using a method such as a long short-term memory (LSTM), a gated recurrent unit (GRU), or the like.

Also, the encoder 340 generates each of at least one feature value corresponding to each of the plurality of vectors. The encoder 340 uses at least one of the mean (μ), standard deviation (σ), and noise (ε) of the natural language vector to convert the natural language vector to at least one feature value (a ⁱ ) having a distribution on a normal distribution, respectively. create

11 is a diagram illustrating an operation configuration of a discriminator according to an embodiment of the present invention.

The discriminator 360 processes classification for each of a sequence and a segment based on at least one of source feature data, natural language feature data, target feature data, and the like, so that behavior recognition of an object is performed. The discriminator 360 according to the present embodiment includes a first discriminator 372 and a second discriminator 374 .

The first discriminator 372 performs an operation of processing classification for a sequence using the target feature data and the source feature data. The first discriminator 372 may receive the target characteristic data and the source characteristic data, and determine whether the target characteristic data is authentic or not. Specifically, the first discriminator 372 compares source feature data obtained by concatenating a plurality of source feature values including sequential information with target feature data obtained by combining a plurality of target feature values including sequential information to obtain a target A first learning result obtained by learning whether the feature data is authentic or not is output. Here, the first learning result may be expressed as a value between [0, 1]. As a result of determining the authenticity of the target feature data by the first discriminator 372, the closer to a value of 0, the more it is classified as a fake signal, and the closer it is to a value of 1, the more it is classified as a real signal.

The second discriminator 374 performs an operation of processing classification for a segment using the target combination data and source feature data that are combined with the natural language feature data and the target feature data. The second discriminator 374 may receive the target combination data and the source feature data, and determine whether the target combination data is authentic or not.

Specifically, the second discriminator 374 outputs a second learning result obtained by comparing the segment unit of the source feature data and the segment unit of the target combined data to learn whether the object combined data is authentic. Here, the second learning result may be expressed as a value between [0, 1]. As a result of the determination of the authenticity of the target binding data by the second discriminator 374, the closer to a value of 0, the more it is classified as a fake signal, and the closer it is to a value of 1, the more it is classified as a real signal.

The second discriminator 374 compares the segment-unit data of the source feature data and the segment-unit target-combined data in which the feature value of the natural language feature data and the feature value of the target feature data are combined (Concatenation) for the segment. classification can be handled.

The above description is merely illustrative of the technical idea of the embodiment of the present invention, and those of ordinary skill in the art to which the embodiment of the present invention pertains may make various modifications and changes within the scope not departing from the essential characteristics of the embodiment of the present invention. transformation will be possible. Accordingly, the embodiments of the present invention are not intended to limit the technical spirit of the embodiment of the present invention, but to explain, and the scope of the technical spirit of the embodiment of the present invention is not limited by these embodiments. The protection scope of the embodiment of the present invention should be interpreted by the following claims, and all technical ideas within the scope equivalent thereto should be construed as being included in the scope of the embodiment of the present invention.

100: behavior recognition device
110: input unit 120: output unit
130: processor 140: memory
150: database
310: image acquisition unit 320: pre-processing unit
322: image feature value processing unit 330: natural language vector acquisition unit
340: encoder 342: first feature value processing unit
350: constructor 352: second feature value processing unit
360: Discriminant

Claims (15)

A behavior recognition learning method performed by a computing device comprising one or more processors and a memory storing one or more programs to be executed by the processor, the computing device comprising:
a natural language acquisition step of acquiring a natural language vector;
a natural language processing step of generating natural language feature data including at least one feature value by inputting a natural language vector;
a generation processing step of generating source feature data of a source image and target feature data for classification based on the natural language feature data; and
A discrimination processing step of processing classification for each of a sequence and a segment based on at least one of the source feature data, the natural language feature data, and the target feature data so that the behavior recognition of the object is performed
Behavior recognition learning method, characterized in that performing. According to claim 1,
The natural language processing step is
Behavior recognition learning method, characterized in that by giving sequential information to the natural language vector, expanding it into a plurality of vectors, and generating each of the at least one feature value corresponding to each of the plurality of vectors. 3. The method of claim 2,
The natural language processing step is
Behavior recognition learning method, characterized in that by using at least one of the mean, standard deviation, and noise of the natural language vector to generate each of the at least one feature value having a distribution on a normal distribution of the natural language vector. 3. The method of claim 2,
The natural language processing step is
Behavior recognition learning method, characterized in that the natural language vector is expanded to the plurality of vectors based on a recursive neural network, and each of the plurality of vectors including the sequential information is generated based on a vector generated at a previous time. According to claim 1,
The generating process step is
Behavior recognition learning method, characterized in that the target feature data for a fake image is generated based on the natural language feature data and a pre-generated random variable. 6. The method of claim 5,
The generating process step is
Behavior recognition learning method, characterized in that generating the target characteristic data through convolutional neural network (CNN) learning. 6. The method of claim 5,
The generating process step is
Behavior recognition learning method, characterized in that generating the target feature data in the same number of segments as the natural language feature data. According to claim 1,
The discrimination processing step is
a first discrimination processing step of processing classification for a sequence using the target feature data and the source feature data; and
A second differentiation processing step of processing classification of a segment using the target combination data obtained by combining the natural language characteristic data and the target characteristic data and the source characteristic data
Behavior recognition learning method comprising a. 9. The method of claim 8,
The first discrimination processing step is
A first learning result obtained by learning whether the target feature data is authentic by comparing the source feature data in which a plurality of source feature values including sequential information are combined with the target feature data in which a plurality of target feature values including sequential information are combined Behavior recognition learning method, characterized in that outputting. 10. The method of claim 9,
The first discrimination processing step is
Feedback information is transmitted to the step of generating the target characteristic data based on the first learning result, and the source characteristic data is compared with the target characteristic data and repeated until the target characteristic data corresponds to a true signal. A behavior recognition learning method, characterized in that it learns whether the target feature data is authentic or not. 9. The method of claim 8,
The second discrimination processing step is
Behavior recognition learning method, characterized in that by comparing the segment unit of the source feature data and the segment unit of the target combined data, and outputting a second learning result of learning whether the object combined data is authentic. 12. The method of claim 11,
The second discrimination processing step is
Feedback information is transmitted to the step of generating the target characteristic data based on the second learning result, and the source characteristic data is compared with the target combination data and repeated until the target combination data corresponds to a true signal. A behavior recognition learning method, characterized in that it learns whether the object binding data is authentic or not. 9. The method of claim 8,
The second discrimination processing step is
Behavior recognition, characterized in that the classification of the segment is processed using the segment-unit data of the source feature data, the segment-by-segment combination data obtained by combining a feature value of the natural language feature data and a feature value of the target feature data How to learn. As a device for recognizing unseen behavior,
one or more processors; and
a memory storing one or more programs executed by the processor, wherein the programs, when executed by the one or more processors, in the one or more processors;
a natural language acquisition step of acquiring a natural language vector;
a natural language processing step of generating natural language feature data including at least one feature value by inputting a natural language vector;
a generation processing step of generating source feature data of a source image and target feature data for classification based on the natural language feature data; and
A discrimination processing step of processing classification for each of a sequence and a segment based on at least one of the source feature data, the natural language feature data, and the target feature data so that the behavior recognition of the object is performed
Behavior recognition device, characterized in that it performs operations comprising a. A behavior recognition method performed by a computing device comprising one or more processors and a memory storing one or more programs executed by the processors, the method comprising:
The computing device is
Receives a source image that has never been seen as an input, determines a behavior by applying a first learning result of learning sequence feature data and a second learning result of learning segment feature data to the source feature data of the source image data, and recognizes the determined behavior A behavior recognition method comprising outputting a result.

Images