KR20230080804A - Apparatus and method for estimating human pose based AI - Google Patents

Abstract

An apparatus and method for estimating human pose based AI are provided. An extraction part extracts the area containing a person (hereinafter referred to as an 'original person image') from an original image. A scaling part scales the original person image extracted from the extraction part to have different resolutions and generates multiple target person images. A posture estimation part can estimate a human pose by inputting multiple target person images generated by the scaling part into different networks of artificial intelligence models previously trained to estimate the human pose.

Description

Apparatus and method for estimating human pose based AI}

The present invention relates to an artificial intelligence-based human posture estimation apparatus and method, and more particularly, to an artificial intelligence-based human posture estimation apparatus and method capable of estimating a human posture by inputting multi-resolution images to an artificial intelligence model. It is about.

2D human pose estimation has received a lot of attention in the field of vision, and its performance has been greatly improved by the recent introduction of artificial intelligence-based deep learning. Human posture estimation is used in various fields such as surveillance systems and autonomous driving, and helps in recognizing actions and gestures. MS COCO and MPII datasets, which are the most used posture estimation datasets, contain a lot of high-resolution image training data. Accordingly, the existing artificial intelligence-based human posture extraction model has excellent performance in high-resolution images, but has a problem in that its performance rapidly deteriorates when a person is photographed at a small distance or when the resolution of the photographed image is low.

The technique proposed in the preceding paper (Sun, Ke, et al. "Deep high-resolution representation learning for human pose estimation." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2019.) The region is converted to a fixed image size, and it is learned by inputting it to the model, and the input image size is set to what works well on average in the test dataset. However, there is a problem of showing low performance for small-sized images in the test dataset.

In addition, the technique proposed in Korean Patent Publication No. 10-2019-0034380 includes a process of estimating a person's posture from an input image and correcting the motion, but the part that enables robust operation for low-resolution images not considering

However, for practical application in the industrial field, it is useful to estimate the posture of a nearby person and at the same time estimating the posture of a distant person. Therefore, a model capable of detecting a posture in a low-resolution image is essential.

In addition, the existing technology converts images of all sizes into one fixed size and inputs them to an artificial intelligence learning model. In this process, artifacts occur when up-scaling a low-resolution image, resulting in low posture estimation accuracy, and loss of information when reducing the size of a high-resolution image.

Korean Patent Publication No. 10-2019-0034380

In order to solve the above problems, the technical problem to be achieved by the present invention is an artificial intelligence-based human posture that operates robustly in images of various sizes and can estimate a human posture in an image including a small-sized person or a low-resolution image. It is to present an estimation device and method.

The problems of the present invention are not limited to those mentioned above, and other problems not mentioned will be clearly understood by those skilled in the art from the description below.

As a means for solving the above technical problem, according to an embodiment of the present invention, an artificial intelligence-based human posture estimation apparatus extracts a region including a person from an original image (hereinafter, referred to as 'original human image'). an extraction unit to; a scaling unit configured to generate a plurality of target human images by scaling the original human image extracted by the extraction unit to have different resolutions; and a posture estimator for estimating a human posture by inputting a plurality of target human images generated by the scaling unit to different networks of pre-learned artificial intelligence models to estimate a human posture.

The scaling unit may generate the plurality of target human images based on an input image resolution set in the artificial intelligence model.

A guiding channel adding unit which selects one or more of the plurality of target human images generated by the scaling unit and adds a guiding channel meaning a reference image, wherein the posture estimating unit performs the added guiding A human posture may be estimated by recognizing and referring to one or more target human images similar to the original human image based on the channel.

The posture estimator, when the plurality of target human images are input to a predetermined stem network among a plurality of stem networks built on an artificial intelligence model, generates a feature map while downscaling the resolution of the plurality of target human images, The generated feature maps may be input to a corresponding subnetwork among a plurality of subnetworks, and a human posture may be estimated by fusion while selectively changing or maintaining the resolution of the feature maps.

Meanwhile, according to another embodiment of the present invention, an artificial intelligence-based human posture estimation method includes (A) an electronic device extracting a region including a person from an original image (hereinafter, referred to as 'original human image') step; (B) generating, by the electronic device, a plurality of target human images by scaling the original human image extracted in step (A) to have different resolutions; and (C) estimating, by the electronic device, a human posture by inputting the plurality of target human images generated in the step (B) to different networks of pre-learned artificial intelligence models for human posture estimation. can include

In the step (B), the plurality of target human images may be generated based on the input image resolution set in the artificial intelligence model.

After the step (B), (D) selecting, by the electronic device, one or more of the plurality of target human images generated in the step (B) and adding a guiding channel indicating a reference image; In the step (C), the human posture may be estimated by recognizing and referring to one or more target human images similar to the original human image based on the guiding channel added in the step (D).

The step (C) may include: (C1) inputting the plurality of target human images to a predetermined stem network among a plurality of stem networks built on an artificial intelligence model; (C2) each of the plurality of stem networks generating feature maps while downscaling the resolution of a target human image, and inputting the generated feature maps to a corresponding subnetwork among a plurality of subnetworks; and (C3) estimating a human posture by fusing the plurality of sub-networks while selectively changing or maintaining the resolution of the feature maps input in the step (C2).

According to the present invention, it is possible to provide excellent posture estimation performance by operating robustly in images of various resolutions, and in particular, to provide high posture estimation performance even in an image including a small-sized person or a low-resolution image.

In addition, according to the present invention, 2D human posture estimation performance can be improved through the fusion of multi-scale input features and a guiding channel without the addition of a Super Resolution module, which is computationally expensive.

The effects of the present invention are not limited to those mentioned above, and other effects not mentioned will be clearly understood by those skilled in the art from the description below.

1 is a block diagram showing an artificial intelligence-based human posture estimation apparatus 100 according to an embodiment of the present invention;
2 is a diagram showing an operation in which multi-resolution input and guiding channels are input to a network of an artificial intelligence model and output features are output;
3 is a diagram showing a framework of HRNet according to an embodiment of the present invention;
4 is an exemplary view showing a modified structure of various artificial intelligence models according to an embodiment of the present invention;
5 is an exemplary view showing the result of scaling images of a COCO verification set to a size of 24Х18, and
6 is a flowchart illustrating a method for extracting a human posture based on artificial intelligence according to an embodiment of the present invention.

The above objects, other objects, features and advantages of the present invention will be easily understood through the following preferred embodiments in conjunction with the accompanying drawings. However, the present invention is not limited to the embodiments described herein and may be embodied in other forms. Rather, the embodiments introduced herein are provided so that the disclosed content will be thorough and complete and the spirit of the present invention will be sufficiently conveyed to those skilled in the art.

In some cases, it is mentioned in advance that parts that are commonly known in describing the invention and are not greatly related to the invention are not described in order to prevent confusion for no particular reason in explaining the present invention.

In this specification, when terms such as first and second are used to describe components, these components should not be limited by these terms. These terms are only used to distinguish one component from another.

In addition, it should be understood that, unless otherwise specified, the component may be implemented in any form of software, hardware, or both software and hardware.

In addition, terms used in this specification are for describing embodiments and are not intended to limit the present invention. In this specification, singular forms also include plural forms unless specifically stated otherwise in a phrase. The terms 'comprises' and/or 'comprising' used in the specification do not exclude the presence or addition of one or more other elements.

Also, in this specification, terms such as 'unit', 'platform', and 'device' may be intended to refer to functional and structural combinations of hardware and software driven by or driving the hardware. For example, the hardware herein may be a data processing device including a CPU or other processor. Also, software driven by hardware may refer to a running process, an object, an executable file, a thread of execution, a program, and the like.

In addition, the terms may mean a predetermined code and a logical unit of hardware resources for executing the predetermined code, and do not necessarily mean physically connected codes or one type of hardware in the present invention. can be easily deduced to the average expert in the art.

Hereinafter, with reference to the accompanying drawings for the specific technical content to be carried out in the present invention will be described in detail.

Studies on 2D human posture estimation are largely divided into bottom-up and top-down methods.

In the case of the bottom-up method, joint candidates are obtained for the input image, and then the human posture is estimated by analyzing the correlation between each joint. In the case of the top-down method, a person is first detected in the input image, and then joints are estimated in the bounding box containing the detected person. An embodiment of the present invention uses a top-down method showing higher performance for posture estimation.

More specifically, an embodiment of the present invention is based on the structure of HRNet, which represents the top-down method, while showing high posture estimation performance even in low-resolution images, artificial intelligence that can operate robustly in images of all sizes (ie, all resolutions) Intelligence models are available.

The networks of the existing model scale the original image to a fixed size for posture estimation and use it as an input of the artificial intelligence model. In this method, when scaling a high-resolution image to a low resolution, excessive information loss occurs, and when scaling a low-resolution image to a high-resolution image, many artifacts are generated in the image, increasing the possibility of a posture estimation error.

Therefore, in an embodiment of the present invention, a region including a person in an original image may be scaled in multi-resolution, and a plurality of target images generated by the scaling may be used as an input of an artificial intelligence model for posture estimation. In addition, an embodiment of the present invention can inform the artificial intelligence model which target image is most similar to the resolution (ie, size) of the original image, that is, to recognize and pose the most similar target image in the network of artificial intelligence models. Guiding information can be provided for reference in estimation. As a result, one network can operate robustly and effectively even in images of various resolutions.

To this end, an embodiment of the present invention can improve 2D human posture estimation performance by adding a plurality of light convolution layers without a computationally expensive module such as a super-resolution module.

1 is a block diagram showing an artificial intelligence-based human posture estimation apparatus 100 according to an embodiment of the present invention, and FIG. 2 shows output characteristics ( feature) is a diagram showing the output operation.

Referring to FIG. 1 , an artificial intelligence-based human posture estimation apparatus 100 according to an embodiment of the present invention includes an extraction unit 110, a scaling unit 120, a channel addition unit 130, and a posture estimation unit 140. can include

The extractor 110 may extract a region including a person (hereinafter referred to as 'original person image', I) from the original image. The extractor 110 may detect a person by analyzing an input original image, and set and crop a bounding box to include the detected person. Hereinafter, an image corresponding to the cropped bounding box is referred to as an original human image I. The original image may be one frame of a moving image or a still image.

The scaling unit 120 may generate a plurality of target human images I ₁ and I _n by scaling the original human image I extracted by the extractor 110 to have different resolutions. In detail, the scaling unit 120 converts the original human image I consisting of height h Х width w to target human images I ₁ , I ₂ , ..., I _n , ... which are multi-resolution input images. , I _N ). Here, the sizes of the scaled images (I ₁ , I ₂ , …, I _n , …, I _N ) are (h ₁ Хw ₁ ), (h ₂ Хw ₂ ), … , (h _N Хw _N ). The target human images I ₁ , I ₂ , ..., I _n , ..., I _N may be generated by bicubic interpolation (higher order interpolation) of the original human image I.

The scaling unit 120 may generate a plurality of target human images I ₁ and I _n based on a plurality of different input image resolutions set in the artificial intelligence model.

Accordingly, the resolution of the target human images I ₁ and I _n may be smaller or larger than that of the original human image I. In the case of FIG. 2 , the resolution of the target human image I ₁ is greater than that of the original human image I, and the resolution of the target human images I _n is smaller than that of the original human image I.

The resolution to be scaled by the scaling unit 120 may be determined as an optimal value through an experiment while learning a plurality of datasets in the process of constructing an artificial intelligence model. Therefore, the scaling unit 120 interlocks with the artificial intelligence model to check a plurality of resolutions set in the artificial intelligence model in advance to generate a plurality of target human images, or a resolution input from a user through a user interface (not shown) ( That is, a plurality of target human images may be generated at a resolution set in the artificial intelligence model).

The channel adder 130 may include one or more target human images I ₁ and I _n generated by the scaling unit 120 that are closest to the resolution of the original human image I or belong to a predetermined resolution range. A target human image may be selected and a guiding channel indicating a reference image may be added.

The guiding channel is a channel (C _G1 ) added so that the network (or artificial intelligence model) recognizes and references a target human image having the most similar size to the original human image, and a channel indicating that the target human image is not a reference image (C _G0 ). Accordingly, the target human image to which the guiding channel C _G1 is added may be used as a reference image for estimating a human posture in the network.

Referring to FIG. 2 , one target human image I ₁ is composed of an R channel (C _R ), a G channel (C _G ), and a B channel (C _B ). The channel adder 130 may add a guiding channel (C _G0 or C _G1 ) to the three RGB channels. The guiding channel is a channel filled with binary values when N>2, and may be added to the fourth channel of the target human image. For example, the display factor '0' is displayed for each pixel in the guiding channel (C _G0 ), and the display factor '1' is displayed in the guiding channel (C _G1 ).

In detail, the channel adder 130 creates and adds a guiding channel (C _G1 ) having a display factor of '1' when the resolution of the target human image corresponds to a predefined range, and adds it when the resolution of the target human image is out of the range A guiding channel (C _GO ) having a display factor of '0' may be added. Resolution conditions for adding a guiding channel will be described later with reference to [Table 1].

Alternatively, the channel adder 130 compares the resolution of each of the plurality of target human images (I ₁ , I _n ) with the resolution of the original human image (I), and selects one target human image having the smallest resolution difference ( For example, a guiding channel (C _G1 ) having a display factor of '1' is created and added to In), and a guiding channel (C _G0 ) having a display factor of '0' is added to the remaining target human images. can

Alternatively, when there are multiple candidate target human images within the threshold value, the channel adding unit 130 extracts and displays target human images in the order closest to a predetermined number (1 or more) of the plurality of candidate target human images. A guiding channel (C _G1 ) having a factor of '1' may be added to the extracted target human image, and a guiding channel (C _G0 ) having a display factor of '0' may be added to the remaining target human images.

In addition, the channel adder 130 determines that the resolution of the target human image I ₁ is greater than that of the original human image I, the resolution of the target human images _{I n} is smaller than that of the original human image I, and that I When the resolution difference between ₁ and I and the resolution difference between I _n and I are the same, if the original human image I has a higher resolution than the preset standard, the target human image I ₁ has a display factor '1' A guiding channel may be added, and a guiding channel having a display factor of '0' may be added to _the target human image In if the resolution is smaller than the reference.

According to the above description, the channel adder 130 may add a guiding channel for inducing multi-resolution images to selectively affect estimation of a person in a network according to the resolution of an original human image. Thereby, when the original person image is of low resolution, the network may be guided to estimate the posture by further referring to the target person image (or the feature of the target person image) scaled at a small magnification with little degradation. On the other hand, if the original human image has a high resolution, it may be influenced by a target human image (or a feature of the target human image) of a large size with almost no loss of information and may be induced to estimate the posture.

The above-described operation of the channel adder 130 may be omitted, and whether or not to be omitted may be determined by the user.

The posture estimator 140 may estimate a human posture by inputting a plurality of target human images generated by the scaling unit 120 to different networks of pre-learned artificial intelligence models for human posture estimation.

In addition, when the guiding channel is added to the target human images by the channel adding unit 130, the posture estimator 140 selects an original image among a plurality of target human images I ₁ and I _n based on the guiding channel. A human posture may be estimated by recognizing and referring to one or more target human images closest to the human image I.

In addition, when a plurality of target human images I ₁ and I _n are input to a predetermined stem network among a plurality of stem networks built on an artificial intelligence model, the posture estimator 140 generates a plurality of target human images ( While downscaling the resolutions of I ₁ and I _n , feature maps are created, and the generated feature maps are input to a corresponding subnetwork among a plurality of subnetworks to selectively change or maintain the resolution of the feature maps while fusion (fusion). ) to estimate the human posture. The feature map or features of the feature map may include information used to shape a person or to estimate a posture, such as a person's outline, shape, contrast, and color.

Referring to FIG. 2 , an operation of the posture estimating unit 140 estimating a posture using an artificial intelligence model will be described in detail.

The artificial intelligence model shown in FIG. 2 may include a plurality of STEM networks and HRNet. The stem network includes at least two strided convolution layers that extract image features from the target human image In, and the HRNet may include a plurality of sub-networks. Image features of multi-resolution images, ie, multiple target human images I ₁ and I _n , may be extracted through a stem network and used to estimate human joint positions. HRNet includes a plurality of sub-networks, and image features extracted from the n-th sub-network may be added to downscaled features from the (n-1)-th sub-network and inserted into the next network. The operation of HRNet will be described later with reference to FIG. 3 .

In FIG. 2 , a target human image I ₁ is input to a first stem network SN ₁ , and a target human image I _n is input to an n th system network SN _n . The first stem network (SN ₁ ) extracts features from the input target human image (I ₁ ), and the features of the extracted target human image (I ₁ ) are transferred to the first stage (S ₁₁ ) of the first subnetwork of HRNet. , and features are extracted again, and the extracted features may be downscaled or maintained by the convolution layer.

In addition, the nth-stem network (SN _n ) extracts features from the input target human image (I _n ), and the features of the extracted target human image (I _n ) are transferred to the first stage (S _n1 ) of the nth sub-network. is entered as At this time, the features output from the middle layer of the upper sub-network (eg, the first sub-network) are fused with the features extracted from the n-stem network, and the features from the upper sub-network are matched to match the dimensions of each feature. Maps can be downscaled and then fused.

Assuming that the feature map output from the (n-1)th subnetwork is fused with the input of the nth subnetwork, the input feature of S _n1 can be expressed as [Equation 1].

Referring to [Equation 1], stride=2 is the feature output from the (n-1)th stage of the (n-1)th subnetwork after the resolution of the feature map is reduced by 50% by the convolution layer, It means that it is fused with the feature output from the nth stem network.

3 is a diagram illustrating a framework of HRNet according to an embodiment of the present invention.

Referring to FIG. 3 , in a 2D pose estimation network, HRNet configures layers in series and parallel to provide high-performance pose estimation. In HRNet, multiple high-to-low subnetworks are connected in parallel, and multiple stages are connected in series in each high-to-low subnetwork. The kth stage of the nth subnetwork is denoted by S _nk .

Each subnetwork may receive target human images of different resolutions.

Each stage involves a multi-scale fusion process that aggregates features from different sub-networks.

The input features of S _nk are S _1(n+k-2), S _2(n+k-2), … , we can aggregate the output features of S _(n+k-2)1 . For example, the features output in S ₁₃ , S ₂₂ and S ₃₁ are combined and inserted into stage S ₂₃ .

Strided convolution or up-sampling is used to match the size of the feature maps to be integrated. In order to generate the input feature of S _nk , in an embodiment of the present invention, after reducing the size of the output feature of S (n-1) k by a predetermined ratio (eg, 50%) by strided convolution, the lower subnetwork can be output as The size of the output feature of stage S _(n+1)(k-2) is doubled through nearest neighbor upsampling and 1Х1 convolution. At S _n(k-1), the feature map passes through the convolutional layer without changing its size. The input feature of S _nk is created by adding feature maps of the same size.

Heat maps M corresponding to the number of joints for regressing joint positions may be obtained from output features of the first subnetwork. Heat maps corresponding to actual values may be generated by applying 2D Gaussian with a standard deviation of 1 pixel at the measured pixel position of each joint. Additionally, the output representation can be trained to use mean square error (MSE) loss to approximate ground truth heatmaps.

Through the above-described operation, the posture estimator 140 inputs multi-resolution images (I ₁ , _In ) to an artificial intelligence model including a plurality of stem networks and HRNet, and calculates 17 joints in which human postures are estimated. It is possible to output 17 heat maps (M) to be displayed, and to output a resultant image showing attitude estimation results from the 17 heat maps (M).

4 is an exemplary view showing a modified structure of various artificial intelligence models according to an embodiment of the present invention.

Referring to (a) to (c) of FIG. 4 , all resolutions of I ₁ are set to 256Х192. In addition, a scaled I ₂ or I ₃ with a lower resolution (74Х48, 32Х24 or 128Х96) is inserted into the sub-network. Accordingly, when the artificial intelligence model has the structure of (a), the scaling unit 120 scales the original input image to target human images I ₁ and I ₂ having resolutions of 256Х192 and 64Х48, respectively. The posture estimator 140 may estimate the posture by changing the structure to one of a plurality of modified artificial intelligence models shown in FIG. 4 according to the resolution of the original human image I.

[Table 1] shows the guiding channel conditions applied in the artificial intelligence models (a) to (c) shown in FIG. 4.

I_n
(height x width) Stem Output
(height x width) Guiding Channel
(condition) OURS-(a)

(256Х192)

(256Х192) 64Х48 0

(74Х48) 32Х24 1, if 96 96 < h w
0, otherwise OURS-(b)

(256Х192) 64Х48 0

(32Х24) 16Х12 1, if 96 96 < h w
0, otherwise OURS-(c)

(256Х192) 64Х48 0

(128Х96) 32Х24 1, if 64 64 < h w < 128 128
0, otherwise

(74Х48) 16Х12 1, if h w < 64 640, otherwise

Referring to (a) of FIG. 4 and [Table 1], since the resolution of the target person image I ₂ is smaller than 96Х96, the channel adder 130 converts the guiding channel having the display factor '1' to the target person. It can be created and added as a fourth channel of the image I ₂ to indicate that it is a reference image. In addition, a guiding channel having a display factor '0' may be added to the target human image I ₁ to indicate that it is not a reference image.

As an example, the guiding channel condition for each resolution described in [Table 1] can be changed according to the structure of the artificial intelligence model or the posture estimation performance.

5 is an exemplary view showing a result of scaling images of a COCO verification set to a size of 24Х18.

In FIG. 5, the conventional method refers to a method of estimating a posture by scaling an input image to one fixed size, and the present invention is a method of estimating a posture using a multi-resolution image according to an embodiment of the present invention described above. method, and the actual value means the pose estimated from the actual original image included in the COCO verification set.

Referring to FIG. 5 , posture estimation according to an embodiment of the present invention estimates joint positions more accurately than a baseline estimated by the conventional method, and the skeleton also expresses a posture that is more identical or almost similar to the actual one compared to the baseline. .

6 is a flowchart illustrating a method for extracting a human posture based on artificial intelligence according to an embodiment of the present invention.

An electronic device for performing the artificial intelligence-based human posture extraction method shown in FIG. 6 may be the human posture estimating apparatus 100 described with reference to FIGS. 1 to 5 , and thus a detailed description thereof will be omitted.

Referring to FIG. 6 , the electronic device may extract an original human image including a person from an input original image (S610).

The electronic device may generate a plurality of target human images by scaling the original human image extracted in step S610 to have different resolutions (S620).

The electronic device selects one or more target human images that are closest to the resolution of the original human image or within a predetermined resolution range among the plurality of target human images generated in step S620, and selects a guiding channel or reference image, which means a reference image. It is possible to add a guiding channel indicating that it is not (S630).

The electronic device may estimate a human posture by inputting a plurality of target human images to which a guiding channel is added to a network of different layers of a pre-learned artificial intelligence model for human posture estimation (S640).

In step S640, when a plurality of target human images are input to a predetermined stem network among a plurality of stem networks built on the artificial intelligence model, a feature map is generated while the resolution of the plurality of target human images is downscaled, and the generated features A human posture may be estimated by inputting the maps to a corresponding subnetwork among a plurality of subnetworks and fusion while selectively changing or maintaining the resolution of the feature maps.

7 is a block diagram illustrating a computing system performing an artificial intelligence-based human posture estimation method according to an embodiment of the present invention.

Referring to FIG. 7 , a computing system 700 includes at least one processor 710, a memory 730, a user interface input device 740, a user interface output device 750, and a storage connected through a bus 720. 760 , and a network interface 770 , which may be the electronic device described in FIG. 6 or the apparatus 100 for estimating a human posture described in FIG. 1 .

The processor 710 may be a central processing unit (CPU) or a semiconductor device that processes commands stored in the memory 730 and/or the storage 760 . The memory 730 and the storage 760 may include various types of volatile or non-volatile storage media. For example, the memory 730 may include a read only memory (ROM) 731 and a random access memory (RAM) 732 .

Accordingly, steps of a method or algorithm described in connection with the embodiments disclosed herein may be directly implemented as hardware executed by the processor 710, a software module, or a combination of the two. A software module resides in a storage medium (i.e., memory 730 and/or storage 760) such as RAM memory, flash memory, ROM memory, EPROM memory, EEPROM memory, registers, hard disk, a removable disk, a CD-ROM. You may. An exemplary storage medium is coupled to the processor 710, and the processor 710 can read information from, and write information to, the storage medium. Alternatively, the storage medium may be integral with the processor 710 . The processor and storage medium may reside within an application specific integrated circuit (ASIC). An ASIC may reside within a user terminal. Alternatively, the processor and storage medium may reside as separate components within a user terminal.

On the other hand, although the above has been described and illustrated in relation to preferred embodiments for illustrating the technical idea of the present invention, the present invention is not limited to the configuration and operation as shown and described in this way, and departs from the scope of the technical idea. It will be apparent to those skilled in the art that many changes and modifications can be made to the present invention without modification. Accordingly, all such appropriate alterations and modifications and equivalents are to be regarded as falling within the scope of the present invention. Therefore, the true technical protection scope of the present invention should be determined by the technical spirit of the attached claims.

100: human posture estimation device 110: extraction unit
120: scaling unit 130: channel addition unit
140: posture estimation unit

Claims (8)

an extraction unit for extracting a region including a person from the original image (hereinafter, referred to as 'original human image');
a scaling unit configured to generate a plurality of target human images by scaling the original human image extracted by the extraction unit to have different resolutions; and
a posture estimator for estimating a human posture by inputting the plurality of target human images generated by the scaling unit to different networks of pre-learned artificial intelligence models to estimate a human posture;
An artificial intelligence-based human posture estimation device comprising a.
According to claim 1,
The scaling unit,
The artificial intelligence-based human posture estimation device, characterized in that for generating the plurality of target human images based on the input image resolution set in the artificial intelligence model.
According to claim 1,
a guiding channel adding unit selecting one or more of the plurality of target human images generated by the scaling unit and adding a guiding channel indicating a reference image;
Including more,
The posture estimation unit,
and estimating a human posture by recognizing and referring to one or more target human images similar to the original human image based on the added guiding channel.
According to claim 1,
The posture estimation unit,
When a plurality of target human images are input to a predetermined stem network among a plurality of stem networks built on an artificial intelligence model, a feature map is generated while the resolution of the plurality of target human images is downscaled, and the generated feature map is generated. An artificial intelligence-based human posture estimating device, characterized by estimating a human posture by inputting them into a corresponding subnetwork among a plurality of subnetworks and selectively changing or maintaining the resolution of feature maps and fusion them.
(A) extracting, by an electronic device, a region including a person (hereinafter, referred to as an 'original person image') from an original image;
(B) generating, by the electronic device, a plurality of target human images by scaling the original human image extracted in step (A) to have different resolutions; and
(C) estimating, by the electronic device, a human posture by inputting a plurality of target human images generated in step (B) to different networks of pre-learned artificial intelligence models for human posture estimation;
Artificial intelligence-based human posture extraction method including.
According to claim 5,
In step (B),
An artificial intelligence-based human posture extraction method, characterized in that for generating the plurality of target human images based on the input image resolution set in the artificial intelligence model.
According to claim 5,
After step (B),
(D) selecting, by the electronic device, one or more of the plurality of target human images generated in the step (B) and adding a guiding channel indicating a reference image;
Including more,
In the step (C), the human posture is estimated by recognizing and referring to one or more target human images similar to the original human image based on the guiding channel added in the step (D). How to extract human posture.
According to claim 5,
In step (C),
(C1) inputting the plurality of target human images to a predetermined stem network among a plurality of stem networks built on an artificial intelligence model;
(C2) each of the plurality of stem networks generating feature maps while downscaling the resolution of a target human image, and inputting the generated feature maps to a corresponding subnetwork among a plurality of subnetworks; and
(C3) estimating a human posture by fusing the plurality of sub-networks while selectively changing or maintaining the resolution of the feature maps input in step (C2);
Artificial intelligence-based human posture extraction method comprising a.

Images