KR101931220B1 - Apparatus and Method for Body Components Detection - Google Patents

Abstract

The present invention provides a method and an apparatus for measuring a human body part. The method comprises the steps of: identifying a group sub image associated with a body part in an image to be measured; Assigning a confidence factor for a human body part in each subimage of the group subimage based on a basic visual feature of the subimage and an extended feature of the peripheral region of the subimage; And merging the sub-images having a confidence coefficient higher than the threshold value to measure the position of the human body part. The method and apparatus according to the present invention can improve the performance of body part measurement more efficiently.

Description

TECHNICAL FIELD [0001] The present invention relates to an apparatus and a method for measuring a body part,

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a computer vision and mode identification technology, and more particularly, to a method and apparatus for measuring a human body based on a multi-part context descriptor.

Human body posture estimation is a fundamental task of computer vision technology, and the technique is widely used in various fields such as human-machine interaction, gay, movie, real-role animation, and 3D. Recently, based on his technical and commercial values, research on human body posture has become the focus of computer vision. Particularly, human body part measurement is a basic technique of human body posture estimation technology. Human body part measurement technology can provide important observation information to guess the body posture.

In human body part measurement, the object measurement can be divided into two types. That is, human body part measurement based on an object and human body part measurement based on a pixel. The object-based body part measuring method regards each pixel in the object area as a positive sample, and the relationship between the pixel in the object area and its surrounding pixels is typically used as a descriptor. Pixel-based human body part measurement methods can measure an object using surrounding information, but do not use the internal and contour features of the object. The pixel-based human body region measuring method regards the entire object region as a positive sample and describes the region using some of the features. The body part measurement method based on the object uses the internal features and contour features of the object but does not use the peripheral information of the object.

In order to improve the accuracy of human body part measurement, there is a need for a human body part method and apparatus that combines the advantages of an object-based human body measurement and a pixel-based human body measurement method.

According to one aspect of the present invention, a method for measuring a body part is provided. The method comprises the steps of: identifying a group sub image associated with a body part in an image to be measured; Assigning a confidence factor for a human body part in each subimage of the group subimage based on a basic visual feature of the subimage and an extended feature of the peripheral region of the subimage; And merging the sub-images having a confidence coefficient higher than the threshold value to measure the position of the human body part.

According to one aspect of the present invention, the step of assigning a confidence factor for a human part in each sub-image of the group sub-image comprises defining a multi-part context descriptor for each sub-image, The site context descriptor includes a basic descriptor and an extension descriptor; And assigning a confidence factor for each subimage based on the similarity between the multi-site context descriptor of each subimage and the multi-site descriptor of a previously trained human body region. The basic descriptor describes basic visual characteristics of the human body part of the sub image, and the expansion descriptor describes the spatial structure relationship of the human body part and the surrounding human body part of the sub image.

According to one aspect of the present invention, the sub-image is an image area having different geometric shapes, positions, and sizes in the image to be measured.

According to one aspect of the present invention, a plurality of sub-images in the sub-image group overlap each other.

According to one aspect of the invention, the method further comprises preprocessing the image to be measured before identifying the image to be measured.

According to an aspect of the present invention, the pre-processing step includes: performing quantization on data of an image to be measured; Measuring image 3D information or image depth information for the image to be measured; Dividing the image to be measured; And extracting a foreground of the image to be measured.

According to one aspect of the present invention, the basic visual characteristics include an ordinal characteristic; Binary feature; Haar features; Histogram of Gradient description Feature; Contour feature; And at least one of gray histogram features.

According to one aspect of the present invention, the step of merging the sub-images comprises one of direct merging and weighting merging, and the threshold value depends on the selection and combination of multi-region context descriptors.

According to an aspect of the present invention, the direct merging step comprises merging the positions of the selected sub-images directly by statistical and geometric methods, selecting sub-images having a confidence coefficient higher than a predetermined value .

According to one aspect of the present invention, the weighted merging step comprises merging the positions of the subimages based on different trust factor thresholds or merging the positions of the subimages according to a clustering algorithm.

According to another aspect of the present invention, there is provided a body part measuring apparatus. The apparatus comprising: a sub image identification unit for identifying a group sub image associated with a human body part in an image to be measured; A confidence coefficient is assigned to a human body part in each subimage of the group subimage based on a basic visual characteristic of the subimage and an extended characteristic of a peripheral region of the subimage, And a human body part identification part for merging the images to measure the position of the human body part.

According to another aspect of the present invention, the human body part measuring device further includes a training part for performing training on the sample image to obtain a multi-part context descriptor of the human body part. Wherein the multi-region context descriptor includes a basic descriptor and an extension descriptor, the basic descriptor describes a basic visual feature for a human body portion of the sub-image, and the extension descriptor describes a spatial structure relationship between a human body portion of the sub- Lt; / RTI > Wherein the human body part identification unit defines a multi-region context descriptor of each sub-image and determines a trust for each sub-image based on the similarity between the multi-region context descriptor of each sub-image and the multi-site descriptor of the human body part trained in the training unit Quot;

According to another aspect of the present invention, the human body part measuring apparatus further includes a preprocessing unit for performing preprocessing on an image or a sample image to be measured.

Hereinafter, embodiments will be described with reference to the accompanying drawings. The above and other objects and advantages will be more clearly understood.
1 is a flowchart illustrating a training process of a human body part measuring method according to the present invention.
2A-2C illustrate a multi-part context descriptor in a depth image according to an embodiment of the present invention.
3 is a flowchart illustrating a body part measuring method according to an embodiment of the present invention.
FIG. 4 illustrates a human body part measuring system according to an embodiment of the present invention.

Hereinafter, embodiments will be described in detail with reference to the accompanying drawings.

1 is a flowchart illustrating a training process of a human body part measuring method according to the present invention.

As shown in FIG. 1, the training process of the human body part measuring method according to the present invention includes constructing training image data 110; A preprocessing step 120; A training sample preparation step 130 and a human body sorter training step 140. Hereinafter, each of the above steps will be described in detail.

First, in the training image data building step 110, training image data can be constructed by combining image data or real image data.

The preprocessing step 120 then performs quantization and background removal on the training image data. The quantization processing is a representation of the original image data as quantized depth image data to perform subsequent processing. For example, a gray image is a representative quantization depth image, and the value of its image data is quantized from 0 to 255. By quantization, the original image data can be quantized to reduce noise or reduce computation. In general, background removal processing is performed based on restrictions on depth values, related area extraction, and motion capture.

Specifically, the processing on the training image data comprises: performing quantization on the training image data; Measuring image 3D information or image depth information for the training image data; Dividing the training image data; And extracting a foreground with respect to the training image data.

로 표시할 수 있고,

및

는 부위 영역의 주변 영역의 중심 위치를 의미하고, R는 주변 영역의 사이즈를 의미하고, M는 주변 영역의 분할된 형상을 의미한다. 분할된 형상은 사각형, 원형 등이 될 수 있다. 네거티브 샘플은 측정될 인체 부위를 포함하는 서브 이미지와 반대의 서브 이미지인 것이다. 즉, 측정될 인체 부위를 포함하지 않는 서브 이미지인 것이다. 일반적으로 샘플 트레이닝 단계에서 두 기지 타입의 네거티브 샘플을 사용한다. 하나는 인체가 포함되지 않는 배경 이미지인 것이고, 또 하나는 해당 트레이닝 인체 부위가 이미 추출된 인체 이미지인 것이다. 네거티브 샘플의 트레이닝에 의하여 서브 이미지에 측정될 인체 부위가 포함되는지 구분할 수 있다. Then, in the training sample preparation step 130, three types of training samples are prepared. A site sample, a site context sample, and a negative sample. The site sample focuses on the site object itself. The site-context sample focuses on the surrounding area of the site. The segmentation for the region around the site

As shown in FIG.

And

Denotes the center position of the peripheral region of the region, R denotes the size of the peripheral region, and M denotes the divided shape of the peripheral region. The divided shape may be a square, a circle, or the like. A negative sample is a subimage opposite to the subimage containing the body part to be measured. That is, it is a sub-image that does not include the human body part to be measured. Generally, two sample types of negative samples are used in the sample training phase. One is a background image that does not contain a human body, and the other is a human image that has already been extracted. By training the negative sample, it is possible to distinguish whether the sub-image contains the body part to be measured.

Next, in the human body classifier training step 140, a human body classifier training is performed to measure a human body part. Human body classifier training includes describing a sample; Improving training; Building a meter involves three steps.

In describing the sample, a sample is described based on the basic characteristics of each subimage and the extended features of the peripheral region of the subimage. According to an embodiment of the present invention, a basic feature of each sub-image and an extended feature of a peripheral region of the sub-image can be expressed using a multi-region context descriptor. The multi-region context descriptor according to an embodiment of the present invention not only describes the basic visual characteristics of the part target of each sub-image but also describes the context information of the neighboring parts adjacent to the part, that is, the extended feature. For example, embodiments of the present invention can be regarded as a basic feature of a multi-site context descriptor using the features of multi-size ordinal mode (MSOP).

Based on the multi-site context descriptor of the MSOP, it can be expressed as Equation (1) as follows:

Equation (1)

로 표현할 수 있다.

는 MSOP 모드의 파라미터를 의미하고 MSOP 모드의 파라미터는 모드의 위치, 크기 및 유형을 포함할 수 있다. 본 발명의 실시예에 따른 MSOP 모드는 서브 이미지에 포함된 부위의 기본적인 특징을 기술하기 위한 MSOP 모드 및 주변 부위의 확장 특징을 기술하기 위한 MSOP 모드를 포함할 수 있다. 그러나 MSOP 모드에 근거하는 특징 기술자는 본 발명의 실시예에 따른 멀티 부위 콘텍스트 기술자의 한 예시로써 본 발명이 이에 의하여 제한되지 않는다. 본 발명은 기타의 시각 특징 기술 방식으로 멀티 부위 콘텍스트 기술자를 형성할 수 있다. 예를 들어, 이미지의 Haar 특징; 기울기 히스토그램(HOG) 특징; 바이너리(binary) 특징; 윤곽 특징; 그레이(gray) 히스토그램 특징을 이용하여 멀티 부위 콘텍스트 기술자의 기본적인 기술자로 간주한다.In the above equation (1), b (x) means a Boolean function, and b (x) = 1 if x> 0 and b (x) = 0 if x <0. gi denotes the pixel value in the grid in the MSOP mode, i denotes the index of the grid, gc denotes the average pixel value of the subimage and the grid around the subimage,

.

Means the parameter of the MSOP mode, and the parameters of the MSOP mode may include the position, size and type of the mode. An MSOP mode according to an embodiment of the present invention may include an MSOP mode for describing a basic feature of a part included in a sub image and an MSOP mode for describing an extended feature of a surrounding part. However, the feature descriptor based on the MSOP mode is an example of a multi-region context descriptor according to an embodiment of the present invention, and thus the present invention is not limited thereto. The present invention may form a multi-site context descriptor in other visual feature description schemes. For example, Haar features of images; Tilt histogram (HOG) feature; Binary feature; Contour feature; We consider the gray histogram feature to be the basic descriptor of the multi-site context descriptor.

Hereinafter, modes of the multi-region context descriptor according to an embodiment of the present invention will be described with reference to FIGS. 2A to 2C.

를 도시한 것이다. 도 2C는 본 발명의 실시예에 따른 멀티 부위 콘텍스 기술자에 있는 부위 주변 영역의 콘텍스트 정보를 기술하는 확장 기술자의 모드 파라미터

를 도시한 것이다. 상기 기술 방식에 있어서, x및y 는 기술자의 위치 파라미터를 의미하고, w및 h는 기술자의 형상 파라미터를 의미하고, t는 기술자의 유형을 의미한다. 상기 기술 방식으로, 주변 영역의 크기 및 형상은 측정될 인체 부위를 포함하는 서브 이미지와 일정 관계를 구비하는 것으로 정의될 수 있다. 예를 들어, 주변 영역의 위치는 서브 이미지의 영역을 중심으로 하는 더 큰 영역이 될 수 있고(도2B의 블록205 및 207을 참조), 서브 이미지 영역의 위쪽, 아래쪽, 왼쪽, 오른쪽 영역이 될 수도 있다. 주변 영역의 크지는 서브 이미지 영역의 2배, 3배, 3.5배등이 될 수 있다. 주변 영역의 형상은 서브 이미지와 같은 형상이 될 수 있고 고정된 사각형, 원형 등도 될 수 있다. 2A-2C illustrate mode parameters of a multi-region context descriptor according to an embodiment of the present invention. Referring to FIG. 2B, blocks 201 and 203 refer to a sub-image including a human body part. The body part included in the block 201 is the upper arm of the human body, and the body part included in the block 203 is the human body's head. Blocks 205 and 207 denote peripheral regions corresponding to blocks 201 and 203, respectively. FIG. 2A is a block diagram of a mode descriptor of a basic descriptor describing a visual feature of a region in a multi-region context descriptor according to an embodiment of the present invention.

FIG. FIG. 2C illustrates a mode parameter of an extension descriptor describing context information of a region around a region in a multi-region context descriptor according to an embodiment of the present invention.

FIG. In the above description, x and y mean the position parameter of the descriptor, w and h mean the shape parameter of the descriptor, and t mean the type of the descriptor. In the above description, the size and shape of the peripheral region can be defined as having a certain relationship with the sub-image including the human body region to be measured. For example, the position of the peripheral region may be a larger region centered on the region of the sub-image (see Blocks 205 and 207 of FIG. 2B), and may be the upper, lower, left, It is possible. The size of the surrounding area may be 2 times, 3 times, 3.5 times, or the like of the sub image area. The shape of the peripheral region may be the same as the shape of the sub-image, or may be a fixed rectangle, a circle, or the like.

With respect to the head, the multi-site contex descriptor according to an embodiment of the present invention not only describes the characteristics of the head (e.g., oval contour features), but also describes information about the neck or shoulder area around the head . With respect to the upper arm, the multi-site contex technician according to the embodiment of the present invention can not only describe the shape of the upper arm, but also information on the upper body around the upper arm. Accordingly, the multi-region context descriptor according to the embodiment of the present invention not only includes the internal features and the contour structure of the region object, but also includes the context information of the surrounding region. Accordingly, the multi-site context descriptor according to the embodiment of the present invention has higher stability as compared with the local descriptor of the present technology.

In addition, as described above, although the example of the human body region by the head and the upper arm is described in FIG. 2, it should be understood that the human body region can be more precisely classified by those skilled in the art. For example, the human body may be divided into two parts: head, left upper arm, left lower arm, left hand, right upper arm, lower right arm, right hand, left thigh, left calf, right thigh, So that more precise multi-site context descriptor training can be performed for each of the regions.

Referring again to FIG. 1, in the step of improving training, all types of classifiers are trained to measure a human body part, and each type of classifier means one characteristic of a sub-image. For example, the classifier is trained through algorithms such as SVM, Forest, and Boosting. In the embodiment of the present invention, the classifier training is performed using the AdaBoost algorithm. The multi-region context descriptor based on AdaBoost can be expressed as Equation (2) as follows.

Equation (2)

는 부위 대상 자체의 특징을 기술하기 위한 분류기를 의미하고,

는 부위 주변 영역에 대한 콘텍스트 정보의 특징을 기술하기 위한 분류기를 의미하고,

는

분류기의 개수를 의미하고,

는

분류기의 개수를 의미하고, F(x)는 최종의 분류기를 의미한다.In the above equation (2), x denotes a sample,

Means a classifier for describing a characteristic of a site object itself,

Quot; means a classifier for describing a feature of context information about a region around a region,

The

Means the number of classifiers,

The

Denotes the number of classifiers, and F (x) denotes the final classifier.

Then, various kinds of classifiers trained in the step of constructing the human body part are cascadeed to improve the performance of human body part measurement.

Hereinafter, a method of measuring a human body part according to an embodiment of the present invention will be described with reference to FIG.

First, in step 310, the group subimage associated with the human body part is identified in the image to be measured. Specifically, in the process of identifying, the sub-image may be searched based on the specified position and size of the depth image and it may be determined whether the sub-image is a human body part to be measured.

로 표현할 수 있고,

는 각각 서브 이미지를 검색할 때 사용된 가장 작은 사이즈, 가장 큰 사이즈, 및 step size 를 의미하고,

는 각각 서브 이미지를 검색할 때 사용된 이니셜 검색 포인트 위치, 종점 위치 및 searching step size를 의미한다. 다리 말하면, 서브 이미지는 측정될 이미지에 서로 다른 기하 형상, 위치 및 크기를 구비하는 이미지 영역이 될 수 있고 서브 이미지 그룹에 있는 복수의 서브 이미지가 서로 오버랩이 될 수 있다. 구체적으로 상기 검색 과정은 이미지에서 순환으로 서브 이미지 영역을 추출하는 과정으로 간주될 수 있다. 우선, 이니셜 사이즈 Smin부터 서브 이미지 영역의 크기를 결정하고, 그 다음, 이미지의 이니셜 검색 포인트 위치Pstart로부터 서브 이미지 영역의 중심 위치를 결정하고 종점 위치 Pend에 도작할 때까지 Ps를 step로 하여 순서대로 이미지 영역의 중심 위치를 이동하고, 트래버스(traverse)된 각 위치에서 서브 이미지를 추출한다. 그 다음, 가장 큰 사이즈Smax 가 될 때까지 step Sstep에 의하여 순서대로 서브 이미지 영역의 크기를 증가한다. 각 영역 사이즈마다 한번씩 서브 이미지의 중심으로 Pstart로부터 Pend까지 트래버스하여 검색할 모든 서브 이미지를 획득한다.The search process

Can be expressed as,

Quot; means the smallest size, the largest size, and step size used when searching for the sub image, respectively,

Respectively denote the initial search point position, the end point position, and the searching step size, which are used when the sub image is searched. In short, a sub-image may be an image area having a different geometry, position, and size in an image to be measured, and a plurality of sub-images in the sub-image group may overlap each other. Specifically, the searching process can be regarded as a process of extracting a sub-image region from a circulation in an image. First, the size of the subimage area is determined from the initial size Smin, then the center position of the subimage area is determined from the initial search point position Pstart of the image, and Ps is stepped Moves the center position of the image region and extracts the subimage at each traversed position. Then, the size of the subimage area is increased in order by the step Sstep until it becomes the largest size Smax. And traverses from Pstart to Pend at the center of the subimage once for each region size to obtain all the subimages to be searched.

It is desirable to extract the human body in the image to be measured through the foreground extraction technique in the identification process. With the foreground extraction, only human foreground regions can be measured and the search volume for subimage locations can be reduced. In addition, the depth of the extracted foreground object can be measured to reduce the search range for the sub image. In addition, sub-images can be identified by image 3D information and image segmentation.

Then, at step 320, a confidence factor is assigned to the human body portion of each sub-image in the group sub-image based on the basic feature description of each sub-image and the extended feature description of the sub-image surrounding region. In other words, based on the trained multi-site context descriptor, it is determined whether the sub-image is a human part. Specifically, confidence coefficients are calculated for sub-images in the form of a final classifier F (x) (see equation (2)) finally obtained by human body classifier training. First, we calculate the corresponding classifier f according to the characteristics of the human body classifier training part, overlay the calculated values for each classifier by a combination of F (x), and output the overlaid output value to the corresponding sub- It is regarded as a confidence coefficient.

Then, in step 330, the position of the human body part is measured by merging the sub-images whose confidence coefficients are higher than the threshold value. Here, the threshold value depends on the selection and combination of the types of the multi-region context descriptors, and according to the embodiment of the present invention, the merging step may include both direct merging and weighting merging methods. In the direct merging step, a sub-image having a confidence coefficient higher than a predetermined value is selected and the position of the human body part is finally acquired by merging the positions of the selected sub-images directly by the statistical and geometric methods. Specifically, the average center position and the average size size of the sub image matching all the reliability coefficient requests can be calculated by the average value method to be the final human body part position. In addition, the region where the sub images corresponding to all the reliability coefficient requests are most concentrated can be regarded as the final output position of the human body region.

In the weighted merge step, the positions of the sub images are merged based on the different confidence coefficient thresholds, or the positions of the sub images are merged according to the clustering algorithm. Clustering algorithms can include mean-shift, k near-neighbor. Specifically, in a weighted merger, the confidence factor for the subimage matching the confidence factor requirement can be maintained at the same time and determines the importance of the subimage in the merger. For example, the weighted average center position and the weighted average size size of the subimage may be the positions of the final human body region.

Hereinafter, the operation principle of the human body part measuring system according to the embodiment of the present invention will be described with reference to FIG. 4, the human body part measuring system according to the embodiment of the present invention includes an image obtaining device 410 and a human body part measuring device 420. Although the image acquiring device 410 and the body part measuring device 420 are installed separately in FIG. 4, the two devices can be realized as one facility.

In addition, devices and facilities such as a Primesence mapping device, a Time-of-Flight (ToF) camera, and a multi-view camera can be regarded as the image acquisition device 410. The image data obtained by the image acquisition device 410 may be used as training image data and may be used as image data to be measured.

The human body part measuring device 420 includes an image preprocessing part 421, a training part 422, a sub image identifying part 423, and a human body part identifying part 424.

An image preprocessing section 421 is used for preprocessing the image data. Here, the image preprocessing section 421 preprocesses the image data obtained by the image acquisition apparatus 410 or the sample image data stored in the training sample image data base. For example, the image preprocessing section 421 can perform quantization on the image data, and the training section 422 and the subimage identification section 423 can perform foreground extraction processing on the image data, Value measurement, image 3D information measurement, image segmentation, and so on.

The training section 422 trains a human body part sorter (i.e., a multi-site context descriptor) that prepares training samples from the training sample image data and uses the training samples to measure the human body area, and uses the human body part sorter Build. The training sample image data may be image data that is stored in advance in the image database and may be image data obtained by the image acquisition device 410. [

The sub image identification unit 423 identifies a group sub image related to the human body part in the image to be measured. The human body part identification unit 424 assigns a confidence coefficient to a human body part in each subimage of the group subimage based on the basic visual characteristic of each subimage and the extended characteristic of the peripheral area of the subimage, The sub-images higher than the threshold value are merged to measure the position of the human body part. In the exemplary embodiment of the present invention, the human body part identification unit 424 defines a multi-region context descriptor of each sub-image, and based on the similarity between the multi-region context descriptor of each sub-image and the multi-site descriptor of the trained human body region, A confidence coefficient is assigned to the subimage. In addition, the human body part identification unit 424 outputs the measured result to an external device, processes it by an external device, and displays the final human body part identification result.

Each module of the human body part measuring apparatus 420 described above can be divided into more modules or can be merged into a smaller number of modules by way of example.

Compared with the human body part measurement technique based on the currently used object and the pixel part based human body part measurement technique, the multi-part context descriptor according to the present invention can describe the internal and contour features of the human body part itself But also the characteristics of the peripheral region of the human body part target can be described. Therefore, the performance of the human body part measuring device can be improved by using the surrounding information of the human body part, and the object of the classifier training can be achieved more quickly, so that the complexity of the classifier can be reduced. Therefore, the human body part measuring method according to the present invention can improve the performance of human body part more efficiently.

While the present invention has been particularly shown and described with reference to exemplary embodiments thereof, it is to be understood that the invention is not limited to the disclosed embodiments, but, on the contrary, Do. Therefore, the scope of the present invention should not be limited to the described embodiments, but should be determined by the equivalents of the claims, as well as the claims.

Claims (17)

Generating a plurality of subimages based on an input image including a human body portion;
Determining a confidence coefficient of the plurality of sub-images based on a basic characteristic of the sub-image and an extended characteristic of the peripheral region of the sub-image; And
And combining the sub-images having a confidence coefficient higher than the threshold value when the at least two sub-images of the plurality of sub-images have a confidence coefficient higher than a threshold value to detect the human body part and,
Wherein the step of determining the confidence factor comprises:
Determining a confidence coefficient of the sub-images based on a similarity between a multi-site context descriptor for the sub-images and a multi-site descriptor pre-trained for the body part,
Wherein the multi-region context descriptor comprises:
A basic descriptor corresponding to the basic feature, and an extended descriptor corresponding to the extended feature
Method of measuring body part. The method according to claim 1,
The basic engineer
Describing the basic features for the body part of the sub-images,
The extension descriptor
Wherein the spatial structure relationship between the human body region of the sub images and the peripheral regions of the human body region is described. The method according to claim 1,
A step of preprocessing the input image
Further comprising the steps of: The method of claim 3,
Wherein the step of preprocessing the input image comprises:
Detecting image 3D information or depth information for the input image; And
Extracting a foreground from the input image based on the 3D information or the depth information
/ RTI &gt; The method according to claim 1,
The above-
Ordinal features; Binary feature; Haar features; Histogram of Gradient description Feature; Contour feature; And at least one of gray histogram features. The method according to claim 1,
The extended feature
And a multi-size audio pattern (MSOP) feature based on a parameter indicative of the peripheral region. The method according to claim 1,
The step of detecting the human body part comprises:
Merging the sub-images based on at least one of a direct merging technique and a weighting merging technique
/ RTI &gt; The method according to claim 1,
The threshold value may be,
Wherein the at least one basic descriptor in the multi-region context descriptor and the at least one extension descriptor in the multi-region context descriptor. 8. The method of claim 7,
In the direct merging technique,
Selecting a subimage having a confidence coefficient higher than a predetermined value and merging the selected subimages using at least one of a statistical algorithm and a geometric algorithm. 8. The method of claim 7,
The weighted merging technique includes:
Merging positions of sub-images based on different confidence coefficient thresholds, or merging positions of sub-images based on a clustering algorithm. 8. The method of claim 7,
Calculating an average size size and an average center position of the sub images matching the confidence factor requirement higher than the threshold and using at least one of the average size size and the average center position to determine a final position of the body part step
Further comprising the steps of: 12. The method of claim 11,
Weighting the average size size and the average center position based on each confidence factor to determine a final position of the body part
Further comprising the steps of: The method according to claim 1,
The confidence coefficient of the sub-images is
Wherein the body part is calculated using a final classifier obtained by training a plurality of human body classifiers. A sub image identification unit for generating a plurality of sub images based on an input image;
And determining a confidence coefficient of the plurality of sub-images based on a basic characteristic of the sub-images and an extended characteristic of a peripheral region of the sub-images, wherein at least two of the plurality of sub-images have a threshold value, A human body part identification part for detecting a human body part by merging sub images having a confidence coefficient higher than the threshold value,
Containing
Wherein the human body part identification unit comprises:
Determining a confidence coefficient of the sub-images based on a similarity between a multi-region context descriptor for the sub-images and a preset multi-region context descriptor for the human body region,
Wherein the multi-region context descriptor comprises:
A basic descriptor corresponding to the basic feature, and an extended descriptor corresponding to the extended feature
Human body part measuring device. 15. The method of claim 14,
The threshold value
A plurality of basic descriptors in the multi-region context descriptor, and at least one extension descriptor in the multi-
The human body part identification part
Calculating an average size size and an average center position of the sub images matching the confidence factor requirement higher than the threshold and using at least one of the average size size and the average center position to determine a final position of the body part , A body part measuring device. 16. The method of claim 15,
The human body part identification part
And weighting the mean size size and the mean center position based on each confidence coefficient to determine a final position of the body part. 15. The method of claim 14,
The sub image identification unit
A sub-image based on at least one of a predetermined position and a size of the depth image, and determines whether the sub-image identified in the search corresponds to a human body portion of the detected image.

Images