KR20160121481A - Object recognition system and method the same - Google Patents

Abstract

The present invention relates to an object recognizing system and an object recognizing method thereof which extract right and left feature vectors from a stereo image generated from an object, find a robust feature vector coexisting in the extracted right and left feature vectors, and compare information of the right and left feature vectors and the robust feature vector with information stored in a database to extract information on the name of the object so as to recognize the object. Therefore, the object recognizing system provides improved object recognizing performance.

Description

BACKGROUND OF THE INVENTION 1. Field of the Invention [0001] The present invention relates to an object recognition system,

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an object recognition system and an object recognition method thereof, and more particularly, to an object recognition apparatus and an object recognition method for recognizing an object from left and right image information given from a stereo camera.

In general, object recognition in the field of computer vision is performed by establishing a database of objects to be recognized offline, tracing through a training process, recognizing registered objects through a feature extraction algorithm and a recognizer, It has been developed to judge that there is no object.

When the 3D database model is used to construct the object database, it is not possible to utilize it in terms of the actual robot service because the manual operation of the human is necessary. In recent years, object feature extraction and registration techniques that do not require manual operation have been developed, but they have limitations in terms of object recognition that is robust to environmental changes.

The environmental changes considered in object recognition are largely divided into the aspects of photometric invariance caused by illumination changes and noise, and the geometric invariance associated with changes in camera angles and distances to objects. The Invariance factor is important because the types and characteristics of the objects used by each family and individual are different from the service aspect of the robot. It is not realistic to register it, and users have to register themselves. In other words, it is troublesome to collect a large number of images corresponding to each change in each object in the sense that a user must directly learn an object on-line or off-line. Therefore, it is important to determine how many changes can be accommodated in one registration and a stable recognition rate can be maintained .

However, existing object recognition systems have limitations on object recognition performance due to the existence of systems suitable for a single camera.

One aspect of the present invention provides an object recognition system for extracting texture-based features from a stereo image to improve object recognition performance, and an object recognition method for the object recognition system.

The object recognition system includes a stereo camera for outputting left and right images of the same object; Extracting left and right feature vectors from left and right images input from a stereo camera, matching the extracted left and right feature vectors with each other, extracting a plurality of common feature vectors common to left and right images, And an object recognizing device for matching a plurality of common feature vectors common to right and left images for each object stored in the database and recognizing an object in left and right images based on a matching result.

In addition, the object recognition apparatus may match a plurality of common feature vectors common to right and left images with a plurality of common feature vectors stored in a database unit, and may perform a search based on Euclidean distance values for matching feature vectors of common feature vectors. A score is assigned to each object, and the added points are added together to recognize the object having the highest sum as an object in the left and right images.

The object recognition apparatus may further include: a feature extraction unit that extracts a plurality of feature vectors from a plurality of feature points of the left image and extracts a plurality of feature vectors from the plurality of feature points of the right image; A feature vector matching unit for matching a plurality of feature vectors extracted from a left image and a plurality of feature vectors extracted from a right image to extract a plurality of common feature vectors common to left and right images; And an object recognition unit for recognizing an object in the left and right images by matching the extracted common feature vectors and a plurality of common feature vectors stored in the database unit.

In addition, the feature extraction unit detects vectors in a plurality of regions centered on a plurality of feature points, and extracts a set of detected vectors as a feature vector.

The object recognition apparatus also includes a plurality of left and right feature vector information extracted from left and right images, a plurality of common feature vector information common to left and right images, and three-dimensional position information of a plurality of common feature vectors to information Each of the matching results is given a score, and the added points are summed to recognize the object having the highest sum as an object in the left and right images.

In addition, the object recognizing apparatus recognizes an object and then verifies the recognized object using at least one of two-dimensional homography and a three-dimensional Euclidean transformation on a plurality of common feature vectors matched .

The object recognizing method includes: inputting different left and right images of the same object; Extract left and right feature vectors from input left and right images; Extracting a plurality of common feature vectors common to left and right images by matching the extracted left and right feature vectors with each other; Matching the extracted common feature vectors with a plurality of common feature vectors common to left and right images for each object stored in the database unit; And recognizes the objects in the left and right images according to the matching result.

The recognition of an object is performed by assigning a score to each of the plurality of matching common feature vectors according to the Euclidean distance value and summing the assigned points to obtain an object having the highest total sum as an object .

The method further includes verifying the recognized object using at least one of two-dimensional homography and a three-dimensional Euclidean transform on a plurality of common feature vectors after recognizing the object .

According to the embodiment of the present invention described above, the invariance performance is improved and the recognition performance is stable through the registration of one database using the left and right image information inputted through the stereo camera, It is possible to provide intelligent robots and provide practical services, and can also be used for security surveillance using a stereo camera in addition to robots.

1 is a schematic control block diagram of an object recognition system according to an embodiment of the present invention.
2 is a control flowchart for an object recognition method according to an embodiment of the present invention.
FIG. 3 is a diagram for explaining feature vector matching of left and right images in FIG.
FIG. 4 is a diagram showing the matching result of FIG. 3. FIG.
FIG. 5 is a control flowchart for explaining an object recognition process of FIG. 2. FIG.
FIG. 6 is a control flowchart for explaining a process of recognizing an object in a manner different from that described in FIG.
FIG. 7 is a diagram for explaining weighting between three-dimensional distances between matched feature vectors in the object recognition system and three-dimensional distances between feature vectors stored in the database unit according to an embodiment of the present invention.

Hereinafter, embodiments of the present invention will be described with reference to the accompanying drawings.

1 is a schematic control block diagram of an object recognition system according to an embodiment of the present invention.

As shown in FIG. 1, an object recognition system according to an embodiment of the present invention includes a stereo camera 10 and an object recognition device 20.

The stereo camera 10 outputs a stereo image. Stereo images are images of different viewpoints as different right and left images including the same scene. The stereo camera 10 for acquiring such a stereo image can be largely divided into a binocular and a simplex manner. A single-frame camera captures images at different points of view using one camera. A bi-plane camera captures images at different points of view using two cameras.

The object recognition apparatus 20 extracts a plurality of feature vectors from a plurality of feature points in the left and right images input from the stereo camera 10, matches each of the plurality of feature vectors extracted from the left and right images, And recognizes the object of the left and right images using the plurality of feature vectors and the matched feature vectors.

The object recognition apparatus 20 includes a system management unit 21, an image storage unit 22, a first feature extraction unit 24, a second feature extraction unit 25, a feature vector matching unit 26, 23, an object recognition unit 28, and a database unit 27.

The system management unit 21 outputs a video output command to output the stereo image to the stereo camera 10 and outputs a video storage command to store the video in the video storage unit 22. [ Further, the system management unit 21 outputs a feature vector extraction instruction to the feature vector management unit 23 for feature vector extraction. Also, the system management unit 21 outputs an object recognition command to the object recognition unit 28 and receives the object recognition result.

The image storage unit 22 receives the left and right images output from the stereo camera 10, and transmits the left and right images to the feature extraction units 24 and 25.

The first feature extraction unit 24 extracts a feature vector from the left image information received from the image storage unit 22. [ The feature vector of the left image is output to the feature vector matching unit 26 and the feature vector management unit 23.

The second feature extraction unit 25 extracts a feature vector from the right image information received from the image storage unit 22. The feature vector of the right image is output to the feature vector matching unit 26 and the feature vector management unit 23.

The feature extraction units 24 and 25 detect vectors in a plurality of regions centered on a plurality of feature points, and extract a set of detected vectors as feature vectors. The feature extraction units 24 and 25 select the feature points from the corners of the image information in the left and right images. The feature extracting units 24 and 25 extract the interested feature points from the left and right images, detect the vectors in a plurality of regions centered on the respective feature points, and extract the set of detected vectors as feature vectors. At this time, edges and corners are usually used as feature points.

As a method for extracting a feature vector, there are a Scale Invariant Feature Transform (SIFT) method and a Speed Up Robust Feature (SURF) method.

Both of these methods have a common feature that they are invariant to scale change and rotation because they use scale space for feature vector extraction. Since the scale space is adjusted by adjusting the size of the image in various ways, the feature vector that is invariable to the scale change can be obtained since both the large acquired image and the small acquired image are considered. A method of obtaining a scaled-down image is to use a Gaussian kernel. When the convolution of the image is performed, the larger the variance of the gau time kernel, the smaller the image is produced. In the case of shift (SIFT), the size of the original target is reduced and the convolution with the Gaussian kernel is performed when the variance increases to some extent. Then, we calculate the difference (Difference of Gaussian, DoG) between neighboring images. The local extrema of DoG in the scale space is selected as the feature point and feature vector. Laplacian is calculated and the pole is found while varying the dispersion. The SURF simplifies the extraction process of the feature vector to overcome the long computation time which is a disadvantage of the shift (SIFT).

The feature vector management unit 23 transmits a feature extraction instruction to the feature extraction units 24 and 25 and receives and stores a plurality of feature vector information of each image from the feature extraction units 24 and 25. Also, the feature vector matching unit 26 receives and stores information of feature vectors common to right and left images according to the matching result. In addition, the feature vector management unit 23 extracts and stores three-dimensional position coordinates of feature vectors commonly present in the left and right images using the camera calibration data of the stereo camera 10. In addition, the feature vector management unit 23 transmits to the object recognition unit 28 a plurality of feature vector information of each of the left and right images, information of the matched feature vectors, and three-dimensional position information of the matched feature vectors.

The feature vector matching unit 26 performs stereo matching on the feature vectors output from the first feature extracting unit 24 and the second feature extracting unit 25 to find feature vectors that exist in the left and right images. The matched feature vectors are transmitted to the feature vector management unit 23. The feature vector matching unit 26 finds a feature vector that exists in the left and right images based on the similarity between the feature vectors of the left image and the feature vectors of the right image. In this method, a normalized cross correlation (NCC) method using a normal cross-correlation value between feature vectors of left and right images, a Euclidean distance using an Euclidean distance value between feature vectors of left and right images, distance method.

The object recognition unit 28 receives the plurality of feature vector information of each of the left and right images received from the feature vector management unit 23, the information of the matched feature vectors, and the three-dimensional position information of the matched feature vectors, ) To recognize the object in contrast with the information stored for each object. That is, the object recognizing unit 28 recognizes the presence or absence of an object and the name information of existing objects.

A plurality of feature vector information of the left image, a plurality of feature vector information of the right image, information of the matched feature vectors, three-dimensional position information of the matched feature vectors, and the like are stored in advance in the database unit 27 for each object. In addition, the object recognition unit 28 verifies the result of recognizing the object using two-dimensional homography and three-dimensional Euclidean Transform.

The object recognition apparatus 20 according to the embodiment of the present invention extracts characteristics that are robust against change through the same object or the left and right images provided from the stereo camera 10 at the same time. In addition, feature groups common to the left and right images are separately selected to constitute a feature group resistant to environmental changes. In this case, when performing the matching operation to construct a common feature vector group, the matching for the feature vector group from the right and left images is performed only for the adjacent scale spaces in consideration of the calculation time. In addition, the object recognition apparatus 20 performs matching of the feature vectors derived from the right and left images when recognizing the object, and at the same time, recognizes the reliability of the object recognition using the relationship between the three- .

2 is a control flowchart for an object recognition method according to an embodiment of the present invention.

Referring to FIG. 2, the object recognition apparatus 20 according to the embodiment of the present invention calls the stereo camera 10 for object recognition when the system is started, so that the stereo camera 10 can recognize different right and left And outputs an image.

The object recognition apparatus 20 receives the left and right images from the stereo camera 10 (100).

The object recognition apparatus 20 stores the left and right images input from the stereo camera 10 in the image storage unit 22 (101). At this time, the object recognition apparatus 20 transmits the left image of the left and right images stored in the image storage unit 22 to the first feature extraction unit 24, and transmits the right image to the second feature extraction unit 25.

The object recognition apparatus 20 extracts the feature vectors of the left image through the first feature extraction unit 24 and extracts the feature vectors of the right image through the second feature extraction unit 25, .

After extracting the feature vectors of the left and right images, the object recognition apparatus 20 transmits the feature vectors of the left and right images to the feature vector matching unit 26 to perform stereo matching on the left and right images (104). By the stereo matching of the left and right images, feature vectors (common feature vectors) that are common to the right and left images and correspond to each other among the feature vectors of the right image and the left image are extracted.

By this stereo matching, the object recognition apparatus 20 extracts information (information of matched feature vectors) of the feature vectors common to both the left and right images (105). At this time, in addition to the matched feature vector information, the three-dimensional positions of the matched feature vectors are extracted together using the camera calibration data having the left and right image parameter information of the stereo camera 10.

The object recognition apparatus 20 acquires the information of each of the plurality of feature vectors of the left and right images, the information of the feature vectors common to both the left and right images, and the three-dimensional position information of the matched feature vectors, And recognizes the object by comparing it with the information of the object stored so as to correspond thereto. At this time, the object recognition apparatus 20 recognizes the Euclidean distance value, which is a virtual distance between the feature vectors of the right and left images and the feature vectors of the object stored in the database unit 27, The object can be correctly recognized by using the three-dimensional coordinate distance between vectors alone or both. That is, the object recognition performance is improved by using the relationship between the three-dimensional coordinates between the feature vectors matched with the right and left images.

FIG. 3 is a diagram for explaining feature vector matching of left and right images in FIG. 2, and FIG. 4 is a diagram showing matching results of FIG.

As shown in FIG. 3, when stereo matching is performed in the object recognition apparatus 20 according to the embodiment of the present invention, feature vectors that are matched in feature vectors extracted from left and right images, And excludes unmatched feature vectors. At this time, not all feature vectors are compared with each other, but feature vectors existing on a similar scale space are compared with each other for efficiency of computation time. The feature vectors extracted through the matching operation acquire the three-dimensional coordinate information of each feature vector using the camera calibration data.

및 우측 영상의 특징 벡터군

는 다음과 같다.As shown in FIG. 3, the feature vector group of the left image

And the feature vector group of the right image

Is as follows.

및 우측 영상의 ii번째 스케일 공간의 특징 벡터군

은 다음과 같다.Also, the feature vector group of the ii-th scale space of the left image

And the feature vector group of the ii-th scale space of the right image

Is as follows.

의 i 번째 특징 벡터(D차원)

및

의 i 번째 특징 벡터(D차원)

는 다음과 같다.Also,

The i-th feature vector (D-dimensional)

And

The i-th feature vector (D-dimensional)

Is as follows.

및

은 다음과 같다.In addition,

And

Is as follows.

FIG. 5 is a control flowchart for explaining an object recognition process of FIG. 2. FIG.

및 데이터베이스부(27) 내의 상의 물체별 좌측 영상의 특징 벡터군

를 매칭하는 과정(200), 이 둘 간을 순차적으로 매칭시켜 매칭 결과에 점수를 부여하는 과정(201), 우측 영상의 특징 벡터군

및 데이터베이스부(27) 내의 물체별 우측 영상의 특징 벡터군

를 매칭하는 과정(202), 이 둘 간을 순차적으로 매칭시켜 매칭 결과에 점수를 부여하는 과정(204), 그리고 좌우 합산 점수에 따라 물체를 판단하는 과정(204)을 수행함으로써 좌우 영상의 특징 벡터들을 데이터베이스부(27) 내에 물체별로 저장된 좌우 영상의 특징 벡터들과 순차적으로 매칭시키고, 좌우 영상을 매칭했을 때 유클리디언 거리값이 가까워 일치 정보가 높은 특징 벡터에 대해서 큰 점수를 주는 방식을 이용하여 좌우 영상의 특징 벡터들의 점수를 합산하여 물체를 정확히 인식한다.As shown in Fig. 5, the object recognition apparatus 20 recognizes a feature vector group

And the characteristic vector group of the left side image of the object in the database unit 27

(200), a step (201) of sequentially matching the two with each other to give a score to the matching result, a step

And the feature vector group of the right side image per object in the database unit 27

A step 204 of sequentially matching the two to assign a score to the matching result, and a step 204 of determining an object according to the left and right summation scores, Are successively matched with the feature vectors of the left and right images stored in the database unit 27 for each object, and when the left and right images are matched, the Euclidean distance values are close to each other and a large score is given to the feature vector having high matching information And adds the scores of the feature vectors of the left and right images to accurately recognize the object.

FIG. 6 is a control flowchart for explaining a process of recognizing an object in a manner different from that described in FIG.

The object recognition apparatus 20 recognizes an object accurately by using a distance similarity between three-dimensional coordinates, which is an actual distance between feature vectors matched in left and right images. This method is called pair-wise feature matching and voting. This pair-wise feature matching and boating is performed, for example, in such a manner that coordinates of two arbitrary feature vectors f1 and f2 of the right and left images are (x1, y1, z1), (x2, y2, (X3, y3, z3), (x4, y4, z4), and f1 and f3 match and f2 and f4 match, assuming that the coordinates of two feature vectors f3 and f4 corresponding to the object in By using the fact that the three-dimensional coordinate distance and the three-dimensional coordinate distance between f1 and f2 should be similar, accuracy and reliability of object recognition are improved by increasing the voting score when the distance is less than a certain reference condition. In the case of a stereo camera, since the accuracy of the three-dimensional coordinates deteriorates as the distance from the object increases, the condition ratios such as the following equations [1] to [3] are used instead of absolute threshold values.

식 [1]

Equation [1]

식 [2]

Equation [2]

식 [3]

Equation [3]

는 영상의 두 특징 벡터 사이의 3차원 거리이고,

는 DB상의 대응되는 두 특징 벡터 사이의 거리이고,

는

의 유사도 결정 조건 상수값이고, alpha는 거리에 따른 가중치를 주기 위한 기준 함수이다.here,

Is the three-dimensional distance between two feature vectors of the image,

Is the distance between two corresponding feature vectors on the DB,

The

, And alpha is a reference function for giving a weight according to the distance.

As shown in FIG. 7, when the camera and the feature vector are equal to or less than z_min, the reference value alpha is the same as the dist_ratio, but the reference value increases from z_min to z_max in the form of a linear function. Further, from the specific distance z_max or more, the same reference value is used without increasing the reference value any more.

와

를 매칭하는 과정(300), 데이터베이스부(27) 내에 물체별로 저장된 매칭된 좌우 영상의 특징 벡터군

에서 임의의 두 개의 특징 벡터간의 거리 dist _DB를 계산하는 과정(301), 현재의 좌우 영상의

에서 임의의 두 개의 특징 벡터간의 거리 dist_scene를 계산하는 과정(302), 위의 계산된 거리값들을 이용하여 위의 식 [1]을 계산하는 과정(303), 위의 계산된 식 [1]을 이용하여 식 [2]를 계산하는 과정(304), 식 [3]의 조건을 만족하는지를 판단하는 과정(305), 위의 식 [3]을 만족하지 않으면 작동 모드 300으로 리턴하여 이하의 작동을 수행하고, 위의 식 [3]을 만족하는 경우 물체의 점수를 증가시키는 과정(306), 물체의 점수 증가 후 모든 매칭이 완료했는지를 판단하는 과정(307), 모든 매칭이 완료되지 않은 경우 작동 모드 300으로 리턴하여 이하의 작동을 수행하고, 모든 매칭이 완료된 경우, 점수에 따라 물체를 판단하는 과정(308)을 수행한다.Accordingly, the object recognition apparatus 20 recognizes the feature vectors of the matched left and right images

Wow

A feature vector group of the matched left and right images stored for each object in the database unit 27,

(Step 301) of calculating a distance dist_DB between arbitrary two feature vectors in the current left /

A process 302 for calculating a distance dist_scene between arbitrary two feature vectors, a process 303 for calculating the above equation [1] using the calculated distance values, and the above calculated equation [1] A step 305 for calculating whether the condition of the equation [3] is satisfied or not, and if the condition [3] is not satisfied, the operation mode 300 is returned and the following operation is performed A step 306 for increasing the score of the object when the above equation [3] is satisfied, a step 307 for determining whether all matching is completed after the increase of the score of the object, Mode 300 to perform the following operation, and if all matching is completed, a step 308 of judging an object according to the score is performed.

Meanwhile, the object recognition apparatus 20 can recognize an object using a graph method instead of the above-described pair-wise feature matching and voting method.

The graphing method is a method of generalizing pair-wise feature matching. In this method, a matching graph is searched by integrating the relationship between matching feature loads and minutiae points. The computation time is a fairness It takes longer than pair-wise feature matching, but has better recognition performance.

In the above graph method, the graph model (G) of the object can be expressed as follows.

는 스테레오 매칭된 3차원 특징점들의 집합이고,

는 3차원 특징점들 간의 에지(Edge) 조합들의 집합이고,

는 3차원 특징점들의 디스크립터 집합(특징 벡터)이고,

는 3차원 특징점들 간의 3차원 유클리디언 거리이다.here,

Is a set of stereo-matched three-dimensional feature points,

Is a set of edge combinations between three-dimensional feature points,

Is a descriptor set (feature vector) of three-dimensional feature points,

Dimensional Euclidian distance between three-dimensional feature points.

If the model of each object in the database unit 27 is G1, G2, G3 ... GM and the model of the input image is Gs (having A three-dimensional feature points), then an arbitrary object model Gm (M, s, m) can be defined as follows.

Here, M _ai = {0,1}

At this time, the a-th 3-dimensional feature point of the input image and the matching value M _ai of the 3-dimensional feature point of the i-th object model have a value of 1 when they are matched with each other and a value of 0 when they are not matched.

).The basic concept of the graph method is to find the above matching matrix M (s, m) for each model (ie, find M (s, 1), M (s, 2) And selecting M (s, m) having the best matching score (

).

)가 최소가 되는 Mai 해를 구한다.To find the above M (s, m) matrices, the following function

) Is minimized.

이고, α는 조절 파라미터이고,

는 '임의의'를 나타내는 기호이고, here,

, ≪ / RTI > is an adjustment parameter,

Is an arbitrary symbol,

이다.

to be.

)의

와

를 다음과 같이 제안한다.In the embodiment of the present invention,

)of

Wow

Is proposed as follows.

)를 최소화하는 해를 찾기 위해서는 M_ai에 대해 미분한 다음의 식(Q_ai)을 이용한다.The above function (

), We use the following equation (Q _ai ) that is differentiated for M _ai .

)는 다음과 같이 표현할 수 있다.Using the above function,

) Can be expressed as follows.

이고, here,

ego,

According to the above method, since only one matching node is required in any column or row of M (s, m), a multi-correspondence check process is applied.

)가 최대값이 되는 매칭 행렬 M(s,m)을 찾은 후 다음 조건을 만족하는 물체 후보들을 찾고, 이 물체 후보들 중에서 매칭 점수가 가장 높은 물체를 좌우 영상에 있는 물체로 인식한다.The above function (

Finds a matching matrix M (s, m) having a maximum value, and finds object candidates satisfying the following conditions, and recognizes an object having the highest matching score among the object candidates as an object in the left and right images.

On the other hand, the object recognition apparatus 20 uses the matching feature vectors to verify the recognized object using two-dimensional homography and three-dimensional Euclidean Transform And further derives the three-dimensional position information.

The object recognition apparatus 20 first performs a verification operation using the Euclidean transformation on the three-dimensional plane. At this time, if there is not enough feature vector with matching 3-dimensional information, 3-D based operation is omitted and 2-dimensional homography based verification work is performed. In addition, when a proper rotation matrix (hereinafter referred to as R) and a translation matrix (hereinafter referred to as T) can not be obtained even when a feature vector having three-dimensional information is sufficient, Perform the operation.

= (x1, y1, z1)이고, 현재 영상의 물체에 있는 3차원 좌표를

= (x2, y2, z2)라고 하면, 4개 이상의 매칭 쌍을 통해 다음과 같은 R,T를 구한다.When the three-dimensional coordinates in the object in the database unit 27 are

= (x1, y1, z1) and the three-dimensional coordinates in the object of the current image are

= (x2, y2, z2), the following R and T are obtained through four or more matching pairs.

은

와 같아야 한다.If the obtained R and T are ideal,

silver

.

Accordingly, the coordinate values are substituted into the feature points in the object in the database 27, which are matched with the remaining R and T obtained from the arbitrary four points, and the error between the coordinate values on the actually matched image is constant Increases the number of inliers that are small matching pairs that are below the condition (TH). An example of obtaining the inlier counting condition is as follows.

The following method can be used to determine whether R and T are appropriate at the end.

If the following conditions are satisfied, R and T are judged to be appropriate. In this case, the object recognition apparatus 20 determines that the object recognition is correctly performed.

On the other hand, if the object recognizing apparatus 20 can not pass the above process, the object recognizing apparatus 20 performs a 2D homography-based verification operation to determine again whether the object recognition is correctly performed.

10: stereo camera 20: object recognition device
21: system management unit 22: image storage unit
23: feature vector management unit 24: first feature extraction unit
25: second feature extracting unit 26: feature vector matching unit
27: Database part 28: Object recognition part

Claims (9)

A stereo camera for outputting left and right images of the same object; And
Extracting left and right feature vectors from left and right images input from the stereo camera, extracting a plurality of common feature vectors common to the right and left images by matching the extracted left and right feature vectors with each other, And an object recognizing device for matching the plurality of common feature vectors commonly present in the left and right images with respect to the vectors and the objects stored in the database, and recognizing an object in the left and right images based on the matching result. The method according to claim 1,
Wherein the object recognition apparatus matches a plurality of common feature vectors common to the left and right images and a plurality of common feature vectors stored in the database unit and generates an Euclidean distance corresponding to the matched feature vectors of the common feature vectors, Wherein the score is given for each object in accordance with the value of the sum of the scores and the sum of the assigned scores to recognize the object having the highest sum as the object in the left and right images. The method according to claim 1,
The object recognition apparatus comprises:
A feature extraction unit for extracting a plurality of feature vectors from the plurality of feature points of the left image and extracting a plurality of feature vectors from the plurality of feature points of the right image;
A feature vector matching unit for matching a plurality of feature vectors extracted from the left image and a plurality of feature vectors extracted from the right image to extract a plurality of common feature vectors common to the left and right images; And
And an object recognition unit for recognizing an object in the left and right images by matching the extracted common feature vectors with a plurality of common feature vectors stored in the database unit. The method of claim 3,
Wherein the feature extraction unit detects vectors in a plurality of regions centered on the plurality of feature points, and extracts the set of detected vectors as a feature vector. The method according to claim 1,
The object recognition apparatus includes a plurality of left and right feature vector information extracted from the left and right images, a plurality of common feature vector information common to the left and right images, and three- For each of the matching results, a score corresponding to each of the matching results, and for summing the assigned scores, and for recognizing an object having the highest summed score as an object in the left and right images, Recognition system. The method according to claim 1,
The object recognition apparatus recognizes the object, and then, using at least one of a two-dimensional homography and a three-dimensional Euclidean transformation, The object recognition system. Receiving different left and right images taken of the same object;
Extracting left and right feature vectors from the input left and right images;
Extracting a plurality of common feature vectors common to the left and right images by matching the extracted left and right feature vectors with each other;
Matching the extracted common feature vectors with a plurality of common feature vectors common to left and right images for each object stored in the database unit;
And recognizing an object in the left and right images according to the matching result. 8. The method of claim 7,
The recognition of the object may be performed by assigning a score to each of the plurality of common feature vectors according to an Euclidean distance value for each object and summing the given points so that the object having the highest summed score An object recognition method comprising recognizing an object in a video. 8. The method of claim 7,
And verifying the recognized object using at least one of two-dimensional homography and a three-dimensional Euclidean transform on the matched plurality of common feature vectors after recognizing the object Comprising:

Images