KR101060321B1 - Vehicle detection method and system - Google Patents

Abstract

The present invention relates to a vehicle detection method and system. More particularly, the present invention relates to a vehicle detection method and system capable of quickly and accurately detecting a vehicle area from the captured image information by analyzing the captured image information. The present invention includes the steps of: (a) detecting a vehicle candidate region for each frame constituting the received image; And (b) a polynomial function extracting a feature vector from the vehicle candidate region for each frame and reducing the amount of computation by reducing the number of parameters calculated according to the input dimension of the feature vector or the feature vector and receiving the feature vector. And a vehicle region verification step of verifying whether the vehicle candidate region for each frame is an actual vehicle region by using a discrimination function that classifies the vehicle candidate region for each frame into a vehicle region or a non-vehicle region. According to the present invention, in the manual method of detecting the vehicle area by analyzing the input image information among the vehicle detection methods, it is possible to accurately and quickly detect the area of the vehicle and the non-vehicle area from the received image information to improve the processing speed. Has the possible effect.

Vehicle detection, classifier, normalization, preprocessing

Description

Method and system for detecting a vehicle

The present invention relates to a vehicle detection method and system. More particularly, the present invention relates to a vehicle detection method and system capable of quickly and accurately detecting a vehicle area from the captured image information by analyzing the captured image information.

In the autonomous driving system, detecting the other vehicle located in front of or behind the autonomous vehicle during autonomous driving is essential for maintaining the safety of the vehicle occupant by preventing traffic accidents. In order to prevent traffic accidents, it is very important to prevent traffic accidents. Therefore, various methods of detecting vehicles from captured images are used.

In general, a vehicle detection method can be largely divided into an active method and a passive method.

First, the active method uses a method of detecting a vehicle by using an active sensor such as a laser or a radar for detecting a vehicle, and has an advantage of directly receiving peripheral information because the active sensor is used. The problem is that the resolution is low, the price is high, and interference with other active sensors occurs.

In particular, in the case of the active method, it is possible to directly receive the surrounding information, but it is difficult to accurately determine what the input surrounding information means.

Next, the passive method analyzes the input image after receiving the captured image information using the optical sensor which is cheaper than the active sensor for detecting the vehicle. Therefore, the utilization is not limited to the vehicle detection field. , Pedestrian detection, or black box has the advantage that can be extended to various uses.

However, in the case of the passive method, various information can be analyzed by analyzing the received image after receiving the captured image information. However, in the case of the conventional passive method, the information is analyzed in proportion to the amount of information requiring analysis. Since the calculation speed required is slow, there is a problem that it is not suitable as a vehicle detection method for an automatic driving system that places importance on the processing speed.

An object of the present invention is to provide a vehicle detection method and system capable of quickly and accurately detecting a vehicle region from the photographed image information by analyzing image information that has been devised to solve the above problems.

According to an aspect of the present invention, there is provided a vehicle detection method comprising: (a) detecting a vehicle candidate region for each frame constituting an input image; And (b) a polynomial function extracting a feature vector from the vehicle candidate region for each frame and reducing the amount of computation by reducing the number of parameters calculated according to the input dimension of the feature vector or the feature vector and receiving the feature vector. And a vehicle region verification step of verifying whether the vehicle candidate region for each frame is an actual vehicle region by using a discrimination function that classifies the vehicle candidate region for each frame into a vehicle region or a non-vehicle region.

In addition, the vehicle detection system according to the present invention includes a vehicle candidate region detection unit for detecting a vehicle candidate region for each frame constituting the received image; And a polynomial function extracting a feature vector from the vehicle candidate region for each frame and reducing the amount of computation by reducing the number of parameters calculated according to the input dimension of the feature vector or the feature vector, and receiving the specific vector as a vehicle. And a vehicle region verification unit for verifying whether the vehicle candidate region for each frame is an actual vehicle region by using a discrimination function that classifies the candidate region into a vehicle region or a region other than the vehicle.

According to the present invention, in the manual method of detecting the vehicle area by analyzing the input image information among the vehicle detection methods, it is possible to accurately and quickly detect the area of the vehicle and the non-vehicle area from the received image information to improve the processing speed. Has the possible effect.

Hereinafter, exemplary embodiments of the present invention will be described in detail with reference to the accompanying drawings. First, it is to be noted that the components of each drawing have the same reference numerals as much as possible even though they are shown in different drawings. In the following description of the present invention, a detailed description of known functions and configurations incorporated herein will be omitted when it may make the subject matter of the present invention rather unclear. In addition, preferred embodiments of the present invention will be described below, but the technical idea of the present invention is not limited thereto, but may be implemented by those skilled in the art.

1 is a block diagram of a vehicle detection system according to a preferred embodiment of the present invention.

As shown in FIG. 1, the vehicle detection system 1 according to an exemplary embodiment of the present invention includes an image input unit 10, an image preprocessor 20, a vehicle candidate region detector 30, and a vehicle candidate region normalizer 40. ), A vehicle region verification unit 50, a vehicle region revalidation unit 60, and a vehicle template database 70.

The image input unit 10 receives an image photographing the front or rear region of the driving vehicle. In this case, the photographed front or rear region image of the driving vehicle may be input in the form of a moving image or a still image, and the photographing of the front or rear region of the driving vehicle is performed using a wide angle camera to secure a wider field of view. Can be.

The image preprocessor 20 performs preprocessing for each frame constituting the received image.

At this time, histogram equalization is performed to increase the identification of each frame of the input image as a preprocessing for each frame of the input image, and Gaussian filtering is performed to remove noise for each frame of the input image. can do.

As such, the reason for performing the preprocessing for each frame of the input image by the image preprocessor 20 is that the image input unit 10 minimizes the influence of ambient light such as illumination that can be included in the input image. This is to prevent a drop in accuracy when detecting a vehicle candidate area for each frame.

The vehicle candidate region detector 30 detects the vehicle candidate region for each frame of the image preprocessed by the image preprocessor 20.

In this case, to detect a vehicle candidate region for each frame of the preprocessed image, a knowledge-based method using an image symmetry, color, or shadow information, a stereo using Inverse Perspective Mapping (IPM) or Dsipari Map A stereo-based method or a motion-based method using a motion of an object such as an optical flow may be used. Since the methods are already known, the detailed description thereof will be omitted.

As such, the reason for detecting the vehicle candidate area for each frame of the preprocessed image by using the above methods is to detect the vehicle candidate area including all regions having a similar shape to the vehicle for each frame of the preprocessed image. This is for the vehicle detection system of the present invention to have a high detection rate.

The vehicle candidate region normalization unit 40 performs normalization for each frame of the vehicle candidate region detected by the vehicle candidate region detection unit 30.

In this case, a detailed configuration of the vehicle candidate region normalization unit 40 will be described below with reference to FIG. 2A.

As such, the reason for performing normalization on the vehicle candidate region for each frame detected by the vehicle candidate region detector 30 is detected by the vehicle candidate region detector 30 so as to have a high detection rate. This is to remove an unnecessary background area included in the vehicle candidate area for each frame.

The vehicle region verification unit 50 extracts a feature vector from each of the normalized vehicle candidate regions for each frame, and reduces the computational amount by reducing the number of parameters calculated according to the input dimension of the feature vector or the feature vector. It is verified whether the vehicle candidate region for each frame is an actual vehicle region by receiving a vector as a discrimination function for classifying the vehicle candidate region for each frame into a vehicle region or an area other than a vehicle.

Here, the reason for using the polynomial function which reduces the calculation amount as the discriminant function for vehicle area verification is that when the polynomial function is used as the discriminant function for vehicle area verification, the amount of computation of the polynomial function is increased as the input dimension of the feature vector is increased. This is to prevent the vehicle speed of the vehicle area verification from decreasing significantly.

In this case, a detailed configuration of the vehicle area verification unit 50 will be described below with reference to FIGS. 3A and 3B.

The vehicle region revalidation unit 60 performs vehicle region revalidation on an unverified region among the vehicle candidate regions for each frame on which the vehicle region verification is performed, and the vehicle region template database 70 revalidates the vehicle region. The vehicle template image for performing the above is stored in advance.

In this case, the vehicle region re-validation may be performed by matching the previously stored vehicle template image with an unverified region of the vehicle candidate region for each frame on which the vehicle region verification is performed.

2A is a detailed block diagram of a vehicle candidate region normalization unit according to a preferred embodiment of the present invention, and FIG. 2B is a reference diagram for a normalization process of a vehicle candidate region according to a preferred embodiment of the present invention.

As shown in FIG. 2A, the vehicle candidate region normalization unit 40 according to an exemplary embodiment of the present invention may include an edge image converter 42, an edge histogram generator 44, a local maximum value region calculator 46, And a resizing unit 48.

The edge image converter 42 converts the vehicle candidate region (a) of FIG. 2B of each frame of the received image detected by the vehicle candidate region detection unit 30 into an edge image. (B of FIG. 2B). ))

The edge histogram generator 44 generates a horizontal edge histogram and a vertical edge histogram from the vehicle candidate region of each frame of the input image converted into an edge image. (FIG. 2B (c)).

The local maximum value area calculator 46 calculates a region having a local maximum value among the horizontal edge histogram and the vertical histogram generated from the vehicle candidate area for each frame of the input image.

The resizing unit 48 may be configured according to the horizontal coordinates of the region having the local maximum value among the horizontal edge histograms calculated from the vehicle candidate regions for each frame of the input image and the vertical coordinates of the region having the local maximum value among the vertical edge histograms. Resizing of the detected vehicle candidate region for each frame is performed ((d) of FIG. 2B).

As such, the vehicle candidate region normalizer 40 performs normalization on the vehicle candidate region for each frame of the image detected by the vehicle candidate region detector 30 to include the frame in the vehicle candidate region for each frame of the received image. Since the surrounding image or the wrongly detected part is removed, performance degradation of the vehicle region verification unit 50 and the vehicle region revalidation unit 60 can be prevented.

3A is a detailed block diagram of a vehicle region verification unit according to a first embodiment of the present invention.

As shown in FIG. 3, the vehicle region verification unit 50a according to the first embodiment of the present invention includes a feature vector extractor 52a, a cross validation performer 54a, a multivariate polynomial function calculator 56a, and An area verification unit 58a is included.

The feature vector extractor 52a extracts a basis vector for the vehicle candidate region for each frame, and then extracts a feature vector for the vehicle candidate region for each frame by projecting the vehicle candidate region for each frame onto the basis vector. .

In this case, as one of methods for extracting the feature vector, Principal Component Analysis (PCA) may be used.

Principal component analysis extracts a basis vector that can characterize a change in the vehicle candidate region for each normalized frame from the vehicle candidate region for each normalized frame, and then uses the basis vector to determine the vehicle candidate region for each normalized frame. The approximation can be made as follows.

First, when the number of normalized vehicle candidate regions for each frame is M, the average of the vehicle candidate regions for each frame may be obtained as shown in Equation 1 below.

은 상기 정규화된 각 프레임별 차량후보 영역의 평균,

(n=1,...M)는 상기 정규화된 각 프레임별 차량후보 영역을 의미한다.From here,

Is an average of the normalized vehicle candidate area for each frame,

(n = 1, ... M) means the normalized vehicle candidate area for each frame.

Next, the covariance of the normalized vehicle candidate region for each frame and the normalized vehicle candidate region for each frame calculated in Equation 1 is obtained as in Equation 2 below.

(n=1,2,...M), 및

는

의 전치행렬을 의미한다.Where C is the covariance,

(n = 1,2, ... M), and

Is

Means transpose of.

In this case, the covariance calculated in Equation 2 is a value representing a change in the normalized vehicle candidate region for each frame, and a basis vector is obtained as shown in Equation 3 below to divide these values into basis vectors.

는 기저 벡터, 및

는 고유값을 의미한다.Where C is the covariance,

The basis vector, and

Means eigenvalues.

In this case, the total M candidate frames for each normalized frame may be represented by a linear combination of total M basis vectors, that is, the linear projection of the normalized vehicle candidate regions for each normalized frame into the base space where the basis vector is located. It can be expressed as shown in Equation 4 below.

는 상기 정규화된 각 프레임별 차량 후보영역,

는 기저 벡터,

는 =(k=1,2....,M)인 상기 정규화된 각 프레임별 차량 후보영역의 투영 계수를 의미한다.From here,

Denotes a vehicle candidate region for each normalized frame;

Vector,

Denotes the projection coefficient of the normalized vehicle candidate region for each frame of which = (k = 1,2 ...., M).

=[x1, x2, ...., xM]과 같이 나타낼 수 있다.Accordingly, the feature vector extracted from the M candidate regions for each normalized frame calculated by Equation 4 is

It can be expressed as = [x1, x2, ...., xM].

The cross validation performing unit 54a performs cross-validation on the extracted feature vectors in order to reduce an input dimension of the feature vectors extracted from each normalized vehicle candidate region for each frame.

In this case, the reason for performing cross-verification on the feature vector extracted from the normalized vehicle candidate region for each frame is as follows.

In case of setting a multivariate polynomial function of r-order having l input dimension as a discriminant function for vehicle region verification of the normalized vehicle candidate region for each frame, the multivariate polynomial function of r-order having l input dimension The total number of parameters of is equal to l ^r , and as a result, the amount of computation of the multivariate polynomial increases greatly as the input dimension or order increases, thereby reducing the input dimension of the feature vector.

In this case, the cross-validation of the feature vector extracted from the normalized vehicle candidate region for each frame may be performed as follows.

=[x1 x2]이고 상기 다변량 다항 함수의 산출을 위해 다항 전개된 다변량 다항 함수의 입력 벡터

는

=[1 x₁ x₂ x₁₂ x₁x₂ x₂₂]^T 로써 총 6개의 항으로 이루어진다.For example, a multivariate polynomial function with input dimension l = 2 and order r = 2 has a feature vector with input dimension l = 2 which is an input of the multivariate polynomial function.

input vector of the multivariate polynomial function multiplied for the calculation of the multivariate polynomial function == x1 x2]

Is

= [1 x ₁ x ₂ x ₁₂ x ₁ x ₂ x ₂₂ ] ^T, consisting of a total of six terms.

=[1 x₁ x₂ x₁₂ x₁x₂ x₂₂]^T의 6개의 항에 대한 교차 검증을 수행하여 6개의 항중 2개의 항을 선택할 수 있으며, 상기 교차 검증의 결과 선택된 2개의 항에 대한 파라미터를 추정하여 다변량 다항 함수를 산출할 수 있다.At this time, to reduce the input dimension of the feature vector

= [1 x ₁ x ₂ x ₁₂ x ₁ x ₂ x ₂₂ ] Two terms of six terms can be selected by performing a cross-validation on six terms of ^T , and as a result of the cross-validation, The parameter can be estimated to yield a multivariate polynomial function.

Using the method described above, since two terms are used for an equation that originally requires six terms, a multivariate polynomial function that is faster and does not deteriorate due to a small amount of calculation can be calculated.

Further, in addition to the method of performing the cross-validation on the extracted feature vector, the cross-validation of the features of the input data is performed to reduce the number of dimensions, which is required for calculating the multivariate polynomial function. Reducing the number of parameters can also be used.

For example, if one feature of the input data has a size of 700 × 1 and you want to use 200 of those features, you can perform cross-validation on the 700 feature values to generate 200 of the 700 feature values. The multivariate polynomial is calculated by selecting the characteristic values and then obtaining the parameters for the selected 200 characteristic values.

By using the method as described above, the calculation amount of the multivariate polynomial is reduced, so that it is possible to calculate the multivariate polynomial which has a fast calculation speed and no deterioration in performance.

The multivariate polynomial function calculation unit 56a calculates a multivariate polynomial function that takes the cross-validated feature vector as an input.

Here, the multivariate polynomial function is used widely in the fields of optimization, signal processing, and pattern classification because it provides a method to effectively express input and output relations in the form of polynomial when the input and output of the function are nonlinear. It can be expressed as Equation 5.

는 다변량 다항 함수, n1,n2,....nl은 n1+n2+.....+nl≤ r(이때, r은 다변량 다항식 함수의 근사치 계산을 위한 차수)인 조건을 만족시키는 음수가 아닌 상수,

는 다변량 다항 함수 산출을 위해 추정될 파라미터 벡터,

는 상기 각 프레임별 차량 후보 영역으로부터 추출된 특징 벡터, x₁,x₂,...x_l은 상기 각 프레임별 차량 후보 영역으로부터 추출된 투영 계수, l은 특징 벡터의 입력 차원의 개수, 및 K는 다변량 다항식의 총 항수를 의미한다.From here,

Is a non-negative that satisfies the condition that is a multivariate polynomial function, n1, n2, .... nl is n1 + n2 + ..... + nl≤ r (where r is an order for approximation of the multivariate polynomial function) a constant,

Is the parameter vector to be estimated for calculating the multivariate polynomial function,

Is a feature vector extracted from the vehicle candidate region for each frame, x ₁ , x ₂ , ... x _l is a projection coefficient extracted from the vehicle candidate region for each frame, l is the number of input dimensions of the feature vector, and K is the total number of terms in the multivariate polynomial.

For example, a multivariate polynomial function having an order of 2 for approximation calculation and the number of input dimensions of a feature vector is 2 (r = 2, l = 2) may be expressed as Equation 6 below.

는 다변량 다항 함수,

는

=[α₁ α₂ α₃ α₄ α₅ α₆]인 다변량 다항 함수 산출을 위해 추정될 파라미터 벡터,

는 상기 각 프레임별 차량 후보 영역으로부터 추출된 특징 벡터,

=[1 x₁ x₂ x₁₂ x₁x₂ x₂₂]^T 는 입력차원 2를 갖는 특징벡터

=[x₁ x₂]가 다변량 다항 함수 산출을 위해 다항 전개된 다변량 다항 함수의 입력벡터, 및

는 의 전치 행렬을 의미한다.From here,

Is a multivariate polynomial function,

Is

the parameter vector to be estimated for calculating the multivariate polynomial function = [α ₁ α ₂ α ₃ α ₄ α ₅ α ₆ ],

Is a feature vector extracted from the vehicle candidate region for each frame,

= [1 x ₁ x ₂ x ₁₂ x ₁ x ₂ x ₂₂ ] ^T is a feature vector with input dimension 2

= [x ₁ x ₂ ] is the input vector of the multivariate polynomial function polynomial expanded to yield the multivariate polynomial function, and

Means transpose of.

In addition, since the number of input dimensions of the feature vector is l = 2, a minimum squared error minimization for reducing the error of the multivariate polynomial function is given when m data points having a total number of terms K = 6 and m> K are given. The method can be applied to calculate a parameter vector for the calculation of multivariate polynomials.

In this case, the least-squares error minimization method may be represented by Equation 7 below.

는 최소 자승 오차 최소화(least-squares error minimization) 값,

는

=[y₁,....y_m]^T (

∈R^m)인 트레이닝 데이터(Training data)를 통하여 알 수 있는 추정 벡터(Inference vector), 및 P(P∈R^mxK)는

의 자코비안 행렬(Jacobian matrix)를 의미한다.From here,

Is the least-squares error minimization value,

Is

= [y ₁ , .... y _m ] ^T (

The inference vector and P (P∈R ^mxK ), which can be known from the training data of ∈R ^m ,

Means the Jacobian matrix.

를 아래의 수학식 8과 같이 나타낼 수 있다.In addition, using the equation (7) is a parameter vector for multivariate polynomial calculation

May be expressed as in Equation 8 below.

는 =[y₁,....y_m]^T (

∈R^m)인 트레이닝을 통하여 알 수 있는 인 퍼런스(inference) 벡터, P(P∈R^mxK)는

의 자코비안 행렬(Jacobian matrix)로써 아래의 수학식 9와 같이 나타낼 수 있다.From here,

== y ₁ , .... y _m ] ^T (

The inference vector, P (P∈R ^mxK ), obtained from training ^트레이닝 R ^m is

The Jacobian of (Jacobian matrix) can be expressed as shown in Equation 9 below.

In this case, in the component x _{j , k} (j = 1,2, k = 1, ...., m) of the Jacobian matrix P, j is the number of inputs and k is the number of instances included in the input. From the parameter vector calculated by Equation 8, a multivariate polynomial function having the feature vector as an input can be calculated.

The area verification unit 58a performs the vehicle area verification on the vehicle candidate area for each frame by using the multivariate polynomial function calculated by the multivariate polynomial function calculation unit 56a as a discrimination function.

At this time, the vehicle region verification for the vehicle candidate region for each frame is performed by plotting the projection coefficient of the vehicle candidate region for each frame calculated in Equation 4 in the feature space where the discrimination function is located. It is classified into two classes of vehicle area and non-vehicle area, and the discrimination function is set to '1' which is Negative (non-vehicle area) and 'Other' which is positive (area area). Divided by 0 ', it is classified as Equation 10 below.

Here, α ^T g _PM (x) means a multivariate polynomial that is a discrimination function.

3B is a detailed block diagram of the vehicle region verification unit according to the second embodiment of the present invention.

As shown in FIG. 3B, the vehicle region verifying unit 50b according to the second embodiment of the present invention includes a feature vector extracting unit 52b, a reduced multivariate polynomial function calculating unit 54b, and an area verifying unit 56b. It includes.

The feature vector extracting unit 52b extracts a basis vector for the vehicle candidate region for each frame, and then extracts a feature vector for the vehicle candidate region for each frame by projecting the vehicle candidate region for each frame onto the basis vector. .

In this case, the method of extracting the feature vector for the vehicle candidate region for each frame by the feature vector extractor 52b has been described above, and thus will be omitted.

The reduced multivariate polynomial function calculation unit 54b receives the feature vector and calculates a reduced multivariate polynomial function in which the number of parameters calculated according to the input feature vector is reduced.

Here, the reduced multivariate polynomial function is a modification of the existing multivariate polynomial function in order to reduce the number of parameters of the multivariate polynomial function, which greatly increases according to the input dimensions l and the constant r. .

는 축소된 다변량 다항 함수,

는 축소된 다변량 다항 함수 산출을 위해 추정될 파라미터 벡터,

는 상기 각 프레임별 차량 후보 영역으로부터 추출된 특징 벡터,

는 특징 벡터

가 축소된 다변량 다항 함수 산출을 위해 다항 전개된 축소된 다변량 다항 함수의 입력, x_j(j=1,2..k,l)는 특징 벡터 의 계수,α₀, α_kj, α_{rl +j}, α_j 는 축소된 다변량 다항 산출을 위해 추정될 파라미터 벡터의 계수, l은 입력 차원의 개수, 및 r은 축소된 다변량 다항 함수의 근사치 계산을 위한 차수를 의미한다.From here,

Is a reduced multivariate polynomial function,

Is the parameter vector to be estimated for the reduced multivariate polynomial function,

Is a feature vector extracted from the vehicle candidate region for each frame,

Feature vector

For inputting the multinomial reduced multivariate polynomial function, x _j (j = 1,2..k, l) is the coefficient of the feature vector, α ₀ , α _kj , α _{rl + j} and α _j is the coefficient vector of parameters to be estimated for the reduced multivariate polynomial calculation, l is the number of input dimensions, and r refers to the order of approximation for the calculation of the reduced multivariate polynomial function.

는 아래의 수학식 12와 같이 나타낼 수 있다.At this time, the parameter vector for calculating the reduced multivariate polynomial function

May be expressed as in Equation 12 below.

는 파라미터 벡터, b는 정규화 계수(Regulization factor),

(y∈Rm)는 트레이닝 데이터(Training data)를 통하여 알 수 있는 추정 벡터(Inference vector), 및 P는

의 자코비안 행렬(Jacobian matrix)로써 상기수학식 9와 같이 나타낼 수 있다.From here,

Is the parameter vector, b is the regulization factor,

(y∈Rm) is an inference vector that can be known through training data, and P is

The Jacobian matrix can be represented by Equation (9).

를 이용하여 상기 특징 벡터를 입력으로 하는 축소된 다변량 다항 함수를 산출할 수 있다.In this case, in the component x _{j , k} (j = 1,2, k = 1, ...., m) of the Jacobian matrix P, j is the number of inputs and k is the number of instances included in the input. , A parameter vector obtained by using Equation 12

A reduced multivariate polynomial function using the feature vector as an input can be calculated.

In the case of the reduced multivariate polynomial function, the total terms K = l + r + l (2r-1) (K is the total terms of the reduced multivariate polynomial function, l is the input dimension, and r is the order). Since the linear increase with respect to the number of terms, the total number of terms can be significantly reduced compared to the multivariate polynomial function that increases nonlinearly with the number of parameters that increase by l ^r for the l and r values. It is possible to increase speed while maintaining performance.

The area verification unit 56b performs vehicle area verification on the vehicle candidate area for each frame by using the reduced multivariate polynomial function calculated by the reduced multivariate polynomial calculation unit 54b as a discrimination function.

In this case, the vehicle region verification for the vehicle candidate region for each frame is performed by plotting the projection coefficient of the vehicle candidate region for each frame calculated in Equation 4 in the feature space where the discrimination function is located. It is classified into two classes of vehicle area and non-vehicle area, and the discrimination function is set to '1' which is Negative (non-vehicle area) and 'Other' which is positive (area area). It is classified as Equation 13 below by dividing by 0 '.

Here, α ^T g _RM (x) means a reduced multivariate polynomial function that is a discriminant function.

The vehicle detection system of the present invention receives a front or rear image of a photographed vehicle and performs preprocessing, vehicle candidate region detection, and normalization for each frame constituting the received image.

The vehicle candidate region for each frame is classified into a vehicle region or a non-vehicle region by changing a multivariate polynomial function according to a multivariate polynomial model so as to specialize in vehicle verification. versus

A vehicle region verification is performed, and a vehicle region revalidation is performed using a template based method on an unverified region among the vehicle candidate regions for each frame on which the vehicle region verification is performed.

Therefore, it is possible to increase the processing speed while maintaining the vehicle detection performance on the captured image.

4 is a flow chart of a vehicle area detection method according to a preferred embodiment of the present invention.

As shown in FIG. 4, the vehicle area detection method according to the preferred embodiment of the present invention includes the following steps performed in time series in the vehicle area detection system 1.

In S10, the image input unit 10 receives an image of photographing a front or rear region of the driving vehicle. In this case, the photographed front or rear region image of the driving vehicle may be input in the form of a moving image or a still image, and the photographing of the front or rear region of the driving vehicle is performed using a wide angle camera to secure a wider field of view. Can be.

In S20, the image preprocessor 20 performs preprocessing for each frame of the image input in S10.

At this time, histogram equalization for enhancing the identification of each frame of the input image as a preprocessing for each frame of the input image in S20 and Gaussian filtering to remove noise for each frame of the input image. ) Can be performed.

In S30, the vehicle candidate area detection unit 30 detects the vehicle candidate area for each frame of the received image that has been preprocessed in S20.

In this case, a knowledge-based method, an inverse perspective mapping (IPM), or a dsipari map using symmetry, color, or shadow information of the image to detect the vehicle candidate region for each frame of the received image in S30. A stereo-based method or a motion-based method using a motion of an object such as an optical flow may be used. Since the methods are already known, a detailed description thereof will be omitted.

In S40, the vehicle candidate region normalization unit 40 normalizes the vehicle candidate region for each frame of the received image detected in S30.

In this case, a detailed flowchart of S40 will be described below with reference to FIG. 5.

In S50, the vehicle region verifying unit 50 extracts a feature vector from each vehicle candidate region normalized in S40, and reduces the calculation amount by reducing the number of parameters calculated according to the input dimension of the feature vector or the feature vector. A function is input to the feature vector to determine whether the vehicle candidate region for each frame is classified as a vehicle region or a non-vehicle region to verify whether the vehicle candidate region for each frame is an actual vehicle region.

In this case, a detailed flowchart of S50 will be described below with reference to FIGS. 6A and 6B.

If the second vehicle region verification unit 60 performs a vehicle region revalidation on an unverified region among the vehicle candidate regions for each frame on which the vehicle region verification is performed, the termination is performed.

At this time, the vehicle region re-validation in S60 is performed by the vehicle region revalidation unit 60 among the vehicle template images previously stored in the vehicle template database 70 and the vehicle candidate regions for each frame in which the vehicle region verification is performed. It can be done in a way to match the unverified areas.

5 is a detailed flowchart of the S40.

In S42, the edge image converter 42 converts the vehicle candidate region for each frame of the input image detected in S30 into an edge image.

In operation S44, the edge histogram generator 44 generates a horizontal edge histogram and a vertical edge histogram for the vehicle candidate region of each frame of the input image converted into an edge image.

In S46, the local maximum value area calculator 46 calculates an area having a local maximum value among the horizontal edge histogram and the vertical edge histogram for each frame candidate vehicle of the received image.

In operation S48, the resizing unit 48 detects each frame according to the horizontal coordinates of the region having the local maximum value among the horizontal edge histograms of each frame of the input image and the vertical coordinates of the region having the local maximum value among the vertical edge histograms. Termination is performed when resizing is performed for each vehicle candidate region.

6A is a detailed flowchart of S50 according to the first embodiment of the present invention.

In S52a, the feature vector extracting unit 52a extracts a basis vector of the vehicle candidate region for each frame of the input image normalized in the S40, and then projects the vehicle candidate region for each frame onto the basis vector for each frame. The feature vector of the star vehicle candidate region is extracted.

In S54a, the cross validation performing unit 54a performs cross validation on the extracted feature vector to reduce an input dimension of the extracted feature vector.

In S56a, the multivariate polynomial function calculating unit 56a calculates a multivariate polynomial function that takes a feature vector on which the cross-validation is performed.

In S58a, when the area verification unit 58a performs the vehicle area verification for the vehicle candidate area for each frame using the calculated multivariate polynomial function as the discrimination function, the termination is completed.

At this time, since the detailed process for the S52a to S58a has been described above, it will be omitted.

6B is a detailed flowchart of S50 according to the second embodiment of the present invention.

In S52b, the feature vector determiner 52a extracts a basis vector of the vehicle candidate region for each frame of the received image normalized in S40, and then projects the vehicle candidate region for each frame onto the basis vector to determine each frame. The feature vector of the star vehicle candidate region is extracted.

In operation S54b, the reduced multivariate polynomial function calculation unit 54b calculates a reduced multivariate polynomial function using the determined feature vector as an input.

In S56b, when the area verification unit 56b performs the vehicle area verification for the vehicle candidate area for each frame using the calculated reduced multivariate polynomial function as the discrimination function, the termination is completed.

At this time, since the detailed process for the S52b to S56b has been described above, it will be omitted.

The vehicle detection method of the present invention receives a front or rear image of a photographed vehicle and performs preprocessing, vehicle candidate region detection, and normalization for each frame constituting the received image.

The vehicle candidate region for each frame is classified into a vehicle region or a non-vehicle region by changing a multivariate polynomial function according to a multivariate polynomial model so as to specialize in vehicle verification. Vehicle area verification is performed, and vehicle area revalidation is performed using a template based method on an unverified area among vehicle candidate areas for each frame on which the vehicle area verification is performed.

Therefore, it is possible to increase the processing speed while maintaining the vehicle detection performance on the captured image.

The above description is merely illustrative of the technical idea of the present invention, and various modifications, changes, and substitutions may be made by those skilled in the art without departing from the essential characteristics of the present invention. It will be possible. Therefore, the embodiments disclosed in the present invention and the accompanying drawings are intended to illustrate and not to limit the technical spirit of the present invention, and the scope of the technical idea of the present invention is not limited by these embodiments and the accompanying drawings . The protection scope of the present invention should be interpreted by the following claims, and all technical ideas within the equivalent scope should be interpreted as being included in the scope of the present invention.

According to the present invention, in the manual method of detecting the vehicle area by analyzing the input image information among the vehicle detection methods, it is possible to accurately and quickly detect the area of the vehicle and the non-vehicle area from the received image information to improve the processing speed. Because of this, it can be used for the automatic driving system using the manual method.

1 is a block diagram of a vehicle detection system according to a preferred embodiment of the present invention;

2A is a detailed block diagram of a vehicle candidate region normalization unit according to a preferred embodiment of the present invention;

2B is a reference diagram for a normalization process of a vehicle candidate region according to an exemplary embodiment of the present invention;

3A is a detailed block diagram of a vehicle region verification unit according to a first embodiment of the present invention;

3B is a detailed block diagram of a vehicle region verification unit according to a second embodiment of the present invention;

4 is a flow chart of a vehicle area detection method according to a preferred embodiment of the present invention;

5 is a detailed flowchart of the S40, and

6A is a detailed flowchart of S50 according to the first embodiment of the present invention, and

6B is a detailed flowchart of S50 according to the second embodiment of the present invention.

<Brief description of the main parts of the drawings>

(1): vehicle detection system (10): video input unit

20: Image preprocessing unit 30 Vehicle candidate area detection unit

40: vehicle candidate region normalization unit 42: edge image conversion unit

44: edge histogram generator 46 local maximum value area calculator

(48): resizing unit (50a, 50b): vehicle area verification unit

(52a, 52b): feature vector extracting unit (54a): cross-validating unit

(54b): Reduced multivariate polynomial function calculation unit

56a: multivariate polynomial function calculating section 56b, 58a: region verifying section

60: vehicle area revalidation unit 70: vehicle template database

Claims (10)

(a) detecting a vehicle candidate region for each frame constituting the received image; And (b) a polynomial function extracting a feature vector from the vehicle candidate region for each frame and reducing the amount of computation by reducing the number of parameters calculated according to the input dimension of the feature vector or the feature vector, and receiving the feature vector from each frame Vehicle region verification step of verifying whether the vehicle candidate region for each frame is an actual vehicle region by using a discrimination function for classifying each vehicle candidate region into a vehicle region or a background region. Wherein the feature vector is extracted by extracting a basis vector capable of characterizing a change in vehicle candidate region for each frame from the vehicle candidate region for each frame and then projecting the vehicle candidate region for each frame onto the basis vector. Vehicle detection method, characterized in that. The method of claim 1, In step (b), (b1) extracting a basis vector for the vehicle candidate region for each frame, and then extracting a feature vector for the vehicle candidate region for each frame by projecting the vehicle candidate region for each frame onto the basis vector; (b2) performing cross-validation on the extracted feature vectors to reduce an input dimension of the extracted feature vectors; (b3) calculating a multivariate polynomial function using the feature vector on which the cross-validation is performed; And and (b4) performing vehicle region verification on the vehicle candidate region for each frame using the calculated multivariate polynomial function as a discrimination function. The method of claim 1, In step (b), (b1) extracting a basis vector for the vehicle candidate region for each frame, and then extracting a feature vector for the vehicle candidate region for each frame by projecting the vehicle candidate region for each frame onto the basis vector; (b2) calculating a reduced multivariate polynimial function with the feature vector as an input and reducing the number of parameters calculated according to the input feature vector; And and (b3) classifying the vehicle candidate region for each frame into a vehicle region or a background region using the calculated reduced multivariate polynomial function as a discrimination function to perform vehicle region verification. The method of claim 1, Following step (a), (a1) further comprising normalizing the vehicle candidate region for each frame; Step (a1), (a11) converting the detected vehicle candidate regions for each frame into edge images; (a12) generating a horizontal edge histogram and a vertical edge histogram for the vehicle candidate region for each frame converted into the edge image; calculating a region having a local maximum value of a horizontal edge histogram and a vertical edge histogram for each of the generated vehicle candidate regions for each frame; And (a14) The detected vehicle for each frame according to the calculated horizontal coordinate of the region having the local maximum value in the horizontal edge histogram for the vehicle candidate region for each frame and the vertical coordinate of the region having the local maximum value in the vertical edge histogram. And resizing the candidate area. The method of claim 1, Following step (b), (c) performing a vehicle region revalidation by matching an unverified vehicle candidate region among previously stored vehicle candidate regions of each frame in which the vehicle region verification is performed and a previously stored vehicle template image; Detection method. A vehicle candidate region detector for detecting a vehicle candidate region for each frame constituting the input image; And The vehicle candidate for each frame by receiving the feature vector is a polynomial function that extracts a feature vector from the vehicle candidate region for each frame and reduces the amount of computation by reducing the number of parameters calculated according to the input dimension of the feature vector or the feature vector. Vehicle region verification unit for verifying whether the vehicle candidate region for each frame is an actual vehicle region by using a discrimination function that classifies the region into a vehicle region or a background region. Wherein the feature vector is extracted by extracting a basis vector capable of characterizing a change in vehicle candidate region for each frame from the vehicle candidate region for each frame and then projecting the vehicle candidate region for each frame onto the basis vector. Vehicle detection system, characterized in that. The method of claim 6, The vehicle area verification unit, A feature vector extracting unit extracting a feature vector for the vehicle candidate region for each frame by extracting a basis vector for the vehicle candidate region for each frame and then projecting the vehicle candidate region for each frame onto the basis vector; Cross-validation unit for performing cross-validation on the extracted feature vector in order to reduce the input dimension of the feature vector, Multivariate polynomial function using the feature vector on which the cross-validation is performed And a region verifying unit configured to perform vehicle region verification on the vehicle candidate region for each frame using the calculated multivariate polynomial function as a discrimination function. The method of claim 6, The vehicle area verification unit, A feature vector extractor for extracting a feature vector for the vehicle candidate region for each frame by extracting a basis vector for the vehicle candidate region for each frame and then projecting the vehicle candidate region for each frame to the base vector; A reduced multivariate polynomial calculator for calculating a reduced multivariate polynomial function in which the number of parameters calculated according to the input feature vector is reduced, and the reduced multivariate polynomial function is determined And a region verifying section for performing vehicle region verification as a function. The method of claim 6, Further comprising a vehicle candidate region normalization unit for normalizing the vehicle candidate region for each frame, The vehicle candidate area normalization unit, An edge image converter for converting the detected vehicle candidate region for each frame into an edge image; an edge histogram generator for generating a horizontal edge histogram and a vertical edge histogram from the vehicle candidate region for each frame converted into the edge image; A local maximum value area calculator for calculating a region having local maximum values of the horizontal edge histogram and the vertical edge histogram of each vehicle candidate region for each frame; and a local maximum among the calculated horizontal edge histograms of the vehicle candidate region for each frame. And a resizing unit for resizing the detected vehicle candidate region for each frame according to the vertical coordinates of the region having the local maximum value among the horizontal coordinates and the vertical edge histogram of the region having the value. The method of claim 6, And a vehicle region revalidation unit configured to perform vehicle region revalidation on an unverified region of the vehicle candidate regions for each frame on which the vehicle region verification is performed.

Images