KR20140118157A - System and Method for alarming collision of vehicle with support vector machine - Google Patents

Abstract

The present invention relates to a vehicle collision warning system by a support vector machine which includes a vehicle position recognition unit which can recognize a position of a vehicle; a sensor unit which uses ultrasonic waves to detect an obstacle; an image device obtaining an image data of the obstacle; and an obstacle pattern classification unit which determines a type of the obstacle based on information collected from the sensor and the image device. Here, the obstacle pattern classification unit recognizes a pattern and a type of the obstacle extracted by the image device by the support vector machine based on the image data collected by the image device. Accordingly, by using the support vector machine, which is combining a main component analysis method and an independent component analysis method, and using a residual error value as a feature value which can express a feature of an obstacle and a similar obstacle, an obstacle learning performance and an obstacle detection performance can be remarkably improved.

Description

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a vehicle crash alarm system and method,

The present invention relates to a vehicle crash alarm system and method. More particularly, the present invention relates to a vehicle crash alarm system and method using a machine learning technique in which a reference of an obstacle pattern is established by using a machine learning technique on an automatic parking system, and an obstacle classification speed is improved and reliability is improved.

The problem of parking due to the increase in the demand of the vehicle is emerging as a social issue, and in many cases, the driver has difficulty in parking the vehicle in a narrow space. Even if you are a good driver, if you want to park in a small space, you have to go back and forth several times back and forth and repeat the parking. Many people have to guide someone to park their car properly. Accordingly, there is an increasing demand for vehicles equipped with an automatic parking system.

A typical automatic parking system is a system that receives information from various sensors of a vehicle, judges a parking space by using the received information, calculates a parking target point of the vehicle, calculates a locus, and carries the parking by controlling the vehicle.

In order to obtain accurate parking by using this conventional method, an accurate search for a parking space is required, and this information is searched using information about an obstacle around the parking space.

Referring to FIG. 1, the automatic parking system is composed of application software and base software in software. The application software includes a system management unit 10 and an HMI (Human Machine Interface) 20. The system management unit 10 performs a function of integrally managing each block constituting the automatic parking system, that is, the vehicle position tracking unit 11, the parking space determination unit 12, and the parking path generation / tracking unit 13 . The vehicle position tracking unit 11 receives the vehicle information, determines the state and position of the vehicle, and transmits the state and position to the parking path generation / tracking unit 13. The parking space determination unit 12 determines the parking space by using the information received from the ultrasonic sensor, calculates the parking target point of the vehicle, and transmits the calculated parking target point to the parking path generation / tracking unit 13. The parking path generation / follow-up unit 13 calculates the parking induction locus by receiving the vehicle state and position information transmitted from the vehicle position tracking unit 11 and the parking target point information transmitted from the parking space determination unit 12, The parking space follow-up is induced by using the calculated parking induction locus and the obstacle information around the vehicle transmitted from the space determining unit 12. [

In particular, the parking space determination unit 12 detects obstacles in the front, rear, and right and left areas of the vehicle, which are perceived as squares in the driver's seat using an obstacle detection device mounted on the vehicle, and notifies the driver of the obstacles. That is, as shown in FIG. 2, two sensors are installed on the front bumper and the rear bumper of the vehicle, respectively, to detect an obstacle. Here, the sensor detection area is a hemispherical shape with the sensor as a center, and when an obstacle is detected in the sensor detection area, an alarm sounds at a predetermined interval through the control unit.

After securing the candidate region of the obstacle from the camera image, the presence or absence of the obstacle is determined.

To do this, decision boundaries are learned from obstacle sample patterns and applied to obstacle recognition. For example, there is MLP (Multi Layer Perceptron) that minimizes only errors from given sample data in an exemplary manner.

However, there is a problem that an object which is located lower than the sensor, such as a stone on the ground, can not be recognized by simply mounting the sensor, thereby causing an accident. That is, there is a problem that obstacle detection in the vehicle vertical direction is difficult.

In addition, there is a problem that the type of the obstacle can not be determined in the case of detecting the obstacle through the sensor.

In the case of the MLP method, the accuracy can be guaranteed in the learning data, but the accuracy is low in the non-learning data and thus the reliability can not be secured.

SUMMARY OF THE INVENTION The present invention has been made to solve the above-mentioned problems of the conventional art, and it is an object of the present invention to provide a vehicle crash alarm system and method by building a reference of an obstacle pattern by using a machine learning technique, The purpose is to provide.

According to the present invention, the object is achieved by a vehicle position recognition apparatus comprising: a vehicle position recognition unit for recognizing a position of a vehicle; A sensor unit for detecting an obstacle using ultrasonic waves; An image device for acquiring image data of an obstacle; And an obstacle pattern classifying unit that determines the type of the obstacle based on the information collected through the sensor and the video apparatus.

Here, the obstacle pattern classification unit recognizes patterns and types of obstacles extracted through the image device through a machine learning technique (Support Vector Machine) based on the image data collected through the image device.

The obstacle pattern classifier may generate a machine learning technique module using the image data collected through the imaging device as learning data of the machine learning technique.

Here, the obstacle pattern classification unit collects patterns classified as obstacles or similar obstacles through the image data, and performs a principal component analysis (PCA) and an indepependent component analysis (ICA) A base vector extraction unit for extracting a base vector of each obstacle; A feature vector extracting unit for extracting a feature vector from the basis vector output from the basis vector extracting unit; A clustering unit for classifying the feature vectors of the feature vector extracting unit into subclasses; And a SVM module unit for recognizing the type of the obstacle using the machine learning technique module by using the feature vector classified by the clustering unit.

The SVM module may include an SVM learning module that generates a support vector using the feature vector classified by the clustering unit; And an SVM classification module for recognizing the type of the obstacle by comparing the similarity between the support vector and the feature vector obtained through the image data of the obstacle obtained through the imaging device.

The obstacle classifying unit may further include an obstacle type output unit for recognizing the type of the obstacle according to the determination of the obstacle pattern classifying unit.

The alarm unit may further include an alarm unit for recognizing the presence and the type of the obstacle according to whether the obstacle type is an obstacle type of the obstacle kind output unit and at least one of the electric seat belt and the steering actuator So that it is recognized by using one.

In the case of an obstacle that may cause human injury, the alarm unit may be made to recognize by using at least one of an indicator and a buzzer.

The object is achieved according to the present invention by a method comprising: recognizing a position of a vehicle; Detecting an obstacle; Recognizing a type of the detected obstacle using an obstacle pattern classifying unit for determining the type of the obstacle; Outputting the type of the obstacle according to the recognition of the obstacle type; And recognizing the obstacle in accordance with the type of the obstacle.

Here, the obstacle pattern classifying unit may classify the obstacle pattern extracted through the imaging device through a machine learning technique (Machine Learning Method) based on the image data collected through the imaging device that detects the obstacle in the detection step And recognizes the type.

In the step of recognizing the obstacle, in the case of an obstacle capable of causing physical damage, at least one of a power seat belt and a steering actuator is used to recognize the obstacle.

Further, in the step of recognizing the obstacle, in the case of an obstacle which may cause human injury, at least one of an indicator lamp and a buzzer may be further used to recognize the obstacle.

The vehicle crash alarm system using the machine learning method according to the present invention having the above configuration combines a machine learning method, that is, a principal component analysis method and an independent component analysis method, and calculates a residual error value The performance of obstacle learning and obstacle detection can be greatly improved.

In particular, the complexity is reduced by classifying by using a machine learning algorithm (SVM) that extracts feature vectors of data.

By learning not only the obstacle image but also the similar obstacle image as the learning data of the SVM to learn the obstacle pattern classification, it is judged that the illumination or pose is not an obstacle when an obstacle image different from the learned data comes in, It is possible to greatly reduce the error that judges that a pattern similar to the obstacle is an obstacle.

In addition, since the obstacle is automatically detected and the type is recognized, and the obstacle is informed by using the tactile information, the indication, the buzzer or the like according to the degree of danger of the obstacle, it is convenient that the driver does not need to operate it separately.

In addition, safety can be improved by changing the warning means according to the risk group of the obstacle.

1 is a block diagram showing a conventional automatic parking system,
2 is a view showing the position and detection range of a sensor installed in a conventional vehicle,
FIG. 3 is a block diagram illustrating a vehicle collision alerting system based on a machine learning technique according to the present invention,
4 and 5 are a block diagram and a flowchart showing the obstacle pattern classifying unit of the present invention,
6 is a flowchart showing the basis vector extracting unit of the present invention,
7 is a flowchart showing the feature vector extracting unit of the present invention,
FIG. 8 is a diagram illustrating a one-against-all scheme of the SVM module unit of FIG. 4,
9 is a flowchart illustrating a vehicle collision alerting system using a machine learning technique according to the present invention.

Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings in order to clarify the solution to the technical problem of the present invention. In the following description of the present invention, however, the description of related arts will be omitted if the gist of the present invention becomes obscure. In addition, the terms described below are defined in consideration of the functions of the present invention, and may be changed depending on the intention or custom of the designer, the manufacturer, and the like. Therefore, the definition should be based on the contents throughout this specification. In addition, parts denoted by the same reference numerals throughout the specification represent the same elements.

Hereinafter, a vehicle crash alarm system 1 according to the present invention will be described.

3, the vehicle collision alerting system 1 according to the present invention includes a vehicle position recognizing unit 100, a sensor unit 200, a video device 300, an obstacle pattern classifying unit 400 An obstacle kind output unit 500, and an alarm unit 600.

The vehicle position recognition unit 100 recognizes the position of the vehicle based on the vehicle information, as shown in Fig. Here, the vehicle information includes at least some information of a steering angle, a wheel pulse, a yaw rate, and temperature information.

The sensor unit 200 detects obstacles on the sides and front and back of the vehicle based on the sensing values obtained using the ultrasonic sensors. That is, the presence or absence of an obstacle is determined through the sensing value.

The video apparatus 300 acquires image data values by a camera or the like. The obtained image data is used as learning data of the obstacle pattern classifying unit 400, and the type of the obstacle is recognized based on the learning data. Depending on the type of the obstacle, the warning unit 600 informs the driver Recognize obstacles.

The obstacle pattern classifying unit 400 classifies SVMs of the machine learning techniques based on the vehicle position information by the vehicle position recognizing unit 100, the sensed value from the sensor unit 200, Support Vector Machine) module, and recognizes an obstacle pattern and a type by analyzing a recognized obstacle by using the SVM module.

4 to 8, the obstacle pattern classifying unit 400 may include a basis vector extracting unit 410, a feature vector extracting unit 420, a clustering unit 430, and an SVM module unit 440. have.

6, the base vector extraction unit 410 includes an obstacle DB 411 and a similar obstacle DB 412, a PCA basis vector extraction unit 413, a first obstacle DB 414, And a first similar obstacle DB 415, an ICA basis vector extractor 416, a second obstacle DB 417, and a second similar obstacle DB 418.

The obstacle DB 411 is constructed from obstacle images reflecting various lights and pose. In addition, patterns that may cause errors due to similar patterns of obstacles even if they are not obstacles are also collected. At this time, an image pattern having a predetermined value or less is collected using the average obstacle image pattern and the Eculidean distance to construct a similar obstacle DB 412. Here, patterns that are not obstacles are not collected as sample data because they are not meaningful for modeling because they are various kinds. In the case of a pattern determined to be an obstacle as a result of the obstacle pattern classification, the pattern is collected as similar obstacle sample data, stored in the similar obstacle DB 412, and the learning is repeatedly performed.

The PCA basis vector extractor 413 extracts a PCA basis vector by performing principal component analysis (PCA) on an image input from an obstacle DB and a similar obstacle DB.

In addition, the PCA basis vector extracting unit 413 normalizes the position and size of the obstacle image through the coordinates of the image input from the obstacle DB 411 and the similar obstacle DB 412, respectively. Then, principal component analysis is performed to extract PCA basis vectors of space representing obstacles and similar obstacles.

That is, by using the principal component analysis method, the feature space is linearly projected in the direction of maximizing the variance expressed by the image feature vector, thereby reducing the dimension. Then, the eigenvalue and the feature vector are obtained by using the variance. In addition, the feature vectors are arranged in order of magnitude of the eigenvalues, and feature vectors corresponding to the predetermined dimensions are extracted. Generally, k is selected in which the cumulative contribution (cumulative variance) accounts for 99% of the total eigenvalue sum.

Then, the PCA basis vectors are accumulated in the first obstacle DB 414 and the first similar obstacle DB 415, respectively.

Base vectors having a large eigenvalue among the base vectors stored in the first obstacle DB 414 and the first similar obstacle DB 415 are output to the ICA basis vector extraction unit 416.

The ICA basis vector extractor 416 extracts an ICA basis vector by performing an independent component analysis (ICA) on the output base vectors. The extracted ICA basis vectors are accumulated in each of the second obstacle DB 417 and the second similar obstacle DB 418, respectively.

The ICA basis vectors accumulated in the second obstacle DB 417 and the second similar obstacle DB 418 are output to the feature vector extracting unit 420 and used for extracting the feature vector of the obstacle and the similar obstacle image.

7, the feature vector extracting unit 420 extracts a feature vector from an obstacle base vector extracted from the base vector extracting unit 410 and an obstacle base vector stored in the obstacle DB 411 and the similar obstacle DB 412, And similar obstacle images are respectively projected to extract feature vectors.

For example, N feature values (projection coefficients) are extracted by projecting one obstacle image on N obstacle base vectors extracted from the obstacle images, and N similar obstacle base vectors extracted from the similar obstacle images And extracts N feature values (projection coefficients) to obtain a total of 2N feature values.

In addition, the residual error value is calculated and used as a feature value. The residual error means a distance between the restored image vector and the actual image vector before projection when the image vector is multiplied by the feature vector obtained by projecting each image on N base vectors.

Therefore, when the obstacle image is restored by using N obstacle base vectors, the residual error value is small, but when the similar obstacle image is restored by using N obstacle base vectors, the residual error value becomes large.

In summary, the feature vector extractor 420 extracts a feature vector using 2N + 2 feature values including 2N feature values and 2 residual error values obtained from one image.

Then, the feature vector for the extracted obstacle image is stored in the obstacle feature vector DB 421, and the feature vector for the similar obstacle image is stored in the similar obstacle feature vector DB 422, and then extracted into the clustering unit.

The clustering unit 430 measures the Euclidean distance between the obstacle image feature vectors according to the k-means clustering algorithm, classifies the obstacle feature vectors into a predetermined subclass by grouping closest Euclidean distances, . Then, the obstacle feature vector closest to the vector mean of the obstacle feature vectors stored in each sub class is selected as the central obstacle feature vector representing each sub class. Here, similar obstacle image feature vectors are also performed in the same manner.

The SVM module 440 may include an SVM learning module 441 and an SVM classification module 442 as shown in FIG.

8, multiple SVMs are designed by applying one-against-all technique because obstacle and similar obstacle classifications are multi-class classification. Accordingly, the SVM learning module 441 generates a support vector through SVM learning, and the SVM classification module 442 recognizes the obstacle type by comparing the similarity between the support vector and the feature vector of the obstacle or similar obstacle.

First, the preprocessed data is used as learning data of the SVM. Then, SVMs capable of binary classification are generated for each subclass, and a total of M SVMs are designed. The SVM learning then depends on the training data set and the internal kernel function used for learning and the C parameter, the adjustment parameter. Here, the SVM kernel function serves to increase the operation speed of the SVM by linearly converting the nonlinear input values of high dimensional. Therefore, the SVM classification module 442 is designed by setting optimal parameters through experiments.

When the actual obstacle detection and classification are recognized, test data is automatically entered into the SVM classification module 442 generated in the learning stage. Classify obstacle and similar obstacle image feature vectors into subclasses corresponding to SVM according to output value.

4, in the case of an image of the obstacle input by the imaging device 300, the feature vector extraction unit 420 and the clustering unit 430 are used to extract a feature vector and a subclass Classify. Thereafter, the SVM classification module 442 compares the similarity between the support vector of the SVM learning module 441 and the feature vector of the image of the obstacle input by the imaging device 300, thereby recognizing the type of the obstacle.

According to the obstacle classification of the obstacle pattern classification unit 400, the obstacle type output unit 500 recognizes the type of the obstacle in real time to the driver by sound, character, or image, thereby avoiding collision with the obstacle.

The alarm unit 600 generates a stop signal to the system to inform the driver that an obstacle is detected. For example, when the distance from the obstacle is about 50 cm, the alarm unit 600 blinks the indicator and continuously sounds the buzzer. Also, as approaching the obstacle, the blinking cycle and the buzzer cycle are made faster, and when the distance from the obstacle is less than about 20 cm, the indicator lights continuously and the buzzer sounds continuously.

Also, the alarm unit 600 can be recognized by the first alarm means 610 and the second alarm means 620 according to the possibility of human loss and material loss.

The first alarm means 610 recognizes the sensed obstacle through the tactile information of the driver by using either the electric seat belt or the steering actuator when the detected obstacle belongs to the first group obstacle danger group. Here, a group 1 dangerous obstacle means an obstacle such as an obstacle such as a box, a square column, and a cylinder, which is a risk group that can only cause damage to the vehicle without injuring the driver, resulting in physical damage only.

When the detected obstacle belongs to the second group obstacle risk group, the second warning unit 620 recognizes the detected tactile information to the driver through the tactile information. At the same time, the second warning unit 620 informs the driver of the obstacle belonging to the group 2 And that it exists. Here, the Group 2 dangerous obstacle group means, for example, a curb, a car, a wall, a person, and the like, as well as an injury to a driver and an accidenter.

5, the obstacle pattern classifying unit 400 classifies the obstacle DB 411 through the vehicle position recognizing unit 100, the sensor unit 200, and the video device 300, And a similar obstacle DB 412 are constructed (S10). Then, as described above, the base vector is extracted through the PCA basis vector extraction unit 413 and the ICA basis vector extraction unit 416 (S11). Then, the feature vector extraction unit 420 extracts the obstacle feature vector and the similar obstacle feature vector based on the basis vector, calculates the residual error value, and calculates the feature vector for the extracted obstacle image using the obstacle feature vector DB 421, The feature vector for the similar obstacle image is stored in the similar obstacle feature vector DB 422 (S12). Then, the clustering unit 430 classifies the feature vectors into subclasses (S13).

An image for real obstacle detection and type recognition is input (S20). A feature vector is extracted from the input image (S21), and a residual error value is calculated (S22). Accordingly, the subclass to which the feature vectors of the input image belong belong is classified (S23), and the obstacle is detected and the type is recognized by the SVM of the subclass (S24).

Hereinafter, with reference to FIG. 9, a method of alarming an obstacle to a driver by using the automobile crash alarm system 1 by the machine learning technique according to the present invention will be described.

The vehicle collision alerting method S1 according to the present invention includes a step S100 of recognizing a position of a vehicle, a step S200 of detecting an obstacle, a step S300 of recognizing the type of the obstacle, (S500), a first alarm generating step (S600), and a step (S700) of determining whether the second dangerous obstacle group is included in the second dangerous obstacle group (S400) And a second alarm means generation step (S800).

In step S100 of recognizing the position of the vehicle, the position of the vehicle is recognized based on the above-described vehicle information. In operation S200, an obstacle is detected by acquiring image data of an obstacle and an ultrasonic sensor of the sensor unit 200. [

The step of recognizing the type of the obstacle (S300) recognizes the obstacle pattern and the type by analyzing the recognized obstacle using the SVM module of the obstacle pattern classifying unit (400).

The step S400 of outputting the type of the obstacle allows the obstacle kind output unit 500 to recognize the type of the obstacle in real time so as to avoid the collision with the obstacle in advance.

Accordingly, the automobile crash alarm system 1 determines whether the obstacle is included in the first dangerous obstacle group according to the obstacle type signal transmitted by the obstacle type output unit 500 (S500) It is recognized by the driver that the first alarm means 610 is an obstacle causing only physical damage (S600). Here, the first alarm means 610 recognizes the detected obstacle by using the tactile information of the driver using either the electric seat belt or the steering actuator when the detected obstacle belongs to the first group obstacle danger group.

If it is not an obstacle included in the first dangerous obstacle group, it is determined whether it is an obstacle included in the second dangerous obstacle group (S700). If the obstacle is included in the second dangerous obstacle group, the second warning means 620 The driver is informed that it is an obstacle causing both physical damage and human injury (S800). Here, if the detected obstacle belongs to the second-group obstacle risk group, the second warning unit 620 recognizes the detected obstacle to the driver through the tactile information, and informs the driver through at least one of the indicator and the buzzer It informs that the obstacle exists.

It should be understood that the various embodiments according to the present invention can solve various technical problems other than those mentioned in the specification in the related technical field as well as the related art.

The present invention has been described with reference to the embodiments. It will be apparent, however, to one skilled in the art that the present invention may be embodied in various other forms without departing from the spirit or essential characteristics thereof. Therefore, the disclosed embodiments should be considered in an illustrative rather than a restrictive sense. That is, the true technical scope of the present invention is indicated in the appended claims, and all differences within the scope of equivalents thereof should be construed as being included in the present invention.

1: Automobile collision warning system by the machine learning method according to the present invention
100: Vehicle position recognition unit 200: Sensor unit
300: Imaging apparatus 400: Obstacle pattern classifying unit
500: Obstacle type output unit 600: Alarm unit

Claims (11)

A vehicle position recognizing unit for recognizing a position of the vehicle;
A sensor unit for detecting an obstacle using ultrasonic waves;
An image device for acquiring image data of an obstacle;
And an obstacle pattern classifying unit for determining the type of the obstacle based on the information collected through the sensor and the video apparatus,
Wherein the obstacle pattern classification unit recognizes a pattern and a type of the obstacle extracted through the imaging device through a support vector machine based on the image data collected through the imaging device. Vehicle crash alarm system by. The method according to claim 1,
Wherein the obstacle pattern classifier generates a machine learning technique module by using the image data collected through the imaging device as learning data of the machine learning technique. 3. The method according to claim 1 or 2,
The obstacle pattern classifying unit
A pattern classified as an obstacle or a similar obstacle is collected through the image data and a basis vector of each obstacle or similar obstacle is extracted by performing principal component analysis (PCA) and independent component analysis (ICA) A base vector extractor;
A feature vector extracting unit for extracting a feature vector from the basis vector output from the basis vector extracting unit;
A clustering unit for classifying the feature vectors of the feature vector extracting unit into subclasses;
And a SVM module unit for recognizing the type of the obstacle using the machine learning technique module based on the feature vector classified by the clustering unit. The method of claim 3,
The SVM module unit
An SVM learning module for generating a support vector using the feature vectors classified by the clustering unit;
And an SVM classification module for recognizing the type of the obstacle by comparing the similarity between the support vector and the feature vector obtained through the image data of the obstacle obtained through the imaging device. . The method according to claim 1,
And an obstacle type output unit for recognizing the type of the obstacle according to the determination of the obstacle pattern classifying unit. 6. The method of claim 5,
Further comprising an alarm unit for recognizing existence and type of an obstacle according to whether the obstacle type is an obstacle type of the obstacle type output unit,
Wherein the alarm unit recognizes an obstacle capable of causing physical damage using at least one of a power seat belt and a steering actuator. The method according to claim 6,
Wherein the alarm unit is adapted to recognize at least one of an indicator and a buzzer in the case of an obstacle that may cause human injury. Recognizing the position of the vehicle;
Detecting an obstacle;
Recognizing a type of the detected obstacle using an obstacle pattern classifying unit for determining the type of the obstacle;
Outputting the type of the obstacle according to the recognition of the obstacle type;
And recognizing the obstacle in accordance with the type of the obstacle that is output. 9. The method of claim 8,
The obstacle pattern classifying unit may classify patterns and types of obstacles extracted through the imaging device through a machine learning technique based on image data collected through an imaging device that detects an obstacle in the detection step of the obstacle, Of the vehicle collision warning. 10. The method according to claim 8 or 9,
Wherein the step of recognizing the obstacle is performed by using at least one of a power seat belt and a steering actuator in the case of an obstacle capable of causing physical damage. 12. The method of claim 11,
Wherein the step of recognizing the obstacle is performed by using at least one of an indicator and a buzzer in the case of an obstacle capable of causing personal injury.

Images