KR101101372B1 - Parking lot management method based on learning using sample elimination method - Google Patents

Abstract

PURPOSE: A parking surface recognizing method based on the education of a sample eliminating method is provided to enable a user to effectively manage a parking surface within a parking lot by periodically updating data. CONSTITUTION: An adaboost learning theory is applied at a sample image related to a parking surface(S100). A strong classification apparatus is used for plural stages using a cascade type and eliminates a negative sample image(S200). The adaboost learning theory is applied at an input image using a weak classification apparatus and a predetermined weight(S300). A single positive area is recognized as a vehicle based on the overlapping rate of the detected positive area(S400).

Description

Parking Lot Management Method Based on Learning Using Sample Elimination Method

The present invention relates to a vehicle recognition system of a parking lot, and more particularly, to a system for recognizing a vehicle from an image of a camera installed in a parking lot, determining whether a vehicle exists on a parking surface, and managing whether a parking surface can be used.

There are various ways to find out whether a car is parked in a parking lot, such as a physical method using a sensor such as a laser and a software method of analyzing an image acquired by a camera. A typical software method is an algorithm that detects a parked car by storing a background (empty parking space) in advance and using a difference from an input image. This method is simple to implement and apply, but the disadvantage is that the object detection result is too sensitive to changes in the surroundings due to the characteristics of the input image.

Recently, many learning-based object detection algorithms have been used. Among the learning-based algorithms, the AdaBoost method is one of the widely used methods because of its simple implementation and fast detection speed. However, this method also has a disadvantage in that the object detection rate is lowered when the characteristics of the sample image used for learning and the input image for actual detection are different. Another problem is that the training time is too long when using a large number of training samples.

Accordingly, the present invention has been made to solve the above problems, an object of the present invention is to recognize the parking surface based on the learning of the sample removal method that can determine the presence of the vehicle in the parking surface using the learning-based vehicle recognition method A method and a recording medium thereof are provided.

The object of the present invention as described above,

Obtaining a weak classifier (h (x)) and a strong classifier (S (x)) by applying the Adaboost learning theory to the sample image photographing the parking surface (S100);

Removing the negative sample image by using the strong classifier S (x) in a cascade manner over a plurality of stages (S200);

Detecting a positive area in which the vehicle is parked by applying the Adaboost learning theory using the weak classifier h (x) and a predetermined weight α to the input image captured by the camera (S300); And

Recognizing a united positive area as a vehicle based on the mutual overlap ratio of the detected plurality of positive areas (S400) can be achieved by a parking surface recognition method based on the learning of the sample removal method. .

And, in the negative sample image removal step (S200),

In the strong classifier S (x) for each stage, a plurality of weak classifiers h (x) may be provided.

In addition, the number of the weak classifier h (x) is preferably provided to increase as the stage passes.

In addition, when applying the Adaboost learning theory using the weak classifier (h (x)) and the weight α in the positive region detection step (S300), a strong classifier including a positive sample more than a predetermined ratio (K%) ( Only the detection threshold value Th _D , which is a value of S (x), can be determined and detected as a positive area.

Here, the predetermined ratio (K%) preferably has a value in the range of 1% to 5%.

In addition, the recognition step S400 may further include determining and removing an area that is less than or equal to an overlap ratio as an error.

The steps S100 to S400 are preferably performed according to the parking plane positions in the sample image and the input image, respectively.

An object of the present invention as described above, the step of obtaining a weak classifier (h (x)) and a strong classifier (S (x)) by applying the Adaboost learning theory to the sample image photographing the parking surface as another category (S100) );

Removing the negative sample image by using the strong classifier S (x) in a cascade manner over a plurality of stages (S200);

Detecting a positive area in which the vehicle is parked by applying the Adaboost learning theory using the weak classifier h (x) and a predetermined weight α to the input image captured by the camera (S300); And

It may also be achieved by a computer readable recording medium performing step S400 of recognizing a united positive area as a vehicle based on the mutual overlap ratio of the detected plurality of positive areas.

According to the parking surface recognition method and the recording medium based on the learning of the sample removal method according to the present invention, since the Adaboost theory is applied, the implementation is very simple and can be easily applied to the parking image.

It is also convenient to select a feature for a very large set of features.

In addition, since the Adaboost theory is applied, there is an advantage that the vehicle can be recognized more effectively than the conventional method. In addition, the proposed sample elimination method reduces the learning process time, so that data can be updated frequently and the accuracy is improved.

This, in turn, plays an important role in efficiently managing the parking surface in the parking lot.

The following drawings, which are attached in this specification, illustrate the preferred embodiments of the present invention, and together with the detailed description thereof, serve to further understand the technical spirit of the present invention, and therefore, the present invention is limited only to the matters described in the drawings. It should not be interpreted.
1 is a graph showing an experimental data distribution and a threshold value when K = 5% as an embodiment of the present invention.
FIG. 2 is an image showing a positive region before unification in step S400 of one embodiment of the present invention.
FIG. 3 is an image in which one positive area is displayed for each parking surface by subjecting the image of FIG. 2 to be unified.
4 is a configuration diagram of the strong classifiers 10a, 10b, and 10c constituting the stage in a cascade manner according to an embodiment of the present invention.
FIG. 5 is a diagram illustrating a process of recognizing an input image using the strong classifiers 10a, 10b, and 10c of FIG. 4.
6 is a flowchart illustrating a parking surface recognition method based on learning of a sample removing method according to an embodiment of the present invention.

Hereinafter, a parking surface recognition method based on learning of a sample removing method according to an embodiment of the present invention will be described in detail with reference to the accompanying drawings.

< Adboost ( Adaboost ) Algorithm>

Hereinafter, the AdaBoost algorithm known to those skilled in the art will be described. Adamboost is an algorithm for designing strong classifiers from weak classifiers.

Where x is defined as a weak classifier with sample, h _t (x): X → {-1, +1},

This is defined as a strong classifier, where H (x) = sign (f (x)).

A given set of data is represented by _{(x 1, y 1),} (x 2, y 2), ..., (x m, y m), where x and ∈X _i, y _i is y _i ∈ { -1, +1} or y _i ∈ {0, 1}.

Calculate the weak classifier by calculating as shown in [Equation 2].

If εt ≥ 1/2, stop. Otherwise, set as [Equation 3] and update as [Equation 4].

Where Z _t is the normalization coefficient,

Calculate a strong classifier (H (x)) as shown in Equation 5 below.

First, the initial weight D ₁ (i) = 1 / m, where (x ₁ , y ₁ ), (x ₂ , y ₂ ), ..., (x _m , y _m ) is the set of data given Express position and target value. And the initial weight of each data is set to 1 / m.

When t = 1, ..., T, repeat from 1 to T after the data set and initial setup. The repetitive tasks are described as follows.

i) First, find the value of h _t .

Here, a straight line called h _i is obtained, in which the value of epsilon representing the error rate is the smallest. The line h _i represents one straight line drawn from a large hypothetical space containing all the straight lines H. Dt calculates the total error value by multiplying the weights of the errors that are incorrectly recognized. If the error value (ε _t ) is greater than 0.5, it is determined that the error value is ridiculous and the operation is stopped.

ii) If the error value ε _t is less than 0.5, the operation is continued. At this time, the weight α _t is calculated as shown in [Equation 3]. Equation 3 is a formula set to minimize Zt. In the present specification, the proof will be briefly described. First, in the upper bound theorem, y _i f (x _i ) <= 0 if the data is misclassified. This implies that exp (-y _i f (x _i ))> 1. In other words, the result of measuring the wrong error out of m should be less than the product of 1 to t of the value of Zt. Therefore, if the value of Zt is smaller, the error must be smaller than that value. After all, to minimize the overall error, you can minimize the value of Zt.

In order to minimize the Zt value, the weight (a _t ) should be selected well, and the most ideal equation may be referred to as [Equation 3].

iii) The weight (a _t ) is set and the value of D _{t + 1} is updated for all data sets as shown in [Equation 4]. That is, higher weights are used for the poorly matched parts and correct matched data lowers the weights. Because the sum of Dt (i) must be 1.

iv) Then, the process is repeated from step i).

v) It repeats t times you specified and ends or the error is greater than 0.5, and either way. When the work is done, we linearly combine the weak classifier values calculated in each iteration to get the strong classifier values we want. . By obtaining a strong classifier value, the Adboost algorithm is completed.

<Learning course>

Hereinafter, a specific embodiment of recognizing a vehicle by applying an AdaBoost learning method to a sample image photographing a parking surface will be described with reference to the accompanying drawings.

First, in the present embodiment, in order to recognize the vehicle parked on the parking surface of the parking lot, it is assumed as follows for convenience of description.

i) It is assumed that the camera is fixedly installed in a direction facing the parking surface.

ii) See two to four parking planes as a block, and the camera assumes to shoot one block.

AdaBoost learning method, which is a widely used recognition algorithm, is widely used in the field of face recognition. This method selects the weak classifiers (h (x)) that best classify the learning samples, and then combines them to create a strong classifier. The cascaded strong classifier is then used in cascade.

4 is a configuration diagram of the strong classifiers 10a, 10b, and 10c in which a stage is configured in a cascade manner according to an embodiment of the present invention. As shown in Fig. 4, for example, three strong classifiers 10a, 10b and 10c are configured in series (divided into three stages). Each of the strong classifiers 10a, 10b, and 10c includes a plurality of weak classifiers 20. For example, the first strong classifier 10a, which is the first stage, includes five weak classifiers 20, and the second strong classifier 10b includes 20 weak classifiers 20, and the third strong classifier 10 10c) includes 100 weak classifiers 20. That is, as the stage progresses, each of the strong classifiers includes a large number of weak classifiers 20.

In the learning process, a sample image is input to the input side of the first strong classifier 10a.

First, we find n weak classifiers (h (x)) by adboost learning. x represents a sample, and in the present embodiment, represents a sample image of the parking surface photographed by the camera. The weak classifier h (x) has information such as the location and size of the feature in the sample image. Then, a strong classifier (S (x), corresponding to H (x) in the general Adaboost theory described above) for all the positive and negative samples is calculated using Equation 6 (S100).

Α represents a weight of h (x). In this embodiment, the weight α may have a value of 0 to infinity, but may have a value of 0 when the error rate ε is 1/2 (50%) and about 2 when the error rate ε is 1%. And, when the error rate (ε) is 0.1% can have a value of about 3.

Also it detects the S (x) values comprising a positive sample by K% arbitrary ratio, such as first threshold value (Th _D) in the positive and the minimum value of S (x) values of the negative samples and the detection threshold value (Th _D) The median of is determined as the sample removal threshold Th _E. Here, K may have a value in the range of 0% to 100%, and the experiment showed the best results especially in the range of 1% to 5%. In each step, the negative sample whose value is smaller than the sample removal threshold Th _E is determined as a positive negative and removed from the training data (S200). Repeat the above procedure for the maximum number of pre-defined steps, if no negative sample exists. In this process, as the detection threshold Th _D decreases, the recognition rate increases, but the number of necessary steps increases. In this way, by eliminating certain negative samples at each step, the next step gives more weight to the still uncertain sample, which improves the final recognition rate.

<Recognition process>

FIG. 5 is a block diagram illustrating a process of recognizing an input image using the strong classifiers 10a, 10b, and 10c of FIG. 4. As illustrated in FIG. 5, an image (input image) captured by a camera (not shown) is input to the input of the first strong classifier 10a. At this time, the first strong classifier 10a is performed on all detection areas of the input image. When it is determined that the result is positive (the vehicle is present), the second strong classifier 10b is performed. If it is determined as positive by the second strong classifier 10b, the third strong classifier 10c is performed. Then, the third strong classifier 10c finally determines to be positive (S300).

If it is determined that the first, second, and third strong classifiers 10a, 10b, and 10c are negative, the corresponding area is determined to be not a vehicle. The reason for doing this is that by first applying a small number of weak classifiers 20, it is possible to reliably remove the non-vehicle at high speed.

When the object is detected, the boost is detected using the weak classifier (h (x)) and the weight α of each step obtained in the learning process on the input image being photographed by the camera. In this case, the final discriminant used in the proposed method uses Equation 7 in which the threshold value of the existing Ad Boost method is changed to the detection threshold Th _D. As shown in [Equation 7], the strong classifier in which the threshold is changed to the detection threshold Th _D is called H (x).

The following step discrimination is performed on the positively detected regions having a value of 1 through Equation 7 above. This process is the same as the conventional cascade detection method. The location of the vehicle detected using the above-described cascade ad-boost method is that many overlapping regions (corresponding to a plurality of overlapping red squares in FIG. 2) also appear for the same vehicle (see FIG. 2). In order to solve this problem, there is a need for a method in which overlapping areas are combined into one and determined as one vehicle. This method consists of calculating the overlapping area over a certain percentage of the total area to determine the most overlapping area, and removing the overlapping area until there is no more overlapping area (unification of FIG. 3). 2 red squares). The overlapping area below a predetermined ratio among the final areas is determined as an error and is removed (S400).

In addition, according to the location of the parking surface in the vehicle recognition process can be distinguished to the left, right, and center, and even when performing the learning separately can increase the recognition rate.

As described above, those skilled in the art will understand that the present invention can be implemented in other specific forms without changing the technical spirit or essential features. Therefore, it should be understood that the above-described embodiments are to be considered in all respects as illustrative and not restrictive. The scope of the present invention is shown by the following claims rather than the detailed description, and all changes or modifications derived from the meaning and scope and equivalent concept of the claims are included in the scope of the present invention. Should be interpreted.

10a: the first strong classifier,
10b: second strong classifier,
10c: third strong classifier,
20: weak classifier.

Claims (8)

Obtaining a weak classifier (h (x)) and a strong classifier (S (x)) by applying the Adaboost learning theory to the sample image photographing the parking surface (S100);
Removing the negative sample image by using the strong classifier S (x) over a plurality of stages in a cascade manner (S200);
Detecting a positive area in which a vehicle is parked by applying the Adaboost learning theory to the input image captured by the camera using the weak classifier (h (x)) and a predetermined weight α (S300); And
Recognizing a united positive area as a vehicle based on the mutual overlap ratio of the detected plurality of positive areas (S400),
Steps S100 to S400 are performed according to the learning of the sample removing method, characterized in that performed according to the location of the parking surface in the sample image and the input image, respectively. The method of claim 1,
In the negative sample image removal step (S200),
And a plurality of weak classifiers (h (x)) are provided in the strong classifier (S (x)) for each stage. The method of claim 2,
The number of the weak classifier (h (x)) is a parking surface recognition method based on the learning of the sample removal method, characterized in that it is provided to increase as the stage passes. The method of claim 1,
In the positive region detection step S300, when the weak booster h (x) and the weight α are applied to the adaboost learning theory,
Based on the learning of the sample removal method, the positive sample is detected and detected as a positive area only for the detection threshold value Th _D , which is the value of the strong classifier S (x) including the predetermined ratio (K%) or more. Parking surface recognition method. The method of claim 4, wherein the predetermined ratio (K%) has a value ranging from 1% to 5%. The method of claim 1,
The recognition step (S400),
And determining to remove an area having an overlapping ratio of the unified positive area equal to or less than a predetermined ratio as an error and removing the detected area. delete delete

Images