JP2011013924A - Detection system and detection method for vehicle in parking lot - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a vehicle detection method of a parking lot, having a high recognition rate even outdoors.SOLUTION: A garage is cut out from a photographic image by a photographing camera; a plurality of representative values are acquired from cut-out image data and are normalized; image data composed of the normalized representative values are vectorized; a dimension of vectorized vector data is compressed; Mahalanobis distance to a cluster center previously obtained in training data of the compressed vector data is obtained; the Mahalanobis distance is substituted for a membership function (u*) represented by expression (1); a function value thereof is substituted for expression (2) to obtain a membership (tilde u), and it is decided whether a vehicle is present inside the garage by use of the obtained membership. In the expression, αis a mixture ratio of a cluster j inside a cluster q; S is a variance-covariance matrix; and πis a mixture ratio of a class q.

Description

  The present invention relates to a vehicle detection system and a detection method capable of determining whether a vehicle is present in a parking room of a parking lot, that is, a parking state or an empty vehicle state.

When managing a parking lot, it is necessary to grasp the parking situation, and a parking situation measuring method that can measure the parking situation has already been proposed (for example, see Patent Document 1).
This parking situation measurement method is a parking situation measurement method in which the parking rate of the entire parking lot is estimated using a neural network from the number of parking lots in the specific parking area of the parking lot. This is a method of calculating the parking rate for each vehicle type by inputting it into the neural network for type and estimating the parking rate for the entire parking lot by inputting each calculated parking rate into the overall neural network.

  By the way, in this measurement method, a television camera is used to detect the presence or absence of a vehicle, that is, a parking lot is photographed with the television camera, image processing is performed on the photographed image, and the vehicle is parked in the parking lot. It was judged whether or not.

JP 2000-182187 A

  However, image processing is used to detect the vehicle. Usually, edge extraction or the like is performed, and the presence of the vehicle is determined by comparing the extracted image with reference image data. .

  By the way, the vehicle is extracted from the captured image, but when the parking lot is indoors, it is not so much affected by the sun rays, but when the parking lot is outdoors, Due to the influence, that is, the influence of the weather, there is a problem that the recognition rate is lowered.

  Then, an object of this invention is to provide the detection system and detection method of the vehicle in a parking lot with a high recognition rate even if it is a picked-up image in an outdoor parking lot.

In order to solve the above-described problem, the vehicle detection system in the parking lot according to claim 1 of the present invention performs fuzzy clustering on the image data obtained by photographing the parking lot, and the vehicle compartment is in the parking state or in the empty state. A vehicle detection system for determining whether
A normalization unit that cuts out the passenger compartment from the captured image taken by the camera for photography and obtains and normalizes a plurality of representative values from the image data for each of the extracted passenger compartments, and obtained by this normalization unit A vectorization unit that vectorizes image data consisting of representative values, a dimension compression unit that compresses the dimension of the vector data vectorized by the vectorization unit, and a vector data compressed by the dimension compression unit for training in advance The Mahalanobis distance calculation unit for obtaining the Mahalanobis distance with respect to the cluster center obtained from the data, and substituting the Mahalanobis distance obtained by the Mahalanobis distance calculation unit into the membership function (u ^* ) shown in the following equation (1) Membership calculator that calculates the membership (tilde u) by substituting the value into the following formula (2), and this member A vehicle presence determination unit that performs clustering on either the parking state or the empty vehicle class using the membership obtained by the top calculation unit to determine whether or not a vehicle exists in the vehicle interior. is there.

但し、上記式中、α_ｑｊはクラスｑ内のクラスターｊの混合比率、Ｓは分散共分散行列、π_ｑはクラスｑの混合比率である。

In the above formula, α _qj is the mixing ratio of cluster j in class q, S is the variance-covariance matrix, and π _q is the mixing ratio of class q.

  A vehicle detection system in a parking lot according to claim 2 uses the principal component analysis method in the detection system according to claim 1 when the dimension is compressed by the dimension compression unit.

Moreover, the vehicle detection system in parking according to claim 3 is the detection system of claim 1, the parameters at (1) the membership function shown in equation (u ^*) (m, gamma, [nu, The particle swarm optimization method is used in the optimization of α).

Further, according to a fourth aspect of the present invention, there is provided a method for detecting a vehicle in a parking lot, wherein fuzzy clustering is performed on image data obtained by photographing the parking lot to determine whether the passenger compartment is in a parking state or an empty state. Detection method,
A vehicle compartment is cut out from a photographed image taken by the camera for photographing, and a plurality of representative values are obtained from the image data for each cut out compartment, normalized, and image data consisting of the normalized representative values is vectorized. The dimension of the vectorized vector data is compressed, the Mahalanobis distance to the cluster center obtained from the training data of the compressed vector data in advance is obtained, and the obtained Mahalanobis distance is expressed by the following equation (3) Substituting into the membership function (u ^* ) shown below and substituting this function value into the following equation (4) to determine the membership (tilde u), and using this calculated membership, there is a vehicle in the passenger compartment This is a method for determining whether or not to do so.

但し、上記式中、α_ｑｊはクラスｑ内のクラスターｊの混合比率、Ｓは分散共分散行列、π_ｑはクラスｑの混合比率である。

In the above formula, α _qj is the mixing ratio of cluster j in class q, S is the variance-covariance matrix, and π _q is the mixing ratio of class q.

Moreover, the detection method of the vehicle in the parking lot which concerns on Claim 5 is the detection method which concerns on Claim 4,
About each parameter of membership function (u ^* ) shown in (3) formula, it determines beforehand based on training data,
And in making this decision,
First, by performing semi-hard clustering for each class that shows parking and empty conditions using training data, the center and membership of each cluster in each class is obtained,
Next, using the evaluation data, the membership function (u ^* ) shown in the expression (3) and the membership (tilde u) shown in the expression (4) are used so that the membership belongs to the correct class (3 ) Is a method for optimizing each parameter (m, γ, ν, α) in the equation.

Furthermore, the detection method of the vehicle in the parking lot according to claim 6 is the detection method according to claim 4,
This is a method of using the particle swarm optimization method when optimizing each parameter (m, γ, ν, α) in the membership function (u ^* ) expressed by equation (3).

  According to the detection system and the detection method, the vehicle compartment is cut out from the captured image taken by the photographing camera, and a plurality of representative values are obtained from the image data for each of the cut out vehicle compartments, and the image data is normalized. Vectorize, compress the dimension of the vectorized vector data, calculate the Mahalanobis distance to the cluster center obtained from the training data of the dimension-compressed vector data in advance, and use the obtained Mahalanobis distance as a membership function And the membership value is calculated based on this function value, and it is determined whether there is a vehicle in the passenger compartment using the calculated membership value. It is possible to judge parking and empty conditions with high accuracy and high identification rate. .

1 is a schematic perspective view of a parking lot in which a vehicle detection system according to an embodiment of the present invention is used. It is a block diagram which shows the schematic whole structure of the detection system. It is a block diagram which shows schematic structure of the vehicle interior setting part in the detection system. It is a block diagram which shows schematic structure of the data creation / adjustment part in the detection system. It is a block diagram which shows schematic structure of the image compression coefficient calculation part in the detection system. It is a block diagram which shows schematic structure of the parameter adjustment part in the detection system. It is a block diagram which shows schematic structure of the parking / vacancy judgment part in the detection system. It is explanatory drawing of the fuzzy clustering in the detection system.

DESCRIPTION OF EMBODIMENTS Hereinafter, a vehicle detection system and a detection method in a parking lot according to an embodiment of the present invention will be described based on specific examples.
The parking lot which concerns on a present Example is demonstrated as what is provided outdoors.

  As shown in FIG. 1, this parking lot has a large number of parking spaces (parking places, hereinafter referred to as vehicle compartments) S arranged on the ground. At the same time, a shooting camera (for example, a CCD camera is used) 2 is fixed to a predetermined column 1, and whether or not the vehicle C exists in the passenger compartment S using a shot image from the shooting camera 2. In other words, a detection system that efficiently detects the vehicle C by determining whether the vehicle is in a parking state or an empty state is provided.

Hereinafter, a vehicle detection system and a detection method will be described with reference to FIGS.
By the way, a fuzzy clustering method is used for this vehicle detection system and detection method (these are also referred to as a detection system or the like).

  First, the fuzzy clustering method used in this detection system and the like will be briefly described. First, classification is performed into a parking state (referred to as class 1) and an empty state (referred to as class 2) using training data. That is, in addition to classification, for each class, for example, it is classified into two clusters, and using the data such as cluster center position (hereinafter referred to as cluster center), variance-covariance matrix, and evaluation data obtained at this time The parameters of the membership function (described later) are optimized. It should be noted that the training data is image data of the parking state and the empty vehicle state of the vehicle imaged in advance, and the evaluation data is also image data of the vehicle room imaged by the imaging camera.

Next, the fuzzy clustering method (FCM: fuzzy c-means) will be described in detail.
Since this fuzzy clustering method uses an iteratively weighted least square method, the objective function J _i is set as shown in the following equation (1).

但し、（１）式中、Ｄはマハラノビス距離で、下記（２）式にて表わされる。

However, in Formula (1), D is Mahalanobis distance and is represented by the following Formula (2).

Ｓ_ｉは分散共分散行列（ファジィ分散共分散行列）、ｖ_ｉはデータの平均値つまりクラスター（ｉ）の中心で、それぞれ下記（３）式および（４）式にて表わされる。なお、ｋはデータ番号である。

S _i is a variance-covariance matrix (fuzzy variance-covariance matrix), and v _i is an average value of data, that is, the center of cluster (i), and is represented by the following formulas (3) and (4), respectively. Note that k is a data number.

ここで、クラスター（ｉ）の混合比率αを下記（５）式のように表わす。

Here, the mixing ratio α of the cluster (i) is expressed by the following equation (5).

上記（２）式〜（５）式を（１）式で示す目的関数値が収束するまで繰り返すのがＦＣＭクラスタリング法である。

The FCM clustering method repeats the above equations (2) to (5) until the objective function value represented by equation (1) converges.

  In this vehicle detection system, the cluster phase of the Mahalanobis distance and the first phase in which training data is clustered for each class to obtain the variance-covariance matrix, and optimization of membership function parameters And a second phase for classifying (identifying) the evaluation data into classes.

By the way, the membership (value) (tilde u _qk ) to the class q of the k-th data x _k is expressed by the following equation (6). In the formula, a tilde symbol (˜) is added to the head of u, and this “tilde u” is expressed as “* u” in the text. The membership function is expressed as “u ^* ” as shown below.

（６）式中、ｕ_ｑｋｊ ^＊はクラスターｊに対するメンバーシップ関数で、下記（７）式で表わされる。

In the equation (6), u _qkj ^* is a membership function for the cluster j and is represented by the following equation (7).

π _q is a mixing ratio of class q (mixing ratio of training data), that is, a prior probability, c is the number of clusters j for each class, and c = 2 in this embodiment.

上記（７）式のメンバーシップ関数ｕ_ｑｋｊ ^＊には複数のパラメータが含まれており、これらのパラメータは、上述したように、訓練用データおよび評価用データを用いて最適化が行われる。

The membership function u _qkj ^* of the above equation (7) includes a plurality of parameters, and these parameters are optimized using the training data and the evaluation data as described above.

  When performing clustering, semi-hardware clustering is performed in order to shorten the calculation time. In this semi-hardware clustering, discrete values (for example, 0.6, 0.8, etc.) are used as membership.

  In this semi-hardware clustering, as described above, there are two classes, that is, the parking state is class 1, the empty state is class 2, and each of these classes is further divided into two clusters. This is because experience has shown that dividing each class into two clusters performs efficient clustering. The number of clusters is not limited to two, and may be one or three or more, for example. For example, when the discrete value is 0.5, the number of clusters is 1, and when the discrete value is 1.0, two hard (not fuzzy) clusters are obtained.

In the first phase, semi-hardware clustering is performed for each class, that is, for the image data in the parking state and the image data in the empty state, using the training data.
In order to perform semi-hard clustering that is intermediate between hard and fuzzy, membership u _ki is set as shown in the following equation (8).

以下、具体的な手順について説明する（クラスター数（ｉ）は２個とする）。なお、βは上述した離散値で指定される。

Hereinafter, a specific procedure will be described (the number of clusters (i) is two). Note that β is specified by the discrete value described above.

First, the initial membership (u _k1 , u _k2 ) for each cluster is set as in the following equations (9) and (10).

なお、上記式中、ｆ_ｋは１つのクラスの画像データｘ_ｋの主成分得点である。

In the above formula, f _k is a main component score of one class of image data x _k .

  When membership is given by the above equation, v is obtained from equation (4), S is obtained from equation (3), and Mahalanobis distance D is obtained from equations (2) and (3), (8) Update membership u with formula. This update is repeated until membership has converged.

Next, clustering is performed for each class (class 1 → parking state; class 2 → empty state) using the training data. That is, each class is divided into two clusters.
At this time, semi-hardware clustering is performed using the objective function J.

That is, the membership u for each training data that minimizes the objective function J is obtained. Thus, the center v of each cluster and its membership u are obtained.
Then, based on the obtained membership u, clustering is performed for each class.

In the second phase, identification (classification) is performed using the evaluation data.
That is, the Mahalanobis distance D is obtained for the center v of each cluster obtained in the first phase. At this time, the membership u obtained in the first phase is used.

Then, by applying (using) the obtained Mahalanobis distance D to the membership function u ^* shown by the predetermined equation (6), each parameter (m, γ, ν) in the membership function u ^* is used. , Α; these are also called free parameters or hyper parameters). In this optimization, a particle swarm optimization method (PSO) is used.

Here, this particle swarm optimization method will be briefly described.
In this particle swarm optimization method, the cluster mixing ratio α, which is the same parameter as the three parameters (m, γ, ν) in the membership function u ^* , is optimized.

  That is, in this particle group optimization method, the search for the best position of the particle group is performed by the following update formulas (11) and (12).

上記式中、Paraは粒子の位置を示すパラメータで、α，ｍ，γ，νからなるベクトルである。Veloは粒子の速度ベクトルである。Randは「０」と「１」との間の乱数の対角行列である。ｗ_０，ｃ_１，ｃ_２はスカラー定数である。pbset とgbest はそれぞれpbset とgbestの位置ベクトルである。

In the above formula, Para is a parameter indicating the position of the particle and is a vector composed of α, m, γ, and ν. Velo is the velocity vector of the particle. Rand is a diagonal matrix of random numbers between “0” and “1”. w ₀ , c ₁ , and c ₂ are scalar constants. pbset and gbest are position vectors of pbset and gbest, respectively.

These free parameters are determined so as to minimize the misidentification rate of the evaluation data.
That is, optimization is performed so that the membership * u using the membership function u ^* belongs to the correct class, in other words, the correct class membership * u is larger.

In the above-described preparation process, the membership function u ^* is obtained.
And the outline of the judgment process which actually judges the presence or absence of parking is demonstrated.
That is, the Mahalanobis distance D with respect to the center v of each cluster obtained in the preparation step of image data that has been photographed and subjected to predetermined processing (described later) is obtained, and this Mahalanobis distance is expressed by the above equation (6). The membership function value u ^* obtained by substitution is substituted into the above equation (5) to obtain the membership * u for the cluster in each class. Then, membership * u obtained for each cluster is added to obtain memberships * u ₁ and * u ₂ for each class.

Then, by comparing the membership * u ₂ membership * u ₁ and Class 2 Class 1, the larger the value is, it can be determined that the class to which the image data belongs. In other words, if the membership * u ₁ of the class 1 is large, it is determined that the parking state. On the other hand, when the membership * u ₂ of Class 2 is large, it is determined that the unladen state.

Hereinafter, a vehicle detection system that performs the above-described preparation process and determination process will be described.
In the following description, the free parameters (m, γ, ν, α), the cluster center v, the variance-covariance matrix S, the class mixture ratio π, and the like are factors involved in the identification, and are therefore referred to as determination parameters. explain.

  As shown in FIG. 2, the detection system (as a program) is roughly divided into a vehicle interior setting unit (vehicle interior setting program) 11 and creation of necessary data such as an image compression coefficient (described later) and a vehicle. A data creation / adjustment unit (data creation / adjustment program) 12 for adjusting determination parameters used to determine the presence or absence of a vehicle, and parking to determine whether the set cabin S is parked or empty An empty determination unit (parking / empty determination program) 13 and a database unit 14 for storing basic data, various parameters, and the like are provided. The passenger compartment setting unit 11 and the data creation / adjustment unit 12 are for preparation, and the parking / vacancy determination unit 13 is for actual operation. Of course, the database unit 14 is used for both preparation and actual operation.

  As shown in FIG. 3, the passenger compartment setting unit 11 inputs image data captured by the imaging camera 2 fixed to the column 1, for example, and inputs the vehicle data into the image data (that is, in the imaging screen). The vehicle interior position setting unit 21 that sets the vehicle interior S for determining (determining) the presence / absence, and the vehicle interior that acquires the vehicle interior location information set (position determined) by the vehicle interior location setting unit 21. The vehicle location information acquired by the vehicle location information acquisition unit 22 is stored in the database unit 14.

  As shown in FIG. 4, the data creation / adjustment unit 12 is already photographed in the same manner as an image compression coefficient calculation unit 31 that calculates an image compression coefficient for compressing image data based on already photographed image data. A parameter adjustment unit 32 that adjusts a determination parameter based on the image data, and the image compression coefficient obtained by the image compression coefficient calculation unit 31 and the determination parameter adjusted by the parameter adjustment unit 32 are stored in a database unit. 14 is stored.

  As shown in FIG. 5, the image compression coefficient calculation unit 31 reads an image reading unit 31 a that reads (captures) previously captured image data, and the vehicle interior 1 from the image data read by the image reading unit 31 a. A vehicle compartment cutout portion 31b that is cut out (for example, cut out in a rectangular shape such as a parallelogram), a normalization portion 31c that normalizes the vehicle compartment 1 cut out at the vehicle compartment cutout portion 31b, The principal component analysis unit that compresses the data by reducing the dimension of the data (for example, from about 1024 dimensions to about 50 dimensions) by inputting the normalized image data in the conversion unit 31c and performing the principal component analysis 31d and an image compression coefficient acquisition unit 31e that acquires the principal component vector (base vector) obtained by the principal component analysis unit 31d and the average vector of the image data as image compression coefficients. There. Note that the normalization described above is, for example, on the captured image data, because the size of the screen that is determined to be a passenger compartment differs depending on the position of the passenger compartment (especially, the difference in perspective). Pixel data (the same number of data). Specifically, color image data is converted into black and white data, that is, the image data converted into luminance values (for example, numerical values of 0 to 255) is divided into, for example, 32 in both vertical and horizontal directions to obtain 1024 (= 32 × 32) pieces. Get the pixel value (normalization). Actually, the luminance value of one pixel (for example, the central pixel of the divided region) in each of the 1024 divided regions (plural regions) is used as the representative value (the average luminance value of each divided region is used as the representative value). Can also). Thereby, 1024-dimensional data is obtained. Therefore, the normalization unit 31c is configured by a data acquisition unit that inputs image data and acquires representative values of the image data at a plurality of locations (for example, 1024 locations). The normalization unit 31c includes a data region dividing unit that inputs image data and divides it into a plurality of regions, and a data acquisition unit that acquires representative values of image data in the divided regions divided by the data region dividing unit. It can be said that it is composed of

  Further, as shown in FIG. 6, the data etc. creating / adjusting unit 32 reads (captures) previously captured image data, and the image compression acquired by the image compression coefficient acquisition unit 31e. The dimension of the image data read by the image reading unit 32a using a coefficient is reduced to a low dimension, that is, a dimension compression unit 32b that performs dimension compression, and the image data that has been dimensionally compressed by the dimension compression unit 32b is used as a training image. As a training image storage unit 32c, and a parameter acquisition unit 32d that acquires the judgment parameters by performing training by inputting the image data stored in the training image storage unit 32c. The determination parameter acquired by the data creation / adjustment unit 32 is used as an initial parameter during operation, and also as an initial value during training.

Next, the parking / empty determination unit 13 executed during actual operation will be described.
Since the parking / air determination unit 13 operates independently, the image data is read and the components such as dimensional compression are provided with the same components as described above. This component will also be described again.

As shown in FIG. 7, the parking / air determination unit 13 includes an image reading unit 41 that reads image data that is a captured image taken by the shooting camera 2, and image data that is read by the image reading unit 41. A vehicle compartment cutout unit 42 that cuts out (specifies) the vehicle compartment from the vehicle, and a normalization unit that normalizes the image data of the vehicle compartment cut out by the vehicle compartment cutout unit 42, that is, divides the image data into 1024 43, a vectorizing unit 44 that converts the image data normalized by the normalizing unit 43 into a 1024-dimensional vector, and a dimension for compressing the dimension of the vector data vectorized by the vectorizing unit 44 to, for example, 50 dimensions A compression unit 45; a Mahalanobis distance calculation unit 46 for obtaining a Mahalanobis distance with respect to a center of a cluster obtained in advance of the image data dimensionally compressed by the dimension compression unit 45; The membership calculation section 47 above (7) and the equation (6) is provided to determine the membership functions u ^* and membership * u based on the Mahalanobis distances determined by Nobis distance calculator 46, the membership calculation Based on the membership * u obtained by the unit 47, the image data is constituted by a vehicle presence determination unit 48 that determines which class, that is, a parking state or an empty state. In the calculation by the vehicle compartment cutout section 42, the dimension compression section 45, the Mahalanobis distance calculation section 46, the membership calculation section 47, and the vehicle presence determination section 48, appropriate vehicle compartment position information and image compression coefficients are obtained from the database section 14. The determination parameters are read and used. Note that, in actual operation, the read determination parameter is a value optimally adjusted by training or the like. By the way, when the parking lot is large, the range of photographing with the photographing camera is naturally limited, so that grouping is performed. Therefore, since fuzzy clustering is performed for each group, the determination parameter is also adjusted for each group. Also, the image compression coefficient is obtained over a slightly wide range. For example, it is calculated | required for every parking lot instead of every group.

Specifically, the Mahalanobis distance calculation unit 46 calculates the Mahalanobis distance D with respect to the cluster center obtained in advance of the image data, and the membership calculation unit 47 calculates the membership function u ^* (function) based on the Mahalanobis distance D. Value), and membership * u is obtained based on this membership function u ^* . That is, membership * u ₁ , * u ₂ is obtained for each cluster. Then, for each class, membership is added, and memberships * u _c1 and * u _c2 in the class are obtained.

Then, the vehicle presence determination unit 48 compares the obtained memberships * u _c1 and * u _{c2 in} each class, and determines that the image data belongs to the larger one. .

For example, the membership in each class at the time of this determination is shown in FIG. 8, the two Class 1 and Class 2, in which respectively form two clusters 1 and cluster 2, respectively, the class 1 side, there is a cluster 2 clusters 1 and v ₁₂ of v _11, also On the class 2 side, there is also a cluster 1 of v ₂₁ and a cluster 2 of v ₂₂ . In FIG. 8, if the point of X is image data to be judged, the membership of X for one cluster 1 in class 1 is u ₁₁ , the membership of X for the other cluster 2 is u ₁₂ , and the class 2, the membership of X to class 1 is u ₂₁ , and the membership of X to the other cluster 2 is u ₂₂ , the membership u ₁ to class 1 of X is u ₁₁ + u ₁₂ , membership _{u 2} for the class _2, the _u 21 ₊ u _22. That is, the total value of membership of each cluster in each class (which is the sum of membership heights for both clusters) becomes the membership for each class. Of the memberships u ₁ and u ₂ , they belong to the larger class. If it belongs to class 1, it is in a parking state, and if it belongs to class 2, it indicates an empty state.

  Note that free parameters are searched so that the misidentification rate is minimized by the particle swarm optimization method described above. For example, when the cluster mixing ratio α is changed, the cluster region (shown by contour lines) in FIG. Will change, and therefore the boundaries of the two classes will also change non-linearly.

Here, an outline process from installation of the entire system to start of operation will be described.
First, there are two cases in installing the detection system, as described below. Hereinafter, a procedure for the installation will be schematically described, and then a specific description will be given based on one case.

  The first is when the local (parking lot) camera image, that is, image data can be obtained before system operation (when the existing camera is diverted), and the second is when the local image data cannot be obtained before system operation ( (When a camera is newly established).

In the first case;
Using the image data obtained in advance, the setting of the passenger compartment position, the calculation of the image compression coefficient, and the generation (training) of the determination parameters are performed, and the system is operated using these.

Even after the system is operated, the accuracy of the identification can be improved by collecting local image data and continuing the training of the determination parameters.
Note that the image compression coefficient and the determination parameter used in the parking lot are registered in the database unit 14 (also used in another parking lot).

In the second case;
When the photographing camera is installed and image data can be captured, the image data is captured, and the vehicle interior position is temporarily set.

  However, even if the image data can be captured, the vehicle is not parked until the parking lot is opened, so it is not possible to collect on-site image data. Unable to train parameters.

Therefore, when the parking lot is initially opened, the system is operated using the image compression coefficient and the determination parameter of the parking lot whose conditions are similar to the parking lot in the database unit 14.
At the same time, image data at this parking lot is collected, the position of the passenger compartment is set in detail using this image data, and a new calculation of image compression coefficient and generation (training) of judgment parameters are performed. To do.

Then, when the identification accuracy based on the determination parameter becomes better than the determination parameter of the other parking lot currently used, the determination parameter is replaced.
After that, we continue to collect local image data and train the judgment parameters to improve the identification accuracy. Of course, also in this case, the image compression coefficient and the determination parameter used in the parking lot are registered in the database unit (also used in another parking lot).

Next, the first case will be described more specifically.
That is, the database unit 14 stores an image compression coefficient and determination parameters obtained from image data of an existing parking lot.

  As described above, the image compression coefficient is a principal component vector (base vector) when image data is compressed by principal component analysis and an average vector of the image data. There are as many principal component vectors as the number of dimensions to be compressed. That is, if the number of dimensions is m (for example, 50) and the number of pixels is n (for example, 32 × 32 = 1024), the image compression coefficient is n × (m + 1 (for the average vector). ) Will exist.

The determination parameter is data for obtaining the Mahalanobis distance for each cluster that is initially obtained, and is, for example, the center of each cluster.
In a new parking lot detection system, image data is collected by a newly installed shooting camera to improve the accuracy of determination parameters.

That is, a new parking lot is photographed by the photographing camera 2, and the position of the passenger compartment S is specified based on the photographed image data.
When the position of the passenger compartment S is specified, image data is read from the photographing camera 2, the image data is normalized and vectorized, and then the dimension is determined based on the image compression coefficient stored in the database unit 14. Compress from 1024 dimensions to 50 dimensions. Then, the Mahalanobis distance is obtained for the dimensionally compressed image data, the membership function is obtained using the Mahalanobis distance, and the membership for each class is obtained using the obtained membership function value. .

  These obtained memberships are added for each class to obtain the membership for each class, and it is determined that the membership belongs to the class with the larger membership. By repeating this process for all image data, the determination parameter is corrected.

Next, the determination as to whether or not the vehicle is in a parking state will be specifically described.
That is, the image data of the passenger compartment S photographed by the photographing camera 2 is read into the image reading unit 41, and this image data is, for example, 1024-dimensional by the passenger compartment cutout unit 42, the normalizing unit 43, and the vectorizing unit 44. After being vectorized into data, the dimension is compressed by the dimension compression unit 45. In the dimension compressing unit 45, the dimension is compressed from 1024 dimensions to 50 dimensions based on the image compression coefficient stored in the database unit 14. Next, the dimensionally compressed image data is obtained after the Mahalanobis distance is obtained by the Mahalanobis distance calculation unit 46 and the membership function is obtained by the membership calculation unit 47 using the Mahalanobis distance. Membership for each class is determined using the membership function value. Then, among these obtained memberships, membership for each class is obtained by adding for each class.

  Next, the membership for each class is input to the vehicle presence determination unit 48, and it is determined to which class the membership related to the image data belongs. That is, it is determined whether the captured image is in a parking state or an empty state.

Note that most of the components such as the data creation / adjustment unit, the various calculation units, the compression unit, and the determination units in the above description are implemented by programs.
In the above description, the parking lot is described as being outdoors, but the detection system can naturally be applied to indoors with good shooting conditions.

Here, an experimental result in which this detection system is actually applied will be described.
The parking situation of 27 compartments was identified from the image of the surveillance camera (which is a camera for photographing) installed in the rooftop parking lot. In the experiment, 2600 images were selected and used from image data for 3 weeks under various conditions with different weather and time zones.

The procedure for performing the performance evaluation will be described below.
First, the camera number, the number of images to be used, the number of passenger compartments and their grouping were specified, and then the cutout location of the passenger compartment was specified on the screen.

  First, 50 principal component vectors were obtained from the first 100 (2700) image data. The training data was selected from 1000 sheets in the first week and 600 sheets in the 3rd week, and evaluated as 1000 tests (27000) in the 2nd week. By selecting 20 images (50 jumps) from the first week and inputting parking / vacancy (class) of each parking space, training data was collected and learning / training of the classifier was performed.

Next, the image of the first week with misidentification was added to the training data, and finally the classifier was learned and trained again. Furthermore, training data was added in the same manner for the third week.
In the lower part of (Table 1) and (Table 2) shown below, the misidentification rate evaluated with 1000 sheets (27000 cases) in the second week is shown. In addition, each table | surface shows the misidentification rate at the time of dividing day and night.

  That is, a part of the sky was cut out from the camera image, and day / night was determined based on the threshold of the average brightness. (Table 1) selects the training data for the first week from the 1000 sheets of 111 sheets (2997 cases), and (Table 2) adds the first week from the 600 sheets for the first week to 44 sheets (1188 cases). A total of 155 sheets (4185 cases) from the week and third week data were used as training data. When training data for the third week was added (Table 2), compared to the case where it was not added (Table 1), the daytime decreased from 2.53% to 1.42%, and the whole day and night was 2.04. %, The misidentification rate was practical. For evaluation of misidentification, 27000 test data of the second week containing correct class information were used.

このように、屋外では１台の撮影用カメラで多くの駐車スペースをカバーできるため、室内用に比べて低コストの検知システムを実現することができる。また、センサー方式に比べても安価であり、既設駐車場への適用も容易であり、上記のように、性能評価実験では２．０４％という実用化可能な誤識別率となった。

In this way, since a large number of parking spaces can be covered with a single camera for shooting outdoors, a low-cost detection system can be realized as compared with indoor use. Moreover, it is cheaper than the sensor method and can be easily applied to an existing parking lot. As described above, in the performance evaluation experiment, a misclassification rate of 2.04% that can be put into practical use was obtained.

  As described above, according to the detection system and the detection method, the vehicle interior is cut out from the captured image captured by the imaging camera, that is, the image data, and a plurality of representative values are acquired from the image data for each of the extracted vehicle compartments. Normalize, vectorize the image data as the normalized representative value, compress the dimension of the vectorized vector data, and apply the compressed vector data to the cluster center previously obtained from the training data. Calculate the Mahalanobis distance, substitute the calculated Mahalanobis distance into the membership function, determine the membership based on this function value, and use this calculated membership to determine whether there is a vehicle in the passenger compartment. So, even if the parking lot is outdoors, it is efficient, that is, high identification rate Have been (in misidentification less state), it is possible to determine the parking state and the unladen state.

  Further, since the dimensions of the image data are compressed by the principal component analysis method, the pixel data can be reduced while maintaining the image characteristics. In other words, it is possible to detect a parking state or an empty state with high accuracy by using an imaging camera that is an inexpensive configuration.

  Furthermore, among the determination parameters, the membership function free parameters are optimized by the particle swarm optimization method, so that the optimization can be achieved quickly and efficiently. That is, the misidentification rate can be reduced. As shown in the performance evaluation experiment, the misidentification rate can be further greatly reduced by re-optimizing by adding training data.

DESCRIPTION OF SYMBOLS S Parking space 1 Support | pillar 2 Camera for imaging | photography 11 Vehicle compartment setting part 12 Data etc. preparation / adjustment part 13 Parking / vacancy judgment part 14 Database part 21 Vehicle compartment position setting part 22 Vehicle compartment position information acquisition part 31 Image reading unit 31b Vehicle compartment extraction unit 31c Normalization unit 31d Principal component analysis unit 31e Image compression coefficient acquisition unit 32 Parameter adjustment unit 32a Image reading unit 32b Image compression unit 32c Training image storage unit 32d Parameter acquisition unit 41 Image reading unit 42 Vehicle compartment extraction part 43 Normalization part 44 Vectorization part 45 Dimension compression part 46 Mahalanobis distance calculation part 47 Membership calculation part 48 Presence determination part

Claims (6)

A vehicle detection system that performs fuzzy clustering on image data obtained by photographing a parking lot and determines whether the passenger compartment is in a parking state or an empty state.
A normalization unit that cuts out the passenger compartment from the captured image taken by the camera for photography and obtains and normalizes a plurality of representative values from the image data for each of the extracted passenger compartments, and obtained by this normalization unit A vectorization unit that vectorizes image data consisting of representative values, a dimension compression unit that compresses the dimension of the vector data vectorized by the vectorization unit, and a vector data compressed by the dimension compression unit for training in advance The Mahalanobis distance calculation unit for obtaining the Mahalanobis distance with respect to the cluster center obtained from the data, and substituting the Mahalanobis distance obtained by the Mahalanobis distance calculation unit into the membership function (u ^* ) shown in the following equation (1) Membership calculator that calculates the membership (tilde u) by substituting the value into the following formula (2), and this member And a vehicle determination unit that performs clustering on either a parking state or an empty vehicle class using the membership obtained by the top-up calculation unit and determines whether or not a vehicle exists in the passenger compartment. A vehicle detection system in a parking lot.

In the above formula, α _qj is a mixing ratio of cluster j in cluster q, S is a dispersion covariance matrix, and π _q is a mixing ratio of class q.   The system for detecting a vehicle in a parking lot according to claim 1, wherein a principal component analysis method is used when the dimension is compressed by the dimension compression unit. The particle swarm optimization method is used when optimizing each parameter (m, γ, ν, α) in the membership function (u ^* ) represented by the equation (1). The vehicle detection system in the parking lot described in 1. A vehicle detection method for performing fuzzy clustering on image data obtained by photographing a parking lot and determining whether a vehicle compartment is in a parking state or an empty state,
A vehicle compartment is cut out from a photographed image taken by a camera for photography, and a plurality of representative values are acquired from the image data for each cut out vehicle compartment, normalized, and image data composed of the normalized representative values is vectorized. The dimension of the vectorized vector data is compressed, the Mahalanobis distance to the cluster center obtained from the training data of the compressed vector data in advance is obtained, and the obtained Mahalanobis distance is expressed by the following equation (3) Substituting into the membership function (u ^* ) shown below and substituting this function value into the following equation (4) to determine membership (tilde u), and using this calculated membership, there is a vehicle in the passenger compartment A method of detecting a vehicle in a parking lot, characterized by determining whether or not to do so.

In the above formula, α _qj is a mixing ratio in the cluster q, S is a dispersion covariance matrix, and π _q is a mixing ratio of class q. About each parameter of membership function (u ^* ) shown in (3) formula, it determines beforehand based on training data,
And in making this decision,
First, by performing semi-hard clustering for each class that shows parking and empty conditions using training data, the center and membership of each cluster in each class is obtained,
Next, using the evaluation data, the membership function (u ^* ) shown in the expression (3) and the membership (tilde u) shown in the expression (4) are used so that the membership belongs to the correct class (3 5. The method for detecting a vehicle in a parking lot according to claim 4, wherein each parameter (m, γ, ν, α) in the equation (5) is optimized. 6. The particle swarm optimization method is used when optimizing each parameter (m, γ, ν, α) in the membership function (u ^* ) expressed by equation (3). The detection method of the vehicle in the parking lot of description.

Images