JP2010224922A - Class discriminator generating device - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a class discriminator generating device for generating a class discriminator having good detecting accuracy while reducing the amount of features to be learned as much as possible. <P>SOLUTION: Class discriminator candidates are generated by learning a weak discriminator (i.e., feature quantity) using a plurality of positive images for learning and a plurality of negative images for learning (S120). Using the class discriminator candidates generated, it is detected whether or not there is the subject of detection among a group of images for evaluation (S130-S190), and if the detection accuracy thereof is less than a predetermined value (S200:NO), images causing worsening among the group of positive images for learning are replaced with unused positive images. Using the group of new positive images for learning that have replaced the images causing worsening, and the group of negative images for learning, the weak discriminator is learned and new class discriminator candidates are generated (S120). Such new class discriminator candidates include those generated without learning the feature quantities required to identify adverse influencing factors. <P>COPYRIGHT: (C)2011,JPO&INPIT

Description

  The present invention relates to a class discriminator generating device that generates a class discriminator used for detecting a detection target reflected in an input image.

  Conventionally, a pedestrian or a vehicle (hereinafter referred to as a detection target) located in front of the host vehicle is detected based on an image input from an imaging device (hereinafter referred to as an input image), A driving support device that executes vehicle control is known.

In such a driving support device, it is important to accurately detect the detection target reflected in the input image.
As a method for accurately detecting this detection target, weak classifiers having Haar Like features (that is, feature quantities) are learned by boosting theory (for example, Adaboost), and a series of weak classifiers each having a weight assigned thereto are arranged. Further, there is known a technique using a cascade-connected strong discriminator (hereinafter also referred to as a class discriminator) (see, for example, Patent Document 1). Hereinafter, a device that collates such a strong discriminator with an input image from the imaging device and detects a detection target from the input image is referred to as a conventional discriminating device.

  In the conventional identification device, weak classifiers are sequentially applied to the partial areas cut out from the input image in the cascade connection order, and the detection target is reflected in the partial areas based on the identification results by the individual weak classifiers. A likelihood value representing the possibility of being present is obtained. If the weighted linear sum of the obtained likelihood values is equal to or greater than a predetermined threshold set during the weak classifier learning, the detection target is detected as being detected in the partial region. is doing.

  The weak classifier learning is performed using a learning image prepared in advance. The learning image includes a plurality of (for example, 5000) learning positive images in which only one detection target is reflected and a plurality of (for example, 5000) learning negative images in which the detection target is not reflected. Become.

  In specific learning for each stage, each of a plurality of weak classifiers prepared in advance is compared with each of the learning images, and an arrangement of weak classifiers that can identify whether or not they are detection targets, A weight for the weak classifier and a specified threshold value used when detecting a detection target are set. However, in this learning, from the set of learning positive images, the detection target reflected in those images is detected from the learning positive images having a specified rate or more, and the set of learning negative images is set. The weak classifiers that constitute the stage are added until it is correctly detected that the detection target is not reflected in those images from the learning negative images that are higher than the rate. The specified rate is set to a value larger than the set rate, and for example, the specified rate is set to 99.5% and the set rate is set to 50%.

JP 2008-20951 A

  By the way, since the set of learning positive images is arbitrarily selected by the creator of the class discriminator, it is different from a general feature representing a detection target (hereinafter referred to as a heterogeneous feature. If it is a pedestrian, it may include an image that includes a detection target that has individual characteristics such as the pedestrian's orientation and body shape, and features that appear depending on the type of belongings the pedestrian has. There is sex.

  When learning is performed using such a set of positive images for learning, a weak classifier is set so that a detection target having a foreign feature can also be detected. The discriminator does not always match the general characteristics of the detection target.

  In other words, in learning in such a case, over-learning is performed to identify heterogeneous features, and if a detection target reflected in the input image is detected using a class discriminator generated by the learning, detection accuracy decreases. There was a problem that could be.

  In particular, a class discriminator generated in such an overlearned state increases the amount of features (that is, the number of weak discriminators) necessary for detecting a detection target, and is reflected in the input image. A problem arises in that the amount of processing when detecting a detection target that is included increases.

  Therefore, an object of the present invention is to provide a class discriminator generating device that generates a class discriminator with high detection accuracy while suppressing the feature quantity to be learned as much as possible.

  The present invention made to achieve the above object relates to a class discriminator generating device for generating a class discriminator for detecting a detection target reflected in an input image from an imaging device.

  In the classifier generating apparatus of the present invention, a plurality of positive images are stored in the positive image storage means, and a plurality of negative images are stored in the negative image storage means. However, the positive image referred to here is each image in which the detection target is reflected, and the negative image is each image in which the detection target is not reflected.

  The discriminator candidate generation unit includes a positive image group including a part of positive images stored in the positive image storage unit and a negative image including a part of negative images stored in the negative image storage unit. A classifier learning process for generating classifiers learned using groups as class classifier candidates is executed.

  Further, the new candidate generation means uses the deterioration factor image in the positive image stored in the positive image storage means in the previous learning in the positive image group used for learning in the classifier candidate generation means. Then, the classifier candidate generating unit is caused to execute a classifier learning process by replacing the unused classifier with the unused one, thereby causing the classifier candidate generating unit to generate a new class classifier candidate.

  Then, the class discriminator determining means is for evaluation positive images prepared separately from the positive images stored in the positive image storage means and the negative images stored in the negative image storage means. The classifier candidates generated by the classifier candidate generation unit are collated with each negative image group. As a result of the collation, the class is based on an evaluation value that increases as the detection rate of the detection target from the evaluation positive image group increases and increases as the detection rate of the detection target from the evaluation negative image group decreases. Determine the classifier.

  That is, in the class discriminator generation device of the present invention, when detecting a detection target from an input image, adverse effects (for example, reducing the detection accuracy or increasing the amount of processing until a detection result is derived). A positive image that becomes a factor to be affected (for example, a detection target having a special feature or large noise (ie, a heterogeneous feature), hereinafter referred to as an adverse effect factor) is set as a deterioration factor image, and the deterioration factor image is set as another positive image And the classifier learning process is executed. Thereby, a class identifier candidate learned using a positive image group constituted by different positive images is newly generated.

  Among the class identifier candidates newly generated in this manner, those generated without learning the feature amount necessary for identifying the adverse effect factor are included. That is, the class discriminator candidate includes only learned feature quantities for identifying general features to be detected.

  Therefore, according to the class discriminator generation device of the present invention, it is possible to reduce the number of feature quantities to be learned for generating a class discriminator. Then, by using the class discriminator generated by the class discriminator generating apparatus of the present invention, if the detection target reflected in the input image is detected, the detection accuracy can be improved, and the input image can be reflected. It is possible to reduce the amount of processing when detecting the detection target.

In other words, according to the class discriminator generation device of the present invention, it is possible to generate a class discriminator with good detection accuracy while suppressing the feature quantity to be learned as much as possible.
By the way, the class discriminator may be SVM (Support vector machine) or backpropagation, but a plurality of weak discriminators using an integral value obtained by integrating pixel values in a rectangular area set in an input image as a feature amount. It is desirable that the device is configured by cascading devices.

  When such a class discriminator is used, the discriminator candidate generating unit detects a detection target from positive images that form a positive image group and having a specified rate or more, as described in claim 2, It is configured to perform learning as a discriminator learning process by learning a new weak discriminator from a negative image forming a negative image group until a detection target is not detected from a negative image having a set rate or higher. It is desirable. That is, it is desirable that the discriminator candidate generating means execute learning using so-called boosting theory as the discriminator learning process.

However, the specified rate here is less than 1 and a value larger than the set rate.
In the class discriminator generating device according to claim 2, the deterioration factor image that is replaced with another positive image by the new candidate generating unit is the discriminator by the discriminator candidate generating unit as described in claim 3. A positive image in which the detection target is not detected when the learning process is executed may be used. Further, as described in claim 4, the deterioration factor image can detect a detection target by learning more than a set number of weak classifiers when the classifier learning process is executed by the classifier candidate generation unit. It may be a positive image.

  By replacing these deterioration factor images with other positive images, the classifier generated by the classifier generator of the present invention has a weak classifier required to identify adverse factors. Reflection can be reduced more reliably.

  By the way, in this invention, a class discriminator determination means may be comprised so that a class discriminator candidate with the largest evaluation value may be used as a class discriminator. The class discriminator determining means, as described in claim 6, performs class discrimination by classifier candidate generating means for collating the class discriminator candidate with each of the evaluation positive image group and the evaluation negative image group. This is performed every time a classifier candidate is generated. If the evaluation value is equal to or greater than a specified value, the generated class classifier candidate may be configured as a class classifier.

  According to the class discriminator determining means configured as in the former, it is possible to determine, as a class discriminator, a plurality of generated class discriminator candidates that have the best detection target detection performance. As a result, it is possible to accurately detect the detection target reflected in the input image.

  Further, in the class discriminator generating device in which the class discriminator determining means is configured as the latter, the discriminator is configured to end the execution of the discriminator learning process when the class discriminator determining unit determines the class discriminator. If the candidate generation unit and the new candidate generation unit are configured, it is not necessary to generate class classifier candidates more than necessary. As a result, the processing time required until the class discriminator is determined can be shortened.

It is the block diagram which showed the functional structure of the class identifier production | generation apparatus. It is the block diagram which showed schematic structure of the driving assistance system which detects a detection target using a class identifier. It is the flowchart which showed the process sequence of the class identifier production | generation process. It is the flowchart which showed the process sequence of the discriminator learning process. It is drawing which illustrated the positive image list for learning. It is explanatory drawing which showed the concept at the time of replacing the learning positive image in a classifier production | generation process. It is drawing which illustrated the Haar Like feature used as a weak discriminator.

Embodiments of the present invention will be described below with reference to the drawings.
<About the driving support system>
FIG. 2 is a block diagram showing a schematic configuration of an in-vehicle driving support system using a class discriminator generated by a class discriminator generating device to which the present invention is applied.

The class discriminator used in the driving support system 1 (that is, generated by the class discriminator generating device) represents a feature of a preset detection target (for example, a pedestrian). This class discriminator is configured by further cascading each stage composed of a plurality of weak discriminators having Haar Like features as feature quantities. Each stage constituting this class discriminator is formed by cascading weak discriminators each having a weight (discriminator weight α _t described later) in cascade, and the number of connection stages t of the weak discriminators is determined in advance. It is the set number of settings.

  The Haar Like feature is, as shown in FIGS. 7A to 7C, an integral value obtained by integrating the pixel values in the rectangular area (that is, the sum of the pixel values) as a feature amount. In the example illustrated in FIG. 7, a difference between integral values in a shaded area and a non-shaded area is a feature amount.

  And the driving assistance system 1 collates a class identifier with the imaging | photography mechanism (for example, CCD camera) 3 which acquires the image which image | photographed the front of the own vehicle as an input image, and the input image acquired by the imaging | photography mechanism 3. By detecting the presence / absence of a detection target (for example, a pedestrian or a vehicle) reflected in the input image (hereinafter referred to as target presence / absence determination), and according to the detection result, to support the driver, It is a known device including a control device 6 that controls a control target device group 9 including a plurality of in-vehicle devices (for example, monitors and speakers).

  In other words, the control device 6 sequentially applies weak classifiers to the partial areas cut out from the input image in the cascade connection order, and calculates a likelihood value indicating the possibility that the detection target is reflected in the partial areas. Ask. Furthermore, a weighted linear sum is derived from the likelihood value obtained for each weak classifier, and the derived weighted linear sum is obtained when the weak classifier is learned (that is, when the class classifier is generated). If the threshold value is equal to or greater than the set threshold th, the area in which the detection target exists is detected as the detection target being reflected in the partial image.

When the detection target is detected from the input image, the control device 6 displays the input image itself on the monitor and surrounds the marker (for example, the detection target) at the position where the detection target exists on the input image. A large frame). Further, the speaker is made to output a sound indicating that the detection target exists in front of the host vehicle.
<Configuration of classifier generator>
FIG. 1 is a block diagram showing a functional configuration of a class discriminator generation device to which the present invention is applied.

  In this class discriminator generating device 20, the class discriminator is generated by learning a weak discriminator by a boosting theory (for example, AdaBoost) using a learning image prepared in advance. An image used for learning (that is, a learning image) includes a plurality of positive images in which only one detection target is reflected and a plurality of negative images in which the detection target is not reflected.

  It should be noted that all positive images and negative images (here, including an image for evaluation described later in addition to a learning image) are normalized so that the image size is uniform. At the same time, each learning image is associated with an image index for distinguishing each learning image.

  Hereinafter, each positive image that may be used for learning, that is, each positive image as a learning image, each positive image for learning, and each negative image that may be used for learning, that is, each negative image as a learning image Is called a negative image for learning.

  As described above, the class discriminator generation device 20 that performs the weak discriminator learning to generate the class discriminator includes a plurality (for example, 10,000 in this embodiment) of positive images for learning and a plurality of (this embodiment). In the embodiment, there is a learning image storage unit 21 that stores, for example, 10,000 negative learning images. Further, the class discriminator generation device 20 is a part of all learning positive images stored in the learning image storage unit 21 (in this embodiment, for example, 5000) positive learning images (hereinafter, 5000). A learning positive image group) and a part of all the learning negative images (in this embodiment, for example, 5000) negative learning images (hereinafter referred to as a learning negative image group). The classifier learning unit 23 learns weak classifiers and generates class classifier candidates.

  Further, the class discriminator generation device 20 includes a plurality of positive images prepared separately from the learning positive images (hereinafter, each positive image is referred to as an evaluation positive image, and a set of evaluation positive images is referred to as an evaluation positive image. And a plurality of negative images prepared separately from the learning negative image (hereinafter, each negative image is referred to as an evaluation negative image, and a set of evaluation negative images is referred to as an evaluation negative image group). Has an evaluation image storage unit 24 for storing. Further, the class discriminator generation device 20 generates class discriminator candidates generated by the discriminator learning unit 23 for each of the evaluation positive image group and the evaluation negative image group stored in the evaluation image storage unit 24. And classifying the generated class discriminator candidate as a class discriminator according to the evaluation result, or a part of learning positive images (hereinafter, worsened) forming a learning positive image group. Whether to generate a new class classifier candidate by the classifier learning unit 23 by replacing the factor image) with a learning positive image (hereinafter referred to as an unused positive image) that has not been used at the time of previous learning. And a discriminator evaluation unit 26 for determination.

  The class discriminator generation device 20 is a so-called personal computer having an auxiliary storage device (for example, HDD), an operation unit, and a display unit, centering on a known microcomputer having at least a ROM, a RAM, and a CPU. It is a computer.

  That is, the class discriminator generation device 20 generates a class discriminator candidate by learning a weak discriminator (that is, a feature amount) using a learning image according to a processing program stored in a ROM or a RAM, A class that evaluates the generated class classifier candidate using the evaluation positive image group and the evaluation negative image group, and determines whether the generated class classifier candidate is appropriate as a class classifier The discriminator generation process is configured to be executed. Note that the learning positive image, the learning negative image, the evaluation positive image, and the evaluation negative image used in the class discriminator generation process are all stored in the auxiliary storage device constituting the class discriminator.

Accordingly, the auxiliary storage device functions as the learning image storage unit 21 and the evaluation image storage unit 24, and the CPU that has executed the class discriminator generation process functions as the discriminator learning unit 23 and the discriminator evaluation unit 26. To do.
<Classifier generation processing>
Next, a class discriminator generation process executed by the CPU constituting the class discriminator generation device will be described.

Here, FIG. 3 is a flowchart showing the processing procedure of the class discriminator generation processing.
As shown in FIG. 3, when this class discriminator generation process is started, first, in S110, all learning positive images and all learning negative images stored in the auxiliary storage device are respectively displayed. , Expand to the learning image list.

  Specifically, the learning image list is prepared in advance for each of the learning positive image and the learning negative image. The learning image list is associated with the image ranking by associating with the image index the index description field in which the image index is described in association with the list order for each image, and the correctness of detection of the detection target at the time of weak classifier learning. A stage detection availability description column to be described is provided.

  Note that the list order represents the order of learning images used to collate the weak classifiers when learning the weak classifiers. That is, the list order is a learning positive image group and a learning negative image group (hereinafter referred to as a learning positive image group and a learning negative image group) according to a learning positive image and a learning negative image corresponding to 1 to 5000. The image group is collectively referred to as a learning image group). Further, the learning positive image and the learning negative image whose list order is 5001 or more are respectively an unused positive image and an unused negative image (hereinafter referred to as a set (ie, a group) of unused positive images and an unused image). A set (group) of use negative images is collectively referred to as an unused image group).

Subsequently, in S120, classifier learning processing for generating class classifier candidates is executed.
In this classifier learning process, each of the plurality of weak classifiers prepared in advance is sequentially checked against each of the learning images, and the weak classifiers that can detect the partial features to be detected are arranged and their weaknesses. A weight for the discriminator (that is, discriminator weight α _t , where the subscript t is the number t of connection stages of the weak discriminator) and a prescribed threshold th used when detecting the presence or absence of a detection target are set. At this time, among the learning positive images forming the learning positive image group, the detection target reflected in those images is detected from the learning positive images having a specified rate or more, and the learning negative image group is formed. Among the learning negative images to be detected, the weak classifiers constituting each stage are correctly detected from the learning negative images that are equal to or higher than the set rate until it is correctly detected that the detection target is not reflected in those images. Added (ie, the number of connection stages of the weak classifier increases). The specified rate in the present embodiment is set to a value larger than the set rate, for example, set rate is set to 99.5% and set rate: 50%.

That is, by executing the classifier learning process, cascade-connected class classifier candidates are further generated for each stage obtained by cascading weak classifiers to which classifier weights α _t are assigned, respectively.

In S130, one evaluation positive image is extracted from the evaluation positive image group stored in the auxiliary storage device.
In S140, the classifier candidate generated in S120 of the current cycle (that is, the previous classifier learning process) is collated with the evaluation positive image extracted in S130, and is reflected in the evaluation positive image. It is output whether or not the detected object can be detected. The output in S140 (that is, the output value) is “1” if a detection target is detected from the evaluation positive image, and is “0” if no detection target is detected from the evaluation positive image. In addition, since the process which collates a classifier candidate with respect to the image for evaluation and determines the presence or absence of a detection target is the same as the determination of the presence or absence of a target in the driving support system 1, detailed description here is omitted.

  Further, in S150, it is determined whether or not the collation in S140 has been performed for all evaluation positive images. As a result of the determination, if all the positive images for evaluation have not been collated in S140, the process returns to S130, and the evaluation positive images that have not been collated are extracted from the positive images for evaluation. , The verification in S140 is executed. When the collation in S140 is completed for all the positive images for evaluation, the process proceeds to S160.

In S160, one evaluation negative image is extracted from the evaluation negative image group stored in the auxiliary storage device.
Subsequently, in S170, the class discriminator candidate generated in the previous discriminator learning process is collated against the evaluation negative image extracted in S160, and the detection target reflected in the evaluation negative image can be detected. Output whether or not there is. The output (ie, output value) in S170 is “0” if a detection target is detected from the evaluation negative image, and “1” if no detection target is detected from the evaluation negative image.

  Further, in S180, it is determined whether or not the collation in S170 has been performed on all evaluation negative images. As a result of the determination, if all of the negative evaluation images have not been collated in S170, the process returns to S160 to extract an unverified negative evaluation image from the negative evaluation images. , Collation in S170 is executed. When the collation in S170 is completed for all evaluation negative images, the process proceeds to S190.

In S190, the evaluation value SW for the class classifier candidate generated in S120 of the current cycle is derived based on all of the collation result in S140 and the collation result in S170. The evaluation value SW includes a first evaluation value SW ₁ that is a sum of output values in S140 and a second evaluation value SW ₂ that is a sum of output values in S140.

That is, the first evaluation value SW ₁ increases as the detection rate of the detection target from the evaluation positive image forming the evaluation positive image group increases. On the other hand, the second evaluation value SW ₂ becomes larger as the detection rate of the detection target from the evaluation negative image forming the evaluation negative image group is lower.

Subsequently, in S200, whether or not each of the evaluation values SW derived in S190 (that is, both the first evaluation value SW ₁ and the second evaluation value SW ₂ ) is equal to or greater than a preset setting threshold th. judge. As a result of the determination, if both the first evaluation value SW ₁ and the second evaluation value SW ₂ are equal to or greater than the set threshold th, the detection accuracy of the detection target using the classifier candidate generated in S120 of the current cycle Is determined to be equal to or greater than a predetermined value, and the process proceeds to S210.

In S210, the class classifier candidate generated in the classifier learning process of the current cycle is output as a class classifier, and then the class classifier generation process is terminated.
On the other hand, if at least one of the first evaluation value SW ₁ and the second evaluation value SW ₂ is less than the set threshold th as a result of the determination in S200, the class identifier candidate generated in S120 of the current cycle is determined. Assuming that the detection accuracy of the used detection target is less than the predetermined value, the process returns to S120.

  In S120, a classifier learning process is executed. By the discriminator learning process executed at this time, the deterioration factor image included in the learning positive image group is replaced with an unused positive image according to the learning image list, and a newly formed learning positive The weak classifier is learned using the image group. Therefore, a new class classifier candidate is generated.

Then, evaluation value SW respectively for newly generated class classifier candidate (i.e., a first evaluation value SW _1, second both evaluation value SW ₂₎ until the set threshold value th or more, the class identifier generating process Repeated.
<About classifier learning processing>
Next, the classifier learning process will be described.

Here, FIG. 4 is a flowchart showing a processing procedure of weak classifier learning processing.
When the weak classifier learning process is started in S120 of the class classifier generation process, first, a weight for each learning image (hereinafter referred to as image weight D _t (i)) is initialized (S1210). . That is, D _t (i) = 1 / N, N is the number of learning images forming the learning image group (10000 in this embodiment. However, the subscript i here is for each of the learning images. The index t is the number of connection stages of the weak classifier determined in the current cycle.

Subsequently, one weak classifier h _j is extracted from all weak classifiers prepared in advance (S1220). The entity of the weak classifier h _j is an evaluation function h _j for evaluating the Haar Like feature.

The extracted weak classifier h _j is collated with one of the learning images forming the learning image group (hereinafter, extracted image x _i ) to derive an output value y _i (S1230). This output value y _i is derived for each region in the extraction region where the weak classifier h _j is arranged. At this time, the output value y _i is “1” if the weak classifier h _j is compatible (that is, the feature amount is detected), and if the weak classifier h _j is not compatible, the output value y _i is “−1”. However, the subscript j is an index for distinguishing weak classifiers.

Further, it is determined whether or not the weak classifier h _j extracted in S1220 is collated with respect to all the learning images forming the learning image group (S1240). As a result of the determination, if all the learning images have not been collated, the process returns to S1230, the uncollated learning image is read as the extracted image x _{i + 1} , and the weak classifier h _j is Match.

On the other hand, the result of the determination in S1240, if the check completed the weak classifier h _j for all learning images, processing proceeds to S1250.
In the S1250, according to the following equation (1), to derive a detection error e _j (S 1270). It should be noted that h _j (x _i ) ≠ y _i = ca in the expression (1) was an erroneous determination when the weak classifier h _j was collated with the learning image x _i identified by the image index i. Means that.

That is, the detection error e _j is the sum of weak classifiers h _j misjudgment was given to learning image x _i image weights D _t (i) at S1230. Therefore, the detection error e _j is the image weighting Dt (i) is the learning image is a large value, a large value enough to determine weak classifier h _j erroneous.

However, where an erroneous judgment and says, that the weak classifier h _j is collated learning image x _i is equal positive image for learning, the weak classifier h _j becomes non-compliance (the so-called, False Positive ), and it is as long as the weak classifier h _j is collated learning image x _i is negative image for learning, the weak classifier h _j-matched (so-called False negative).

Subsequently, it is determined whether or not all weak classifiers are collated with respect to the extracted image (S1260). As a result of the determination, if collation has not been completed for all weak classifiers, the process returns to S1220, and one unmatched weak classifier h _{j + 1} is selected from all weak classifiers. Proceed to S1230.

On the other hand, if the result of determination in S1260 is that all weak classifiers have been collated, processing proceeds to S1270. In this case, the detection error e _j derived at S1250 in a series of processing routines in S1260 from S1220 is a total amount of weak classifiers h _j.

Then, in S 1270, in accordance with the detected error e _j, 1 single weak classifiers constituting the stage class classifier candidates (hereinafter, referred to as configuration weak classifier h _t) is determined and the weak and before connection stage t-1 Add to the classifier (ie, the combination of weak classifiers). That is, in the S 1260, among the parameters constituting the stage class classifier candidates, the number of connection stages is disposed of t th configuration weak classifier h _t is determined.

Specifically, the series of processing routine from S1220 to S 1260 (hereinafter, simply referred to as processing routine) of all the detected error e _j derived at, weak learners corresponding to those that value is minimum The unit h _j is determined as the constituent weak classifier h _t . Hereinafter, the detection error e _j of the structure weak classifier h _t and detection error e _t.

Furthermore, on the basis of the detection error e _t of the structure weak classifier h _t determined in S 1270, by the following equation (2), to derive the classifier weights alpha _t (S 1280).

The discriminator weight α _t represents the reliability with respect to the constituent weak discriminator h _t, and becomes a larger value as the detection error _et is smaller. That is, in the S 1280, among the parameters constituting the stage class classifier candidates, the number of connection stages is classifier weights alpha _t is determined for t th configuration weak classifier h _t.

Subsequently, based on the discriminator weight α _t derived in S1280, the image weights D _t (i) for all the learning images are updated according to the following equation (3) (that is, the image weights D _{t + 1).} (I) is determined). However, y _{i ht} (x _i ) in the expression (3) represents an output value y _i when the structural weak classifier h _t is collated with the learning image x _i specified by the image index i.

The image weight D _t (i) is set so that the sum of the image weights D _t (i) for all the learning images is 1. For this reason, Z _t in the equation (3) is a normalization factor for deriving the sum of the image weights D _t (i), and is expressed by the following equation (4).

Subsequently, using the class discriminator candidate stage configured at the time of proceeding to S1300, it is determined whether or not the prescribed threshold th can be set so as to satisfy an appropriate condition from the learning image group ( S1300). However, this appropriate condition is the first correct answer for a learning positive image having a specified rate or higher in the learning positive image group, and in the learning negative image group, a learning negative image having a setting rate or higher. Is the second correct answer.

  And the 1st correct answer said here is detecting correctly the detection target reflected in the positive image for learning, and the 2nd correct answer said here cannot detect a detection target in the negative image for learning. It is.

  As a result of the determination in S1300, if the specified threshold th satisfying the appropriate condition cannot be set, the presence / absence of the detection target is detected from the learning image using the class classifier candidate stage generated at the time of proceeding to S1300. When it is determined that the detection accuracy is less than the set value, the process returns to S1220, and the steps from S1220 to S1300 are repeated until a stage in which the detection accuracy is equal to or higher than the set value is generated.

That is, by repeating the steps from S1220 to S1300, learning based on the Viola & Jones algorithm (so-called supervised learning in which Adaboost is extended) is executed. Then, by repeating these steps, the configuration weak classifier _ht is added until the proper condition is satisfied.

  On the other hand, as a result of the determination in S1300, if it is possible to set the specified threshold th satisfying the appropriate condition, the specified threshold th satisfying the appropriate condition is set, and the process proceeds to S1310. That is, one stage forming a class classifier candidate is generated.

  In S1310, the detection target cannot be detected even if the target presence / absence determination is performed using the class discriminator candidate stage generated by setting the specified threshold th so as to satisfy the appropriate condition in S1300 of the current cycle. The learning positive image is recorded as a deterioration factor image in the learning image list.

  Subsequently, it is determined whether or not the number of stages constituting the class discriminator candidate has been generated by a preset number, that is, whether or not the number of stages connected to the stage has reached the set number (S1320). If the number of connected stages of the stages does not reach the set number as a result of the determination, the process proceeds to S1330. In S1330, a learning negative in which a detection target cannot be detected (that is, a correct answer) is determined by the target presence / absence determination using the class classifier candidate stage generated in the current cycle in the learning negative image group. After changing all the images to unused negative images, the process returns to S1220. That is, all of the 2500 or more learning negative images that are correctly detected that the detection target is not reflected by the stage constituting the class classifier candidate generated in the current cycle are replaced with unused negative images. The learning for generating a new stage is executed.

  On the other hand, as a result of the determination in S1320, if the number of stages connected reaches the set number, the process proceeds to S1340. In S1340, all the deterioration factor images recorded in the learning image list in S1310 are set to be replaced with unused positive images when the classifier learning process is started next time. That is, in the learning image list for learning positive images, the list order for all learning positive images recorded as deterioration factor images is the number of the deterioration factor images (hereinafter referred to as replacement number). The order is changed from the lowest order to increase the order of unused positive images by the number of replacements.

Concretely, it demonstrates using FIG. 5 which illustrated the learning image list about the positive image for learning.
In the example shown in FIG. 5, the column described as stage 1 detectability,..., Stage m detectability corresponds to the stage detectability description column, and “x” described in these columns indicates the corresponding learning. This means that the positive image for use is a deterioration factor image, and “◯” means that the corresponding positive image for learning is not a deterioration factor image.

  Therefore, in the example shown in FIG. 5, every time each stage constituting the class classifier candidate is generated in S1300, the deterioration factor image for the generated stage is displayed in the detectability description column corresponding to that stage. To be recorded. When all the stages, that is, class classifier candidates are generated, the learning positive image recorded as the deterioration factor image is replaced with the unused positive image in all the detection possibility description columns. It is done.

  That is, in the class discriminator generation process of the present embodiment, a learning image composed of a plurality of (5000 images in this embodiment) learning positive images and a plurality (5000 images in this embodiment) learning negative images is used. Then, weak classifiers (that is, feature quantities) are learned to generate class classifier candidates. As a result of detecting the presence / absence of a detection target from the evaluation positive image group or the evaluation negative image group using the generated class discriminator candidate, when the detection accuracy is less than a predetermined value, as shown in FIG. In addition, the deterioration factor image included in the learning positive image group is replaced with an unused positive image. Further, weak classifier learning is performed using a new learning positive image group and a learning negative image group formed by replacing the deterioration factor images.

By repeating this, a new class classifier candidate is generated in which the weak classifier is learned using each of the learning positive image groups having different constituent contents of the learning positive image.
Among the newly generated class discriminator candidates, those generated without learning feature quantities necessary for identifying adverse factors are included. However, the adverse effect factor mentioned here is a feature that causes a decrease in detection accuracy when detecting a detection target from an input image or an increase in a processing amount until a detection result is derived. That is, there is a high possibility that the class discriminator learns only the feature amount for identifying the general feature to be detected.
[Effect of the embodiment]
As described above, according to the class discriminator generation device 20 of the present embodiment, the feature quantity learned to generate the finally generated class discriminator, that is, the weak discriminator constituting the class discriminator. Can be reduced.

  And if the driving assistance system 1 detects the presence or absence of the detection target reflected in the input image using the class discriminator generated by such a class discriminator generation device 20, the detection accuracy is improved. be able to.

  In other words, according to the class discriminator generation device 20 of the present embodiment, for the class discriminator, it is possible to improve the detection accuracy while suppressing the number of feature quantities (that is, weak discriminators) constituting the class discriminator as much as possible. it can.

And when the driving assistance system 1 detects the presence or absence of the detection target reflected in the input image using the class discriminator generated by such a class discriminator generating device 20, the class discriminator is used. Since the number of weak classifiers to be configured is small, the amount of processing can be reduced.
[Other Embodiments]
As mentioned above, although embodiment of this invention was described, this invention is not limited to the said embodiment, In the range which does not deviate from the summary of this invention, it is possible to implement in various aspects.

For example, in the above-described embodiment, the specified rate is 99.5% and the set rate is 50%. However, these specified rate and set rate are not 100%, and the specified rate is higher than the set rate. Any value can be used as long as it is a large value. In the classifier learning process in the above embodiment, even if the target presence / absence determination is performed using the class classifier candidate stage for the learning positive image group, Although the learning positive image in which the detection target cannot be detected is the deterioration factor image, the deterioration factor image is not limited to this. For example, when the classifier learning process is executed, a positive image in which a detection target can be detected for the first time by adding weak classifiers of a predetermined number of stages or more may be used as a deterioration factor image, and the configuration in the process of determining the respective weak classifier h _t, erroneous also may the learning positive image number was often a determination as deterioration factors image may be deteriorated factors image combinations thereof.

  By the way, in the said embodiment, although the replacement of a deterioration factor image and an unused positive image was implemented whenever a class discriminator candidate was produced | generated, this exchange is each stage which comprises a class discriminator candidate. It may be carried out every time is generated. That is, the process executed in S1340 of the classifier generation process may be executed in S1310.

  Furthermore, in the class discriminator generation process in the above embodiment, every time a class discriminator candidate is generated, an evaluation value SW for the class discriminator candidate is derived, and whether or not the class discriminator candidate is used as a class discriminator. However, the timing for determining the class identifier is not limited to this. For example, the classifier learning process is repeatedly executed a predetermined number of times or more, and the class classifier candidate having the highest evaluation value SW among the plurality of class classifier candidates generated during the repetition is determined as a class classifier. It may be determined as

  In this case, among the plurality of generated class discriminator candidates, the one with the best detection performance of the detection target can be determined as the class discriminator. As a result, the driving support system 1 can accurately detect the presence / absence of the detection target reflected in the input image.

  Further, in the above embodiment, a weak classifier composed of Haar Like features cascade-connected is generated as a class classifier. However, the class classifier is not limited to this. For example, SVM (Support vector machine) is used. Or back propagation may be used. In other words, any class identifier may be used as long as it is generated by supervised learning using both the learning positive image group and the learning negative image group.

In the above embodiment, a person is a detection target. However, the detection target is not limited to this, and may be, for example, an automobile or an obstacle that may exist on the road.
[Correspondence between Embodiment and Claims]
Finally, the relationship between the description of the above embodiment and the description of the scope of claims will be described.

  First, the auxiliary storage device (function as the learning image storage unit 21) of the class discriminator generation device 20 in the above embodiment corresponds to the positive image storage unit and the negative image storage unit of the present invention.

  The function obtained by executing S120 (classifier learning process) of the classifier generation process of the above embodiment corresponds to the classifier candidate generation means of the present invention. Furthermore, a function obtained by executing S120 (discriminator learning process) through S1340 of the discriminator learning process corresponds to a new candidate generation unit.

  The function obtained by executing S130 to S210 of the class discriminator generation process corresponds to the class discriminator determining means of the present invention.

DESCRIPTION OF SYMBOLS 1 ... Driving support system 3 ... Imaging | photography mechanism 6 ... Control apparatus 9 ... Control object apparatus group 20 ... Class discriminator production | generation apparatus 21 ... Learning image memory | storage part 23 ... Discriminator learning part 24 ... Evaluation image memory | storage part 26 ... Discriminator Evaluation department

Claims (6)

A class discriminator generating device for generating a class discriminator for detecting a detection target reflected in an input image from an imaging device,
Each of the images in which the detection target is reflected is a positive image, and a positive image storage means in which a plurality of positive images are stored;
Negative image storage means in which each of the images in which the detection target is not reflected is a negative image, and a plurality of negative images are stored;
Identification learned using a positive image group consisting of a part of the positive images stored in the positive image storage means and a negative image group consisting of a part of the negative images stored in the negative image storage means Classifier candidate generation means for executing a classifier learning process for generating a classifier as a class classifier candidate;
Among the positive image group used for learning in the classifier candidate generation unit, each positive image that causes a negative effect when detecting a detection target from the input image using the classifier candidate is deteriorated. A factor image is used, and the deterioration factor image is replaced with a positive image stored in the positive image storage unit that has not been used in previous learning, and the classifier candidate generation unit includes the classifier. New candidate generation means for causing the classifier candidate generation means to generate a new class classifier candidate by executing a learning process;
The evaluation positive image group prepared separately from the positive image stored in the positive image storage unit and the evaluation negative image group prepared separately from the negative image stored in the negative image storage unit, respectively. The classifier candidates generated by the classifier candidate generation unit are collated, and as a result of the collation, the higher the detection rate of the detection target from the evaluation positive image group, the larger the value, and from the evaluation negative image group A class discriminator generation device comprising: a class discriminator determining unit that determines the class discriminator based on an evaluation value that increases as the detection rate of the detection target decreases. The classifier candidate generation means includes:
In order to generate as a class classifier candidate a cascade connection of a plurality of weak classifiers with an integral value obtained by integrating pixel values in a rectangular area set in an input image as a feature amount,
Among the positive images forming the positive image group, the detection target is detected from positive images having a predetermined prescribed ratio or more, and among the negative images forming the negative image group, a preset setting ratio or more is set. The class discriminator generation device according to claim 1, wherein learning of a new weak discriminator is executed as the discriminator learning process until the detection target is not detected from a negative image. The new candidate generation means includes:
3. The class discriminator according to claim 2, wherein when the discriminator candidate generation unit executes discriminator learning processing, the positive image that is not detected as the detection target is used as the deterioration factor image. Generator. The new candidate generation means includes:
When the classifier learning process is executed by the classifier candidate generation unit, the positive image whose detection target can be detected by learning more than a preset number of weak classifiers is determined as the deterioration factor. 4. The classifier generating device according to claim 2 or 3, wherein the classifier generating device is an image. The class discriminator determining means includes
The class discriminator generation device according to claim 1, wherein the class discriminator candidate having the maximum evaluation value is the class discriminator. The class discriminator determining means includes
The classifier candidate is collated with each of the evaluation positive image group and the evaluation negative image group each time the classifier candidate generation unit generates a classifier candidate, and the evaluation value is 5. The class discriminator generating apparatus according to claim 1, wherein the class discriminator candidate generated is a class discriminator if the class discriminator candidate is equal to or greater than a predetermined value.