JP2013117865A - Detection method for body and detection device for body using the method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a detection method and a detection device for body that can detect an object of detection in an image precisely in a short time.SOLUTION: In a detection device 10 for body including a learning mechanism 11 which structures a weak discriminator for discriminating an object of detection from a learning image 13, and a determination mechanism 12 which determines whether the object of detection is present within a search range 14 by RealAdaBoost using a discriminator composed of the weak discriminator while setting the search range 14 in an inspection image, the leaning mechanism 11 includes HOG feature quantity calculating means 16 of setting the search range 14 in the learning image and a plurality of sizes of cells 15 determining the size of a local area in the search range 14, and calculating HOG feature quantities A by the cells 15 of the respective sizes by arranging the cells 15 over entirely in the search range 14 while putting the cells 15 partially one over another so as to find HOG feature quantities B of the search range 14, and weak discriminator generation means 17 of selecting HOG feature quantities effective in discriminating the object of detection from the HOG feature quantities B using RealAdaBoost to obtain the weak discriminator.

Description

The present invention relates to an object detection method for determining whether or not a detection target (for example, a person, a specific object, etc.) exists in a captured image, and an object detection apparatus using the method.

In recent years, for the purpose of ITS (Intelligent Transport System), robot vision, and the like, a technique for detecting an object from a camera image has attracted attention. In particular, the detection of pedestrians on camera images can contribute as a system that prevents pedestrians from being overlooked by the driver's carelessness when driving a vehicle in the field of ITS. Therefore, research on pedestrian detection for detecting pedestrians using IT (Information Technology) has been actively conducted. However, pedestrian detection is not easy due to the influence of body shape, clothing color, obstacles, changes in brightness, overlapping of pedestrians, and the like.

In general, various calculation methods have been proposed for feature amounts representing a detection target (that is, a local region in a camera image). Which feature amount is effective varies greatly depending on the detection target. For example, the HOG feature value is effective for detecting pedestrians because it is a feature value that is robust against changes in illumination and shape (see Non-Patent Document 1, for example). For this reason, the HOG feature amount of the learning image consisting of the image that includes the pedestrian (person) and the image that does not include the pedestrian is calculated, and machine learning by RealAdaBoost of the obtained HOG feature amount is performed. Research has been conducted on pedestrian detection methods that generate weak classifiers by selecting HOG features that are effective for detecting each part of the pedestrian and determine whether there are pedestrians in the evaluation image. (For example, refer nonpatent literature 2).

N. Dalal, B.M. Triggs, “Histograms of Oriented Gradients for Human Detection”, IEEE CVPR, p. 886-893, 2005 Satoshi Yamauchi, Hironobu Fujiyoshi, Takayoshi Yamashita, Human detection based on co-occurrence of features based on Boosting, Symposium on Image Recognition and Understanding (MIRU2008), p. 180-187, 2008

Conventionally, when detecting a pedestrian in an evaluation image, a large number of search areas are arranged in the entire evaluation image, and HOG feature values are calculated for each search area. Here, the HOG feature amount of the search region is obtained by dividing the search region into cells each composed of a plurality of preset pixels, calculating the HOG feature amount for each cell, and adding all of them. . Accordingly, whether or not there is a pedestrian in the search area is determined by using the weak classifier obtained by machine learning for the HOG feature amount calculated for each cell, and the output value of each weak classifier. When the total value of (judgment result) is larger than the reference value obtained by machine learning, it is determined that there is a pedestrian, and when the total value is less than the reference value, it is determined that there is a pedestrian. For this reason, in order to determine whether or not there is a pedestrian in the evaluation image, it is necessary to use a computer that can process a large amount of calculations in a short time, which is a problem in terms of calculation cost. In particular, in the field of ITS, shortening the processing time has become an important development factor as well as improved detection accuracy.
Further, in the conventional pedestrian detection method, since the cell size is fixed, the number of HOG feature amounts obtained from the evaluation image is limited, and there is a problem in improving detection accuracy.

The present invention has been made in view of such circumstances, and an object of the present invention is to provide an object detection method and a detection apparatus capable of accurately detecting a detection target in a given image in a short time.

The object detection method according to the first aspect of the present invention that meets the above-described object constructs a plurality of weak classifiers that identify a detection target using a learning image composed of an image of a detection target and an image of a non-detection target. While setting the search range in the inspection image in the learning process, the presence / absence of the detection object is determined from the output value of the classifier obtained for the image within the search range by the RealAdaBoost algorithm using the weak classifier. In a method for detecting an object having a determination process for determining,
The learning process sets the search range in the learning image, sets a plurality of cell sizes that determine the size of the local region in the search range, and sequentially uses the cells of each size, The cells are arranged over the entire search range while partially overlapping each other, the HOG feature amount A of the image in the arranged cell is calculated, and the HOG for each of the cells arranged in the search range. A first step of combining the feature amount A into the HOG feature amount B of the image within the search range;
A second step of selecting a plurality of HOG feature quantities effective for identifying the detection object from the HOG feature quantity B using the RealAdaBoost algorithm, and using each as the weak classifier.

In the object detection method according to the first aspect of the present invention, in the determination process, when the determination process is performed by repeatedly using the weak classifier for the search range set in the inspection image by the RealAdaBoost algorithm. In addition, when the output value of the discriminator is less than a predetermined determination value for each number of the discriminating processes, the discriminator is terminated for each search range set in the inspection image by terminating the discriminating process. It is preferable to change the number of the weak classifiers constituting the.

In the object detection method according to the first aspect of the present invention, the determination value set in advance for each number of times of the identification process is based on the RealAdaBoost algorithm using the weak classifier for the search range set in the learning image. When the detection target is detected by repeating the learning process, it is preferably the minimum value of the output values output from the discriminator obtained every number of the learning processes.

The object detection apparatus according to the second invention that meets the above object constructs a plurality of weak classifiers that identify the detection object using a learning image composed of an image of the detection object and an image of the non-detection object. While setting the search range in the inspection image with the learning mechanism, the presence / absence of the detection object is determined from the output value of the classifier obtained for the image within the search range by the RealAdaBoost algorithm using the weak classifier. In an object detection apparatus having a determination mechanism for determining,
The learning mechanism sets the search range in the learning image and sets a plurality of cell sizes that determine the size of the local region in the search range, and sequentially uses the cells of each size to The cells are arranged over the entire search range while partially overlapping each other, the HOG feature amount A of the image in the arranged cell is calculated, and the HOG for each of the cells arranged in the search range. HOG feature quantity calculating means for obtaining the HOG feature quantity B of the image within the search range by combining the feature quantities A;
A weak classifier generating unit that selects a plurality of HOG feature quantities effective for identifying the detection target object from the HOG feature quantity B using the RealAdaBoost algorithm and uses the HOG feature quantity as the weak classifier;

In the object detection apparatus according to the second invention, when the determination mechanism repeatedly performs identification processing by RealAdaBoost algorithm using the weak classifier for the search range set in the inspection image. In addition, when the output value of the discriminator is less than a predetermined determination value for each number of the discriminating processes, the discriminator is terminated for each search range set in the inspection image by terminating the discriminating process. It is preferable to have an identification processing management means for changing the number of the weak classifiers constituting the.

In the object detection device according to the second invention, the learning mechanism repeats learning processing by the RealAdaBoost algorithm using the weak classifier with respect to the search range set in the learning image to detect the detection object. A determination value generating means for obtaining a minimum value of an output value output from the discriminator obtained for each number of learning processes when detected;
The determination mechanism preferably includes a determination value setting unit that sets each minimum value obtained by the identification determination value generation unit as the determination value set in advance for each number of identification processes.

In the object detection method according to the first invention and the object detection apparatus according to the second invention, a plurality of cell sizes that determine the size of the local region within the search range are set, and cells of each size are sequentially used. Since the cells are arranged over the entire search range while partly overlapping the cells, the image within the search range using the HOG feature amount A of the images of the local regions of various positions and various sizes within the search range. If a plurality of HOG feature values effective for identifying a detection target object are selected from the HOG feature value B using the RealAdaBoost algorithm, the selected HOG feature value is determined as a detection target object. HOG feature values that are optimal for representing each part of the above are collectively included. For this reason, each part of the detection target (for example, when the detection target is a person, the shape of the arm from the cell arranged in the arm part, and the shape of the head from the cell arranged in the head part) If the classifier that identifies the detection object is configured by using the HOG feature quantity that is most suitable for representing each feature in a batch as a weak classifier, the conventional method (cells having the same size in the search range) The number of weak classifiers constituting the classifier can be greatly reduced as compared to a method in which the classifiers are arranged so as not to overlap. As a result, the time required to identify the search range set in the inspection image is shortened, that is, the search range identification process can be speeded up.

In the object detection method according to the first invention and the object detection apparatus according to the second invention, the identification process by the RealAdaBoost algorithm is repeatedly determined using the weak classifier for the search range set in the inspection image. When the identification process is terminated when the output value of the classifier is less than a predetermined determination value for each number of identification processes, the final determination that the detection target does not exist in the search range of the inspection image is identified. This can be performed at an early stage of processing, and the time required for identifying the search range can be greatly reduced.
Furthermore, since the presence / absence of the detection target is determined using a classifier having a different number of weak classifiers for each search area, for example, the background area is quickly identified and detected by a classifier composed of few weak classifiers. The object region and the region similar to the detection object are identified by a classifier having a large number of weak classifiers, thereby improving the identification accuracy of the detection object.

In the object detection method according to the first invention and the object detection apparatus according to the second invention, a determination value preset for each number of identification processes is a weak classifier for a search range set in a learning image. When the detection object is detected by repeating the learning process using the RealAdaBoost algorithm, the search range identification process ends if the output value is the minimum value output from the discriminator obtained for each number of learning processes. Judgment can be made on the safe side.

1 is a block diagram of an object detection apparatus to which an object detection method according to a first embodiment of the present invention is applied. (A) Arrangement of cells set in the search area by the object detection method according to the first embodiment of the present invention, (B) shows the arrangement of cells set in the search area by the conventional object detection method. It is explanatory drawing. It is a graph which shows the relationship between the number of weak discriminators, and the correct answer rate of a pedestrian detection. It is a block diagram of the detection apparatus of the object to which the detection method of the object which concerns on the 2nd Embodiment of this invention is applied. It is a graph which shows the relationship between the number of weak discriminators and an error rate in an Example and a comparative example.

Next, embodiments of the present invention will be described with reference to the accompanying drawings for understanding of the present invention.
First, the object detection apparatus 10 to which the object detection method according to the first embodiment of the present invention is applied will be described. As shown in FIGS. 1 and 2A, the object detection device 10 includes an image of a pedestrian P that is an example of a detection target and a background that is an example of a non-detection target (background of the pedestrian P). A learning mechanism 11 for constructing a weak classifier for identifying a pedestrian P using a learning image 13 composed of images, and a RealAdaBoost algorithm using a weak classifier while setting a search range 14 in a test image (not shown). And a determination mechanism 12 that determines the presence / absence of the pedestrian P in the search range 14 from the output value of the discriminator obtained for the image in the search range 14. Details will be described below.

The learning mechanism 11 sets the search range 14 in the learning image 13, sets a plurality of sizes of the cells 15 that determine the size of the local region in the search range 14, and sequentially uses the cells 15 of each size to obtain the cell 15. Are arranged over the entire search range 14 while partly overlapping the cells 15, HOG feature amounts A of the images in the arranged cells 15 are respectively calculated, and the HOG for every cell 15 arranged in the search range 14. The HOG feature quantity calculating means 16 for obtaining the HOG feature quantity B of the image within the search range 14 by combining the feature quantity A, and the HOG feature effective for identifying the pedestrian P from the HOG feature quantity B using the RealAdaBoost algorithm. Weak discriminator generating means 17 for selecting a plurality of quantities and using them as weak discriminators.

Here, the learning image 13 uses a pedestrian image taken with a camera (not shown) and an image without a pedestrian, for example, resized to a size of 60 pixels in length and 30 pixels in width. As the learning image 13, an INRIA Person Dataset, which is one of pedestrian data databases, can be used. Further, the size of the search range 14 is set according to the shape of the detection target. For example, as illustrated in FIG. 2A, when the detection target is a pedestrian P, the search range 14 has a size of 60 pixels vertically and 30 pixels horizontally that is the same size as the learning image 13.

The HOG feature amount calculation means 16 has a learning image storage unit 18 having a function of storing the learning image 13, and a search range 14 for extracting learning data by taking out the learning image stored in the learning image storage unit 18. And a plurality of sizes of the cells 15 that determine the size of the local region in the search range 14 are set, and the cells 15 of each size are sequentially used, and the cell 15 The first HOG feature quantity calculation unit 19 that calculates the HOG feature quantity A of the image in the cell 15 is arranged while being partially overlapped, and the HOG feature quantity A for every cell 15 arranged in the search range 14 is combined. (For example, by arranging the HOG feature quantities A for each cell 15 having a constant size in the order of the cells 15 arranged in the search range, in the order of the sizes of the cells 15) And a second HOG feature amount calculation unit 20 for obtaining the HOG features B.
Here, as shown in FIG. 2A, the size range of the cell 15 is set according to the occupancy rate of the detection object within the search range 14. For example, when the detection object is a pedestrian P, The vertical width of the cell 15 is 6 to 80% of the maximum vertical width of the pedestrian P, and the horizontal width of the cell 15 is 13 to 80% of the maximum horizontal width of the pedestrian P (for example, the minimum size of the cell 15 is 4 vertical) 4 pixels wide, maximum size is 48 pixels long and 24 pixels wide). Moreover, the overlap margin at the time of arrange | positioning partially overlapping cells 15 is 2-46 pixels, for example.

In the weak classifier generation means 17, the RealAdaBoost of the RealAdaBoost is selected from the HOG feature quantity B that is a set of the feature quantities of the pedestrian P (the HOG feature quantities that are the feature quantities of the pedestrian P obtained from the cells 15 of all sizes). Using an algorithm, a plurality of HOG feature values effective for identifying the pedestrian P are selected to form a weak classifier. FIG. 3 shows the relationship between the number of weak classifiers and the accuracy rate of pedestrian detection when detecting pedestrians P. From FIG. 3, by selecting a certain number (for example, at least 500) of weak classifiers, the accuracy rate when detecting the pedestrian P can be made a certain value or more (95% or more).
Further, the learning mechanism 11 uses the weak classifier obtained from the weak classifier generation means 17 for the search range 14 set in the learning image 13 to repeatedly determine the classifier obtained by learning using the RealAdaBoost algorithm. Finds the minimum value of the output value (a numerical value indicating that the HOG feature value in the search range 14 set in the learning image 13 is the pedestrian P) output when the pedestrian P is detected. Is a determination value generation means 21 for determining a value when the classifier recognizes the pedestrian P (determines that the pedestrian P exists).

The determination mechanism 12 sets the search range 14 in the inspection image, sequentially uses the cells 15 of each size, and arranges the cells 15 over the entire area of the search image 14 while partially overlapping the cells 15. The HOG feature value of the image in the search range 14 is calculated by calculating the HOG feature value of the image in the cell 15 and combining the HOG feature values of all the cells 15 arranged in the entire area of the search image 14. Discrimination by the RealAdaBoost algorithm using the weak discriminator obtained from the weak discriminator generation unit 17 for the feature amount calculation unit 22 and the image in the search range 14 (the HOG feature amount of the image in the search range 14) The final output value from the discriminator obtained by repeating the processing (the output value from the discriminator obtained as a result of discrimination processing by the RealAdaBoost algorithm using all selected weak discriminators). The numerical value indicating that the HOG feature amount in the search range 14 is a pedestrian P), the final output value of the identification means 23 and the determination value of the identification determination value generation means 21 are compared. When the final output value is greater than or equal to the determination value, it is determined that the pedestrian P exists within the search range 14 set in the inspection image, and when the final output value is less than the determination value, it is set in the inspection image. Determination means 24 for determining that there is no pedestrian P within the search range 14.

Here, the learning mechanism 11 is a program that expresses the functions of the HOG feature quantity calculation means 16, the weak classifier generation means 17, and the discrimination determination value generation means 21, and the determination mechanism 12 is an inspection HOG feature quantity calculation means. 22, the program which expresses the function of the identification means 23 and the determination means 24, respectively, can be comprised by mounting in a computer.

Next, an object detection method according to the first embodiment of the present invention will be described.
The object detection method according to the first embodiment of the present invention is to identify a pedestrian P using a learning image 13 including an image of a pedestrian P and an image of a non-pedestrian (background image of the pedestrian P). Learning process for constructing a plurality of weak classifiers effective for the image, and setting the search range 14 in the inspection image, the RealAdaBoost algorithm using the weak classifier is used for the image (HOG feature) in the search range 14 And a determination process for determining the presence or absence of the pedestrian P from the output value of the discriminator determined in this way. Details will be described below.

As shown in FIG. 2A, in the learning process, a search range 14 is set in the learning image 13, and a plurality of sizes of cells 15 that determine the size of the local region in the search range 14 are set. The cells 15 are sequentially used, and the cells 15 are arranged over the entire search range 14 while partially overlapping the cells 15, and the HOG feature amount A of the image in the arranged cell 15 is calculated, respectively. The first step of combining the HOG feature amount A for all the cells 15 arranged in the cell 15 into the HOG feature amount B of the image within the search range 14 and the HOG feature amount B from the pedestrian P using the RealAdaBoost algorithm And a second step of selecting a plurality of HOG feature values effective for identification as a weak classifier.

Since each cell 15 is arranged in the entire search range 14 so that the cells 15 partially overlap each other using the cells 15 of a plurality of sizes, for example, the cell 15 that accommodates the right side of the head of the pedestrian P, walking A cell 15 arranged to accommodate the left shoulder and upper left arm of the pedestrian P, a cell 15 arranged to accommodate the right shoulder and upper right arm of the pedestrian P, both lower arms and torso of the pedestrian P Of the pedestrian P, the cell 15 disposed to accommodate the upper left side of the pedestrian P, the cell 15 disposed to accommodate the lower left side of the pedestrian P, and the pedestrian P There is a case where a relatively wide part of the pedestrian P is surrounded by one cell 15 such as the cell 15 arranged so as to accommodate the right side of the lower body. For this reason, the HOG feature value B in the search range 14 includes the HOG feature value A indicating the shape of the right side of the head of the pedestrian P, and the HOG feature value A indicating the characteristics of the shape of the left shoulder and the upper left arm of the pedestrian P. HOG feature A indicating the shape of the shape of the right shoulder and upper right arm of the pedestrian P, HOG feature A indicating the characteristics of the lower and upper arms of the pedestrian P, and the left side of the upper body of the pedestrian P This includes the HOG feature value A indicating the feature of the shape, the HOG feature value A indicating the feature of the left side of the lower body of the pedestrian P, and the HOG feature amount A indicating the feature of the shape of the lower right side of the pedestrian P.

Accordingly, when a plurality of HOG feature quantities effective for identifying the pedestrian P are selected by the weak classifier generation means 16 and are respectively used as weak classifiers, the HOG feature quantity for identifying the right side of the head of the pedestrian P, the pedestrian P HOG feature that identifies the left shoulder and upper left arm of the pedestrian, HOG feature that identifies the right shoulder and upper right arm of the pedestrian P, HOG feature that identifies the lower arm and trunk of the pedestrian P, pedestrian P A relatively wide part of the pedestrian P, such as an HOG feature that identifies the left side of the upper body, an HOG feature that identifies the left side of the lower body of the pedestrian P, and an HOG feature that identifies the right side of the lower body of the pedestrian P HOG feature quantities that are collectively identified are selected as weak classifiers. Thereby, the number of weak classifiers used when identifying the pedestrian P can be reduced. As a result, while sequentially setting the search range 14 in the inspection image, by determining the presence or absence of the pedestrian P from the output value of the discriminator obtained for the image (HOG feature B) in the search range 14, When determining the presence / absence of a pedestrian P in the inspection image, the time required to identify an image in one search range 14 is shortened (the number of weak classifiers is reduced, so the number of weak classifiers is increased by one). The total processing time associated with a decrease in the number of identification processes performed while increasing) can be determined in a short time to determine the presence or absence of the pedestrian P in the inspection image.

On the other hand, as shown in FIG. 2B, the conventional HOG feature value, for example, the HOG feature value of Non-Patent Document 1 (original HOG feature value by Dalal and Triggs) is a cell 26 of a constant size in the entire search region 25. Are arranged without gaps, and the HOG feature values obtained for each cell 26 are combined to form the HOG feature value of the image within the search range 25. For this reason, when the pedestrian 27 exists in the search range 25, in order to cover all the outlines of the pedestrian 27 using the cell 26, many cells 16 are needed.

For this reason, when a plurality of HOG feature values effective for identifying the pedestrian 27 are selected from the HOG feature values of the image within the search range 25 using the RealAdaBoost algorithm, for example, the right side of the head of the pedestrian 27, the left shoulder portion. A plurality of HOG feature quantities are selected for identifying features of the upper left arm, right shoulder upper right arm, lower both arms and body part, upper body left side, lower body left side, and lower body right side. For this reason, the number of weak classifiers used for identifying the pedestrian 27 increases.
As a result, when determining the presence or absence of a pedestrian in an inspection image by determining the presence or absence of a pedestrian from the image within the search range 25 while sequentially setting the search range 25 in the inspection image, one search range It takes a long time to determine the image within 25, and the presence or absence of a pedestrian in the inspection image cannot be determined in a short time.

Next, an object detection apparatus 26 to which the object detection method according to the second embodiment of the present invention is applied will be described.
As shown in FIG. 4, the object detection device 26 uses a learning image (not shown) including an image of a pedestrian P that is an example of a detection target and a background image that is an example of a non-detection target. A learning mechanism 27 for constructing a weak classifier for identifying the pedestrian P (each part of the pedestrian P), and a RealAdaBoost algorithm using the weak classifier while setting a search range in an inspection image (not shown) And a determination mechanism 28 for determining the presence or absence of the pedestrian P from the output value of the discriminator obtained for the image within the search range.

Here, the object detection device 26 is different from the object detection device 10 in that a determination value generation unit 29 instead of the identification determination value generation unit 21, a determination value setting unit 30 instead of the identification unit 23 and the determination unit 24. And the identification processing management means 31 are provided. For this reason, only the determination value generation means 29, the determination value setting means 30, and the identification process management means 31 will be described, and the same reference numerals will be given to common means, and description thereof will be omitted.

The determination value generation unit 29 repeatedly performs the learning process by the RealAdaBoost algorithm using the weak classifier obtained from the weak classifier generation unit 17 for the HOG feature amount B of the image within the search range set in the learning image. When the pedestrian P is detected, an output value (a numerical value indicating that the HOG feature amount within the search range set in the learning image is the pedestrian P is output from the discriminator obtained every number of learning processes. ) Is provided.
The determination value setting unit 30 has a function of setting the minimum value for each number of learning processes obtained by the determination value generation unit 29 as a determination value for each number of identification processes.

The identification processing management unit 31 repeats the identification processing by the RealAdaBoost algorithm using the weak classifier obtained from the weak classifier generation unit 17 for the HOG feature amount of the image within the search range set in the inspection image. In this case, the output value of the discriminator obtained for each number of identification processes is compared with the determination value for each number of identification processes set by the determination value setting means 30. When the output value of the discriminator is equal to or greater than the determination value, the discrimination process is continued, and when the number of discrimination processes is performed a preset number of times, the pedestrian P exists within the search range set in the inspection image. Then, it has a function to determine.
Here, since the identification process is performed while increasing the number of weak classifiers one by one, the preset number of times is the same as the number of selected weak classifiers (that is, using all the selected weak classifiers). It is the number of times until identification processing is performed). The accuracy rate of pedestrian detection increases as the number of weak classifiers increases. As shown in FIG. 3, when the number of weak classifiers exceeds a certain number, the improvement in the accuracy rate is very small. Therefore, for example, when the output of the classifier exceeds 90% of the expected saturation value, or when the number of identification processes reaches 80% of the number of weak classifiers selected, The identification process may be terminated.
As a result, in the presence determination of the pedestrian P within the search range that is sequentially set in the inspection image, the pedestrian P is identified using discriminators having different numbers of weak classifiers for each search range.
Note that the determination value generation unit 29, the determination value setting unit 30, and the identification processing management unit 31 can be configured by installing a program that expresses each function in a computer.

Next, an object detection method according to the second embodiment of the present invention will be described.
The object detection method according to the second embodiment of the present invention uses a learning image composed of an image of a pedestrian P and an image of a non-pedestrian (a background image of the pedestrian P) to identify the pedestrian P. The learning process for constructing a plurality of effective weak classifiers, and setting the search range in the inspection image, the RealAdaBoost algorithm using the weak classifier was used to determine the image within the search range (HOG feature) And a determination process for determining the presence or absence of the pedestrian P from the output value of the discriminator.
Here, the learning process sets a search range in a learning image including an image of a pedestrian P and a non-pedestrian image (background image of the pedestrian P), and determines the size of a local region within the search range. A plurality of cell sizes are set, cells of each size are used in sequence, cells are arranged over the entire search range while partially overlapping each other, and the HOG feature amount A for each cell is calculated to calculate the cell size within the search range. A first step of setting an HOG feature amount B of the image, and a second step of selecting a plurality of HOG feature amounts effective for identifying the pedestrian P from the HOG feature amount B using the RealAdaBoost algorithm and using each as a weak classifier And have. For this reason, when the image which shows the pedestrian P exists in the search range, the cell which contains the comparatively wide site | part of the pedestrian P exists.

As a result, when a plurality of HOG feature quantities effective for identifying the pedestrian P are selected by the weak classifier generating means 16 and each is used as a weak classifier, the HOG features that collectively identify relatively wide parts of the pedestrian P are collected. A quantity is selected as the weak classifier. Thereby, the number of weak classifiers used when identifying the pedestrian P can be reduced, and the presence or absence of the pedestrian P within the search range is determined while sequentially setting the search range in the inspection image. Thus, when determining the presence or absence of the pedestrian P in the inspection image, the time required to identify the image within one search range is shortened, and the presence or absence of the pedestrian P in the inspection image can be determined in a short time. .

If the pedestrian P exists in the search range, the output value of the discriminator obtained by repeatedly performing the discrimination process by the RealAdaBoost algorithm using the weak classifier on the image in the search range As the number of times increases, the output value output from the discriminator increases. Therefore, when the output value of the discriminator is less than a predetermined determination value for each number of identification processes, it is determined that there is no pedestrian P within the search range, and the identification process can be terminated. As a result, the image in the search range is obtained from the output value of the discriminator obtained by repeatedly performing discrimination processing by the RealAdaBoost algorithm using the weak discriminator for the images within the search range (HOG features) sequentially set in the inspection image. When the presence / absence determination of the pedestrian P is performed, the number of weak classifiers constituting the classifier is different for each search range, and the time required for the presence / absence determination of the pedestrian P in the search range, particularly within the search range Therefore, the time required for the determination that no pedestrian P is present can be greatly reduced.
On the other hand, when there is a pedestrian P within the search range, the output value output from the discriminator is greater than or equal to the determination value for each number of discriminating processes, so the discriminating process until the final output value is output from the discriminator. Will be done.

Here, when the pedestrian P exists within the search range set in the learning image, the determination value set for each number of times of the identification processing is obtained by repeating the learning processing by the RealAdaBoost algorithm using the weak classifier. This is the minimum value of the output value that is output when the discriminator obtained for each processing count recognizes (detects) the pedestrian P. By determining the determination value as the minimum value of the output value that is output when the classifier recognizes (detects) the pedestrian P, the determination that the pedestrian P exists within the search range can be made on the safe side ( (Suppressing the occurrence of erroneous determination).
Further, the identification using a classifier having a different number of weak classifiers for each search area, for example, the background area is identified at high speed by a classifier composed of a few weak classifiers, and there is a pedestrian P. The area and the area similar to the pedestrian P are identified by a classifier having a large number of weak classifiers. For this reason, the identification accuracy improvement of the pedestrian P is achieved as a result.

Example 1
Using the object detection method according to the first invention, from the INRIA Person Dataset, which is one of the pedestrian data databases, 2416 positive class images (images of human whole body images), negative class images (persons) 6000 images) were selected, and 500 weak classifiers that recognize each part of a person were selected by machine learning using RealAdaBoost. Next, 500 weak classifiers are used to construct a human classifier. From the INRIA Person Dataset, 1126 positive class images and 3000 negative class images are selected as test images, and a human detection experiment is performed. It was. Here, the minimum size of the cells set in the learning image and the inspection image is 4 pixels vertically and 4 pixels horizontally, and the maximum size is 48 pixels vertically and 24 pixels horizontally.
As a result, the correct answer rate indicating the ratio of the positive class image determined as the positive class image and the negative class image determined as the negative class image was 96.3%.
In addition, the processing time required for identification within the search area (60 pixels long and 30 pixels wide) set in the inspection image was 28.96 microseconds for both the Positive class image and the Negative class image.
Further, FIG. 5 shows the relationship between the number of weak classifiers constituting the classifier and the error rate.

(Example 2)
Using the object detection method according to the second invention, a human detection experiment was performed on the same inspection image as in Example 1, and the correct answer rate was obtained. Further, the processing time required for identification within one search area having the same size as that of the first embodiment was obtained. The strong classifier used in the second embodiment is the same as the weak classifier used in the first embodiment. As a result, the correct answer rate of the human detection experiment was 97.1%. The processing time required for identification within one search area was 27.80 microseconds for the Positive class image and 0.99 microseconds for the Negative class image. Further, FIG. 5 shows the relationship between the number of weak classifiers constituting the classifier and the error rate.

(Comparative example)
As a comparative example, in the object detection method using the HOG feature amount of Dalal et al. (Non-Patent Document 1), each part of a person is obtained by machine learning using RealAdaBoost on the same learning image as in Example 1. 500 weak classifiers to be recognized were selected, and 500 weak classifiers were linearly combined to form a strong classifier. Then, a human detection experiment was performed on the same inspection image as in Example 1, and the correct answer rate was obtained. Further, the processing time required for identification within one search area having the same size as that of the first embodiment was obtained. Here, the size of the cell set in the learning image and the inspection image is 5 pixels vertically and 5 pixels horizontally.
As a result, the correct answer rate of the human detection experiment was 94.8%. The processing time required for identification within one search area was 152.51 microseconds for both the Positive class image and the Negative class image. Further, FIG. 5 shows the relationship between the number of weak classifiers constituting the classifier and the error rate.

From the results of Examples 1 and 2 and the comparative example, it was confirmed that the processing time in Examples 1 and 2 can be significantly reduced. In particular, in Example 2, the processing speed was improved by about 154 times compared to the comparative example with respect to the negative class image.
In the first and second embodiments, a large-sized cell is used to quickly eliminate a region with few edges, such as a road in the search region, and then a building, a road tree, and a pedestrian are used for accuracy by using a small-sized cell. Well classified. In this way, by increasing the cell size to reduce the number of weak classifiers in a region where identification is easy, and by decreasing the cell size and increasing the number of weak classifiers in a region where identification is difficult, FIG. As shown in Fig. 2, it is possible to significantly reduce the error rate (a great improvement in the identification accuracy) when the number of weak classifiers is small.

As described above, the present invention has been described with reference to the embodiment. However, the present invention is not limited to the configuration described in the above-described embodiment, and the matters described in the scope of claims. Other embodiments and modifications conceivable within the scope are also included.
Further, the present invention also includes a combination of components included in the present embodiment and other embodiments and modifications.

10: Object detection device, 11: Learning mechanism, 12: Determination mechanism, 13: Learning image, 14: Search range, 15: Cell, 16: HOG feature amount calculation means, 17: Weak classifier generation means, 18: Learning Image storage unit, 19: first HOG feature amount calculation unit, 20: second HOG feature amount calculation unit, 21: identification determination value generation unit, 22: inspection HOG feature amount calculation unit, 23: identification unit, 24 : Determination means, 25: search range, 26: object detection device, 27: learning mechanism, 28: determination mechanism, 29: determination value generation means, 30: determination value setting means, 31: identification process management means

Claims (6)

A learning process for constructing a plurality of weak classifiers for identifying the detection target object using a learning image composed of an image of the detection target object and an image of the non-detection target object, while setting a search range in the inspection image, In an object detection method comprising: a determination process for determining the presence / absence of the detection target from an output value of the classifier obtained for an image within the search range by RealAdaBoost algorithm using a weak classifier,
The learning process sets the search range in the learning image, sets a plurality of cell sizes that determine the size of the local region in the search range, and sequentially uses the cells of each size, The cells are arranged over the entire search range while partially overlapping each other, the HOG feature amount A of the image in the arranged cell is calculated, and the HOG for each of the cells arranged in the search range. A first step of combining the feature amount A into the HOG feature amount B of the image within the search range;
A second step of selecting a plurality of HOG feature values effective for identifying the detection object from the HOG feature value B using the RealAdaBoost algorithm, and using each as the weak classifier. Detection method. 2. The object detection method according to claim 1, wherein, in the determination process, when performing the determination by repeatedly performing identification processing by RealAdaBoost algorithm using the weak classifier for the search range set in the inspection image. When the output value of the discriminator is less than a predetermined determination value for each number of the discriminating processes, the discriminator is terminated for each search range set in the inspection image by ending the discriminating process. An object detection method, wherein the number of weak classifiers to be configured is changed. 3. The object detection method according to claim 2, wherein the determination value set in advance for each number of times of the identification processing is learned by a RealAdaBoost algorithm using the weak classifier for the search range set in the learning image. A method for detecting an object, comprising: a minimum value of an output value output from a discriminator obtained for each number of learning processes when the detection target is detected by repeating a process. A learning mechanism for constructing a plurality of weak classifiers for identifying the detection target object using a learning image composed of an image of the detection target object and an image of the non-detection target object, while setting a search range in the inspection image, In an object detection apparatus having a determination mechanism for determining the presence or absence of the detection target object from an output value of the classifier obtained for the image within the search range by the RealAdaBoost algorithm using a weak classifier,
The learning mechanism sets the search range in the learning image and sets a plurality of cell sizes that determine the size of the local region in the search range, and sequentially uses the cells of each size to The cells are arranged over the entire search range while partially overlapping each other, the HOG feature amount A of the image in the arranged cell is calculated, and the HOG for each of the cells arranged in the search range. HOG feature quantity calculating means for obtaining the HOG feature quantity B of the image within the search range by combining the feature quantities A;
An object comprising: a weak classifier generating unit that selects a plurality of HOG feature quantities effective for identifying the detection target object from the HOG feature quantity B using the RealAdaBoost algorithm and uses the HOG feature quantity as the weak classifier Detection device. The object detection apparatus according to claim 4, wherein the determination mechanism performs a determination by repeatedly performing a recognition process by a RealAdaBoost algorithm using the weak classifier for the search range set in the inspection image. When the output value of the discriminator is less than a predetermined determination value for each number of the discriminating processes, the discriminator is terminated for each search range set in the inspection image by ending the discriminating process. An object detection apparatus comprising: an identification processing management unit that changes the number of the weak classifiers. 6. The object detection apparatus according to claim 5, wherein the learning mechanism detects the detection target object by repeatedly performing a learning process using a RealAdaBoost algorithm on the search range set in the learning image using the weak classifier. A determination value generating means for obtaining a minimum value of the output value output from the discriminator obtained for each number of times of the learning process,
The determination mechanism includes a determination value setting unit that sets each minimum value obtained by the identification determination value generation unit as the determination value set in advance for each number of times of the identification process. Detection device.

Images