JP5220482B2 - Object detection apparatus and program - Google Patents

Description

The present invention relates to a Target object detection apparatus and a program.

  Conventionally, an object detection device that directly and accurately detects a person in a screen of a video without performing processing such as background cutting has been disclosed (see Patent Document 1).

The object detection device of Patent Literature 1 detects an object candidate from an image identified while referring to the object feature defined in the object feature dictionary, and resembles the detection target object and the detection target from the object candidates. A discriminant feature dictionary for identifying a pseudo object that is not a detection target object is created, and only the object that is the detection target is discriminated from the object candidates using the discrimination feature dictionary.
JP-A-10-222678

  Usually, the size of an object to be detected or an object candidate (hereinafter referred to as “object etc.”), that is, the resolution of the object is influenced by the distance to the object. Different. However, in Patent Document 1, the feature of an object is defined by the object feature dictionary without considering the resolution, and the discrimination feature dictionary is created. For this reason, there is a problem that a dictionary with good performance is not created. Further, the object detection device of Patent Document 1 uses the same dictionary created from images with a predetermined resolution for images with different resolutions, and thus has a problem that the performance of the identification function is low.

The present invention has been proposed to solve the problems described above, regardless of the size of the object in the image, the object pair Ru can be detected with high accuracy elephant object detection apparatus and program The purpose is to provide.

The object detection device according to the first aspect of the present invention includes a preprocessing unit that generates a plurality of images obtained by performing a plurality of blurring processes with different degrees of blurring on an input image, and the image A search area setting means for setting a search area for an object, and each time the search area is set by the search area setting means, the size of the search area set by the search area setting means is a predetermined value. When larger than the size, among the plurality of images preprocessed by the preprocessing means, an image whose blurring degree increases as the size of the search region increases is used, and the characteristics of the image and the object are determined. Based on the learning model to be detected, it is detected whether there is an object in the search area, and when the size of the search area set by the search area setting means is equal to or less than the predetermined size, Serial input image, a learning model that indicates the characteristic of the object, on the basis, and a, and an object detecting means for detecting whether there is the object in the search area.

Invention of Claim 2 is the target object detection apparatus of Claim 1 , Comprising: The said pre-processing means performs the said several blurring process and an outline part as said pre-processing with respect to the said input image. The plurality of images that have undergone sharpening processing to be enhanced are generated, and the object detection means has a search area size set by the search area setting means that is less than or equal to a second size smaller than a predetermined size. In some cases, the search region based on the image and the learning model indicating the characteristics of the target object using an image in which the contour portion is emphasized among the plurality of images preprocessed by the preprocessing means. Detect if there is an object in

According to a fifth aspect of the present invention, there is provided an object detection program comprising: a preprocessing unit that generates a plurality of images in which a plurality of blurring processes having different degrees of blurring are performed as preprocessing on an input image; A search area setting means for setting a search area of an object for the image, and a size of the search area set by the search area setting means each time the search area is set by the search area setting means Is larger than a predetermined size, among the plurality of images pre-processed by the pre-processing means, an image whose blurring degree increases as the size of the search area increases is used. Based on the learning model indicating the characteristics of the search area, it is detected whether there is an object in the search area, and the size of the search area set by the search area setting means is the predetermined If it is equal to or less than the size, it functions as an object detection means for detecting whether there is an object in the search area based on the input image and a learning model indicating the characteristics of the object Let

  The object detection device and the program according to the present invention each have a gradation corresponding to the size of the search area among a plurality of images preprocessed so that the gradation difference of the contour portion differs every time the search area is set. Since the object is detected using the difference image, the object can be detected with high accuracy regardless of the size of the object in the image.

  Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the drawings.

[principle]
FIG. 1 is a diagram showing images in search windows of various sizes (hereinafter referred to as “search window images”) and object detection results. As shown in the figure, as the distance to the object increases, the image size of the object (pedestrian in the present embodiment) decreases, so the search window image decreases. Conventionally, when the distance to the object is far or near, the object is not detected from the search window image, and the object is detected from the search window image only in the middle distance. Therefore, when the distance to the object is long or close, pre-processing for detecting the object is necessary.

  FIG. 2 is a diagram showing a search window image and a two-dimensional FFT (Fast Fourier Transform) spectrum pattern in the case of being far, close, or intermediate to the object. FIG. 3 is a diagram illustrating a high-frequency portion of the spectrum pattern of the two-dimensional FFT. FIG. 4 is a diagram showing a two-dimensional FFT spectrum pattern of a search window image having a good detection performance and a two-dimensional FFT spectrum pattern when the 7 × 7 Gaussian filter processing is not performed and when the processing is performed. It is. The blurring process is not limited to the Gaussian filter process, and other processes such as a median filter process may be used.

  As shown in FIG. 4, the high frequency component when the 7 × 7 Gaussian filter processing is not performed is not similar to the high frequency component of the search window image having a good detection performance, but the 7 × 7 Gaussian filter processing is performed. The high-frequency component in this case is similar to the high-frequency component of the search window image having a good detection performance. That is, when the distance to the object is short, it is possible to detect the object from the search window image by performing a blurring process on the search window image. This is due to the following reason.

  Depending on the size of the search window (whether the object is near or far from the camera), the resolution, which is the sharpness of the image, differs. For example, if the object is a person, if the person is close to the camera, that is, if the search window size is large, the image will be clear and not only the rough shape of the person but also the clothes pattern and background Is photographed clearly. Therefore, it is difficult to detect a person using this image as it is.

  Also, if the person is far from the camera, that is, the search window is small in size, the image will be blurred, so the information affecting the detection of objects such as clothing patterns and backgrounds will be reduced, but even the rough shape of the person Difficult to detect. When the size of the search window is between these, the search window image is moderately blurred and is most easily detected.

  Therefore, in the case where the size of the search window is large and small, the object detection performance can be improved by performing image processing on the search window image so that the degree of blur is the same as in the middle case. it can. Note that the object detection performance can be improved by performing the same processing as described above not only when detecting an object but also when generating a learning model.

  FIG. 5 is a diagram for explaining how to search for the optimum preprocessing. First, a plurality of groups are created for each size for all the sample images. This image is subjected to detection processing by a conventional object detection apparatus without preprocessing, and a group of samples having the highest detection rate is set as a reference image. Next, as shown in FIG. 5, all the sample images are processed images, processes 1, 2,..., N (process 1 to N: blurring method, blurring process with different blurring degrees). Prepare an image marked with. Then, the similarity between the FFT high frequency component of the reference image and the FFT high frequency component of each image described above is calculated using, for example, the Euclidean distance.

  For each size group, count the number of samples with a higher similarity after processing than the unprocessed similarity, and process more than the number of samples with a lower similarity after processing than the unprocessed similarity. Use effective processing. Select a pre-process to be used for each size group from the valid processes. As a selection method, there is a method of minimizing the processing to be used for the sample image or selecting a processing having the largest number of samples whose similarity after processing is higher than the similarity without processing. The object detection apparatus according to the embodiment of the present invention performs the pre-processing thus found as follows.

[Configuration of object detection device]
FIG. 6 is a block diagram showing the configuration of the object detection device according to the embodiment of the present invention. The object detection device includes a learning preprocessing unit 11 that performs preprocessing for generating a learning model, a learning unit 12 that generates a learning model, a learning model storage unit 13 that stores the generated learning model, It has.

  The object detection apparatus further includes an imaging unit 21 that captures an object, a detection preprocessing unit 22 that performs preprocessing for detecting the object, and a window image extraction that extracts a window image that is an image in the search window. Unit 23, a detection unit 24 that performs processing for detecting an object, and a result output unit 25 that outputs a detection result.

  The object detection apparatus configured as described above executes the following learning processing routine to generate a learning model in order to detect, for example, a pedestrian as an object, and uses this learning model to detect the object detection routine. Execute.

[Learning objects]
FIG. 7 is a flowchart showing a learning processing routine.

  In step S1, the learning preprocessing unit 11 inputs the learning image (positive image) of the object and the learning image (negative image) of the object that is not the object, and proceeds to step S2. In this embodiment, a pedestrian image is used as a positive image, and an image such as a utility pole or a signboard is used as a negative image.

  In step S2, the learning preprocessing unit 11 preprocesses the learning image based on the size of the target object included in the learning image. In the present embodiment, the learning pre-processing unit 11 performs sharpness processing on the learning image if the size of the object (number of pixels in the vertical direction) is 50 pixels or less, and sets the size of the object to 50 pixels. If it exceeds 90 pixels, no processing is performed. The learning pre-processing unit 11 performs 3 × 3 Gaussian filter processing if the size of the object exceeds 90 pixels and is 140 pixels or less, and 7 × 7 if the size of the object exceeds 140 pixels. Perform Gaussian filter processing. That is, the learning pre-processing unit 11 performs image processing on the learning image as necessary so that the degree of blur increases as the size of the object increases.

  In step S3, the learning preprocessing unit 11 selects an image of a necessary area from the learning image obtained in step S2, specifically, an area image in which an object exists (hereinafter referred to as “object area image”). Cut out with a predetermined aspect ratio and proceed to step S4. For example, when the object is a pedestrian, the learning preprocessing unit 11 cuts out the object region image at a ratio of 2 to 1 in the horizontal direction.

  In step S4, the learning unit 12 performs learning using the normalized object region image and generates a learning model. For example, the learning unit 12 normalizes the luminance and size of the object region image and extracts the feature amount. Examples of the feature amount include a Haar-Like feature, a four-way surface feature, a two-dimensional fast Fourier transform spectrum pattern, and the like. In addition, the learning unit 12 may perform learning such as AdaBoost and support vector machine (SVM).

  As described above, the object detection device increases the degree of blurring as the distance to the object becomes shorter with respect to the learning image, that is, as the size of the object included in the learning image increases. A blurring process is performed, and a learning model for detecting an object is generated using the learning image that has been blurred. Thereby, the target object detection apparatus can generate a learning model for accurately detecting the target object without being affected by the distance to the target object.

[Target detection]
FIG. 8 is a flowchart showing an object detection routine. The target object detection apparatus executes the processes from the next step S11 to step S20.

  In step S11, the imaging unit 21 captures a subject and generates an image. The detection preprocessing unit 22 inputs the image generated by the imaging unit 21, and proceeds to step S12. In the present embodiment, the detection preprocessing unit 22 inputs an image generated in real time by the imaging unit 21, but an image stored in advance in a storage device (not shown) may be input.

  In step S12, the detection preprocessing unit 22 creates a preprocessed image by a plurality of preprocesses to be used, and proceeds to step S13. Here, the detection preprocessing unit 22 performs three preprocessing, specifically sharpness processing (for example, an image obtained by subtracting a Laplacian-processed image) from the image generated by the imaging unit 21, 3 × 3 Gaussian filter processing and 7 × 7 Gaussian filter processing are performed, respectively, and a search window of a predetermined size is set.

  In step S13, the window extraction unit 23 selects an image from the preprocessed images and the unprocessed images according to the size of the search window (the number of pixels in the vertical direction), for example, as follows. select.

(Search window size) → (Preprocessed image to be selected)
50 pixels or less → Sharpness processed image Over 50 pixels and 90 pixels or less → Unprocessed image Over 90 pixels and 140 pixels or less → 3 × 3 Gaussian filtered image Over 140 pixels → 7 × 7 Gaussian filtered image As described above, as the size of the search window increases, the gradation difference of the contour portion of the selected image decreases (the degree of blurring increases). The relationship between the search window size and the preprocessed image is merely an example, and the present invention is not limited to this. Then, the window extraction unit 23 sets a search window for the selected image, extracts a search window image that is an image in the search window, and proceeds to step S14.

  In step S14, the window extraction unit 23 normalizes the size and luminance value of the search window image, and proceeds to step S15.

  In step S15, the detection unit 24 calculates a feature amount from the search window image normalized in step S14, and proceeds to step S16. The feature amount includes, for example, a Haar-Like feature, a four-way surface feature, a two-dimensional fast Fourier transform (FFT) spectrum pattern, and the like.

  In step S16, the detection unit 24 compares the normalized feature amount of the search window image with the learning model (feature amount) stored in the learning model storage unit 13, and proceeds to step S17.

  In step S17, the detection unit 24 calculates the likelihood of the target object as an evaluation value as the comparison result in step S16, and determines whether the search window image is the target object based on the evaluation value. It should be noted that a known technique can be used for detecting an object using a search window image and a learning model. If the search window image is an object, the process proceeds to step S18. If the search window image is not an object, the process proceeds to step S19.

  In step S18, the detection unit 24 saves the size and position of the search window when the object is detected in step S15 in the object detection list, and proceeds to step S19.

  In step S19, the detection unit 24 scans the entire search window to determine whether the entire image has been searched. When the search is completed, the process proceeds to step S20. When the search has not been completed, the search is performed. The window position is shifted by one step and the process returns to step S14. And the process from step S13 to step S19 is repeatedly performed again. When the search window searches the entire image, the process proceeds to step S20.

  In step S20, the detection unit 24 determines whether search windows of all sizes are used. If the determination is affirmative, the process proceeds to step S21. If the determination is negative, the size of the search window is changed and the process returns to step S13. . Then, the processes from step S13 to step S20 are repeated, and when the search is completed in the search windows of all sizes, the process proceeds to step S21.

  In step S21, the result output unit 25 displays the detected object surrounded by a window based on the size and position of the search window in the object detection list as a detection result. The result output unit 25 is not limited to the above example, and may output at least one of information on the presence / absence of an object, an object image, and an object position by sound or an image.

  As described above, the object detection apparatus uses an image with a small gradation difference in the contour portion, that is, with a large degree of blurring as the distance to the object decreases, that is, as the size of the search window increases. Detect objects. Thereby, the target object detection apparatus can accurately detect the target object without being affected by the distance to the target object and without preparing a learning model for each distance to the target object.

  Note that the present invention is not limited to the above-described embodiment, and it is needless to say that the present invention can also be applied to a design modified within the scope of the claims.

  In this embodiment, the Gaussian filter process is used as the blur process, but the blur process is not limited to this, and a median filter process or a reduction process may be used.

  Further, the degree of blurring is not limited to the above-described example, and Gaussian filter processing with different filter sizes may be used. Furthermore, median filtering with a plurality of filter sizes may be used. Here, when the size of the object in the image is large, Gaussian filter processing and median filter processing may be performed. When the size of the object in the image is small, sharpness processing may be performed.

When the object detection device detects an object using the learning model generated through the learning preprocessing unit 11 and the learning unit 12, the preprocessing unit for detection 22 performs preprocessing as in the above embodiment. An image that has been processed may be used, or an object may be detected using an image that has not been preprocessed.
Further, when the object detection device detects an object using an image that has passed through the detection preprocessing unit 22, the learning model generated through the learning preprocessing unit 11 and the learning unit 12 as in the above embodiment. Or an object may be detected using a learning image that has not been preprocessed.

It is a figure which shows the search window image and object detection result of various sizes. It is a figure which shows the spectrum pattern of the search window image and the two-dimensional FFT in the case of being far, close, and intermediate to the object. It is a figure which shows the high frequency part of the spectrum pattern of two-dimensional FFT. It is a figure which shows the spectrum pattern of the two-dimensional FFT of the case where the 7 * 7 Gaussian filter process is not performed, and the case where the process is performed, and the two-dimensional FFT spectrum pattern of the search window image having a good detection performance. It is a figure explaining how to search for the optimal preprocessing. It is a block diagram which shows the structure of the target object detection apparatus which concerns on embodiment of this invention. It is a flowchart which shows a learning process routine. It is a flowchart which shows a target object detection routine.

Explanation of symbols

DESCRIPTION OF SYMBOLS 11 Learning pre-processing part 12 Learning part 13 Learning model memory | storage part 21 Imaging part 22 Detection pre-processing part 23 Window extraction part 24 Detection part 25 Result output part

Claims (6)

Preprocessing means for generating a plurality of images subjected to a plurality of blurring processes with different blurring degrees as preprocessing for the input image;
Search area setting means for setting a search area of an object for the image;
Each time the search area is set by the search area setting means, if the size of the search area set by the search area setting means is larger than a predetermined size, the plurality of preprocessed by the preprocessing means Among the images, an image whose blurring degree increases as the size of the search region increases, and the search region has an object based on the image and a learning model indicating the feature of the object. And when the size of the search area set by the search area setting means is less than or equal to the predetermined size, based on the input image and a learning model indicating the characteristics of the object Object detection means for detecting whether there is an object in the search area;
An object detection apparatus comprising: The preprocessing means generates the plurality of images subjected to the plurality of blurring processes and the sharpness processing for enhancing the contour portion as the preprocessing for the input image,
When the size of the search area set by the search area setting unit is equal to or smaller than a second size smaller than a predetermined size, the object detection unit is configured to store a plurality of images preprocessed by the preprocessing unit. among them, using an image contour portion is emphasized, and the image, the learning model indicating the feature of the object, based on the subject of claim 1 for detecting whether there is the object in the search area Object detection device. For each of the plurality of learning images including the target object, when the size of the target object in the learning image is larger than a predetermined size, the preprocessing is blurred as the target size increases. Preprocessing means for learning that performs blurring processing to increase the degree,
  Learning model generation means for generating a learning model indicating the characteristics of the object based on each learning image preprocessed by the learning preprocessing means;
  The object detection device according to claim 1, further comprising: The learning preprocessing means performs a sharpness process for emphasizing an outline portion as the preprocessing when the size of the object in the learning image is equal to or smaller than a second size smaller than the predetermined size.
  The object detection device according to claim 3. Computer
Preprocessing means for generating a plurality of images subjected to a plurality of blurring processes with different blurring degrees as preprocessing for the input image;
Search area setting means for setting a search area of an object for the image;
Each time the search area is set by the search area setting means, if the size of the search area set by the search area setting means is larger than a predetermined size, the plurality of preprocessed by the preprocessing means Among the images, an image whose blurring degree increases as the size of the search region increases, and the search region has an object based on the image and a learning model indicating the feature of the object. And when the size of the search area set by the search area setting means is less than or equal to the predetermined size, based on the input image and a learning model indicating the characteristics of the object Object detection means for detecting whether there is an object in the search area;
The object detection program to make it function.   The preprocessing means generates the plurality of images subjected to the plurality of blurring processes and the sharpness processing for enhancing the contour portion as the preprocessing for the input image,
  When the size of the search area set by the search area setting unit is equal to or smaller than a second size smaller than a predetermined size, the object detection unit is configured to store a plurality of images preprocessed by the preprocessing unit. Of these, an image with an emphasized contour portion is used to detect whether there is an object in the search region based on the image and a learning model indicating the characteristics of the object.
  The object detection program according to claim 5.