JP2012098983A - Object detection device, object detection method and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To prevent errors in detecting a detection target object.SOLUTION: A luminance gradient calculation unit 3 calculates a luminance gradient indicating a change rate of luminance in pixels of image data. A histogram calculation unit 4 calculates histograms with data sections of respective predetermined angular ranges in the luminance gradient for each image area divided into predetermined sizes in the image data. A shift unit 5 changes the histogram by shifting the data section of the histogram. A similarity calculation unit 6 calculates similarity between the histogram of the image area whose data section has been shifted by the shift unit 5 and the histogram of the other image area calculated by the histogram calculation unit 4. An object detection unit 8 detects a detection target object from the image data by using a feature amount of the similarity.

Description

  The present invention relates to an object detection device, an object detection method, and a program.

A technology for automatically detecting an object from an image taken by a camera or the like is widely used such as an automobile or a human monitoring system.
Features an HOG (histograms of oriented gradients) that divides the image into grids for each image area (hereinafter referred to as a cell) of a predetermined size and uses the luminance gradient angle of each pixel in the cell as a data section A technique for detecting an object using a quantity is known.

  Also, a technique called Joint HOG that performs object detection by combining a plurality of HOG features, or CoHOG (Co-occurrence HOG) that performs object detection using the number of appearances of each pixel pair having different angles as a feature quantity. Techniques are known.

JP 2010-044438 A JP 2010-055195 A

  However, with the conventional technology, when the luminance gradient histogram distribution changes due to the influence of gloss change, the feature quantity that distinguishes the object to be detected from other parts changes, and the object may not be detected correctly. was there.

  According to an aspect of the invention, a luminance gradient calculation unit that calculates a luminance gradient indicating a rate of change in luminance at a pixel in image data, and the image area of the image data divided for each image area of a predetermined size On the other hand, a histogram calculation unit that calculates a histogram with a predetermined angular range of the luminance gradient as a data interval, a shift unit that shifts the data interval of the histogram and changes the histogram, and the shift unit A similarity calculation unit that calculates a similarity between the histogram of the image region in which the data section is shifted and the histogram of the other image region calculated by the histogram calculation unit; An object detection device is provided that includes an object detection unit that detects an object to be detected from the image data using the feature amount.

  According to the disclosed object detection apparatus, object detection method, and program, object detection errors can be reduced.

It is a figure which shows an example of the object detection apparatus of 1st Embodiment. It is a figure which shows an example of a certain two cell in image data. It is a figure which shows an example of the histogram of a brightness | luminance gradient. It is a figure which shows an example of the shift of a histogram. It is a figure which shows an example of the overlapping degree of the histogram of two cells. It is a flowchart which shows the flow of an example of an object detection process. It is a figure which shows an example of the object detection apparatus of 2nd Embodiment. It is a figure which shows the example of a scanning of a detection window area | region. It is a figure which shows an example of the pixel coordinate in a certain image data. It is a figure which shows an example of the luminance value inside and outside the cell in a certain image data. It is a figure which shows an example of the angle of the luminance gradient of a cell. It is a figure which shows an example of the magnitude | size of the brightness | luminance gradient of a cell. It is a figure which shows an example of a histogram. It is a flowchart which shows the flow of an example of a learning process. It is a figure which shows the example of a cell setting. It is a flowchart which shows an example of the update of a weight value, and the selection process of a weak discriminator. It is a figure which shows the example of a storage of a class classification parameter. It is a flowchart which shows an example of the process which discriminate | determines the presence or absence of an object. It is a figure which shows an example of the hardware of the computer used for this Embodiment.

Hereinafter, embodiments of the present invention will be described with reference to the drawings.
(First embodiment)
FIG. 1 is a diagram illustrating an example of the object detection apparatus according to the first embodiment.

The object detection apparatus 1 includes an image acquisition unit 2, a luminance gradient calculation unit 3, a histogram calculation unit 4, a shift unit 5, a similarity calculation unit 6, a storage unit 7, and an object detection unit 8.
The image acquisition unit 2 acquires image data from outside the object detection device 1. For example, the image data is input from an imaging device such as a camera or a video, a recording medium such as a DVD (Digital Versatile Disc), a hard disk drive (HDD).

The luminance gradient calculation unit 3 calculates a luminance gradient that indicates the rate of change in luminance at the pixels in the image data. The luminance gradient is represented by, for example, a luminance gradient direction (gradient angle) and a gradient magnitude.
The histogram calculation unit 4 uses a predetermined angular range of the luminance gradient as a data section in the image data cells partitioned in a grid pattern for each cell having a predetermined size, and counts the number of pixels or the integration of the gradient in frequency. A histogram is calculated. When ignoring the direction of the gradient direction, the gradient angle ranges from 0 degrees to 180 degrees.

FIG. 2 is a diagram illustrating an example of two cells in the image data.
FIG. 3 is a diagram illustrating an example of a histogram of luminance gradient.
In FIG. 2, vehicle image data 10 is shown as an example. FIG. 3 shows an example of a histogram of luminance gradients calculated in the cells 11-1 and 11-2 (for example, 5 × 5 pixels) in the region 11 in the image data 10.

3A shows an example of the histogram in the cell 11-1, and FIG. 3B shows an example of the histogram in the cell 11-2. The horizontal axis indicates the gradient angle (degrees), and the vertical axis indicates the frequency.
Each histogram is represented by the frequency for each 20 degree data section. For example, in the cell 11-1, the frequency at 80 degrees to 100 degrees is the maximum. That is, since the change in luminance in the vertical direction is large, it can be seen that the cell 11-1 includes the most horizontal edges. In addition, since the frequency in the range of 160 to 180 degrees is the maximum in the cell 11-2, it can be seen that the cell 11-2 includes the most vertical edges.

  Note that the data interval is not limited to every 20 degrees, and may be as fine as every 10 degrees or vice versa. The data interval is appropriately set in relation to the calculation time.

The shift unit 5 changes the histogram by shifting the data section of the histogram calculated by the histogram calculation unit 4 a predetermined number of times.
FIG. 4 is a diagram illustrating an example of histogram shift.

  FIG. 4 shows an example in which the histogram of the cell 11-2 shown in FIG. 3 is shifted four times. That is, the data section with the gradient angle of 0 degrees to 20 degrees is shifted to the right of the data section of 160 to 180 degrees, and further, the data section of 20 degrees to 40 degrees, the data section of 40 degrees to 60 degrees, and 60 degrees. The data interval of ˜80 degrees is subsequently shifted.

  The similarity calculation unit 6 calculates the similarity between the histogram of the cell whose data section has been shifted and the histograms of other cells calculated by the histogram calculation unit 4. The similarity calculation unit 6 calculates, for example, the degree of overlap between two histograms (hereinafter referred to as the overlap rate) as the similarity. Further, the similarity calculation unit 6 may use the difference between the frequencies of the two histograms as the similarity.

FIG. 5 is a diagram illustrating an example of the degree of overlap of histograms of two cells. The horizontal axis indicates the data section, and the vertical axis indicates the frequency.
In each data section, the left bar shows an example of the histogram of the cell 11-1 shown in FIG. 3A, and the right bar shows the histogram of the cell 11-2 obtained by shifting the data section shown in FIG. 4 four times. An example is shown.

  The overlap ratio of the histogram is obtained by selecting the smaller value (the white bar in FIG. 5) from the frequency of two cells in the same data section and calculating the sum for all the data sections. .

  The storage unit 7 stores classification parameters obtained from the result of learning using the similarity calculated by the similarity calculation unit 6 as a feature amount. The classification parameter is used to determine whether or not it is considered that there is an object to be detected in the image data from the similarity (feature value) obtained by the similarity calculation unit 6. Processing related to learning will be described later.

  The object detection unit 8 uses the similarity calculated by the similarity calculation unit 6 as a feature amount, and refers to the classification parameter stored in the storage unit 7 to detect the detection target object from the image data. Details of the object detection process will be described later.

Hereinafter, the flow of object detection processing by the object detection apparatus 1 will be summarized.
FIG. 6 is a flowchart illustrating an exemplary flow of object detection processing.
Step S1: The image acquisition unit 2 acquires image data from the outside of the object detection device 1.

Step S2: The luminance gradient calculation unit 3 calculates a luminance gradient indicating the rate of change in luminance at the pixels in the image data.
Step S3: The histogram calculation unit 4 integrates the number of pixels or the magnitude of the gradient using the luminance gradient angle as a data interval in a cell of interest in the image data divided into cells of a predetermined size. A histogram is calculated with the frequency as.

Step S4: The shift unit 5 changes the histogram by shifting the data section of the histogram calculated by the histogram calculation unit 4 a predetermined number of times.
Step S5: The similarity calculation unit 6 calculates a similarity between the histogram of the cell whose data section is shifted by the shift unit 5 and the histogram of another cell calculated by the histogram calculation unit 4.

Step S6: Using the similarity calculated by the similarity calculation unit 6 as a feature quantity, the classification parameter stored in the storage unit 7 is referred to detect an object to be detected from the image data.
As described above, the object detection apparatus 1 detects an object in an image using the similarity between the histogram of the luminance gradient in a certain cell in the image data and the histogram obtained by shifting the data section in another cell as a feature amount. For this reason, even if the luminance gradient histogram distribution changes due to a change in gloss or the like, the feature quantity for discriminating the object to be detected from other parts is unlikely to change. As a result, object detection errors can be reduced.

  In addition, since the similarity between the histogram of the brightness gradient in one cell and the histogram obtained by shifting the data interval in another cell is calculated and used as a feature value, the connection state of edges and the relationship between distant edges are used as feature values. It is possible to detect the object in consideration of these.

(Second Embodiment)
FIG. 7 is a diagram illustrating an example of the object detection device according to the second embodiment.
The object detection device 50 includes an image acquisition unit 51, a detection window scanning unit 52, a luminance gradient calculation unit 53, a histogram calculation unit 54, a shift unit 55, a similarity calculation unit 56, a learning unit 57, a storage unit 58, and an object detection unit 59. have.

Hereinafter, each part will be described in detail.
(Image acquisition unit 51)
The image acquisition unit 51 acquires image data from an input device external to the object detection device 50, for example. The input device may be other than a camera, and the image acquisition unit 51 may acquire an image cut out from a video, a recording medium such as a DVD, or a moving image file stored on the HDD.

  The image data is given by, for example, a rectangle of horizontal 640 pixels × vertical 480 pixels, and one pixel is given by a brightness gradation value (luminance). For example, as in the example shown in FIG. 9 described later, the gradation value of the luminance (I) of the pixel at the coordinates (i, j) indicated by the integers i, j is given as an 8-bit digital value.

(Detection window scanning unit 52)
When performing detection processing of an object, the detection window scanning unit 52 includes a detection window region (a training image at the time of learning processing to be described later and a training image at the time of learning processing) in the input unknown image data (for example, image size = 640 × 480 pixels). Set the same image size. Then, the detection window scanning unit 52 scans the detection window region with an appropriate number of pixels in consideration of the processing time.

FIG. 8 is a diagram illustrating a scanning example of the detection window region.
An example in which a detection window area 61 is set for the image data 60 is shown. The detection window scanning unit 52 moves the detection window region 61 on the unknown image data, for example, in the direction of the arrow in FIG. When the area where the target object is not detected (for example, the area 62 in FIG. 8) is known in advance, the detection window scanning unit 52 can reduce the processing time by excluding it from the scanning target of the detection window area 61. is there.

(Luminance gradient calculation unit 53)
The luminance gradient calculation unit 53 calculates a luminance gradient that is a luminance change rate in an image in the image data. Various methods can be applied as the method of calculating the luminance gradient. Here, it is assumed that the following simplest method is used.

FIG. 9 is a diagram illustrating an example of pixel coordinates in certain image data. The pixel coordinates of a certain area in the image data 70 are indicated by integers i and j.
The luminance gradient angle θ (i, j) at the pixel 71 at the coordinate (i, j) in FIG. 9 is the luminance I (i, j + 1), I of the pixels 72, 73, 74, 75 above, below, left, and right of the pixel 71. Using (i-1, j), I (i, j-1), I (i + 1, j),

  Further, the magnitude η (i, j) of the luminance gradient at the coordinates (i, j) is expressed by the following equation.

FIG. 10 is a diagram illustrating an example of luminance values inside and outside a cell in certain image data. In FIG. 10, a cell 80 having 5 pixels in the horizontal direction and 5 pixels in the vertical direction is shown.
In such a cell 80, the gradient angle of luminance has the following value according to the above equation (1).

FIG. 11 is a diagram illustrating an example of the angle of the luminance gradient of the cell. The numerical values in FIG. 11 indicate the brightness gradient angle (in degrees) at each pixel position.
Further, the magnitude of the luminance gradient at each pixel position of the cell 80 is a value as shown below by the above equation (2).

FIG. 12 is a diagram illustrating an example of the magnitude of the luminance gradient of a cell. The numerical values in FIG. 12 indicate the magnitude of the luminance gradient at each pixel position.
(Histogram calculation unit 54)
The histogram calculation unit 54 uses a luminance gradient angle as a data section in a cell of interest in image data divided for each cell having a predetermined size, and the integration of the number of pixels or the gradient size as a frequency. The histogram to be calculated is calculated. When the direction of the gradient direction is ignored, the gradient angle of the luminance is in the range of 0 degrees to 180 degrees. If the number of data sections is ND, the angle range of one section is 180 / ND degrees.

For example, as shown in FIG. 3, the histogram calculation unit 54 calculates a histogram with the magnitude of the corresponding gradient as the frequency for the gradient angle corresponding to each section.
The following histogram is calculated from the gradient angle and the magnitude of the brightness of the cell 80 as shown in FIGS.

FIG. 13 is a diagram illustrating an example of a histogram. The horizontal axis indicates the gradient angle (degrees), and the vertical axis indicates the frequency.
The histogram shown in FIG. 13 is indicated by the frequency for each data interval of 20 degrees. In the example of this histogram, since the frequency in the data section having the gradient angle of 0 to 20 degrees is the maximum, it can be seen that the cell 80 includes the most vertical edges.

(Shift section 55)
The shift unit 55 changes the histogram by shifting the data section of the histogram calculated by the histogram calculation unit 54 a predetermined number of times. When the number of data sections in the histogram is ND, the number of shifts NS is given as an integer value that satisfies 0 ≦ NS <ND.

For example, as described above, when the histogram shown in FIG. 3B is shifted four times, a histogram as shown in the lower diagram of FIG. 4 is obtained.
(Similarity calculation unit 56)
The similarity calculation unit 56 calculates the similarity between the histogram of the cell whose data section is shifted by the shift unit 55 and the histogram of another cell calculated by the histogram calculation unit 54. Here, the overlapping ratio of histograms shown below is calculated as the similarity.

  The histogram overlap rate S is expressed by the following equation.

Here, H _i is the histogram of the i th cell, h _k ⁽ⁱ⁾ is the frequency of the k th data interval of the histogram of the i th cell, H _j is the histogram of the j th cell, and h _k ^(j) is The frequency of the kth data section of the jth histogram, ND, indicates the total number of data sections.

(Learning part 57)
The learning unit 57 calculates a class classification parameter for discriminating between a positive example image (an image with an object to be detected) and a negative example image (an image without an object to be detected) using the histogram overlap rate as a feature amount. To do. As a class classification learning method, a learning method such as Real Adaboost, support vector machine, or decision tree is applicable.

(Storage unit 58)
The storage unit 58 stores the class classification parameter calculated by the learning unit 57. The class classification parameters stored in the storage unit 58 are referred to when the object detection unit 59 detects an object.

(Object detection unit 59)
The object detection unit 59 uses the similarity calculated by the similarity calculation unit 56 as a feature amount and refers to the class classification parameter stored in the storage unit 58 to detect an object to be detected from the image data.

Hereinafter, the operation of the object detection device 50 will be described. First, processing during learning will be described.
(Learning process)
In the following description, a case where the learning unit 57 performs learning using Real Adaboost as the class classification learning method will be described as an example.

FIG. 14 is a flowchart illustrating an exemplary flow of the learning process.
In order for the learning unit 57 to calculate the classification parameter, for example, hundreds to thousands of image data of positive examples and negative examples are input to the image acquisition unit 51 as training image data, for example. The positive image data is image data obtained by cutting out an area where only the object to be detected is shown. The negative image data is image data obtained by cutting out an image region in which the object to be detected is not shown, and has the same image size as the positive image data.

For example, the image acquisition unit 51 first initializes the number m of the training image data (m = 1) (step S10), and acquires the mth training image data (step S11).
Then, the histogram calculation unit 54 divides the training image data into cell units, and calculates a luminance gradient histogram for all cells based on the luminance gradient calculated by the luminance gradient calculation unit 53 (step S12).

FIG. 15 is a diagram illustrating a cell setting example.
For example, if the image size of the training image data 90 is 45 pixels wide and 45 pixels long and the cell size is 5 × 5 pixels, the total number of cells is 81.

  For example, the histogram calculation unit 54 assigns a cell number ncell to each cell, starting from the upper left cell of the training image data 90. The cell number satisfies the relationship of 0 ≦ ncell <Ncell. Here, Ncell indicates the total number of cells, and in the example of FIG. 15, Ncell = 81.

  The similarity calculation unit 56 selects two cells from these cells and calculates the similarity of the histogram. Note that the case where two cells are the same cell is also allowed. For example, the number of combinations (number of overlapping combinations) for selecting two from 81 cells is 81H2 = 3321.

  Hereinafter, the two cells selected by the similarity calculation unit 56 are referred to as a first cell and a second cell. In addition, the cell number of the first cell is ncell1, and the cell number of the second cell is ncell2.

  The similarity calculation unit 56 initializes the cell number ncell1 of the first cell and the element number k of the feature vector (step S13). Here, the similarity calculation unit 56 sets ncell1 = 0 and k = 0.

  Moreover, the similarity calculation unit 56 initializes the cell number ncell2 of the second cell (step S14). Specifically, the similarity calculation unit 56 sets ncell2 = ncell1. That is, in the initial state, the cell of ncell1 = ncell2 = 0 is selected as the first cell and the second cell.

Subsequently, the shift unit 55 initializes the number of shifts NS to NS = 0 (step S15), and shifts the data section of the histogram of the second cell NS times (step S16).
After that, the similarity calculation unit 56 calculates the overlap rate S (H _ncell1 , H _ncell2 ) between the histogram of the first cell and the histogram of the second cell shifted NS times from the equation (3). Is a feature vector v _m (k) (step S17). H _ncell1 represents a histogram of the first cell, and H _ncell2 represents a histogram of the second cell shifted NS times.

  Thereafter, the shift unit 55 updates the number of shifts NS to NS = NS + 1 (step S18). Then, the shift unit 55 determines whether or not the number of shifts NS has reached the number of data sections ND of the histogram (step S19). If NS <ND, the process from step S16 is repeated, and if NS = ND, the process of step S20 is performed.

  For example, in the case of a histogram as shown in FIG. 3 or FIG. In this case, the shift unit 55 generates nine histograms for the second cell.

  When the number of shifts NS reaches the number of data sections ND, the similarity calculation unit 56 updates the second cell as ncell2 = ncell2 + 1 (step S20). Then, the similarity calculation unit 56 determines whether ncell2 has reached the total number Ncell of cells (step S21). When ncell2 <Ncell, the process from step S15 is repeated, and when ncell2 = Ncell, the process of step S22 is performed.

  When the cell number ncell2 of the second cell reaches the total number Ncell of cells, the similarity calculation unit 56 updates the first cell as ncell1 = ncell1 + 1 (step S22). Then, the similarity calculation unit 56 determines whether ncell1 has reached the total number of cells Ncell (step S23). When ncell1 <Ncell, the process from step S14 is repeated, and when ncell1 = Ncell, the process of step S24 is performed.

  In the case of the example shown in FIG. 15, the dimension number Dim of the feature vector at the time when ncell1 = Ncell (= 81) is 3321 (the first cell and the number of data sections ND of the histogram is 9). The number of overlapping combinations of the second cells) × 9 = 29889. That is, a feature quantity vector having a dimension number of 29889 is calculated for each number of pieces of training image data.

  Next, the image acquisition unit 51 updates the training image data number m to m + 1 (step S24). And the image acquisition part 51 determines whether the number m of training image data is below the total number M of training image data (step S25). If m ≦ M, the process from step S11 is repeated. When the number m of training image data exceeds the total number M of training image data, the process of step S26 is performed.

The feature vector calculated in the process up to step S25 is sent to the learning unit 57, and the following process is performed.
Learning unit 57, feature vector v _m the element number k of (k) (1 ≦ k ≦ Dim) fixed, the maximum value of the feature vector v _m (k) when changing the number m of the training image data V _MAX (k) is calculated (step S26).

The maximum value V _MAX (k) is obtained by V _MAX (k) = MAX (v ₁ (k), v ₂ (k),..., V _M (k)). In this equation, MAX () is a function that outputs the maximum value.
Next, the learning unit 57 uses the maximum value V _MAX (k) to normalize the feature quantity vector and obtain a normalized feature quantity vector (step S27).

Normalized feature vector is expressed as _{[v m (1) / v} MAX (1) ... v m (k) / v MAX (k) ... v m (Dim) / v MAX (Dim)].
Thereafter, the learning unit 57 quantizes the normalized feature vector using the number of data sections P of the histogram to obtain a learning sample x _m (step S28).

The learning sample x _m is expressed by the following formula.

Here, INT () indicates a function that rounds off the decimal part and outputs only an integer.
According to the equation (4), quantization is performed such that x _m (k) = j and 0 ≦ j <P.
Furthermore, the learning unit 57, the training image data of the number m is the case of a positive example, y _m = + 1, in the case of the negative sample determines the label y _m to y _m = -1. Then, the learning unit 57, a learning sample x _m, the data sets for learning by combining the label _{y m {(x 1, y} 1) ... (x M, y M)} to produce a (step S29).

Thereafter, the learning unit 57 performs updating of the weight values of each training image data and selection processing of weak classifiers having high separation performance between positive examples and negative examples (step S30).
FIG. 16 is a flowchart illustrating an example of weight value update and weak classifier selection processing.

The weight value applied to the training image data in the t-th learning is D _t (m), 1 <m ≦ M. The weight value D _t (m) satisfies the following relationship.

First, the learning unit 57 initializes the weight value D ₁ (m) of the training image data such that D ₁ (m) = 1 / M using the total number M of training image data (step S31). Note that the learning unit 57 also sets m = 0, k = 0, t = 0 for the training image data number m, the feature vector element number (hereinafter also referred to as weak classifier number) k, and the learning number t. And initialize.

After that, the learning unit 57 replaces the learning sample x _m (k) with the integer j obtained in the process of step S28 of FIG. 14 (step S32), and 1 element P (number of data sections of the histogram) is 1 The weight value D _t (m) is added to the dimension array W ± [j] (step S33). When the training image data is a positive example (y _m = + 1), the learning unit 57 calculates W ₊ [j] = W ₊ [j] + D _t (m), and the negative example (y _m = −1). In this case, W ₋ [j] = W ₋ [j] + D _t (m) is calculated.

  Next, after updating the training image data number m to m = m + 1 (step S34), the learning unit 57 determines whether the training image data number m is greater than the total number M of training image data ( Step S35). If m> M, the process of step S36 is performed. If m ≦ M, the process from step S32 is repeated.

  When m> M, the learning unit 57 calculates an index Z for selecting a weak classifier using the following equation (step S36).

  The learning unit 57 updates the weak classifier number k to k = k + 1, sets the training image data number m to m = 0 (step S37), and determines whether or not all weak classifiers have been investigated. Therefore, it is determined whether or not k> Dim (step S38). If k> Dim, the process of step S39 is performed. If k ≦ Dim, the process from step S32 is repeated.

When k> Dim, the learning unit 57 selects the weak classifier that gives the smallest index Z, and the weak classifier number k ^{(t) at} that time and the one-dimensional array W ₊ ^(t) [j], W _- a ^(t) [j], as classification parameters of the learning number t, stored in the storage unit 58 (step S39).

  Thereafter, the learning unit 57 updates the weight value with the following equation for the training image data used in the next learning (step S40).

Here, _ht (x) indicates the output of the weak classifier at the learning number t, and is expressed by the following equation.

Here, ε = 10 ⁻⁷ .
After that, the learning unit 57 updates the learning number t to t = t + 1, sets the training image data number m to m = 0, and sets the feature vector element number k to k = 0 (step S41), and then the number of learnings. It is determined whether or not has reached a predetermined number of times (T times) (step S42). If t <T, the process from step S32 is repeated, and if t ≧ T, the learning unit 57 ends the learning process.

Through the learning process as described above, for example, the classification parameter is stored in the storage unit 58 as follows.
FIG. 17 is a diagram illustrating a storage example of class classification parameters.

  Here, a weak classifier number and a one-dimensional array of positive examples and a one-dimensional array of negative examples each having P number of elements are stored as class classification parameters for learning numbers t ranging from 0 to T-1. An example is shown.

The object detection unit 59 performs object detection using a class classification parameter as shown in FIG.
Details of the object detection process will be described below.

(Object detection processing)
When performing the object detection process, the detection window scanning unit 52 adds the detection window region 61 as shown in FIG. 8 to the unknown image data (for example, image size = 640 × 480 pixels) input from the image acquisition unit 51. Set. The detection window area 61 has, for example, the same image size as the training image data, and is 45 × 45 pixels in the example at the time of the learning process described above.

  Thereafter, the image data in the detection window region 61 is processed by the luminance gradient calculation unit 53, the histogram calculation unit 54, the shift unit 55, and the similarity calculation unit 56 by the same processing as that during learning, and the feature vector is calculated. Is done.

When the number of scans by the detection window scanning unit 52 is the mth, the object detection unit 59 is the same as the above-described steps S26 to S28 performed by the learning unit 57 from the feature amount vector calculated by the similarity calculation unit 56. Then, the quantized feature vector x _m (k) is calculated.

Thereafter, the object detection unit 59 performs the following process.
FIG. 18 is a flowchart illustrating an example of processing for determining the presence or absence of an object.
First, the object detection unit 59 initializes the learning number t to t = 0 and the strong classifier output H to H = 0 (step S50). In addition, the object detection unit 59 reads the weak classifier number k ^(t) corresponding to the learning number t from the storage unit 58 (step S51). Thereafter, the object detection unit 59 replaces the quantized feature quantity vector x _m (k ^(t) ) with an integer j (step S52).

Further, the object detection unit 59 reads the positive and negative one-dimensional arrays W ₊ ^(t) [j] and W ₋ ^(t) [j] from the storage unit 58 (step S53). Then, the object detection unit 59, using equation (8), to calculate the output h _t of the weak classifier (step S54). Then, the object detection unit 59, by performing the calculation of H = H + h _t, adds the outputs of the weak discriminators to the strong classifier (step S55).

  Thereafter, the object detection unit 59 updates the learning number t to t = t + 1 (step S56), and determines whether or not the learning number t has reached the learning count T (step S57). If t ≧ T, the determination process of step S58 is performed, and if t <T, the process from step S51 is repeated.

  When t ≧ T, the object detection unit 59 determines whether or not the strong discriminator output H is equal to or greater than a predetermined threshold value Hth (step S58). If H ≧ Hth, the object detection unit 59 outputs a detection result that there is a target object in the detection window region 61 (step S59). If H <Hth, the object detection unit 59 The detection result that there is no object is output (step S60).

  In the object detection process as described above, the object detection in the image is performed using the similarity between the histogram of the luminance gradient in a certain cell in the image data and the histogram obtained by shifting the data section in the other cell as a feature amount. For this reason, even if the luminance gradient histogram distribution changes due to a change in gloss or the like, the feature quantity for discriminating the object to be detected from other parts is unlikely to change. As a result, object detection errors can be reduced.

  In addition, since the similarity between the histogram of the brightness gradient in a certain cell and the histogram obtained by shifting the data section in another cell is calculated and used as a feature amount in the detection window, the connected state of edges and the distance between distant edges Relevance can be expressed as a feature quantity, and object detection taking these into account becomes possible.

  In the above description, the similarity calculation unit 56 calculates the histogram overlap rate using Expression (3), but it may be calculated using the following expression.

Here, H _i is the histogram of the i th cell, h _k ⁽ⁱ⁾ is the frequency of the k th data interval of the histogram of the i th cell, H _j is the histogram of the j th cell, and h _k ^(j) is The frequency of the kth data section of the jth histogram, ND, indicates the total number of data sections. Σ is expressed by the following equation.

max () is a function that outputs the maximum value in parentheses.
As shown in Expression (9), the overlap ratio is normalized to 0 to 1 by normalizing the overlap ratio of the histogram using σ. Thereby, the normalization processing of the feature amount vector as performed in the above-described steps S26 and S27 in the learning unit 57 becomes unnecessary, and the calculation amount can be reduced.

Incidentally, the object detection devices 1 and 50 as described above may be image processing boards or the like, or may be the following computers.
FIG. 19 is a diagram illustrating an example of hardware of a computer used in this embodiment. The computer 100 is entirely controlled by a CPU (Central Processing Unit) 101. A RAM (Random Access Memory) 102 and a plurality of peripheral devices are connected to the CPU 101 via a bus 108.

  The RAM 102 is used as a main storage device of the computer 100. The RAM 102 temporarily stores at least part of an OS (Operating System) program and application programs to be executed by the CPU 101. The RAM 102 stores various data used for processing by the CPU 101.

  Peripheral devices connected to the bus 108 include an HDD 103, a graphic processing device 104, an input interface 105, an optical drive device 106, and a communication interface 107.

  The HDD 103 magnetically writes and reads data to and from the built-in disk. The HDD 103 is used as a secondary storage device of the computer 100. The HDD 103 stores an OS program, application programs, and various data. Note that a semiconductor storage device such as a flash memory can also be used as the secondary storage device.

  A monitor 104 a is connected to the graphic processing device 104. The graphic processing device 104 displays an image on the screen of the monitor 104a in accordance with a command from the CPU 101. Examples of the monitor 104a include a display device using a CRT (Cathode Ray Tube) and a liquid crystal display device.

  A keyboard 105 a and a mouse 105 b are connected to the input interface 105. The input interface 105 transmits signals sent from the keyboard 105a and the mouse 105b to the CPU 101. Note that the mouse 105b is an example of a pointing device, and other pointing devices can also be used. Examples of other pointing devices include a touch panel, a tablet, a touch pad, and a trackball.

  The optical drive device 106 reads data recorded on the optical disc 106a using laser light or the like. The optical disk 106a is a portable recording medium on which data is recorded so that it can be read by reflection of light. The optical disk 106a includes a DVD, a DVD-RAM, a CD-ROM (Compact Disc Read Only Memory), a CD-R (Recordable) / RW (ReWritable), and the like.

  The communication interface 107 is connected to the network 107a. The communication interface 107 transmits / receives data to / from other computers or communication devices via the network 107a.

The processing functions of the present embodiment can be realized by the hardware as described above.
As described above, the processing functions of the present embodiment can be realized by a computer. In that case, a program describing the processing contents of the functions that the object detection apparatuses 1 and 50 should have is provided. By executing the program on a computer, the above processing functions are realized on the computer. The program describing the processing contents can be recorded on a computer-readable recording medium. Examples of the computer-readable recording medium include a magnetic storage device, an optical disk, a magneto-optical recording medium, and a semiconductor memory. Magnetic storage devices include HDDs, flexible disks (FD), and magnetic tapes. Optical discs include DVD, DVD-RAM, CD-ROM / RW, and the like. Magneto-optical recording media include MO (Magneto-Optical disk).

  When distributing the program, for example, a portable recording medium such as a DVD or a CD-ROM in which the program is recorded is sold. It is also possible to store the program in a storage device of a server computer and transfer the program from the server computer to another computer via a network.

  The computer that executes the program stores, for example, the program recorded on the portable recording medium or the program transferred from the server computer in its own storage device. Then, the computer reads the program from its own storage device and executes processing according to the program. The computer can also read the program directly from the portable recording medium and execute processing according to the program. In addition, each time a program is transferred from a server computer connected via a network, the computer can sequentially execute processing according to the received program.

  In addition, at least a part of the above processing functions can be realized by an electronic circuit such as a DSP (Digital Signal Processor), an ASIC (Application Specific Integrated Circuit), or a PLD (Programmable Logic Device).

  As described above, one aspect of the object detection device, the object detection method, and the program according to the present invention has been described based on the embodiments, but these are merely examples, and the present invention is not limited to the above description.

DESCRIPTION OF SYMBOLS 1 Object detection apparatus 2 Image acquisition part 3 Luminance gradient calculation part 4 Histogram calculation part 5 Shift part 6 Similarity calculation part 7 Storage part 8 Object detection part

Claims (5)

A luminance gradient calculating unit for calculating a luminance gradient indicating a rate of change in luminance at a pixel in the image data;
A histogram calculation unit that calculates a histogram having a data interval for each predetermined angle range of the luminance gradient for the image area of the image data divided for each image area of a predetermined size;
A shift unit that changes the histogram by shifting the data section of the histogram;
A similarity calculation unit that calculates a similarity between the histogram of the image region in which the data section is shifted by the shift unit and the histogram of the other image region calculated by the histogram calculation unit;
An object detection unit that detects an object to be detected from the image data using the similarity as a feature amount;
An object detection apparatus comprising:   The object detection apparatus according to claim 1, wherein the similarity is a degree of overlap between the histogram of the image area where the data section is shifted and the histogram of another image area.   The similarity calculation unit is configured to determine the degree of overlap as a sum of a larger value in each data section of the histogram of the image area where the data section is shifted and the histogram of the other image area. The object detection device according to claim 2, wherein the object detection device is normalized by. Calculate a luminance gradient indicating the rate of change in luminance at the pixels in the image data,
For the image area of the image data divided for each image area of a predetermined size, a histogram having a data interval for each predetermined angle range of the luminance gradient is calculated,
Changing the histogram by shifting the data section of the histogram;
Calculating a similarity between the histogram of the image region in which the data section is shifted and the histogram of the other image region;
An object detection method, wherein an object to be detected is detected from the image data using the similarity as a feature amount. On the computer,
Calculate a luminance gradient indicating the rate of change in luminance at the pixels in the image data,
For the image area of the image data divided for each image area of a predetermined size, a histogram having a data interval for each predetermined angle range of the luminance gradient is calculated,
Changing the histogram by shifting the data section of the histogram;
Calculating a similarity between the histogram of the image region in which the data section is shifted and the histogram of the other image region;
A program for executing processing for detecting an object to be detected from the image data, using the similarity as a feature amount.

Images