JP2005209155A - Motion detector and motion detection method - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To easily calculate travel speed, travel direction and position of a featured area by detecting motion of the feature area such as a contour of an object, etc. in continuous images. <P>SOLUTION: A featured area is standardized by extracting the feature area such as an edge from an image picked up by an image pickup part 1 by a feature extraction part 2 and by standardizing edge width by a featured standardization part 3. Count values of a location where the feature area is detected are counted up in a voting part 4, inclination of the counted up count values is calculated in a travel speed detection part 5 and the travel direction, the travel speed and the location of the feature area are calculated based on the inclination of the count values. <P>COPYRIGHT: (C)2005,JPO&NCIPI

Description

  The present invention relates to a motion detection apparatus and a motion detection method for detecting a target object included in continuous images and detecting a moving speed, a moving direction, a position, and the like of the target object.

  In recent years, for example, when mounted on a vehicle, various peripheral monitoring devices have been proposed and put into practical use for the purpose of recognizing a forward vehicle or a road ahead of the vehicle when the vehicle is traveling (for example, the following) (A vehicle periphery monitoring device disclosed in Japanese Patent Laid-Open No. 2003-67752 of Patent Document 1). Such a peripheral monitoring device recognizes changes in the periphery by performing image processing on a moving image captured by a camera. As one method of moving image processing technology, optical flow (camera observation) is used. An analysis method based on a velocity field of each point on an image generated by relative movement between a person and a moving object is known.

  As a conventional optical flow detection method for detecting a moving speed component from an image, a correlation method for determining a motion vector by determining a corresponding point between frames, a spatial and temporal gradient of brightness of each point, or the like. A method such as a gradient method using a relationship between the two is generally used.

First, the correlation method is a method that uses block-to-block matching that divides an image into small regions and finds a region similar to this from successive images. In the correlation method, in a region I (t + 1, x + u, y + v) corresponding to a region continuous with the region of interest I (x, y, t) at time t,
ΣΣ {I (x, y, t) −I (x + u, y + v, t + 1)} ²
ΣΣ | I (x, y, t) -I (x + u, y + v, t + 1) |
(U, v) that minimizes. This (u, v) is the amount of movement of the object between consecutive frames.

  The gradient method then uses the relationship between the spatial and temporal gradients of brightness at each pixel. Assuming that the brightness at the coordinates (x, y) at a certain time t is f (x, y, t), the point (x, y) on the pattern after the time δt moved by δx, δy in the x and y directions. Assume that the brightness does not change. At this time, the following equation holds.

f (x, y, t) = f (x + δx, y + δy, t + δt)
Taylor expansion of the above equation and δt → 0,
δf / δx · dx / dt + δf / δy · dy / dt + δf / δt = 0
Get. Since the motion vector (dx / dt, dy / dt) cannot be determined only by the gradient condition of the motion vector, a motion vector is calculated by further adding a constraint condition to the gradient condition. As a constraint condition to be added, a condition in which a motion vector changes smoothly in space is generally used.

As another technique for detecting the movement of an object, a moving image processing apparatus described in Patent Document 2 below is known. In this moving image processing apparatus, a time series image of a scene is captured by an image input unit, and a flow vector of each pixel between an appropriate reference image image having a moving image sequence and a current image is detected by a flow detection unit. Then, the flow verification unit evaluates the reliability of the flow vector for each pixel, and when the reliability is high, measures the movement time when the pixel movement amount reaches a certain amount in the speed determination unit for the flow vector, The moving speed is output.
Japanese Patent Laid-Open No. 2003-67752 JP 2001-209804 A

  However, in the above-described correlation method, the processing procedure is relatively simple. However, since a region for calculating a correlation value is provided around the region of interest, and a correlation value of in-region luminance is calculated between frames, a template is used. There is a problem that block matching processing such as matching is required, the calculation amount is enormous, and the processing time becomes long. That is, in the technique using block matching, in order to calculate detailed position information by using a pixel group of sub-pixels or less, pixel group data that is a registered template is centered on the target position and its surroundings. Since a plurality of correlation values calculated while being shifted by a predetermined amount are interpolated with a high-order curve and the motion is estimated by calculating the position where the extreme value is given, the calculation amount becomes enormous.

  In addition, the gradient method has a problem that it is weak against changes in luminance of the image and the processing procedure is complicated.

  The present invention has been made to solve such a conventional problem. The object of the present invention is to move from within an image by using a relatively simple processing procedure with a smaller amount of data. An object of the present invention is to provide a motion detection apparatus and a motion detection method that can process detection of velocity components at higher speed.

  In order to solve the above-described object, the present invention provides an imaging unit that captures a peripheral image, a feature extraction unit that extracts a feature region of a region for speed calculation from an image captured by the imaging unit, and the feature extraction unit A feature region normalization unit that normalizes the size of the feature region extracted by the feature region, and a feature region that has not been detected based on the feature region information normalized by the feature region normalization unit. A mask generating means for setting as a mask area, a voting means for resetting the count value of the position of the mask area and counting up the count value of the position of the non-mask area not masked by the mask area, and a continuous frame A moving speed detecting means for calculating the moving direction and moving speed of the feature area from the gradient of the count value accumulated over And it features.

  In the motion detection apparatus according to the present invention, a feature region is extracted by the feature extraction unit (feature extraction step) from the image captured by the imaging unit (imaging step), and the feature region is extracted from the feature region normalization unit (feature region). Normalization step), and a mask generation unit (mask generation step) sets a region where no feature region is detected based on the standardized feature region information as a mask region of the feature region. Step) resets the count value of the position of the mask area, and counts up the count value of the position of the non-mask area that is not masked by the mask area, and the movement speed detection means (movement speed detection step) Based on the slope of the count value accumulated over successive frames, the moving direction and speed of the feature area Therefore, the count value is equivalent to the time measurement when the feature region is observed, and the number of frames required for the feature region to move a predetermined number of pixels is measured from the gradient of the count value, The edge moving direction, speed, and acceleration can be easily obtained without using block matching such as template matching. In addition, since the calculation process is performed only based on the position of the non-mask area, it can be processed with a relatively simple procedure for a smaller amount of data, and the detection of the moving speed component from the image can be performed at a higher speed. it can.

  Hereinafter, first to third embodiments of the present invention will be described with reference to the drawings.

[First Embodiment]
The present invention is applied to, for example, a motion detection apparatus configured as shown in FIG.

[Configuration of Motion Detection Device]
This motion detection device is mounted on a vehicle, for example, and a feature indicating an object when a feature point or region (hereinafter referred to as a feature region) of the object is given in a continuous image obtained by imaging around the vehicle. It detects the movement of the area.

  As shown in FIG. 1, the motion detection apparatus includes an imaging unit 1, a feature extraction unit 2, a feature normalization unit 3, a voting unit 4, and a moving speed detection unit 5.

  The imaging unit 1 is configured by an imaging device such as a CMOS (Complementary Metal-Oxide Semiconductor) image sensor or a CCD (Charge Coupled Device) image sensor. The imaging unit 1 captures an image, extracts the image as a frame at regular intervals, and performs continuous processing. I do. At this time, a frame image is acquired at a frame rate sufficiently higher than the phenomenon (movement speed) of the object to be observed.

  The feature extraction unit 2 inputs image data including an object to be a motion detection target from the imaging unit 1, and performs image processing on the image data to extract a feature amount or a feature region of the object. When extracting this feature point or feature region, the feature extraction unit 2 uses an edge detection filter, a corner edge detection filter, or a feature vector calculation filter. The processing content of the feature extraction unit 2 will be described later.

  The feature normalization unit 3 standardizes the feature size by defining a representative point from the feature region extracted by the feature extraction unit 2 and forming a predetermined pattern based on the representative point. That is, the feature normalization unit 3 sets a standardized pattern at a predetermined position including the feature region obtained by the feature extraction unit 2. Thereby, the feature normalization unit 3 normalizes the feature region detected by the feature extraction unit 2 to a predetermined size.

  The voting unit 4 counts up the count value of the memory address corresponding to the position (non-mask area) where the standardized pattern set by the feature standardizing unit 3 is detected, or the standardized pattern is detected. A voting process, which is one of the processes for resetting the count value of the memory address corresponding to the position (mask area) that did not exist, is performed. As a result, the voting unit 4 stores information on how many frames the extracted feature has been detected at the same position in the frame currently being processed.

  The moving speed detection unit 5 calculates the moving speed and moving direction of the extracted feature region based on the count value stored in the voting unit 14 and the gradient over the plurality of positions of the count value.

  As shown in the schematic diagram of the process in FIG. 2, the motion detection apparatus configured as described above first acquires an image P <b> 1 including the other vehicle 11, which is an object, by the imaging unit 1. Then, a feature image P2 obtained by extracting the feature region 12 indicating the other vehicle 11 from the image P1 is created, the feature region 12 is normalized, and a normalized pattern 13 is set. At this time, the feature normalization unit 3 normalizes the size of the feature region 12, thereby setting a normalized pattern 13 normalized at a position including the feature region 12 in the image P3. In the motion detection device, the voting unit 4 indicates the number of detections of the same standardized pattern 13 by operating the imaging unit 1, the feature extraction unit 2, and the feature normalization unit 3 over successive frames. By counting the count value, a plurality of count values 14 included in the same image position are obtained.

  Next, feature extraction processing and feature normalization processing in the motion detection apparatus described above will be described.

(1) When using an edge detection filter First, the case where the feature extraction part 2 detects a feature using an edge detection filter is demonstrated.

  When the edge information is used as a feature, the feature extraction unit 2 uses a filter that extracts an edge that is a complex portion or a portion whose luminance changes abruptly from a grayscale image, such as a Sobel filter or a Prewitt filter. As a result, the feature extraction unit 2 creates an edge image composed of pixels recognized as edges that are features and pixels that are not recognized as edges.

  The feature normalization unit 3 to which edge information is sent from the feature extraction unit 2 includes a binarization unit 21, a thinning unit 22, an expansion unit 23, and a peak position detection unit 24, as shown in FIG. Composed. Such a feature normalization unit 3 normalizes the edge width as a process for normalizing the size of the feature region.

  That is, when the feature normalization unit 3 acquires the edge image from the feature extraction unit 2 by the binarization unit 21, the binarization is performed by converting the edge image into a binarized image having values “0” and “1”. The thinning process is repeatedly performed by the thinning unit 22 on the edge whose binarized value is “1” (active) until the edge width reaches the predetermined number of pixels, and all detected detected values are processed. The edge width is unified to a predetermined number of pixels (one pixel). The expansion unit 23 expands an edge whose edge width is unified (to 1 pixel) to a predetermined number of pixels (for example, 3 pixels) as necessary.

  By these binarization processing, thinning processing, and expansion processing, the edge image extracted by the feature extraction unit 2 is generated as a binary normalized edge image in which the edge width is unified to a predetermined value (3 pixels). Will be. In order to unify the edge width, the peak position detection unit 24 of the feature normalization unit 3 detects the edge peak position of the edge image output from the feature extraction unit 2, and a width of a predetermined number of pixels at the edge peak position. Alternatively, the edge width may be normalized by generating a binary image having the.

  Thereby, the feature normalization unit 3 can normalize the size of the feature region by performing processing by the binarization unit 21, the thinning unit 22, and the expansion unit 23. Information indicating the position of the feature area in the image is sent to the voting unit 4.

(2) When using a corner edge detection filter Next, the case where the feature extraction unit 2 detects a feature using a corner edge detection filter will be described.

  Here, the edge component in the image usually appears at the image position corresponding to the contour of the object or a part of the complex part. On the other hand, the feature extraction unit 2 uses a corner edge detection filter to detect a corner where the contour is bent from the contour of the object.

  As this corner edge detection filter, for example, a SUSAN operator can be used. By using this SUSAN operator, the feature extraction unit 2 detects a pixel having a density value close to the density value of the target pixel from among pixels included in a predetermined area centered on the target pixel, and detects the detected pixel. A corner edge of the object is detected based on the ratio of the image to the predetermined area. As a result, the feature extraction unit 2 detects the corner edges (a), (b), and (c) of the target object as the feature regions (points) of the input image P11 including the target object, as shown in FIG. To do.

  On the other hand, the feature normalization unit 3 sets a normalization pattern for measuring the movement of the corner edge with reference to the position of the corner edge detected by the feature extraction unit 2. At this time, the feature normalization unit 3 receives the input image P11 in FIG. 4 and has the same size at positions corresponding to the positions of the corner edges (a) to (c) detected by the feature extraction unit 2. Normalization patterns (a ′) to (c ′) consisting of are set. As a result, normalization can be performed so that the size of the feature region is the same, and information indicating the position in the image of the normalized feature region is sent to the voting unit 4.

(3) Case of Using Feature Vector Calculation Filter Next, a case will be described in which features are detected by the feature extraction unit 2 using the feature vector calculation filter.

  When the feature extraction unit 2 generates a feature vector indicating the feature of the object from the input image, for example, the feature extraction unit 2 performs an operation using a higher-order local autocorrelation function using the feature of the edge. In this way, by extracting edge features using a higher-order local autocorrelation function, invariant feature extraction can be performed with respect to the position of an object (object) included in the input image, and features with a low dimensionality of feature vectors Can be obtained. Note that the feature vector is not limited to the higher-order local autocorrelation function, and luminance information, pattern information, and the like may be used.

  Then, the feature normalization unit 3 sets a normalization pattern based on the position of the feature region in the image corresponding to the feature vector detection position. At this time, the feature normalization unit 3 performs the same processing as when the above-described edge detection filter is used or when the corner edge detection filter is used.

  Next, a process for obtaining the moving speed and direction of the target object by obtaining the moving speed and direction of the feature region by the motion detection apparatus described above will be described with reference to FIG. In this description, for the sake of simplicity, it is assumed that the feature region moves in the one-dimensional direction of only the x-axis, and the normalized pattern is set to 9 regions of 3 × 3 centering on the detected feature region. explain. In addition, the count up by the voting unit 4 is described as being incremented by +1.

  First, when an input image acquired at time t is input (step S1), feature extraction is performed by the feature extraction unit 2 using the input image (step S2), and a feature region indicated by shading in FIG. If one pixel in the example is detected (FIG. 6 (a-1)), binarization is performed by the feature normalization unit 3 and, for example, three pixels centered on the feature region by the thinning unit 22 and the expansion unit 23 A standardized pattern consisting of x3 pixels is set (FIG. 6 (b-1), step S3).

Then, the center position area position x-1 and x ₀ are adjacent the center, x + 1 of the vote value in the characteristic region (count value) is counted up (step S4), and as shown in FIG. 6 (c-1) at time t In the next frame at time t + 1, a feature region normalized to the region position x + 1 is observed, and the vote value at the region position is further counted up. Reset voting values at locations other than those observed.

  Next, at time t + m after time t, a feature region is detected by the feature extraction unit 2 (FIG. 6 (a-2)), and when a normalization pattern is set by the thinning unit 22 and the expansion unit 23 ( 6 (b-2)), when the characteristic region is shifted by one pixel in the positive direction of the X axis compared with the time t and observed at the position x + 1, the region position x, x + 1, where the normalization pattern is set The vote value of x + 2 is counted up, and the vote value at the position where the normalization pattern is not set, that is, the vote value at the previous area position x−1 is reset (FIG. 6 (c-3)).

  Similarly, at time t + n after time t + m, a feature region is detected by the feature extraction unit 2 (FIG. 6 (a-3)), and a standardized pattern is set by the thinning unit 22 and the expansion unit 23. (FIG. 6 (b-3)) When the characteristic region is shifted by one pixel in the positive direction of the X axis and observed at the position x + 2 compared to the time t + n, the region positions x + 1 and x + 2 where the normalized pattern is set. , X + 3, and the voting value at the position where the standardization pattern is not set, that is, the voting value at the previous region position x is reset (FIG. 6 (c-3)).

  In this process of repeating the feature extraction by the feature extraction unit 2 and the binarization unit 21, the normalization of the features by the thinning unit 22 and the expansion unit 23, and the voting process by the voting unit 4, the process of resetting the voting value (step As shown in FIG. 7, for example, when an edge is used as a feature, S5) is obtained as a normalized pattern in which the edge width is normalized at time t as shown in FIG. ), An area to be voted (non-mask area, white area in the figure) and an area not to be voted (mask area, black area in the figure) are obtained. Then, the vote value generated at time t-1 (FIG. 7C) is counted by the area where the normalized pattern is detected in FIG. 7B generated from the normalized pattern set at time t. The area where the normalized pattern is not detected is masked, that is, reset to a new vote value (FIG. 7F). Similarly, the normalized pattern is plotted for the feature area obtained at time t + 1. 7 (d) is obtained and counted up by defining the mask area and the non-mask area as shown in FIG. 7 (e), and the vote value in FIG. 7 (f) and the vote in FIG. 7 (e) are counted. A value is calculated.

  Next, the gradient of the count value is obtained by the moving speed detection unit 5 from the vote value (count value) obtained as shown in FIG. 6 (step S6). At this time, the moving speed detection unit 5 detects the inclination of the voting value in the moving direction using the voting value for three pixels, and the inclination is the reciprocal of the moving speed of the characteristic area. The speed is obtained (step S7). Further, the moving speed detection unit 5 obtains the position of the feature region by performing a calculation of pixel position x + (minimum vote value h / difference H between the minimum vote value and the maximum vote value) (step S7).

  The gradient of the count value is determined by the interval at which the region position of the normalized pattern is counted by the motion detection device. When the frame rate of the input image is sufficiently higher than the moving speed of the actual object as compared with the moving speed of the characteristic area extracted by the characteristic extracting unit 2, the characteristic area observed between successive frames or Normalization patterns always have overlapping areas. In this way, by counting up the voting value corresponding to the position where the feature region is observed, the voting value becomes equivalent to the time when the feature region is observed at the same position.

  On the other hand, when the feature region moves, the vote value at the position where the feature region is newly observed is 1, which is the smallest value among the observed feature regions. That is, the vote value in the direction in which the feature region moves is small, and the vote value in the direction opposite to the direction in which the feature region moves is large.

  Therefore, the gradient of the voting value is a value obtained by counting how many frames are continuously observed at the same position before the feature region moves, and the moving speed detection unit 5 determines the feature region based on the gradient of the voting value. The moving speed can be calculated.

Furthermore, when the frame rate of the input image is sufficiently higher than the actual moving speed of the target object, it can be assumed that the target is moving at a constant speed. In the example of FIG. Assuming that it is moving, since two frames are continuously observed at time t + 1, the region position x ₀ is 2 frames / {4 frames / 1 pixel} = 0.5 pixels at time t + 1. It can be calculated that the shift has occurred. Therefore, the moving speed detection unit 5 can obtain, for example, a relative position with respect to the host vehicle from the image position of the feature in the input image, and further, the object speed can be determined from the gradient of the count value of the feature in the input image of the target. The moving speed can be obtained.

[Effect of the first embodiment]
As described above in detail, according to the motion detection device according to the first embodiment to which the present invention is applied, by performing the process of obtaining the characteristics of the input image at a sufficiently high rate with respect to the moving speed of the object, The moving speed of the object can be detected with a simple process. That is, by standardizing the size of the feature, the count value of the position where the feature is detected is counted up, and the count value of the position where the feature is not detected is reset, so from the obtained count value The time when the feature is observed can be obtained, and the number of frames required for the feature to move a predetermined number of pixels can be easily measured from the change in the position of the feature, that is, the gradient of the count value. Therefore, according to this motion detection device, the position and moving speed of the object can be obtained from the gradient of the count value without using block matching such as template matching.

  Further, according to this motion detection device, the width of the edge extracted by the feature extraction unit 2 using the edge detection filter is unified by the feature normalization unit 3, and based on the information of the edge having the unified width, Since the voting unit 4 counts up the count value at the position where the edge is detected and resets the count value at the position where the edge is not detected, the time value at which the edge is observed is measured. The number of frames required for the edge to move by a predetermined number of pixels can be easily measured from the gradient of the count value calculated by the moving speed detection unit 5 as equivalent.

  Furthermore, according to this motion detection device, the size of the detection point or region is unified by the feature normalization unit 3 at the corner edge position extracted by the feature extraction unit 2 using the corner edge detection filter, Since the voting unit 4 counts up the count value of the position where the standardized pattern is detected and resets the count value of the position where the standardized pattern is not detected based on the region information whose length is unified, the count value Is equivalent to measuring the time when the standardized pattern is observed, and the number of frames required for the corner edge to move a predetermined number of pixels can be easily determined from the gradient of the count value calculated by the moving speed detector 5. It can be measured.

  Furthermore, according to this motion detection apparatus, the size of the detected position or region is unified by the feature normalization unit 3 at the position where the feature vector extracted by the feature extraction unit 2 is detected, and the size is reduced. Based on the unified area information, the voting unit 4 counts up the count value of the position where the standardized pattern is detected, and resets the count value of the position where the standardized pattern is not detected. Equivalent to measuring the time when the vector is observed, and easily measuring the number of frames required for the feature vector to move by a predetermined number of pixels from the gradient of the count value calculated by the moving speed detector 5 Can do.

  Furthermore, according to this motion detection device, a position representative of the feature is defined from the feature position feature region extracted by the feature extraction unit 2, and a known pattern is formed according to the extracted feature pattern. Then, based on the standardized pattern that is the formed known pattern, the voting unit 4 counts up the count value of the position where the standardized pattern is detected, and the count value of the position where the standardized pattern is not detected Therefore, the count value is equivalent to the time measurement when the standardized pattern is observed, and the feature moves a predetermined number of pixels from the gradient of the count value calculated by the moving speed detection unit 5. It is possible to easily measure the number of frames required for.

[Second Embodiment]
Next, a motion detection device according to the second embodiment will be described.

  FIG. 8 is a block diagram showing a configuration of an embodiment of a motion detection apparatus according to the present invention. As shown in the figure, this motion detection device includes an imaging unit 101 that captures an image of an object for which motion is detected, and an edge that extracts an edge from an image captured by the imaging unit 101 to generate an edge image. An extraction unit 102, an edge width normalization unit 103 that unifies edge widths of edges in the edge image generated by the edge extraction unit 102 into a predetermined width in the direction of edge movement, and an edge width normalization unit The edge count unit 104 counts up the value of the memory address corresponding to the detected position of the edge normalized in 103, and the edge moving speed and direction based on the count value counted up by the edge count unit 104 And a moving speed detector 105 for calculating the position.

  The imaging unit 101 is configured by an imaging element such as a CMOS or a CCD, captures an image, extracts the image as a frame at regular intervals, and performs continuous processing. At this time, frames are extracted at a frame rate sufficiently higher than the edge moving speed.

  The edge extraction unit 102 extracts an edge from an image captured by the imaging unit 101 by using a Sobel filter or the like to generate an edge image.

  The edge width normalization unit 103 standardizes edges by unifying the edge width of edges in the edge image generated by the edge extraction unit 102 to a predetermined width, for example, a predetermined number of pixels, in the edge moving direction. Turn into. Further, the edge width normalization unit 103 defines the edge width of the edge to be normalized according to the moving speed of the object.

  The edge count unit 104 counts up and accumulates the value of the memory address corresponding to the detected position of the edge normalized by the edge width normalizing unit 103, and the memory address corresponding to the position where the edge is not detected Reset the value of. As a result, information indicating how many frames of the standardized edge are detected at the same position is generated.

  The movement speed detection unit 105 calculates the movement speed, movement direction, and position of the extracted edge based on the count value counted up by the edge counting unit 104 and the inclination of the count value.

  Next, motion detection processing by the motion detection device according to the present embodiment will be described based on the flowchart of FIG.

  When the imaging unit 101 captures an image of an object whose motion is to be detected (S21), the imaging unit 101 extracts the image as a frame at regular intervals and provides it to the edge extraction unit 102 (S22). At this time, however, the frame is taken out so that the frame rate is sufficiently higher than the moving speed of the edge.

  The edge extraction unit 102 extracts an edge of the image using an SOBEL filter or the like from the image for each frame supplied from the imaging unit 101, and generates an edge image (S23).

  In the edge width normalization unit 103, the edge width of the edge in the edge image generated by the edge extraction unit 102 is standardized to a predetermined width in the edge moving direction. Here, the edge width normalization processing in the edge width normalization unit 103 will be described with reference to FIGS.

  FIG. 10 is a block diagram illustrating a detailed configuration of the edge width normalization unit 103. As shown in the figure, the edge width normalization unit 103 binarizes the edge image generated by the edge extraction unit 102 and the edge binarized by the binarization unit 131. It is composed of a thinning unit 132 for thinning the edge width to unify it to a predetermined number of pixels, and an expansion unit 133 for expanding the edge width of the edge thinned by the thinning unit 132 to a predetermined number of pixels. .

  In the edge width normalization unit 103 configured as described above, when the edge image is provided from the edge extraction unit 102, the binarization unit 131 performs binarization processing on the edge image (S24). In this binarization processing, a pixel at a position where an edge is detected is set to 1, and a pixel at a position where an edge is not detected is set to 0, thereby generating a binarized image as shown in FIG.

  Next, a thinning process is performed on the binarized image generated in this manner by the thinning unit 132 (S25). This thinning process is a process of reducing the edge width of the detected edge until a predetermined pixel width is reached. For example, in FIG. 11B, the edge width of the edge is thinned until the predetermined pixel width reaches 1 pixel. However, here, as an example, the case of thinning to one pixel has been described, but thinning may be made to other numbers of pixels. Then, the center position that is the center of the edge is set by thinning the edge until a predetermined pixel width is obtained.

  Next, the expansion unit 133 performs expansion processing for expanding the edge width of the thinned edge (S26). In this expansion process, the edge width is expanded from the center position set by thinning toward the edge moving direction, and the edge width is also expanded from the center position in the opposite direction to the edge moving direction. For example, in FIG. 11C, one pixel is expanded from the edge center position x0 to the edge movement direction (the positive direction of the x axis), and the edge movement direction from the edge center position x0 is opposite to the edge movement direction (of the x axis). 1 pixel is expanded in the negative direction), and the edge width is expanded to 3 pixels.

  By performing the thinning process and the expansion process in this way, the edge width of the edge generated by the edge extraction unit 102 is standardized to a predetermined width in the edge moving direction.

  Further, here, the edge width is unified by performing the thinning process and the expansion process, but the edge peak position in the edge image generated by the edge extraction unit 102 is detected, and a predetermined value is detected from the edge peak position. The edge width may be normalized by providing a width corresponding to the number of pixels in both the edge moving direction and the opposite direction. Furthermore, although the case where it expands to 3 pixels is demonstrated as an example here, you may expand to the number of other pixels.

  Next, the edge count unit 104 performs a count-up process on the edge whose edge width is standardized (S27). In this count-up process, the value of the memory address at the position where the edge is detected is counted up, and the value of the memory address at the position where the edge is not detected is initialized.

  Hereinafter, the count-up process in the edge count unit 104 will be described with reference to FIG. In the figure, for the sake of simplicity, the edge is described as moving in the positive direction of the x-axis, but the edge is also described in the same way when moving in the negative direction of the x-axis, the y-axis, or two-dimensionally. Can do.

  As shown in FIG. 11C, the edge has the center position of the edge at a position x0 in a certain frame, the position x0 + 1 of one pixel in the movement direction of the edge from the center position, and the opposite movement direction of the edge from the center position. In the direction, it is expanded to a position x0-1 of one pixel.

  In such a case, the edge count unit 104 increments the count values of the positions x0-1, x0, and x0 + 1 where the edge is detected by the count-up mask 141 one by one, and the count value of the position where the edge is not detected. Reset. For example, in FIG. 11D, since an edge is detected at positions x0-1, x0, x0 + 1 at time t, the count is incremented by 1 at each position, and the count value at position x0 + 1 is 1, position x0. Is 3, and the count value at position x0-1 is 5. Since the edge does not move at the subsequent time t + 1, the edge is detected at each of the positions x0-1, x0, and x0 + 1, and the edge count unit 104 has positions x0-1, x0-1, The count values of x0 and x0 + 1 are further incremented by one, the count value of position x0-1 is 2, the count value of position x0 is 4, and the count value of position x0 + 1 is 6.

  Further, at the subsequent time t + 2, the edge is shifted by one pixel in the positive direction of the x-axis, and the edge is detected at positions x0, x0 + 1, and x0 + 2. Accordingly, the count values at the positions x0, x0 + 1, and x0 + 2 where the edge is detected are counted up, and the count values at the position x0-1 where the edge is not detected are reset. As a result, the count value at position x0 + 2 is 1, the count value at position x0 + 1 is 3, and the count value at position x0 is 5, as shown in FIG. Further, the count value at the position x0-1 where no edge is detected is reset to zero.

  In this way, the edge count unit 104 counts up the count value at the position where the edge is detected, and resets the count value at the position where the edge is not detected. In FIG. 11, as the position for detecting the count value, the center position (x0) of the edge, the position of one pixel (x0 + 1) from the center position to the edge movement direction, and the direction from the center position to the edge movement direction are opposite. The count value is detected at three positions of the position (x0-1) of one pixel. If the slope of the count value described later is obtained, the number of counts can be counted as long as it is two or more with respect to the edge moving direction. The value may be detected.

  If the frame rate is set sufficiently higher than the speed at which the edge moves, the edge is detected a plurality of times at the same position between consecutive frames. For example, in the case of FIG. 11, two edges at time t and time t + 1 are detected at the position x0. Therefore, when the count value at the position where the edge is detected is counted up, the count value becomes equal to the time (number of frames) during which the edge is detected at that position. In particular, the smallest count value h of the edge count values represents how many frames the edge has been in the same position after moving.

  Next, the moving speed detector 105 calculates the moving speed, moving direction, and position of the edge. The moving speed detector 105 first calculates the inclination of the count value in the moving direction (S28), and calculates the moving direction, moving speed and position of the edge based on this inclination (S29).

Hereinafter, this calculation method will be described with reference to FIG. For example, in the case of FIG. 11E, the count values at positions x0-1, x0, and x0 + 1 are 6, 4, and 2, respectively. Therefore, by subtracting the count value 2 of x0 + 1 from the count value 6 of the position x0-1, the slope of the count value can be calculated as H = (6-2) / 2 = 2. this is,
H = {(time from the edge moving to the position x0-1 to the present)-(time after the edge has moved to the position x0 + 1)} / (2 pixels)
This means that the time (number of frames) required for the edge to pass through one pixel at the position x0 is calculated. Accordingly, the slope H of the count value determines how many frames it takes for the edge to move by one pixel, and the edge moving speed 1 / H can be calculated based on the slope H of the count value. .

  For example, in FIG. 11E, it takes 4 frames to move one pixel, so the edge moving speed can be calculated as 1/4 (pixel / frame). Similarly, since H = (5-1) / 1 = 4 in FIG. 11F, the edge moving speed becomes 1/4 (pixel / frame).

  Further, the edge moving direction can be determined by the magnitude of the count value. The count value at the position where the edge is moved and the edge is newly detected is 1, which is the smallest value among the count values at each position. Therefore, the count value in the direction in which the edge moves is small, and the count value in the direction opposite to the direction in which the edge moves is large, so that the edge moving direction can be determined.

  Furthermore, if the frame rate is set sufficiently higher than the speed at which the edge moves, it can be assumed that the detection target object moves at a constant speed. In addition, the smallest count value h among the count values at the current position represents the time when the edge is detected at the position, that is, how many frames the edge has moved to stay at the same position. ing.

As a result, when the center position of the edge is x0, the edge position = x0 + h / H
It can ask for. For example, in FIG. 11E, the edge speed is ¼ (pixel / frame), and the edge is detected at the same position for two consecutive frames at time t + 1. Therefore, the position of the edge at time t + 1 is It can be calculated that 2 (frame) × {1/4 (pixel / frame)} = 0.5 pixel is moved from the position x0.

  When the movement speed, movement direction, and position of the edge are calculated in this way, the movement speed detection unit 105 transmits the calculated movement speed to the edge width normalization unit 103, and receives the movement speed and performs edge width normalization. The unit 103 changes the edge width of the edge to be normalized based on the moving speed (S30). For example, in FIG. 11C, the edge width after the expansion process is 3 pixels, but when the received moving speed is high, the edge width is changed to be large, and when the moving speed is low Change to reduce the edge width.

  If the edge width to be normalized is changed in this way, the edge width can be normalized so that the edges overlap between successive frames according to the movement speed even if the movement speed of the detection target changes. This makes it possible to expand the range of movement speeds that can be detected.

  As described above, in the motion detection device and the motion detection method according to the present embodiment, the count value of the position where the edge is detected is counted up, and the moving speed and movement of the edge are based on the slope of the counted up count value. Since the direction is calculated, the moving speed and moving direction of the edge can be easily calculated without using block matching such as template matching.

  Further, since the moving speed calculation means 5 obtains the position of the edge based on the calculated moving speed of the edge and the smallest count value among the count values, the position of the edge in the sub-pixel order is further subdivided. Can be calculated.

  Furthermore, since the edge width normalization unit 103 changes the edge width to be standardized based on the movement speed calculated by the movement speed detection unit 105, even if the movement speed of the object changes, the changed movement speed is obtained. Accordingly, the edge width can be normalized so that the edges overlap between successive frames, and thereby the range of the moving speed of the object that can be detected can be expanded.

  As mentioned above, although the motion detection apparatus and method of the present invention have been described based on the illustrated embodiment, the present invention is not limited to this, and the configuration of each part is an arbitrary configuration having the same function. Can be replaced with something.

[Third Embodiment]
Next, a motion detection device according to the third embodiment will be described.

  FIG. 12 is a block diagram of a motion detection apparatus according to an embodiment of the present invention.

  In FIG. 12, the motion detection apparatus of this embodiment includes an image pickup unit 211, an edge extraction unit 212, an edge width normalization unit 213, a mask generation unit 214, a reset timing generation unit 215, a voting unit 216, and a moving speed detection unit 217. It is prepared for.

  The image pickup unit 211 is realized by an image pickup device such as a CMOS or a CCD, and picks up a peripheral image. In addition, the edge extraction unit 212 creates an edge image of the target edge (contour) from the image captured by the imaging unit 211 using a Sobel filter or the like. Here, the Sobel filter is a first-order difference filter, but other difference filters (for example, a Prewitt filter, a horizontal or vertical first-order spatial differential filter, a second-order spatial differential filter, and a non-linear difference filter. Filters etc. may be used.

  The edge width normalization unit 213 normalizes the edge width in the edge image extracted by the edge extraction unit 212 to a predetermined number of pixels.

  Further, the mask generation unit 214 sets, for each edge region normalized by the edge width normalization unit 213, a region where no edge is detected as a mask region for the edge. Specifically, for each edge standardized by the edge width normalization means 213, an area obtained by the logical product of the edge mask area in the previous frame and the area in which the edge is not detected in the current frame is obtained. It is set as a mask area of the edge in the frame. The latter voting means 216 resets the count value of the position of the mask area, and counts up the count value of the position for the non-mask area that is a complement of the mask area.

  The reset timing generation unit 215 sets the reset timing for resetting the count value by the voting unit 216 and the mask area by the mask generation unit 214 based on the positional relationship with the adjacent edge. Specifically, a reset timing is generated when it is determined that the non-mask areas of adjacent edges overlap each other, and the count value by the voting means 216 and the mask area by the mask generation means 214 are reset. Note that an edge for which there is no adjacent edge to be overlapped and the mask area continues to be counted may be automatically reset after a certain period of time has elapsed.

  The voting means 216 counts up the count value of the position of the non-mask area that is not masked by the mask area generated by the mask generation means 214. Specifically, the contents of the memory address corresponding to the non-mask area are counted up, and conversely, the contents of the memory address corresponding to the mask area are reset. By voting such continuous time frames and accumulating the vote values, it is possible to obtain information on how many frames the standardized edge was observed in time series.

  Further, the moving speed detecting means 217 calculates the moving direction, speed, and acceleration of the edge from the inclination of the count value accumulated over successive frames by the voting means 216.

  Next, normalization of the edge width by the edge width normalization means 213, mask generation by the mask generation means 214, voting by the voting means 216, calculation of the edge moving direction, speed and acceleration by the moving speed detection means 217, and resetting The generation of the reset timing by the timing generation unit 215 will be described in detail with a specific example.

  First, with reference to FIGS. 13 and 14, the edge width normalization by the edge width normalization means 213 will be described.

  First, FIG. 13 is a more specific configuration diagram of the edge width normalizing means 213. In FIG. 13, the edge width normalization means 213 repeats the thinning process until the edge width reaches a predetermined number of pixels, and the binarization means 221 that performs binarization processing on the edge image extracted by the edge extraction means 212. And a thinning means 222 for unifying all edge widths to a predetermined number of pixels, and an expansion means 223 for expanding the edge widths unified to a predetermined number of pixels to a predetermined number of pixels.

  Next, FIG. 14 illustrates an example of binarization (FIG. 14 (a)), thinning (FIG. 14 (b)), and expansion (FIG. 14 (c)), which are procedures for normalizing the edge width. FIG. First, as shown in FIG. 14A, the binarizing means 221 performs binarization (0/1) processing on the extracted edge image. In the drawing, a white circle indicates a pixel “0”, and a solid circle indicates a pixel “1” (active). Next, in the thinning means 222, as shown in FIG. 14B, it is repeated until the edge width reaches a predetermined number of pixels (one pixel in the x direction in the specific example of FIG. 14). The edge width is unified to a predetermined number of pixels (one pixel).

  Furthermore, as shown in FIG. 14 (c), the expansion means 223 expands an edge having a uniform edge width (to one pixel) to a predetermined number of pixels (3 pixels in the specific example of FIG. 14) as necessary. ing. For example, when an edge position is observed at the position x0, the pixels at the positions x0-1 and x0 + 1 on both sides of the position x0 are activated (1), and the edge width is unified to 3 pixels.

  By these binarization processing, thinning processing and expansion processing, the edge image extracted by the edge extraction means 212 is generated as a binary normalized edge image in which the edge width is unified to a predetermined value (3 pixels). Will be. In order to unify the edge width, the edge peak position of the edge image output from the edge extraction unit 212 is detected, and a binary image having a predetermined number of pixels at the edge peak position is generated. The width may be normalized.

  Next, mask generation by the mask generation unit 214 will be described with reference to FIG. FIG. 15 is an explanatory diagram for explaining mask generation by the mask generation means 214, and exemplifies normalized edge images and generated masks at time t, time t + m, and time t + n. In the following description, for simplicity, it is assumed that the edge moves in the one-dimensional direction of the x-axis, and the extracted edge moves in the positive direction of the x-axis. In addition, it is assumed that the normalized edge is unified to an edge width of 3 pixels by the edge width normalizing unit 213.

  For each edge standardized by the edge width standardization means 213, the mask generation means 214 calculates an area obtained by a logical product of the mask area of the edge in the previous frame and the area where the edge is not detected in the current frame. It is set as a mask area for the edge in the current frame. This logical operation is a positive logical operation when the mask area and the area where no edge is detected are assumed to be true values.

  In FIG. 15, for the sake of easy understanding, normalized edge images (FIGS. 15A, 15C and 15E) and generated masks (FIGS. 15B, 15D and 15F) are shown. When these logical operations are considered, it is a negative logical operation. That is, the area obtained by the logical sum of the non-mask area of the edge in the previous frame and the area where the edge is detected in the current frame is set as the non-mask area of the edge in the current frame.

  First, FIG. 15A shows a normalized edge image at time t. In this edge image, the edge width is standardized to 3 pixels by the edge width normalizing means 213, and for easy viewing, the center pixel is indicated by a filled circle and the pixels on both sides thereof are indicated by white circles. ing. If the reset timing is generated by the reset timing generation means 215 at this time t, the area where no edge is detected becomes the mask area as it is, and the mask at the time t is filled as shown in FIG. It becomes an area. In FIG. 15B, a white area is a non-mask area.

  Next, if the normalized edge image changes as shown in FIG. 15C at time t + m after time m, an unmasked area (see FIG. 15B) and edge at time t are detected. A region obtained by ORing with the unmasked region is obtained as a non-mask region at time t + m, and as shown in FIG. 15D, a filled region which is a complement of the non-mask region is obtained as a mask region.

  Similarly, if the normalized edge image changes as shown in FIG. 15E at time t + n after time n (n> m), the non-mask area at time t + m (see FIG. 15D). A region obtained by ORing the region where the edge is detected is obtained as a non-mask region at time t + n, and as shown in FIG. 15 (f), a filled region which is a complement of the non-mask region is masked. Obtained as a region.

  Next, voting by the voting means 216 will be described with reference to FIG. FIG. 16 is an explanatory diagram for explaining voting by the voting means 216, and exemplifies non-mask areas and accumulated voting values (count values) at time t, time t + m, and time t + n. In FIGS. 16 (a1), (b1), and (c1), one pixel at the center of the edge is shown by a filled circle for easy viewing, and a set of this filled circle and surrounding white circles is shown. Is a non-mask region. 16 (a2), (b2) and (c2) are the voting values at the positions of the hatched bands in the y-axis direction in FIGS. 16 (a1), (b1) and (c1). Value.

  First, when a non-mask area as shown in FIG. 16 (a1) is given from the mask generation means 214 at time t, the voting means 216 counts up the voting value at each position in the non-mask area. At that time, the voting value of the mask area which is a complement of the non-mask area is reset. When the non-mask area changes as shown in FIG. 16 (b1) at time t + m after time m, and when the non-mask area changes as shown in FIG. 16 (c1) at time t + n after time n. Similarly, the voting value at each position in the non-mask area is counted up, and the voting value in the mask area is reset.

  Next, calculation of the edge moving direction, speed, and acceleration by the moving speed detecting means 217 will be described.

  If the frame rate is sufficiently high compared to the speed at which the edges move, the edges observed between consecutive frames will almost always have overlapping regions (in the illustrated case, they overlap by 2 pixels wide) ) As described above, by counting up the vote value corresponding to the position where the edge is observed by the voting means 216, the accumulated vote value is equivalent to the time when the edge is observed at the same position. Become.

  When the edge moves, the vote value at the position where the edge is newly observed is 1, which is the smallest value among the observed edges. That is, the voting value gradually decreases in the direction in which the edge moves, and conversely, the voting value increases in the direction opposite to the direction in which the edge moves. The slope of this vote value is a value obtained by counting how many frames the edge has been observed at the same position before the edge moves, and the profile of this vote value (the value of the first and second derivatives) ), The moving speed and acceleration of the edge can be calculated. In the case of FIG. 16C2, it is possible to detect the moving speed and acceleration of the edge of interest from the gradient of the vote value from the position xj-1 to xj + 3.

  For corner edges having orthogonal edge components, the processing in the x-axis direction described above is also performed in the y-axis direction, and the moving speed vectors in the x-axis direction and the y-axis direction are combined to obtain the edge. And the moving speed in the moving direction are obtained, and the acceleration in the moving direction of the edge is obtained by combining the acceleration vectors in the x-axis direction and the y-axis direction.

  In addition, in this embodiment, since the non-mask area (mask) is generated based on the time-series normalized edge with the edge width normalized, it is possible to suppress the influence of the singular edge instantaneously, and more Since the edge moving direction, velocity and acceleration are detected using highly reliable edge information, more accurate velocity and acceleration can be observed.

  Next, generation of reset timing by the reset timing generation means 215 will be described with reference to FIG. FIG. 17 is a time chart illustrating the frame acquisition timing and the reset timing.

  When the time for observing the movement of the edge is short, there are edges that appear or disappear instantaneously. In this case, the voting value is small, which reduces the reliability of measuring the moving speed and acceleration of the edge. . On the other hand, when the edge observation time is too long, it becomes difficult to fit the relationship between time and speed with a curve or the like if the moving speed of the edge becomes faster or slower, so the speed profile is analyzed accurately. It will be difficult to do. That is, it is desirable to define the cycle for measuring the moving speed of the edge in accordance with the dynamics of the target image.

  When the frame rate is sufficiently high, it can be assumed that there is almost no movement of the target edge between frames, and that the motion is constant. On the other hand, as described above, the measurement accuracy may be lowered due to the influence of an edge that appears or disappears instantaneously or a small vote value. Therefore, it is desirable to observe the movement of the target edge within a range of time variation that is such that the vote value by the voting means can be determined to be significant and does not have a complicated speed profile. .

  Also, if there are adjacent edges, the masks are generated in time series based on the normalized edges for the adjacent edges as well, and if the non-mask areas of the adjacent edges are overlapped, Since the noise is added to the vote value at the position of, the reset timing generation means 215 generates the reset timing when the non-mask regions of the adjacent edges overlap each other, and the vote value and mask by the vote means 216 The area needs to be reset.

  Next, normalization of the edge width by the edge width normalization means 213 described above, mask generation by the mask generation means 214, voting by the voting means 216, calculation of the edge moving direction, speed and acceleration by the movement speed detection means 217, In addition, the flow of reset timing generation processing by the reset timing generation means 215 will be described with reference to the flowchart shown in FIG.

  First, in step S101, imaging is performed by the imaging unit 211. In step S102, the edge extraction unit 212 extracts an edge image. Next, in step S103, binarization processing is performed on the edge image extracted in step S102, thinning processing is performed in step S104 to unify the edge width to a predetermined number of pixels (one pixel), and then in step S105. Dilation processing expands to a predetermined number of pixels (3 pixels), and unifies the edge width to a predetermined width (3 pixels).

  Next, in step S106, the mask in the current frame is obtained from the non-mask area obtained by the logical sum of the normalized edge normalized in the processing from step S103 to step S105 and the non-mask area of the edge in the previous frame. Generate.

  Next, in step S107, the reset timing generation means 215 compares the number of effective pixels of the mask in the previous frame and the mask in the current frame. It is determined that the non-mask area is superimposed, a reset timing is generated, and the process proceeds to step S109 to reset the accumulated vote value and mask. In this case, the process returns to step S105, a mask is generated based only on the normalized edge normalized by the processing from step S103 to step S105, and the process proceeds to step S108. If it is determined in step S107 that the non-mask area of the adjacent edge is not superimposed, the process proceeds to step S108.

  Next, in step S108, the voting value of the memory address corresponding to the effective pixel (the pixel in the non-mask area) based on the mask generated in step S106 is counted up, and the voting value of the memory address corresponding to the invalid pixel in the mask area is counted. To reset.

  Further, in step S110, the gradient of the accumulated voting value is obtained, and based on the gradient of the voting place, the moving direction, moving speed, and acceleration of the edge are calculated in step S111.

  As described above, in the motion detection apparatus according to the present embodiment, an edge is extracted by the edge extraction unit 212 (edge extraction step) from the image captured by the imaging unit 211 (imaging step), and the width of the edge is set to the edge. Normalization is performed by the width normalization unit 213 (edge width normalization step), and an area where no edge is detected based on the standardized edge information is set as a mask region of the edge by the mask generation unit 214 (mask generation step). The voting means 216 (voting step) resets the count value of the position of the mask area, counts up the count value of the position of the non-mask area not masked by the mask area, and moves the movement speed detection means 217. (Moving speed detection step), the accumulated data over consecutive frames Based on the inclination of the cement value, calculates the moving direction of the edge, velocity and acceleration.

  As a result, the count value is equivalent to measuring the time when the edge is observed. Therefore, the number of frames required for the edge to move a predetermined number of pixels is measured from the gradient of the count value, and the edge value is measured. The moving direction, speed, and acceleration can be easily obtained without using block matching such as template matching. Further, it is possible to prevent pixels at other edges from being taken in by the mask area, and to observe more accurate speed and acceleration without affecting each other between the edges. Furthermore, since the calculation process is performed only based on the position of the non-mask area, it can be processed with a relatively simple procedure for a smaller amount of data, and the detection of the moving speed component from the image can be processed at a higher speed. (Effects of claims 1 and 5).

  In the motion detection apparatus of this embodiment, the reset timing generation unit 17 (reset timing generation step) resets the count value and the mask area by the voting unit 216 (voting step) based on the positional relationship with the adjacent edge. More specifically, the timing is set, and more specifically, when the non-mask areas of adjacent edges overlap each other, the reset timing is generated and the count value and the mask area by the voting means 216 (voting step) are reset. The time to observe each edge speed and acceleration can be secured until the non-mask areas for observing the speed and acceleration between adjacent edges overlap, improving the time-series speed change detection S / N (Effects of claims 2, 4, 6 and 8) .

  Further, in the motion detection apparatus of this embodiment, the mask of the edge in the previous frame for each edge normalized by the edge width normalizing means 213 (edge width normalizing step) by the mask generating means 214 (mask generating step). Since the area obtained by the logical product of the area and the area where the edge is not detected in the current frame is set as the mask area of the edge in the current frame, it is calculated by the moving speed detecting means 217 (moving speed detecting step). From the count value gradient, the number of frames required for the noted edge to move a predetermined number of pixels can be counted. That is, since it is possible to observe the time change of how many frames the target edge is observed at the same position, it is possible to improve the detection speed / acceleration S / N of the target. In addition, the non-mask area (mask) is generated based on the normalized edge with standardized edge width, so that the influence of singular edges can be kept low instantaneously and more reliable edge information is used. Thus, the moving direction, speed and acceleration of the edge are detected, and more accurate speed and acceleration can be observed (effects of claims 3 and 7).

  The above-described embodiment is an example of the present invention. For this reason, the present invention is not limited to the above-described embodiment, and various modifications can be made depending on the design and the like as long as the technical idea according to the present invention is not deviated from this embodiment. Of course, it is possible to change.

  This is extremely useful as a method for easily calculating the moving speed, moving direction and position of an object by detecting features such as the contour movement of the object in continuous image processing.

It is a block diagram which shows the structure of the motion detection apparatus to which this invention is applied. It is a block diagram for demonstrating operation | movement of the motion detection apparatus to which this invention is applied. It is a block diagram which shows the structure in the case of using an edge detection filter by the feature normalization part of the motion detection apparatus to which this invention is applied. It is a figure which shows operation | movement in case the corner edge detection filter is used by the characteristic normalization part of the motion detection apparatus to which this invention is applied. It is a flowchart which shows the process sequence which detects the position and speed of the target object by the motion detection apparatus to which this invention is applied. It is a figure for demonstrating the operation | movement which obtains a count value from a normalization pattern with the motion detection apparatus to which this invention is applied. It is a figure for demonstrating the mask produced | generated by the motion detection apparatus to which this invention is applied. It is a block diagram which shows the structure of the motion detection apparatus which concerns on embodiment of this invention. It is a flowchart for demonstrating the motion detection process by the motion detection apparatus which concerns on this invention. It is a block diagram which shows the detailed structure of an edge width specification part. It is a figure for demonstrating the thinning process and expansion process in an edge width normalization part, and the count-up process in an edge count part. It is a block diagram of the motion detection apparatus which concerns on one Example of this invention. It is a more concrete block diagram of an edge width normalization means. It is explanatory drawing which illustrates the process of binarization (FIG. 14 (a)), thinning (FIG.14 (b)), and expansion (FIG.14 (c)) which are the procedures of normalization of edge width. It is explanatory drawing explaining the mask production | generation by a mask production | generation means, illustrating the normalized edge image and the produced | generated mask at the time t, the time t + m, and the time t + n. It is explanatory drawing explaining the vote by a voting means, illustrating the non-mask area | region and the accumulated voting value in the time t, the time t + m, and the time t + n. It is a time chart which illustrates a frame acquisition timing and a reset timing. It is a flowchart explaining the optical flow detection method in a motion detection apparatus.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1,101 Image pick-up part 2 Feature extraction part 3 Feature normalization part 4 Voting part 5,105 Movement speed detection part 21,131 Binarization part 22,132 Thinning part 23,133 Expansion part 24 Peak position detection part 102 Edge extraction Unit 103 edge width normalization unit 104 edge count unit 141 count-up mask 211 imaging unit 212 edge extraction unit 213 edge width normalization unit 214 mask generation unit 215 reset timing generation unit 216 voting unit 217 movement speed detection unit 221 binarization unit 222 Thinning means 223 Expansion means

Claims (20)

An imaging means for capturing a peripheral image;
A feature extraction unit for extracting a feature region of a region for speed calculation from the image captured by the imaging unit;
Feature region normalization means for normalizing the size of the feature region extracted by the feature extraction means;
Mask generating means for setting, as a mask area of the feature area, an area in which no feature area is detected, based on the feature area information normalized by the feature area normalizing means;
Voting means for resetting the count value of the position of the mask area and counting up the count value of the position of the non-mask area not masked by the mask area;
And a moving speed detecting means for calculating a moving direction and a moving speed of the feature area from the inclination of the count value accumulated over successive frames.   The motion detection apparatus according to claim 1, wherein the feature extraction unit detects a feature region that is an edge component of an object.   The motion detection apparatus according to claim 1, wherein the feature extraction unit detects a feature region that is a corner edge component of an object.   The motion detection apparatus according to claim 1, wherein the feature extraction unit detects a feature region that is a feature vector of an object.   5. The feature area normalization means defines a representative point from the feature area extracted by the feature extraction means, and forms a predetermined pattern with the representative point as a reference. The motion detection device according to claim 1.   5. The movement speed calculation unit calculates the position of the feature region based on a slope of the count value and a minimum count value among the count values. 6. Any one of the motion detection apparatuses.   The motion detection according to any one of claims 1 to 4, wherein the feature region normalizing means defines a feature region to be standardized according to a moving speed of the feature region. apparatus.   5. The apparatus according to claim 1, further comprising reset timing generation means for setting a count value by the voting means and a reset timing for resetting the mask area based on a positional relationship with an adjacent feature area. The motion detection apparatus described in 1.   The mask generation unit obtains, for each feature region normalized by the feature region normalization unit, a logical product of the mask region of the feature region in the previous frame and the region in which the feature region has not been detected in the current frame. The motion detection apparatus according to claim 1, wherein the region is set as a mask region of the feature region in the current frame.   The reset timing generation means generates the reset timing and resets the count value and the mask area by the voting means when non-mask areas of adjacent feature areas overlap each other. The motion detection device according to any one of claims 4 to 8. An imaging step of capturing a peripheral image;
A feature extraction step of extracting a feature region of a region for speed calculation from the image captured by the imaging step;
A feature region normalizing step for normalizing the size of the feature region extracted by the feature region extracting step;
Based on the feature region information standardized by the feature region normalization step, a mask generation step for setting a region where no feature region is detected as a mask region of the feature region;
A voting step that resets the count value of the position of the mask area and counts up the count value of the position of the non-mask area that is not masked by the mask area;
And a moving speed detecting step of calculating a moving direction and a speed of the feature area from the gradient of the count value accumulated over successive frames.   The motion detection method according to claim 11, wherein the feature extraction step detects a feature region that is an edge component of an object.   The motion detection method according to claim 11, wherein the feature extraction step detects a feature region that is a corner edge component of an object.   The motion detection method according to claim 11, wherein the feature extraction unit detects a feature region that is a feature vector of an object.   15. The feature region normalization step defines a representative point from the feature region extracted by the feature extraction unit, and forms a predetermined pattern with the representative point as a reference. The motion detection method according to crab.   15. The position of the feature region is calculated based on the inclination of the count value and the minimum count value among the count values in the moving speed calculation step. The motion detection method according to any one of the above.   The motion detection according to any one of claims 11 to 14, wherein the feature region normalizing step defines a feature region to be standardized according to a moving speed of the feature region. Method.   15. The method according to claim 11, further comprising a reset timing generation step for setting a count value obtained by the voting step and a reset timing for resetting the mask region based on a positional relationship with an adjacent feature region. The motion detection method described in 1.   The mask generation step is obtained for each feature region normalized by the feature region normalization step by a logical product of the mask region of the feature region in the previous frame and the region in which the feature region has not been detected in the current frame. The motion detection method according to claim 11, wherein the region is set as a mask region of the feature region in the current frame.   The reset timing generation step generates the reset timing and resets the count value and the mask region in the voting step when non-mask regions of adjacent feature regions overlap each other. The motion detection method according to any one of claims 14, 18, and 19.

Images