JPH02206883A - Picture processor - Google Patents

Abstract

PURPOSE:To accelerate calculation, and to easily grasp tendency on the whole by connecting a digital differential analysis circuit in series, and storing the calculated result for every window in a storage circuit, and sorting and detecting connection to connect objects to be processed for every window from its contents. CONSTITUTION:The picture element signal of every object point to be processed on an original picture taken in by a camera 11 is inputted through a signal inputting part 12, and is edge-detected 13, and edgized data is inputted to a preprocessing part 17. This processing is repeated whenever picture data is inputted, and the picture data is successively sent to the preprocessing part 17 as digital data. Then, in the preprocessing part 17, a signal whose preprocessing is finished is sent to an initial value arithmetic part 40. Next, the calculation of a rotary motion recurrence formula for obtaining a Hough curve is executed as digital differential analysis DDA arithmetic operation. Then, the filtering 19 processing of the intersection point of these finished curves and the sorting processing of a sorting part 20 are executed, and finally, the approximate straight line to connect the processed objects on the original picture is obtained.

Description

DETAILED DESCRIPTION OF THE INVENTION [Industrial Application Field] The present invention relates to an image processing device, which is used for controlling unmanned vehicles, real-time recognition of the motion of an observed object, and the like.

[Conventional technology]

For example, when controlling unmanned robots or autonomous vehicles, it is necessary to use a camera to capture images of identification lines displayed in advance on the road ahead, or the center line or shoulder line of the road, and process the images in real time.

Figure 22 is for explaining road recognition using images. Figure 22 (a) is an image captured by a camera, Figure 22 (b) is an image captured by a camera.
is the luminance (or its rate of change, etc.) in figure (a)
3 is a diagram showing pixels with high values as black dots. FIG.

As shown in Figure (a), the camera image 1 shows a road 3 extending toward infinity from the horizon 2. Road shoulder lines 4 are drawn on both sides of the road 3, and a center line 5 is drawn in the center. is depicted. Here, the shoulder line 4 and center line 5 of the road 3 have higher brightness than other parts,
Therefore, in the same figure (b), dots 4' and 5 are placed in these parts.
' will appear consecutively. In such a camera image 1, in order to recognize the traveling direction and curved shape of the road 3, an approximate straight line of the curve connecting dots 4', L2. It is sufficient to recognize L3 etc.

Traditionally, as a method for finding such an approximate straight line,
A technique called Hough transform is known (for example, US Pat. No. 3,069,654). This is the second
This will be explained with reference to FIGS. 3 to 25. As shown in FIG. 23(a), when a processing target point P (x, y,) exists in the original image indicated by the x-y coordinate system, a straight line flcf1. fl, etc.) can be drawn infinitely. And this straight line 1 1! , ... and passing through b origin 0 (0,0), the straight line N
l! You can draw one line for each.

a b Here, regarding the straight line passing through the origin 0 (0,0), the length to the straight line IC1, Ib, etc.) is ρ (ρ,.

ρ5, etc.), and the angle between it and the X axis is θ (θ3.θ.

), then the above ρ of the straight line passing through this origin.

θ is expressed as a sine curve, ie, a Hough curve, as shown in FIG.

Here, the origin 0 (0,0) and the processing target point P(x
, y) is the distance ρ of this processing target p p
The above ρ(
ρ3. ρ6. ...) is the longest among them, and becomes 2 2.1/2 ρ1ax-(x+yp, and when θ-0, ρo "" X.

Next, three points P lined up in a straight line as shown in Figure 4(a)
-P, when the Hough transformation shown in Fig. 23 is applied, the above sine curve (Hough
The sine curve number for point P becomes like the dotted line in Figure 24(b), and the sine curve for point P3 looks like the dotted line in Figure 24(b). It looks like the two-dot chain line. Here, the peaks (ρ , θ ), (ρ , θ2) and (
ρ, θ3) are respectively at the origin 0 (0,
0) and the points P , P , P , the distance ρ ~ ρ
and the angle θ to θ3 formed with the X axis.

In Figure 24(b), if we focus on the point where the three Hough curves (sine curves) intersect, we can see that the coordinates here are (
ρ , θ ), which is the same as the ρ , θ of the straight line passing through the origin 0 (0,0), which is perpendicular to the straight line in Figure (a).
is equal to Therefore, if we find the intersection point of such a sine curve, we can find the approximate straight line of the curve between the dots (black dots) drawn in the x-y orthogonal coordinate system of the original image (however, in Fig. 24, this curve and the approximate straight line are ) can be found.

To explain this with reference to Fig. 25, first, in Fig. 25 (a), there are many dots (processing target points) to be Hough-transformed on the x-y coordinate plane (original image plane), and these are lined up on a curve. shall be.

Here, in the same figure (a), there are three approximate straight lines in the curve connecting the dots, L2. Can draw L3.

Therefore, if all of these dots are converted into sine curves (Hough conversion) as shown in Figure 23, we get
As shown in FIG. 23(b), intersection points of the sine curves are obtained centered on three points. The coordinates of this intersection point are shown in FIG. 25(a) (ρ, 1.θ1,), (ρt
2′θ ) and (ρ θ ), so this t
If 2 13' t3 is expressed in the coordinate system of ρ, θ, H, where H is the degree of appearance of the intersection point, the result will be as shown in FIG. 3(b).

Therefore, the approximate straight lines L1 to L of the curve corresponding to the shoulder line 4 of the road 3 as shown in FIG.
This can be determined by the values of ρ and θ at the peak of H (frequency of appearance of crossover points) in FIG.

[Problem to be solved by the invention]

However, the Hough transformation method described above is
It has not been easy to apply it to high-speed and real-time image processing. This is because the coordinate values (x, y) of the processing target point P in the original image given as data in FIG.
Based on p Hough curve (sine curve), when trying to find the distance from the origin to straight line g,
e sin O+ y ”c
os θ −(1)p must be executed. For example, if θ is divided into 512, trigonometric function calculations must be performed 1024 times, multiplications 1024 times, and additions 512 times. If the original image to be calculated is composed of, for example, 512×512 pixels, the total number of calculations will be extremely large, and processing with a normal processor will significantly increase the processing time.

Of course, the sin θ required in the above equation (1)
It is also possible to shorten the time required for calculation by storing the values of and eO8θ in ROM etc. FROM
``Hough conversion hardware using Hough conversion hardware'' 1986 IEICE 70th Anniversary National Conference, No.

1587). However, in the former case, even if the trigonometric function data is stored in ROM, the number of multiplications in the calculation of equation (1) remains the same as before, and considering that the multiplication time is considerably longer than the addition time, the fundamental It can't be a solution.

On the other hand, if we consider the latter document ``Hough conversion hardware using FROM'' from a different perspective, we can see that by parallelizing the calculation section that performs the calculation of equation (1) above, the signal processing as a whole can be improved. We are trying to speed it up. However, in this way, y −5lnθ,X
A memory table (ROM) for determining 'eO8θp is required as many as the number of arithmetic units connected in parallel, and the system becomes extremely large in terms of hardware.

Also, it is not suitable for LSI implementation. Although there is a method of converting ROM into RAM, there is a drawback in terms of the degree of integration.

For this reason, it has been almost impossible to control a vehicle traveling at high speed in real time based on image data, or to recognize an observed object moving at high speed in real time based on image data. Furthermore, when attempting to increase the speed by parallelizing the arithmetic units, it was inevitable that the hardware would become larger. Furthermore, if we examine the recognition of the characteristics of the distribution of processing target points in the original image from the viewpoint of ease of recognition, we find that there are parts of the original image where it is easy to understand the characteristics of the distribution of processing target points and parts where it is difficult to understand. If difficult parts were processed all at once, it would be difficult to grasp the overall, big-picture trends.

Therefore, the present invention not only makes it possible to process image data in real time at high speed, but also makes it possible to downsize the hardware, and also makes it possible to easily determine the global trend of the original image. It is an object of the present invention to provide an image processing device that can grasp the information.

[Means to solve the problem]

A first image processing device according to the present invention is an image processing device that separates and extracts feature points of an original image from a plurality of processing target points on the original image captured by an imaging means, and includes the following elements. Be prepared. That is, a dividing means for dividing the original image into a plurality of windows, and a DDA having the same configuration.
(DigitalDtrrerentta+^naly
sis (digital differential analysis); a DDA calculation means configured by connecting a plurality of calculation elements in series; a storage means for storing calculation results of the DDA calculation means for each window corresponding to the DDA calculation element; The present invention is characterized by comprising means for extracting feature points of the original image for each window based on the stored contents.

Further, a second image processing device according to the present invention is an image processing device that derives an approximate straight line of a curve connecting a plurality of processing target points on an original image captured by an imaging means, and includes the following elements. . In other words, the dividing means divides the original image into multiple windows, and the coordinates of one point on the circumference of an approximate circle drawn in the α-β orthogonal coordinate system are (α, β), and the coordinates of the next point on the circumference are Let the rotation angle to the coordinates (α, β) be ε, and when 141 1+ (where i is a positive integer), αtit””α, β1. ε) α l β −f (α , β1.8) Country β 1 A DDA calculation means that sequentially calculates the rotational motion recurrence formula for each predetermined rotation angle in a pipeline manner, and this DDA calculation means sequentially calculates Of each calculated result,
A storage means for storing at least the value of β1 for each window, and an approximation for deriving an approximate straight line of the above curve for each window from the intersection of the Hough curves for each of a plurality of processing target points obtained based on the storage contents of this storage means. The method is characterized by comprising a straight line deriving means.

[Effect]

According to the configuration of the present invention, a calculation unit can be easily formed by simply connecting DDA calculation circuits (calculation elements) of the same configuration in series, and there is no need to refer to a memory table or the like during calculation. It becomes significantly simpler and smaller.

Further, by configuring the arithmetic circuit to execute the rotational motion recurrence formula in a pipeline manner, it is possible to speed up calculation. Furthermore, since the original image is divided into multiple windows and processed, it is easy to grasp the overall trend.

〔Example〕

Below, based on Figures 1 to 21 of the attached drawings,
The present invention will be described in detail. In addition, in the description of the drawings, the same elements are given the same reference numerals, and redundant description will be omitted.

FIG. 1 is a block diagram showing the overall configuration of an image processing apparatus according to an embodiment. In the figure, a camera 11 captures an original image by capturing an object to be processed (for example, a road, a high-speed moving object, etc.), and this image signal is digitized by a signal input section 12 and sent to an edge detection section 13. . The edge detection unit 13 extracts the edges of the image signal and generates edge data with shading, for example, 512
The signal is sent to the multilevel memory 14 as X512 pixel signals (edged pixel signals). The multilevel memory 14 stores edged data for each pixel, and sends the edged data to the D/A converter 15 each time one screen is scanned, and is applied to the CRT display 16 as an analog signal. Therefore, this edged data is displayed on the CRT display 16.

On the other hand, the edged pixel signal output from the edge detection section 13 is given to the preprocessing section 17, and the edged pixel signal subjected to preprocessing as will be described in detail later is subjected to DDA calculation via the initial value calculation section 40. Section 18. DDA calculation unit 18
is composed of n DDA calculation circuits 18 to 18, and these On-1 circuits are connected in series with each other. Then, the DDA calculation unit 1
A neighborhood filter 19 and a sorting section 20 are provided on the output side of 8.
are connected, thereby performing neighborhood filtering processing and sorting processing (described in detail later). Note that the above circuit elements are connected to the CP via the VME bus 21.
It is connected to U22 to control signal processing operations and synchronize processing timing. In addition, the preprocessing section 17, the DDA
The calculation unit 18 and the neighborhood filter 19 are connected to each other via a VME bus 23, and synchronous control of the transfer of DDA calculation results and the transfer of gray value data is performed.

Next, the detailed configuration of the main parts of the image processing apparatus shown in FIG. 1 will be explained.

FIG. 2 is a configuration diagram thereof, which corresponds to the preprocessing section 17, initial value calculation section 40, DDA calculation section 18, and neighborhood filter 9 in FIG. As shown in the figure, the preprocessing section 17
This is realized by an FO (First-In First-Out) board 17'. PIFO 17' is the coordinate value (X) of the point to be processed on the X-Y plane.

Y) is input as an address signal, and edged gray value data DI is input as a data signal.

Then, as described later, this FIFO 17' performs coordinate conversion from X-Y coordinates to x-y coordinates, multiple window settings, and threshold processing at a predetermined level, and the results are transferred to the FIFO
Outputs sequentially according to the FO method.

The initial value calculation section shown in FIG. 1 has two flip-flops (F/F) 41 and 42 and an initial value calculation circuit 43.
/F41 is the coordinate value (
x, y) are temporarily stored, and the F/F 42 temporarily stores the grayscale value data D1 of the processing target point P. Based on the coordinate values (x, y), the initial value calculation circuit 43 calculates the initial value coordinates (α0°β.) in the αp-β orthogonal coordinate system, thereby making it possible to calculate the rotational motion recurrence formula. do.

The DDA operation circuits 188 to 18 in each stage shown in FIG. 1 each have three flip-flops (
F/F) 31, 32, and 33, F/F 31 temporarily stores address signals α0 to αn-1'β to β, respectively, and F/FOn-1 32 stores gray value data D1. Temporarily store each, F
/F 33 is connected to RAM 34 (RAM
- Temporarily stores the histogram data DMO=DM(n-1) read out from each of the RAMs. DDA37 (DDAo~ODA
) calculates the rotational motion recurrence formula described later for each rotation angle, and inputs address signals α and β and outputs an address signal α1+1°1]β. The adder ADD351+1 (ADD - ADD) inputs data from FIFO17' to O
n-1 gradation value data D1 and histogram data DMo to DH(n-
1), and the output thereof is temporarily stored in the buffer 36 and then sent to each of the RAMs (On-1) through the switching section 91. Flip-flop (F/F) 92 is F
It is used to temporarily store window data from the IFO 17' and send it to the switching section 91. The timing controller 25 controls the timing of signal processing in these circuit elements, and controls the timing pulse φ.
, φ to S a φr. It is connected to a command/status dispute interface (1/F) not shown.

Next, an overview of the overall operation of the image processing apparatus shown in FIGS. 1 and 2 will be explained with reference to FIG. 3.

FIG. 3 is a flowchart explaining this.

First, a pixel signal for each point to be processed on the original image captured by the camera 11 is inputted via the signal input section 12 (step 102), and the edge detection section 13 performs edge detection (step 104) to generate edge data. Input to the preprocessing section 17 (step 106). Above steps 102-1
The processing of 06 is repeated every time a pixel signal is input, and the result (edge data) is sequentially sent to the signal preprocessing unit 17.
is sent as digital data.

The preprocessing unit 17 performs predetermined (described later) preprocessing (step 10
8), and the processed data is sent to the initial value calculation unit 40.
(step 110). This preprocessing is also sequentially repeated each time edged data is provided.

Next, the calculation of the rotational motion recurrence formula for obtaining the Hough curve (sine curve) is executed as a DDA calculation (step 112), as will be explained later, but prior to this calculation, the initial value is executed in step 111.

Then, the calculation in the DDA calculation unit 18 is performed on the pixel signals of one screen (original image plane) to be processed by the preprocessing unit 17.
The processing continues until all the pixels other than those excluded as outside the window or below the threshold are completed (step 114), and once completed, filtering processing (step 116) and sorting processing (118), which will be described later, are carried out regarding the intersection of the Hough curves. is executed by the neighborhood filter 19 and the sorting unit 20, and as a final result, an approximate straight line of a curve connecting the processing target points on the original image is obtained.

Next, the edge detection method and edged data in the edge detection section 13 will be explained with reference to FIG.

The original image captured by the camera 11 is shown in Figure 4 (a).
If we take out the line indicated by reference numeral 8 in the figure using the X' coordinate, we can see that the luminance S is shown in analog form as shown in figure (b). That is, road 3
The brightness of the outer part of the road 3 and the road shoulder is low, but the brightness of the road shoulder line 4 is very high. Here, in Figure 4 (
In the embodiment, the brightness distribution as shown in b) is recognized as digital data of 256 gradations, but understanding the distribution of brightness itself is not sufficient to accurately recognize the shape of the road. .

Therefore, when the brightness S of the pixel signal obtained through the signal input section 12 is differentiated with respect to the coordinate X'(dS/dX'') and understood as the rate of change in brightness, the edges are clearly seen as shown in Figure 4 (C). When this is expressed as an absolute value IdS/dx'l, it becomes as shown in the same figure (d), and the edge detection unit 13 converts the edge data (digital data) to clearly recognize the edge of the road shoulder line 4. This edge detection section 13
In the next step, preprocessing section 17 performs predetermined preprocessing on the edged data.

FIG. 5 is a flowchart for explaining the preprocessing in the preprocessing section 17 (P I FO 17').

First, the edge data obtained as shown in FIG. 4 is input from the edge detection section 13 to the preprocessing section 17 (step 122), and this data is input into any of a plurality of preset windows. It is determined whether (step 124).

Although only one window may be set, for example as shown by reference numeral 9 in FIG. 4(a), in the present invention, a plurality of windows are set. That is, for the original image 1 in FIG. 6(a), there are two windows WD in FIG. 6(b)).
, WD2, or three windows WD - WD3 as shown in FIG. 3(C). Also, the same figure (d
), when windows overlap, one of the windows (WDl in the figure) is given priority.

Here, which window WD the edged data is inside can be determined by the coordinate value (X, Y) p of the data (processing target point P) on the X-Y coordinate plane. The signal indicating whether it is engineering/solidification data is FIFE17' in Figure 2.
is output as window data. The next step 126 is then executed on the edged data within these windows.

In step 126, it is digitally determined whether the edged data within the window is equal to or greater than a predetermined threshold value (threshold level) using, for example, a look-up table (LUT) attached to the FIFO 17'. That is, as shown in Fig. 7, when no LUT is provided, the input data and output data of PIF017' are 1:1.
(as shown in figure (a)), but if you use LUT to set the threshold level to ``th'',
When the input data is less than Ith, the output data becomes 0 (zero). Note that when the gradation value data included in this edged data is grasped in 256 gradations, the threshold level lth can be set arbitrarily between 0 and 255. If this threshold value is shown in analog form, for example, in Figure 4 (d
) is set as shown by the dotted line, so step 12
The data processed in step 6 becomes data (digital data) mainly corresponding to the road shoulder line 4 and center line 5 in the original image. Therefore, only the main data (for example, data corresponding to the shoulder line and center line of the road) need to be subjected to signal processing, which will be described later, so it is not affected by noise components and the overall processing speed is reduced. It can be made faster.

Next, in step 128, a coordinate transformation is performed.

That is, the X-Y coordinate shown in FIG. 4(a) is converted into the X-y coordinate. As a result, the coordinates (X, Y) of the processing target point in the original image are converted to coordinates (x, y) in the Xp-y coordinate system. The above processing completes the preprocessing for Hough transformation.

Note that the order of steps 124 to 128 in FIG. 5 may be different. For example, the coordinate transformation in step 128 may be performed first, but in consideration of the time required for data processing, it is considered most preferable to perform the coordinate transformation in the order shown in FIG.

Next, application of the Hough transform in this embodiment will be specifically explained with reference to FIGS. 8 and 9.

As already explained in Fig. 23, if a Hough curve (sine curve) is obtained for the point P (x, y) shown in Fig. 8 (a), it will become as shown in Fig. 8 (C). be. By the way, it is also easily understood from the theorem of trigonometric functions that the locus of such a sine curve can be replaced with the locus of circular motion as shown in FIG. In other words, Hough about point P (x, y) in figure (a)
Executing the transformation p to obtain the Hough curve shown in FIG. 3(c) is equivalent to obtaining the circumferential locus of circular motion as shown in FIG. 2(b). Here, the radius R of the circle in Figure (b) is 2 2 ) 1/2, (2) y), the initial value θ is θd-π/2-θlaX, where tan θ-y/x...( 3)
Zen ax pp. Also, at this time, ρ-X.

The inventor focused on this fact and applied the recurrence formula of circular motion when drawing the circle shown in FIG. 8(b),
We have found a method that can easily perform Hough transformation of point P (x, y) to p (c) in the same figure. Here, according to the recurrence formula for circular motion above, α−
The coordinates (α , β
) is obtained as αtit”f (α, β, ε) α 11 β −f (α, β, ε) (4) country β 11 when l is a positive integer.

As for the specific content of this equation (4), DDA has been known for a long time, and for example, when the rotation angle ε is set to ε-21 (rad) (however, m-0, 1, 2, -), a++I'' al -2β1 β -2α +β ... (5) ti
There is a method called t tit i. In addition, the inventor has found that the calculation is more accurate and easier, as follows: −2−”” ) 1+l ! Ichirin −3Il +β (−2+ε /6) −m −3vr β −α (2+ε /6)tit
t + β (1-2-"") ■ ... (8) etc. may also be used.

Therefore, if we apply the recurrence formula of equation (7) above, the second
Prior to explaining the specific operation of the circuit shown in the figure, this calculation method will be specifically explained.

FIG. 9 is a flow chart thereof. First, data preprocessed by the FIFO 17' according to FIG. 5 is input (step 132), and initial values (α, β0) for calculation of the rotational motion recurrence formula are determined. In Fig. 8, the Hough curve corresponding to the processing target point P(x,y)p is expressed as an arbitrary θ' on the ρ-θ space.
If we start drawing from (-θo) (rad), then the starting point of circular motion is aom-xpsinθ'+y, cosθ'βo-x, c
osθ'+ypsinθ' (9). This (α, β.) is calculated by the initial value calculation unit 40, and DD
It is given to the A calculation unit 18 as an address signal. The initial value calculation unit 40 stores the values of sin θ' and cos θ' in advance in a ROM (not shown),
The calculation of equation (9) may be performed using an adder and a multiplier (both not shown) while referring to these data. Here, although the calculation of equation (9) requires more calculations and is more complicated than the calculation of the rotational motion recurrence equation, it is necessary to perform one calculation for one processing target point (x, y). Therefore, the overall calculation time does not increase much, and the hardware does not increase too much. Note that when starting the circular motion from θ'-0@90@180°270'' at the processing target point P(x,y), this initial value calculation becomes extremely simple.

Next, β according to the above equation (9). After storing the values in RAMo as addresses according to the window, α and β are determined by equation (7).

These are α and β calculated using equation (9). can be obtained from the output of DDAo by substituting it into equation (7) (step 136), DDA , DDA2゜DDA ,...
The result (β , β , β ,
...) sequentially to RAM1゜RAM, RAM3...
・Memorize the address according to the window.
(Step 138). On the other hand, between step 136 and step 138, the gray value data is accumulated.

That is, the address βl is RAM34 (RAM,
) Histogram data DM1 and -FI read from
Add the gray value data DI from FO17' and add this to R.
AM, it is memorized again (step 137).

Then, step 140) repeats this calculation for each rotation angle ε until 1/2 of the circle is circled, and after 1/2 of the circle, the Hough curve for one processing point on the original image becomes
β0゜β , β , β ,... memorized above
It is not only determined from the values of θ, θ, θ2°, etc. (rotation angle ε), but also the result of weighting using grayscale value data (histogram MOMI M(n-1)). D,...D Hereinafter, when the processing shown in Fig. 9 is executed for all processing target points on the original image, multiple Hough curves weighted by gray value data can be obtained in the ρ-θ coordinate system. These will have an intersection point as shown in FIG. 24(b).

Next, the operation shown in the flowchart of FIG. 9 will be explained in more detail with reference to FIGS. 2 and 10.

First, the data input in step 132 in FIG. 9 involves inputting a ready signal from the FIFO 17' in FIG. 2 to the timing controller 25, and then a read strobe signal from the timing controller 25 to the FIFO 17'
After the coordinates (x, y) of the point P to be processed are stored in the F/F 41p, the gray value data DI at that point P is stored in the F/F 41p.
2, and the window data WD at that point P is stored in F/F 95 in order to match the phase.
This is done by storing it in /F92. Then, in synchronization with the timing pulse φ from the timing controller 25 (X).

p y ) is sent from the F/F 41, and the initial value calculation unit 43 calculates the initial values (α, β.) of the p recurrence formula calculation from the coordinate values (x, y) of the processing target point P.

The calculation of the recurrence formula in step 136 is performed by the initial value calculation circuit 43.
The output from F/F 3 is used as address signals α and β0.
1, and the grayscale value data D of the processing target point P are input to F.
This is done by inputting from /F42 to F/F32. Here, addresses and data input to the F/Fs 31 and 32 are performed in synchronization with timing pulses from the timing controller 25. Then, in synchronization with the rising or falling of the timing pulse φ, the F/F3
Address signals α and β0 of 1 are the first DDA. (37
) is entered.

In this DDAo, the process of step 136 in FIG. 9 is performed. That is, the calculation of the recurrence formula according to the above-mentioned equation (7) is executed, and the calculation results (address signals α, βl) are sent to the DDAI (not shown) through the next F/F 31. Here, the above recurrence formula (7) basically does not include trigonometric function calculations or multiplications,
Further, since there is no need to refer to a memory table (ROM), calculations can be performed easily and quickly. These have sufficient accuracy (few errors) to perform circular motion.

Note that this DDA is DDA ~ DDA and 0
1 n-1 It is similarly configured, and specifically includes four adders 51 to 54 and three inverters 61 to 63 as shown in FIG. For inputs α and β to the circuit of FIG. 10, outputs α and β are 1 l
1+1 1+1
2m-1 β -2α +β (1-2) Country 1, respectively, as shown in equation (7) above.
1.

Address signal β stored in F/F 31. is RAMo
(34), and thereby the histogram data DIIIo stored in RAMo is read out. That is, RAMo has the rotation angle θ with β0 as the address. Histogram data D regarding other processing points corresponding to
Therefore, by giving address signal β0 from F/F 31 to RAMo, RAM
The histogram data DMo is sent from the RAMo to the F/F 33 in synchronization with the timing pulse φb that controls reading/writing of the F/F 33.

Next, in synchronization with the timing pulse φ from the timing controller 25, ADDo (35
), but this AD
D is given the dark p-dark value data DI of the processing target point (x, y) which is the calculation target from the F/F 32. Therefore, in ADDo, the rotation angle θ that had been stored in RAM o.

and address β. Histogram data DM corresponding to
o and the rotation angle θ of the processing target point (x, y) that is being processed. and atopresβ. The gray value data D1 corresponding to is added (step 137 in FIG. 9). Then, this addition result (D
-DMo+D1) is temporarily held in the buffer 36, and then sent to RAMo in synchronization with the timing pulse φd, and the memory according to step 138 in FIG. 9 is stored at address β0 of RAMo corresponding to θ. It will be done against.

What is important in the processing of cycle 1 above is that in the present invention, multiple windows are set for the original image as described above, and therefore the switching control of the storage section of the RAM 34 is performed in parallel. It is. That is, the first
As schematically shown in Figure 1, in this embodiment, the FIFO 17'
has a window data generation section 95 in addition to the coordinate conversion section 93 and LUT 94, and this window data generation section 95 generates a window corresponding to P into WD - WDk from the coordinates (X, Y) of the processing point P. It is designed to output a love data WD. Further, the RAM 34 has storage units corresponding to the windows WD to WDk, and which storage unit to store data in is controlled by the window data WD via the switching unit 91. It has become. Therefore, for example, as shown in FIG. 6(b) (when two windows WD and WD2 are set, there are two types of window data, which correspond to the address signals α, β, and gradation values for each point to be processed. It is transferred in synchronization with the data DI.
Data for each window will be stored in each storage unit.

In this way, in the present invention, data is sorted and stored for each window, so it is possible to extract the global trend of the distribution of processing target points in the original image very easily. For example, when recognizing a road as shown in Figure 6(a), if the entire original image is processed, the horizon, buildings, guardrails, etc. will be included, and many approximate straight lines of these will also be extracted. be done. This makes it difficult to extract the road shoulder line. On the other hand, if only the window W D 2 in FIG. 6(b) is processed, for example, the horizon, buildings, guardrails, etc. are hardly included, and therefore the road shoulder line can be extracted extremely easily.

The above one-cycle processing is performed by the DDA calculation circuit 181 (
1-1, 2+...n-1), the first
It will look like Figure 2. Note that the switching time based on window data is omitted.

First, from the coordinate values (x, y) of the point to be processed, p The obtained address signals (initial values) α and β are α. .

β-α, β-α, β-...α, β and 01122 tt The address signal α1°β as a result of sequential calculation and the gray value I of this processing target point (x, y)
pI) Data DI is input from F/Fs 31 and 32 and held (step 202), and address signals α and β1 are sent from F/F 31 to DDA, t,:.

After being sent, in step 204, the address signal α1. β
The calculation of the recurrence formula based on the above is performed in the DDA. And the resulting address signals α,β
is sent to the F/F 31 provided in front of 1+1 DDA in the next DDA calculation circuit 18 in synchronization with the end of one cycle of processing, tit i+.
do.

On the other hand, reading of the histogram data DM1 using the address signal β1 is executed in step 206. This step 206 corresponds to the address signal β1 in FIG.
is applied from the F/F 31 to the RAM via the switching unit 91, and the histogram data DM at address βl is applied to the switching unit 9.
This is done by storing it in the F/F 33 via 1. Then, in step 208, the gray value data D1 of the histogram data DM1 is added. This step 208 is
This is executed with ADD in FIG. Then step 20
Histogram data DMl-D Ml added in step 8
+D+ is written in the RAM at address β0 in step 210.

(2) To explain the accumulation of histogram data in more detail, refer to FIG. 13.

FIG. 13 shows the concept of the histogram memory using the RAM 34 of FIG. 2, and only one of the windows is illustrated for simplicity of explanation. As shown in the figure, the histogram memory has n RAM -RAM areas, and these are the rotation angles θ (-〇') to θ in the ρ-θ space of the n-n-1Hou transformation. On-1 corresponds to θ. And each RAM area is +511~
It has addresses β (-ρ) from 0 to -512, and each address can store 16-bit histogram data (shade values).

Therefore, in DDAl, the rotation angle θ is determined by equation (7) above.
α and β corresponding to are address signals! +1
1+1 1+1α1. β1
When the address signal β1 is calculated from RAM1
Histogram data D at address β1 given to
is read out. Then, the histogram data (DMi+DI) is added to the gradation value data D1 of the N(1) processing target point and written to the address β1 of the RAM again.

As described above, the calculation of the recurrence formula shown in equation (7) is performed by the Vibrine method by receiving and passing the address signals α and β1 one after another. And from this αl, β
, hiss 1 1+1 during operation to β
Since the 1+1 togerum data D is accumulated (accumulated) using the grayscale value data D1, the time required for the entire processing of one cycle can be shortened.

This one-cycle processing is performed in parallel by the DDA arithmetic circuits 18 to 18 constituting the DDA arithmetic unit 18. That is, when the data "■, empty, ■, empty, empty, ■, ■, empty" is input from the FIFO 17' as shown in FIG.
In the second cycle, it becomes as shown in (C) in the same figure, and in the third cycle, it becomes as shown in (d) in the same figure, and the same process is performed thereafter, and in the eighth cycle, it becomes as shown in (d) in the same figure. (e)
become that way. Here, (α, β) in the same figure (a)
~ (α, β) 01 0f 04 04 is the coordinate value (x y
)1 4 p1'91~(xp4"
F4) respectively corresponding address signals, and DI
t to DI4 are gray value data at each of the processing target points P1 to P4. Further, in (b) to (e) of the same figure, θ to θ are rotation angles from the position (angle) θ' at which the On-1 initial values (α, β.) were obtained, and are the RAM of FIG. 2, respectively.

- Corresponds to RAM.

Next, neighborhood filtering after Hough transformation is completed will be described.

Now, when the intersection of the Hough curves is expressed on the ρ-θ plane, it is assumed that it becomes as shown in FIG. 15(a). Note that (a) in the same figure shows a histogram H in which the intersection points appearing in each unit of the ρ-θ plane are weighted by the gray value data (change rate of brightness) of the pixels (processing target points) in the original image. , which is expressed as a contour scale to make the explanation easier to understand, and does not necessarily match the histogram according to the present invention.

Here, it is assumed that the histogram is high at point p1 in FIG.
There is a point p4. It can be seen that high histogram portions also occur in p5 and the like. However, what is particularly important in image processing is to find points p1 to p3 that are far apart from each other. For example, point p1 is located on the shoulder line of the road, and point p
2 corresponds to the center line, and point p3 corresponds to the shoulder line of the curved road ahead. On the other hand, the point p4 near the maximum histogram point p1. P5 etc. often correspond to partial bends in the road shoulder line, etc.
In terms of image processing, this mainly corresponds to a noise component.

Therefore, the influence of such noise components is reduced by, for example, 8-neighborhood filtering. In other words, Fig. 15 (
Prepare an 8-neighborhood filter like b) L, F −F
Compare the histograms of the intersections of the Hough curves for the area. Then, when F > F - F and F - F hold for the central area F5, the data of this area F5 is set as the data to be detected. Specifically, for example, when one unit (element area) pixel on the ρ-θ plane is assigned to F1 to F9, the number of histograms of intersection points is F 1-6 F 2-8F a -4F4- 2 F
5-14, F6-10°F 7” 7F g-9F
When it becomes 9 and 8, F > F −F SF
Since −F5 1 4 θ 9 holds true, the intersection point of F5 is set as the data to be detected. On the other hand, when F, -8F2-4 F3-3 F -14, F5-10, F6-7 F7-9 F8-8 F9-2, F5<F4, so the area of F5 is Not considered data to be detected.

The above filtering process is performed using a neighborhood filter as shown in FIG. That is, a signal readout circuit 71 is attached to the histogram memory (RAM 34) configured as shown in FIG.
) to a shift register 74 having a Then, the histogram data D of F-F and F6 to F9 of the shift register are transferred to comparators C-C and C-M.
148 and 09, and also inputs the histogram data DH of F5 to all comparators. Then,
When the histogram data in each area of the histogram memory 34 is DM1 to DM9 as shown in the figure,
The value of data DM5 is different from other data D-D, D-D
compared to And MI M4 MB
When M9 DM5 is at its maximum, the AND gate 75 outputs a peak signal as "1", and data DM5 at this time becomes peak data.

By performing the above filtering process, the 15th
In figure (a), point p4. The high points of the second and third histograms, p2., are unaffected by the presence of p5. p
3 can be detected.

That is, if the above filtering were not performed, the next high histogram point after the first high histogram point p1 would be point p4. p5, and the point p2. which is desired to be obtained as the second and third high histogram points. p3 becomes the fourth and fifth high histogram points, making subsequent signal processing extremely difficult. It should be noted that this 8-neighborhood filter does not need to be provided with hardware that corresponds to the above-mentioned windows, and can be used for all windows with the same configuration.

Next, the sorting process shown as step 118 in FIG. 3 will be explained in detail with reference to FIG. 17.

FIG. 17 is a flow chart thereof. First, for sorting processing, a plurality of input memories (transfer memories) M1
□~M+4 and comparison memory (result memory) MMl~MM4
is prepared and initialized (step 152). Next, step 154 is performed to input data into the input memory M11.
), when this input data is determined to be present in step 156, steps 158 to 184 are executed, and when it is determined to be absent, steps 190 to 199 are executed. here,
The processing in steps 190, 196 and 199 is the same as the processing in steps 158 and 160, respectively, and
2. The processing in step 198 is the same as the processing in steps 162 to 168, and the processing in step 194 is the same as that in step 1.
This is the same as the processing in steps 70-176.

In steps 158, 162, 170, and 178, the contents of the corresponding input memory M1 and comparison memory M are compared, and when M1≦MM, the wing transfers the contents of the human memory M1 to the next one (step 16).
4,172,180). On the other hand, when Ml>M, the contents of comparison memory MM are transferred to the next input memory M1 (step 166゜174.1).
82) Together, put the contents of the input memory M1 into the corresponding comparison memory MM (step 168.176.184).

Then, finally, comparison memory MMl=MM. for,
n pieces of input data are held in ascending order.

This is specifically shown in FIGS. 18 and 19. First, as shown in FIG. 18(a), four memories MIt to M+4 are prepared as input memories and four memories MMl to MM4 are prepared as comparison memories, and these are paired to form a four-stage circuit. As shown in FIG. 18(b), each stage of circuitry is composed of a pair of input memory M and comparison memory MM, switching circuits 81 and 82, and a comparator 83 for controlling them. Input memory M! When the histogram data input to is larger than the data stored in the comparison memory MM, the output of the comparator 83 causes the switching circuit 81 to change as shown by the solid line in the figure, and the input data is stored in the comparison memory MM. . at the same time,
Since the switching circuit 82 is also shown as a solid line,
The data stored in comparison memory MM is sent to the next stage. On the other hand, when the input data (input memory M,) is smaller than the data stored in the comparison memory MM, the switching circuits 81 and 82 are controlled by the comparator 83.
is as shown by the dotted line in the figure, the input data is sent as is to the next stage, and the contents of the comparison memory MM do not change.

In such a sorting section, it is assumed that the input data are 10 pieces of data "5, 7, 2, 8, 4.9, 3, 1, 6.8" as shown in FIG. 19(a). Then, after performing the initialization operation as shown in FIG. 5(b), the data stored in the comparison memories MM1 to MM4 change as shown by the arrows in FIG. The contents of Ml-9 to MM2"'8, MM3-8", MM4""7 will be stored. Note that this sorting process also does not require the provision of hardware compatible with windows. Further, this process may be executed by software.

By executing the series of processes as described above, all steps of signal processing by the image processing apparatus according to the present invention are completed. Then, an approximate straight line of a curve connecting the processing target points of the original image is determined for each window based on the values of ρ and θ.

The present invention is not limited to the above embodiments, and various modifications are possible.

For example, the window settings may be predetermined or may be changed depending on the processing results. Further, a window data generation section may be configured outside the FIFO.

When obtaining a histogram of intersection points of Hough curves, by superimposing binarized data instead of gray value data regarding the rate of change in brightness (value of differentiated edged data), only the concentration of intersection points can be obtained as a histogram. It can also be understood as Furthermore, instead of converting the data into edge data through differentiation, the data corresponding to the luminance may be digitally processed as gradation value data, and then the histogram of the intersection of the Hough curves may be obtained.

It is not necessary to calculate the rotational motion recurrence formula for the entire circumference (-circumference) of the approximate circle, but for 1/2 circumference, 1/2 circumference, etc.
It may be 4 laps or 178 laps. For example 1/4
If the calculation is performed by arranging DDAs in series that perform circumference calculations for O≦θ×π/2 and π≦θ×3π/2, then other circumferences (π/2≦θ×π, 3π/2≦θku2
π) can be immediately determined from these. Further, it is not necessarily necessary that the rotation angles are always the same, and it is not impossible to make them partially different. Also, due to the nature of the straight line, 1/2 circle (0° ~ 180") is Ho
It may also be a target of ugh conversion.

The specific configuration of the DDA that calculates the rotational motion recurrence formula is as follows.
It may be as shown in Figure 0. That is, six adders 51 to 56 and four inverters 61 to 63.99 and 2
It consists of 1/6 dividers 65°66. According to this DDA, the above-mentioned equation (8) can be executed. Alternatively, as shown in FIG. 21, it may be configured with two adders 51, 52 and one inverter 61. If you do this,
The calculation of equation (5) above can be executed.

〔Effect of the invention〕

As described above in detail, according to the present invention, a calculation unit can be easily formed by simply connecting DDA calculation circuits with the same configuration in series, and there is no need to refer to a memory table or the like during calculation. becomes significantly simpler and smaller. In addition, by configuring the arithmetic circuit to execute the rotational motion recurrence formula using the Vibrine method, it is possible to speed up the calculation. Furthermore, since the original image is divided into multiple windows, it is extremely easy to grasp global trends.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing the overall configuration of an image processing device according to an embodiment of the present invention, FIG. 2 is a block diagram showing the configuration of the main parts of FIG. 1, and FIG. 3 shows the overall operation. 4 is a diagram illustrating an example of an image and edge detection of an input pixel signal, FIG. 5 is a flowchart illustrating preprocessing of edged data, and FIG. 6 is a diagram illustrating an example of window settings. , FIG. 7 is a diagram explaining a lookup table, FIG. 8 is a diagram explaining Hough transformation in an embodiment of the present invention, FIG. Figure 10 is a circuit diagram showing the specific configuration of the DDA that calculates the rotational motion recurrence formula,
FIG. 11 is a diagram showing a switching configuration based on window data, FIG. 12 is a diagram explaining one cycle processing, FIG. 13 is a conceptual diagram of a histogram memory, and FIG. 14 is a diagram explaining pipeline processing in an embodiment. , FIG. 15 is a diagram explaining the 8-neighbor filtering process, FIG. 16 is a diagram showing the specific configuration of the neighborhood filter, FIG. 17 is a flowchart explaining the sorting process, and FIG. 18 is the specific configuration of the sorting section. FIG. 19 is a diagram specifically explaining the sorting process, and FIGS. Figure 22 is a diagram explaining road recognition.
3 to 25 are diagrams illustrating conventional Hough transformation. 1... Camera image, 2... Horizon, 3... Road,
4... Road shoulder line, 5... Center line, 11.
. . . Camera, 12 . Preprocessing section, 17'... FIFo, 18... DDA operation section and histogram memory, 18-18... DDA operation circuit On-1 path and histogram memory, 19... Neighborhood filter, 2
0... Sorting section, 21.23... VME bus, 22... CPU, 31, 32, 33.92... Flip-flop (F/F), 34-RAMo~RAM
(histogram memory), 35...ADD
-ADD (adder), 36...Nokuraf On
-1 A, 37...DDA ~D D A n-t (
DDA calculation circuit), 40... initial value calculation section, 91
. . . switching section, 95 . . . window data generation section. Patent Applicant: Honda Motor Co., Ltd. Representative Patent Attorney Yoshiki Hase Pre-processing flowchart Figure 5 Input data (a) Input data without LUT (b) Lookup table with LUT (LUT) Figure 7 Function diagram Diagram Diagram? l-7. Concept of togeram memory Fig. 13 Fig. Explanation of sorting Fig. 19 (2) Diagram L2 Road recognition Fig. 22 Fig. Explanation of Hough transformation (II)

Claims (10)

[Claims] 1. In an image processing device that separates and extracts a curve connecting processing target points from characteristics of the distribution of a plurality of processing target points on an original image captured by an imaging means, the dividing means divides the original image into a plurality of windows. a DDA calculation means configured by connecting a plurality of DDA calculation elements of the same configuration in series and performs a predetermined calculation on the processing target point; and a calculation result of the DDA calculation means is transmitted to the window in correspondence with the DDA calculation element. An image processing apparatus comprising: a storage means for storing each window; and a means for classifying and extracting the curve for each window based on the stored contents of the storage means. 2. An image processing device that derives an approximate straight line of a curve connecting a plurality of processing target points on an original image captured by an imaging means, comprising: a dividing means that divides the original image into a plurality of windows; Let the coordinates of one point on the circumference of the approximate circle drawn be (α_i, β_i), and let the rotation angle to the coordinates (α_i_+_1, β_i_+_1) of the next point on the circumference be ε (however, i is positive). ), α_i_+_1=f_α(α_i, β_i, ε) β_i
a DDA calculation means that sequentially calculates a rotational motion recurrence formula such as ___+_1=f_β(α_i, β_i, ε) for each predetermined rotation angle with respect to the processing target point in a pipeline manner; A storage means for storing at least the value of β_i for each window among the respective results calculated in
An image processing apparatus comprising: approximate straight line deriving means for deriving the approximate straight line for each window from the intersection of the ugh curves. 3. 3. The image processing apparatus according to claim 2, wherein the dividing means outputs window data based on the coordinates of the processing target point on the original image. 4. 3. The image processing apparatus according to claim 2, wherein the DDA calculation means is configured by connecting in series a plurality of calculation circuits that calculate the rotational motion recurrence formula for each rotation angle. 5. The rotational motion recurrence formula sets the rotational angle to ε=2^-^
When m, α_i_+_1=α_i(1-2^-^2^m^-^1
)-2^-^mβ_iβ_i_+_1=2^-^mα_
The image processing apparatus according to claim 2, wherein i+β_i(1-2^-^2^m^-^1). 6. 3. The image processing apparatus according to claim 2, wherein the storage means stores at least the value of β_i for each window in correspondence with the rotation angle. 7. 3. The image processing apparatus according to claim 2, wherein the storage means stores gradation value data corresponding to the brightness or the rate of change of the brightness of the processing target point in the original image for each window. 8. 7. The image processing apparatus according to claim 6, wherein the storage means stores the gradation value data of the processing target point in the original image for each window in correspondence with the rotation angle. 9. 3. The image processing apparatus according to claim 2, wherein the approximate straight line deriving means includes a neighborhood filter that performs neighborhood filtering processing on data regarding the intersection of the Hough curves. 10. 8. The image processing apparatus according to claim 7, wherein the approximate straight line deriving means includes a neighborhood filter that performs neighborhood filtering processing on data obtained by superimposing the gradation value data on data related to the intersection of the Hough curves.