JPS58123462A - Computing method for speed pattern - Google Patents

Abstract

PURPOSE:To enable to perform a uniform computation of a speed pattern of a whole surface of a picture, by a method wherein a mathematical operation such as three- dimensional lightness gradients, vector products, is conducted in a partial and a parallel manner on a three-dimensional picture which is constituted by aligning consecutive pictures in order of time. CONSTITUTION:A digital picture consisting of 3NXN picture elements photographed at sufficiently short intervals is stored in a picture memory 1, and the address of the memory cell are P(i, j, k)(i=1-N, j=1-N, k=1-3). A picture element in a cube of 5X5X3 centering around a point of a coordinate value P(i, j, 2) of the picture memory 1 is transferred to a resistor 2 to store it by a formula of Q(l, m, k) P(i-2+l, j-2+m, k). The contents of the resistor 2 are transferred to 8 resistors by a formula of Rst(l', m', k) Q(l'+(s-1), m'+(t-1), k). Three-dimensional differential computers 411-433 compute a lightness gradient vector in connection with contents of resistors 311-333. Vector product computers 51-54 compute a vector product between the outputs of the respective three-dimensional differential computers. A vector product selecting part 6 selects and outputs (Xm, Ym, Tm), Em, having a maximum value, out of the outputs of the vector product computers 51-54. A condition deciding part 7 decides whether the outputs of the vector product selecting part 6 meet the conditions together. A speed computing part computes a speed.

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to a speed pattern calculation method. Conventionally, several methods have been proposed for measuring the movement of an object, which changes from moment to moment, using a plurality of continuously captured images (continuous images). Typical methods include correlation methods and differential methods that perform signal processing on original image information, as well as methods that process multiple images to some extent to extract morphological information such as line drawings, and extract morphological information between individual images. Those who compare information law specialists, etc. are different.

Furthermore, there are methods that calculate optic flow based on temporal changes in image brightness. However, methods that use morphological information take time to extract, and are often difficult for complex objects including curved surfaces. The difference method calculates the difference between pixels located at the same coordinates in two consecutive images, and detects motion based on the fact that the difference image is not zero. This method is simple and can be carried out to a great extent, but in order to find the movement pattern, it is necessary to analyze the difference image and further the morphological information in the two original images. In addition, the correlation method compares two consecutive partial images using the correlation value as an evaluation function, but it has the disadvantage that it is not suitable for high-speed execution because it requires searching for similarities that maximize the evaluation function. . The method using optical flow also requires repeated calculations to determine the flow, making it difficult to speed up the process.

The present invention has been made in view of the above, and does not involve comparing morphological information, searching for similarities, or repeating operations, but directly performs local parallel operations on original continuous images to create images on images. This method calculates the velocity at each point and obtains a velocity pattern. This method can be implemented as a dedicated parallel high-speed arithmetic unit, and has an extremely wide range of applications as a general-purpose motion measuring device.

The details of the invention will be described below.

Consider a five-dimensional space in which N consecutive images are arranged in time order, and each image has two coordinates in the horizontal direction, a V coordinate in the vertical direction, and a t coordinate in the depth direction (FIG. 1). If continuous mis is sampled at short time intervals such that the density in the time direction is sufficiently large, this space can be considered as a continuous space. The brightness of one point P (" r 'II r t ) in this space is I (x, y
, t), then the brightness gradient at point P is arad I = (''/ax, aI/ay,
aVat ) (1). Now, if the minute positional displacement is ΔV=(Δ2.ΔV, Δt), the brightness change Δ
can be approximated to the first order as follows.

II = arad I −X'
(2) If grad II and bar stop are respectively the five-dimensional vector gradI and the projection vector of the source onto the Z-y plane, then the minute displacement (I
2°Δy, 1) in equation (2), ΔI=0. Δt=
By setting it as 1, grad l note 6* = −(ax/at)
(3) is satisfied. Source* is the movement (velocity) that we are aware of on the image. At a certain time t, two neighboring points PL, P! in the same image If the speeds Xl* are equal, then the following holds true, and the total 9 can be found as follows.

Sawa*= (x/T, Y/T) (
s) However, (X, Y, T) Ei7? 'cLd1. Xi7? ”
Gd1. (6) 'l' = det (gra
d engineering, grad II) '< 0 (
7).

This principle is applied to digital 1iir images. Time'O
+ 'I r...*' Mut the image group obtained by N-1
II m...

Let it be lN-1. brightness gradient image grad in time
Engineering 4 is C4-1, ta, jA+I K Okel 5
The Om images Ia-t, Ia, and Ia+t are obtained using an optimal edge detection operator for five-dimensional images. That is, for a point P in IA, 27 lightness values of a small cube of 5 x 3 x 3 points centered on that point and the following matrix M are used.

Time differential image (a'IAt) a is M and engineering 4+! It is determined by the difference between the convolution integral of and the convolution integral of M and I4 i . (aVBye) a, (”/ay)
Each differential image of a is obtained by performing the calculation of (''/at) a by rotating it in space (FIG. 2).

The vector product (X
, Y, T) are obtained from the brightness gradient image as follows.

First, as shown in Figure 5, a set of target points on the same screen centered on point P on In (Q%).

Rm) (m='o~5). brightness gradient image gr
Since the brightness gradients at Qm and flm are known from ad engineering 4, the vector product of both (Xm + Ym +
Tm ) is calculated. Let (Xm, YmrTm) having the maximum absolute value among 0 to 3 be the vector product (X, Y, T) at point PK. The velocity source * at point P is determined by equation (5). At this time, the condition of Too is interpreted as follows.

ITI≧conet, (9) Here, θ is two two-dimensional vectors grad il and grad IF (!: f) Nass angle W and −rtta
', Tm l grad 1. ”1*l grad
It's le sing. That is, the magnitude of T is determined depending on the magnitudes of both two-dimensional vectors and the angle they form. When calculating the speed, the larger θ is, the more reliable the value can be obtained. Therefore, we add the following condition.

Title ≧ at, (to) However, E
=Ig? '4dLI"1ffadI*I C11
const, , conat, are appropriately selected nine thresholds.

In this way, among all the points on the image I1 at a certain time, the velocity component * can be obtained from the equation (5) for those that satisfy the conditions of the equation (9) and (g).

In other words, the speed pattern over time can be determined.

In the above explanation, when calculating the brightness gradient at time t4, five images in 'Al t 'A HeA+* are used.
+ "A + 'Todoroki+1+'Todoroki+!' may be used. Also, although the matrix M shown in equation (8) was used, the brightness gradient of the five-dimensional image in equation (1) can be It is also possible to use something that can be calculated with good approximation, for example, the following M', which is a simplified version of M in equation (8).

An example of the configuration of speed pattern needle #1 that satisfies the method of the present invention as described above is shown in FIG.

The image memory l consists of 95 Nx images taken at sufficiently short time intervals.
It stores a digital image consisting of N pixels,
The address of the memory cell is p <i+irh> (
i-1 to N, j-1 to N, C=1 to 5). The register has a capacity of 5×5×5 pixels, and each cell is Q (j, vx, &) (1=1~5.m=1~5.
k −1 to 5).

Pixels in a 5×5×5 rectangular parallelepiped centered on a point with coordinate values P (s, j, z) in image memory 1 are transferred to a register. That is, store it as follows.

Q(1,m,k)+-p(i-2+l,j-2+fFL
, &) QJ registers 511 to 335 are each 5X5X
It has a capacity of 5 pixels. Let each cell be Rat (l′p
m', #) (11=1~5.t-1~5+
(8+')"q(24)+"=1~s, j'=1~
5. −1 to 5). Transfer the contents of the registers to eight registers as follows.

R, gt(J', m', S+-Q (J'+(Jl-j
), m, '+u-1), &) (The 145-dimensional dimension calculators 411 to 455 calculate the contents of the registers 511 to 553, respectively, using M in equation (8) as shown in FIG. Calculate the brightness gradient vector. Vector product calculator 5
1 to 54 are five-dimensional machine calculation units 411 and 455, respectively.
: 412, 452: 413, 451: 4
Calculate the vector product between each output of 23°421. At this time, in addition to the vector product (Xyn t Ym + Tm), the value Em corresponding to equation (b) is also determined.

The vector product selection unit t has the maximum absolute value among the outputs of the vector product operators 51 to 54 (x''r'1m,
Tm) and Em are selected and output. The condition determination unit 7 determines whether the output of the vector product selection unit B satisfies both the conditions of equations (9) and (A). The speed calculation part is (
5) Calculate the speed using the formula.

If pixels are sequentially read out from the image memory l, the speed of each point can be calculated.

As explained in detail above, the present invention provides continuous ijig
The velocity pattern is obtained by performing mathematical operations such as 5-dimensional brightness gradient and vector product locally and in parallel on a 3-dimensional image constructed by arranging I in time order. This method has extremely advantageous technical effects, such as being able to uniformly calculate the speed pattern over the entire surface of the image without any problems, and also making it possible to increase the speed using a dedicated processing device.

[Brief explanation of the drawing]

Figure 1 is an explanatory diagram of a 5-dimensional intermediate structure constructed by arranging N images taken sequentially in time order, and Figure 2 is an explanatory diagram of a 5-dimensional intermediate.
FIG. 3 is an explanatory diagram showing a calculation to obtain a dimensional brightness gradient from a digital image, FIG. 3 is an explanatory diagram of a set of target points for calculating a vector product, and FIG. 4 is a schematic configuration diagram of an example of a speed pattern measuring device. In the figure, 1 is an image storage, 411 to 435 are five-dimensional mechanical calculators, 51 to 54 are vector product calculators, t is a vector product selection section, 7 is a condition determination section, and 1 is a speed calculation section. Designated Agent Agency of Industrial Science and Technology Figure 3 QoOOP @R. Q1Q2 Q:3

Claims (1)

[Claims] A 5-dimensional gradient vector is calculated on a 6-dimensional image constructed by arranging multiple images taken sequentially in time order, and a 5-dimensional gradient vector is calculated between two neighboring points that are considered to be moving at the same speed. A speed pattern calculation method characterized by evaluating speed by calculating a vector product between degree gradient vectors.