JPS59111576A - Window size determining method in object recognizing system - Google Patents

Abstract

PURPOSE:To execute efficiently a labeling processing in case of an object recognizing processing by deriving a dimension number from a progression for showing a corresponding relation of an inter-feature point distance and a distance rank model, and determining a window size basing on its dimension number. CONSTITUTION:With regard to an object to be detected, a coordinate of each feature point Q1-QN is calculated. A distance rank space of the N-th dimension, which uses each distance rank in a distance rank model as an axis is supposed. Subsequently, with regard to a point Q in this distance rank space, a distance rank vector V is derived. A dimension number of the distance rank vector is initialized to (q) and a partial different value of the distance rank vector between the feature points is calculated. A distance rank matrix is derived by using this partial different value as a criterion, and compared with a unit matrix. As a result, a window size (w) is determined from the distance rank by the dimension number (q).

Description

[Detailed Description of the Invention] <Technical Field of the Invention> The present invention focuses on points (hereinafter referred to as "feature points") that are related to the shape characteristics of an object to be detected, such as corners of an object, and
By associating feature points of the detected object with arbitrary orientations with specific feature points in the detected object model within a window of a predetermined field of view and assigning a label (hereinafter this process is referred to as "labeling"), In relation to an object recognition system that recognizes the position, posture, etc. of a detected object, the present invention particularly provides a window size determining method for setting the size of the window to the necessary minimum size.

<Background of the Invention> In recent years, in this type of object recognition system, a method has been proposed in which, when labeling feature points of a detected object, a window having an appropriate field of view is set. If such a window size is unnecessarily large, data processing becomes complicated and inefficient, while if the window size is small, labeling of feature points becomes impossible. For this reason, it is necessary to set the window size to the minimum necessary, but conventionally, labeling of feature points is executed with an appropriate window size, and when labeling is impossible, resetting the window size etc. The process is re-performed, which causes waste in the labeling process, leading to problems such as a significant drop in processing efficiency.

<Objective of the Invention> The present invention provides a novel window size determination method in which the window size is preset before performing labeling of feature points, thereby improving the efficiency of labeling processing in object recognition processing. With the goal.

<Configuration and Effects of the Invention> In order to achieve the above object, a distance rank model is formed in advance by arranging distances between each feature point of a detected object model and other feature points, and For the distance between each feature point of an object and another feature point, the correspondence relationship with the distance rank model is expressed as a number sequence, and after finding the number of dimensions where all the number sequences are inconsistent with each other, this number of dimensions is The window size is now determined based on this.

According to the present invention, the minimum required window size that is optimal for labeling processing can be set in advance, eliminating the need for resetting the window size or reprocessing labeling as in the past, and improving the efficiency of labeling processing in an object recognition system. It has excellent effects that achieve the purpose of the invention, such as a significant improvement.

<Explanation of Examples> Fig. 1 shows a conveyance line for objects to be detected, in which objects to be detected 2 on a belt conveyor 1 are photographed one by one by a robot 3 installed at the downstream end of the conveyance. , and then transferred to the next treatment or processing step.

Fig. 2 shows a detected object model 2A, which has a total of 18 feature points P1 + P2 +... forming a value angle in a planar shape.
..., has PIB.

The objects 2 to be detected are facing in arbitrary directions on the belt conveyor 1, and the fingertip portion 4 of the robot 30 rotates according to the orientation of each object 2 to be detected, and places the finger at a certain point on the object 2. Piece 4a.

4b, grip the object 2 to be detected, and photograph it.

The robot fingertip section 4 is checked by the object 2 to be detected using the device shown in FIG. 3 according to the present invention, and its rotation direction, angle, etc. are controlled. In FIG. 3, a camera device 5 is disposed upstream of the belt conveyor 1, and images the planar shape of the object 2 to be detected, and detects a plurality of feature points included within a predetermined window.

, In Fig. 4, a window W having a square rectangular field of view is shown by a broken line, and in the illustrated example, a total of two feature points Q, ,
Q5 is included within the field of view of window W. Camera device 5
The output is sent to the arithmetic control means 7 (hereinafter simply l'-CP
The positions of the respective feature points are stored in the image memory 8 as coordinate data. The CPU 7 controls reading and reading of such data, decodes and executes various programs related to object recognition, checks the orientation of the detected object 2, and controls the operation of the robot control device 9 based on this.

In addition to the above programs, the memory 10 in FIG.
All feature points P1 to PLO of the detected object model 2A and reference data regarding distances between the feature points (hereinafter referred to as distance rank model) are stored.

FIG. 5 shows the configuration of the distance rank model.

The distance rank in the figure refers to all the distance values between the feature points on the detected object 2A in descending order of dl.

d2. ...+di+..., dN.

In addition, the label set means a set of feature points associated with each distance rank, and the set is χ(d+).

χ(d2),...,χ(di),...,χ(d
N).

The present invention provides a new method for setting the size (in this embodiment, the length and width) of the window W prior to the feature point labeling process. It is.

First, in step 11, the CPU 7 calculates each feature point Q+ + Q for the detected object 2 imaged by the camera device 5.
2 Calculate the coordinates of +...+Qi+...+QN.

By the way, the distance rank in the distance rank model is d
, , d2. ..., mountain, ..., dN (however, d+<
Assuming d2<...(di (...<dN), an N-dimensional distance rank space having each distance rank as an axis is assumed.

Next, if we consider a distance rank vector ■ (hereinafter expressed as V) for a point Q in this distance rank space, the vector component of ■ is (vl.

V2. ... r vN ). in this case,
Each element vi of the vector component is set to 0 if it does not have a corresponding distance rank in the distance rank model.

Thus, in step 12, the CPU 7 selects each feature point Qi (
i=1.2. ..., N), the distance rank vector Vi is determined. This Vi can be expressed as follows.

Vi= (V(i, 1) l V(i12), ---
-1V(i IN)) Next, in step 13, the number of dimensions of the distance rank vector involved in determining the window size is initialized to 9 (however, 1≦9≦N), and then in step 14
, the CPU 7 calculates the partial difference value D(r, s: q
) is calculated using the following formula.

In the formula 0, v(r, k) and v(s, k) are Vr,
Vector components of Vs (v(r, ■), v(r, 2),
−, v(r, N) )(v(s, 1), v(s, 2),
, v(s+N)).

Also, the sign represents exclusive OR, and the sign - represents the product set of =1 expressions.

Therefore, D(r, s; 3)-1・1・0-0. This means that Vr and Vs have different components under the dimension number 9-3 of the internal a rank.

Thus, in the route of steps 14, 15, and 16, all the feature points Qlr Q2 , . . . +Qr+ ., Q
s l−, QN (however, r+5=L2+ ”・+N)
Accordingly, the partial difference value D(r, S; 9) is calculated. As a result, the determination in step 15 becomes "=YES", and the next distance rank matrix [R(q)] represented by a matrix of N rows and N columns is determined.

Then, in step 17, the CPU 7 checks whether the distance rank matrix CR(q)) satisfies the following equation.

However, [E] indicates a unit matrix.

In the above, when (R(q))': (E), the partial dissimilarity value D(r, s; 9) = 1 (however, rNS) in the matrix
In this case, the judgment in step 17 is "NO"', and in step 18, the number of dimensions is 9 and 1 is included.
After the addition, similar processing to steps 14, 15 and 16 is performed. and (R(q)) −(E
), the process proceeds to the next step 19, and the CPU 7 calculates the window size W using the following equation.

W-2・d9・(1+α)・・Song ■In addition, in the 0 formula,
d9 is the ninth distance rank, and α is the distance rank d
In addition, FIG. 7 shows the relationship between the window size W and the area X9 of distance rank d9.

However, after determining the window size using the formula 0, the CPU 7 executes a singular point labeling process within the field of view of the window.

[Brief explanation of the drawing]

FIG. 1 is an explanatory diagram showing a conveyance line for an object to be detected, FIG. 2 is a plan view of a model of an object to be detected, FIG. 3 is a circuit block diagram of an example of a device implementing the present invention, and FIG. 4 is an illustration of an object to be detected. 5 is an explanatory diagram showing the distance rank model, FIG. 6 is a flowchart showing the window size determination method according to the present invention, and FIG. 7 is a diagram showing the window size and distance rank area. FIG. 2...Detected object 2A...Detected object model W...Window W...Window size 494- Te/times 30 Square σ Figure 1 495= Oq 100 edition

Claims (1)

[Claims] In an object recognition system that assigns a label of a corresponding feature point of a detected object model to a feature point within a window of a predetermined field of view in an image of a detected object, the label of the corresponding feature point of the detected object model is A distance rank model is formed in advance by arranging each distance, and the correspondence relationship between the distance rank model and the distance between the feature points of the detected object and other feature points is calculated using a numerical sequence. 1. A method for determining a window size in an object recognition system, the method comprising determining a window size based on the number of dimensions after determining the number of dimensions in which the number sequences of are mutually inconsistent.