JPH0523463B2 - - Google Patents

Description

[Detailed Description of the Invention] [Field of Industrial Application] The present invention relates to a recognition device that performs image processing on an object to recognize the object, particularly an object in which a plurality of objects touch or overlap each other (hereinafter also referred to as a superimposed object). The present invention relates to a recognition device capable of recognizing even .).

[Conventional technology]

There are various methods for such recognition, one of which is known as the Sternford algorithm developed at Stanford University in the United States. The outline will be explained below.

First, the target object is binarized into white and black levels, and this binarized data is encoded by run length (length of white run or black run).
Next, connectivity analysis is performed on the run-length data to calculate various feature amounts. Furthermore, many more useful feature parameters are calculated based on this basic feature calculation. Then, recognition is performed using these feature amounts and parameters as clues, but prior to this, standard sample learning is performed. In other words, each object is imaged independently using an imaging device such as a television camera, the images are processed to calculate various feature amounts and parameters, and after statistical processing, the average value, standard deviation value, etc. are calculated. Then, an operation is performed to store the data in a predetermined memory as a dictionary. Note that this operation is performed on all target objects. Next, feature quantities and parameters are obtained for the unknown object, and these are compared with those of a standard sample stored in memory, and when the difference is less than a certain tolerance value, the unknown object is identified. At this time, appropriate parameters are selected and used in consideration of the shape and complexity of the target object, the type of the target object, as well as identification accuracy (accuracy), processing time, etc.

[Problem that the invention seeks to solve]

However, since such a method is based on the assumption that individual objects are identified in a state where they are separated from each other, multiple objects may come into contact with each other, or some of them may be partially separated as shown in Figure 16. There was a problem that if they overlap, recognition becomes impossible or misjudgment occurs. Oh, Figure 16 is an example of a case where a circle and a rectangle overlap.

Therefore, an object of the present invention is to enable accurate recognition of unknown objects including superimposed objects.

[Means for solving problems]

A process is performed in advance to trace the outline of each of the plurality of types of objects to be recognized, divide it into primitives such as straight parts and arc parts, and register the attributes of each and the attributes of each vertex as a dictionary in a dictionary memory, After that, the unknown object is divided into primitives like a dictionary, and the unknown object is described using the attributes of the primitives and vertices and stored in a predetermined memory, and when the unknown object is recognized from these data,
Dictionary searching means searches for primitives to be matched with each primitive of an unknown object from a dictionary memory, and extracts all matched sets as a matching set; A correspondence table creation means that creates a matching correspondence table by checking the degree of matching with those of an unknown object, and a means for creating a matching correspondence table that estimates primitives or vertices hidden in the unknown object from this correspondence table and calculates the primitive length. Hidden part estimating means corrects and checks the degree of matching with the dictionary and extracts the optimal one as the estimated object, and removes this estimated object from the unknown object and supplements the remaining part with any primitives and vertices. unknown object updating means for estimating whether the object is to be detected and updating the unknown object based on the result.

[Effect]

FIG. 1 is a block diagram showing the configuration of an embodiment of the present invention. In the figure, 1 is a contour tracking section, 2 is a curvature calculation section, 3 is a primitive description section, 4 is a dictionary creation section, 5 is a dictionary search section, 6 is a matching section, 7 is a hidden part estimation section, and 8 is a In the unknown object update section, M1 is a dictionary data storage section, M2 is an unknown object data storage section, and ID is an estimated object data output section.

First, a dictionary is created (learning process).

That is, image data input via an imaging device (not shown) is converted by the contour tracking unit 1 into a sequence of boundary point coordinates on the contour. Note that the contour tracking is performed by, for example, the well-known Freeman coding. Curvature calculation section 2
calculates the curvature at each boundary point, divides the boundary point sequence at its maximum point (vertex), and extracts it as a primitive. The primitive description unit 3 classifies the primitives obtained in this way according to their average curvature, and those whose average curvature is larger than a certain threshold are classified as arc parts, and those with smaller average curvatures are classified as straight parts. The dictionary creation unit 4 examines the attributes (type, length, radius, center coordinates, starting point vertex number, end point number) of each primitive, and also checks the attributes (type, coordinates, angle, input primitive number) of vertices. ., output primitive No.) and store them in the storage unit M1 as a dictionary.
Here, the starting point vertex number and ending point vertex number of the primitive are used to describe the adjacency relationship of the primitives. Further, an input primitive is a primitive whose end point is the vertex, and an output primitive is a primitive whose start point is the vertex.

Next, the unknown object is divided into primitives in the same manner as described above, and the object is described using primitive and vertex data.
Data regarding the vertices is stored in the storage unit M2.

The dictionary search unit 5 checks in order from the longest primitive among the primitives of the unknown object whether there is a primitive that matches it in the dictionary. Then, the obtained set of the primitive in the unknown object and the primitive in the dictionary is sent to the matching unit (correspondence table creation unit) 6 as a matching set. At this time, if there are a large number of matching primitives in the dictionary for one primitive of the unknown object, there will be matching sets corresponding to that number.

Next, the matching unit 6 associates the dictionary pattern with the primitives and vertices on the unknown object for each of the matching sets sent from the dictionary search unit, and writes the results into a matching correspondence table. At this time, the degree of matching is defined as the error between the data of the dictionary pattern and the data associated with the unknown object, normalized by the data of the dictionary pattern. Among the matching sets, the matching correspondence table of one set having the highest matching degree is sent to the calm part estimating section 7.

In the matching correspondence table, the primitives of the dictionary pattern, the primitives corresponding to the vertices on the unknown object, and the places where there are no vertices are considered to be hidden by other objects, so the hidden part estimation unit 7 hides them. Estimate the portion that has been
Estimation is done one vertex or one primitive at a time. If the vertex is hidden, the position and angle of the vertex are estimated by extending the primitives at both ends of the vertex and finding their intersection, and the lengths of both primitives are corrected. On the other hand, when a primitive is hidden and what is classified as one primitive in the dictionary is hidden by another object in the unknown object and divided into two or more, the extension or connection of both primitives. The primitive length is corrected by . As a result of this estimation, if an object that matches the pattern in the dictionary can be constructed, this is output as the estimated object to the output unit ID. If the estimation does not match the dictionary pattern, the process returns to the matching section 6, which fetches a matching correspondence table of combinations with a high degree of matching, and estimates the hidden portion again.

When the hidden part estimator 7 outputs an estimated object in the dictionary, the unknown object updater 8 functions to remove this estimated object from the unknown objects. The unknown object update unit 8 removes primitives and vertices that have correspondence with the dictionary from the unknown object, estimates what should be filled in the remaining blank space, and updates the vertex angles, primitive lengths, and connection relationships. etc. will be corrected. Then, this corrected or updated object is input as a new unknown object, and the same process is repeated from the dictionary search section 5. Finally, if an unknown object that matches the dictionary remains, the entire process ends.

〔Example〕

FIG. 2 is a block diagram showing a recognition device to which the present invention is applied. In the figure, 10 is a target object,
11 is an imaging device such as a television camera; 12 is a preprocessing circuit including an amplifier and a binarization circuit; 13 is a feature extraction circuit for extracting various feature quantities of an object; 14;
1 is an image memory, 15 is a processing unit (processor) consisting of a microprocessor, etc., 16 is an interface circuit, 17 is a terminal consisting of a keyboard and display, and M is a data memory, which is the storage section M in FIG.
1 and M2. That is, the image signal (video signal) of the target object 10 obtained through the television camera 11 is binarized by the preprocessing circuit 12, and further converted into an image as segment (corresponding to run length) information by the feature extraction circuit 13. The data is written to the memory 14. This data is processed by the processor 15 to create a dictionary and describe the unknown object, and then stores it in the memory M. Thereafter, by performing the processing described in FIG. 1 based on this data, it becomes possible to separate and recognize individual objects from the unknown heavy objects. Note that by inputting keys from the terminal 17 via the interface circuit 16, it is possible to change the binarization threshold value, etc., and it is also possible to view the processing results on the display.

Although the outline of its operation is as explained in FIG. 1, the operation of each characteristic part will be explained in detail below. Note that the dictionary patterns and their data obtained through the contour tracking unit 1, curvature calculation unit 2, primitive description unit 3, and dictionary creation unit 4 in FIG. 1 are shown in FIG. 3 (dictionary 1) and FIG. 4 (dictionary), respectively. 2)
5 (Dictionary 3), and a case where an unknown heavy object as shown in FIG. 6 is identified will be explained as a specific example. Note that Figure 6A is
FIG. 6B is an explanatory diagram for explaining the primitive data shown in the figure, and FIG. 6B is an explanatory diagram for explaining the vertex data.

First, the dictionary search section will be explained.

The dictionary search unit searches the dictionary for a primitive that matches the longest primitive of the unknown object. The matching conditions are
It is as follows.

- For straight line primitives: Primitives whose side lengths and angles at both ends match - For circle or arc primitives: Primitives whose radii match. Matching here includes cases where the difference is within the tolerance range. If a matching primitive is not found for the longest primitive in the unknown object, a search is performed for the next longest primitive. For all the primitives in the dictionary obtained by the search, the matching unit creates a matching set (dictionary number - primitive number in the dictionary, unknown object primitive number (PR No.), matching degree) expressed as follows. send to

The degree of matching M _{i -jk} when the i-th primitive in the unknown object and the k-th primitive of dictionary j are matched is calculated using the following equation.

M _i-jk = {1-|L _i −L _jk |/L _jk } ×{1-|R _i −R _jk |/R _jk } ×{1−|T _i ^- −T _jk ^- |/T _jk ^- } ×{1−|T _i ⁺ −T _jk ⁺ |/T _jk ⁺ } …(1) Here, L _i is the i-th primitive length of the unknown object,
R _i is also its radius, T _i ^- is also its starting point angle,
T _i ⁺ is also its end point angle. Furthermore, L _jk is the k-th primitive length of dictionary j, R _jk is its radius, T _jk ^- is its starting point angle, and T _jk ⁺ is its ending point angle. In equation (1), in the case of a straight line primitive, the second term = 1, and in the case of a circle or arc primitive, the third term = 4th term = 1.

The longest primitive in the example in Figure 6 is length
140 arc primitive 7. If you search for a primitive that matches this in the dictionary,
Primitive 1 (PR1) of dictionary 1 shown in Figure 3 is found, but the matching degree M ⁷ _1-1 at this time is M ⁷ _1-1 = {1-|R ₇ −R ₁₁ |/R ₁₁ } { 1-| _L7 - _L11 |/ _L11 }={1-|30-30|/30} {1-|140-188|/188}=0.74. Since there are no other primitives in the dictionary that match this primitive, the following matching set (dictionary 1 - PR1, PR7, 0.74) is sent to the matching section. In addition, X, Y in Figure 3 (b)
indicate the X and Y coordinates of the center.

FIGS. 7A to 7E are flowcharts for explaining the above-described operations in the dictionary search section.

That is, FIG. 7A is a partial flowchart for searching for the longest primitive. The process is started assuming that the primitive number of the unknown object is "1" and the longest primitive length is "L ₁ ", and the longest primitive is searched. Extract the number I _MAX and its length L _MAX . Also, FIG. 7B is a partial flowchart for explaining the operation when the longest primitive is a circle, FIG. 7C is a partial flowchart for explaining the operation when the longest primitive is an arc, and FIG. 7D is a partial flowchart for explaining the operation when the longest primitive is an arc. FIG. 7E is a partial flowchart for explaining the operation when the longest primitive is a straight line. FIG. 7E is a partial flowchart for explaining the operation when there is a matching set and when there is no matching set.
In these figures, i is the primitive number of the unknown object, j is the dictionary number, k is the primitive number of dictionary j, N is the total number of primitives of the unknown object, m is the total number of dictionaries, and N _j is the total number of primitives in dictionary j. each represents.

Matching section (correspondence table creation section) Here, for each matching set sent from the dictionary search section, we check how well the primitives and vertices in the dictionary are matched with unknown objects, and create a matching correspondence table. create.

Now, from the dictionary search section (dictionary j−PRk, PRi,
Suppose that a matching set M ⁱ _jk ) is sent.
This is the i-th primitive in the unknown object and the dictionary j
This means that the kth primitive of is well matched. The matching correspondence table is a list of corresponding primitives, vertices, and their matching degrees. Note that the definition of matching degree is different between primitives and vertices. That is, the matching degree MP ⁱ _jk of the primitive is MP ⁱ _jk = {1−|R _i −R _jk |/R _jk } ×{1−|L _i −L _jk |/L _jk } …(2) , the matching degree of the vertices is MA ⁱ _jk = {1−|T _i −T _jk |/T _jk } (3). Here, T _i represents the angle of the i-th vertex on the unknown object, and T _jk represents the angle of the k-th vertex of dictionary j.

First, the correspondence between primitives and vertices listed in the matching set is copied to a correspondence table. A matching table is a set of values (matching elements in the dictionary, matching elements on unknown objects, matching degree) for all elements in the dictionary (primitives, the sum of their vertices is called an element). be. If there is no element on the unknown object that corresponds to the dictionary element, enter 0 (zero) in the second and third terms in the matching table. This part will be the subject of estimation in a later step. In other words, in FIGS. 7A to 7E, the dictionary was searched based on the primitive of the unknown object, but here, the dictionary is searched for the unknown object.
The operation is shown in FIGS. 8A to 8K. First, since the i-th primitive of the unknown object and the k-th primitive of dictionary j are matched, it is determined whether dictionary j is a circle or not by setting i ₀ = i, k ₀ = k (see Figure 8A, ○ I). , see ○b).

A: When dictionary j is a circle, copy the contents of the matching group as is into the matching correspondence table. Next, search for an arc primitive whose radius and center match the i-th primitive on the unknown object, and if there is one, calculate the degree of matching between that primitive and the first primitive in dictionary j (a circle has only one primitive on the object). and add it to the matching compatibility table.

B: When the k-th primitive of dictionary j is a straight line B-1 In this case, proceed from Figure 8A to Figure 8B, and first create a matching set of three elements (start point vertex, primitive, end point vertex).
Divide them into 1 and write them in the matching table along with their matching degree (see Figure 8B, ○A, ○B, and ○C).

Now, for example, as a matching group (dictionary j
−PRk, PRi, M ⁱ _jk ) are sent. Here, primitive k of dictionary j is a straight line. Therefore, the fact that primitive k in dictionary j and primitive i on the unknown object are matched means that the three elements, the starting point angle, the primitive length (radius), and the ending point angle, are matched, so ( j−Ak ^- , Ai ^- , MA ^i- _jk ^- ): Starting point angle (j−PRk, PRi, MP ⁱ _jk ): Primitive (j−Ak ⁺ , Ai ⁺ , MA ⁱ⁺ _jk ⁺ ): Ending point angle Write in the matching correspondence table. Here,
Ak ^- and Ai ^- indicate the starting point of the k-th primitive and the starting point of the i-th primitive, respectively, and Ak ⁺ and Ai ⁺ indicate the ending point of the k-th primitive and the ending point of the i-th primitive, respectively. Then, each of k and i sent in the matching group at the beginning is
Leave them as k ₀ and i ₀ . This is used to check the completion of matching (see Figure 8A).

B-2 From the primitives matched by the matching set, check how many matches can be made in the counterclockwise direction. In this case, proceed from FIG. 8B to FIG. 8C.
For the dictionary, the primitive adjacent to the kth primitive in the counterclockwise direction is k=k mod N _j +1. Here, N _j is the total number of primitives in dictionary j. Furthermore, A modB is A mod B.
Represents the remainder when divided by . On the other hand, for the primitive adjacent to the i-th primitive in the unknown object in the counterclockwise direction, if the total number of primitives on the unknown object is N, then i=i modN+1, and the lengths of these two primitives are compared ( (See Figure 8C ○a).

B-2- Case of L _i <L _jk In this case, the process branches from FIG. 8C to FIG. 8D. At this time, the primitive on the unknown object corresponding to the k-th primitive of dictionary j is partially hidden by another object. Therefore, first add the following values to the matching table for L _i and L _jk (see Figure 8D).

(Dictionary j=PRk, PRi, MP ⁱ _jk ) After that, it is checked whether there are any remaining primitives corresponding to the k-th primitive of dictionary j on the unknown object. To check, variable ii=
Set i, update as ii=imodN+1 and trace the unknown object in a counterclockwise direction, searching for a primitive that is on the same straight line as the i-th primitive and whose distance between the starting points is less than or equal to L _jk (8th D Figure○Ro,
○Refer to C).

If it is not found up to a○ ii=i ₀ -1,
The process branches from FIG. 8D to FIG. 8E, and the search in this direction is terminated. And kfwd=
k, proceed to the next backward search B-ii-3) term ().

b○ If found, Γ If the distance between the i-th primitive start point and the ii-th primitive end point <L _jk , add (dictionary j - PRk, PRii, MP ⁱⁱ _jk ) to the matching correspondence table, and from ○ro in Figure 8D. Go back and continue exploring.

Γ When the distance between the i-th primitive start point and the ii-th primitive end point = L _jk , add (dictionary j - PRk, PRii, MP ⁱⁱ _jk ) to the matching correspondence table, set i = ii, and proceed to Figure 8C.

Γ When the distance between the i-th primitive start point and the ii-th primitive end point>L _jk , the process branches from FIG. 8D to FIG. 8E. In this case, the
The ii primitive is considered to be a primitive that corresponds to the k-th primitive of the dictionary j and another primitive overlapped with each other. Therefore,
This is divided and matching is performed only for the portion corresponding to the k-th primitive of dictionary j. Therefore, from the starting point of the i-th primitive
Find the coordinates of a point on the ii-th primitive that is separated by L _jk , set the distance between this point and the start point of the ii-th primitive as L _ii ′, and use this L _ii ′ to calculate the matching degree MP ⁱⁱ ′ _jk = {1− Find |L _ii ′−L _jk |/L _jk } and write (dictionary j−PRk, PRii, MP ⁱⁱ ′ _jk ) in the matching correspondence table. The processing after setting i=ii is the same as b○ in the following section B-2-iii).

B-2-ii When _Li = L _jk a○ Since the k-th primitive of dictionary j and the i-th primitive on the unknown object have been matched, (dictionary j-PRk, PRi, MP ⁱ _jk ) is matched. Add to the table (see Figure 8C, ○B, ○C).

b○ Check whether the vertices at the end point correspond. Before that, if the unknown object and the dictionary j are exactly the same, in order to check whether the correspondence has been completed for the dictionary j, the vertex number of the end point of the kth primitive and the _k0th Check whether the starting point vertex numbers of the primitives are the same (see Figure 8C). If they are equal,
Finish creating the matching correspondence table (
reference). If they are not equal, the end point of the i-th primitive of the unknown object and the k-th primitive of the dictionary j
Check whether the end points of the primitives are equal in type and angle (see Figure 8C).

If Γ is equal, find the matching degree of the vertex, add (dictionary j−k ⁺ , i ⁺ , MA ⁱ⁺ _jk ⁺ ) to the matching correspondence table, and then return to checking the adjacent primitives (8th
(See Figure C).

If Γ is not equal, there is no vertex to be matched to the end point of the k-th primitive of dictionary j, so proceed to Figure 8 via Figure 8C and match (dictionary j−k ⁺ , 0, 0). In addition to the correspondence table, kfwd=k,
Matching in this direction is completed with ifwd=i, and the process proceeds to the next section B-3.

B-2- Case of L _i >L _jk In this case, the process branches from FIG. 8C to FIG. 8E. At this time, it is considered that the i-th primitive on the unknown object is a primitive that corresponds to the k-th primitive of the dictionary j and another primitive overlap.

a○ Matching is performed on the portion of the i-th primitive that corresponds to the k-th primitive of dictionary j. Set MP ⁱ _jk = 1 and add (dictionary j − PRk, PRi, 1) to the matching correspondence table.

b○ To indicate that the i-th primitive on the unknown object is divided into a part that corresponds to the k-th primitive of dictionary j and a part that does not, set a division flag for the unknown object primitive data (Fig. 8E○I). reference).

c○ Since the end point angle of the k-th primitive of dictionary j is angle = 180°, matching cannot be achieved, so add (dictionary j−k ⁺ , 0, 0) to the matching correspondence table, and kfwd = k,
Matching in this direction is completed with ifwd=i, and the process proceeds to the next section B-3.

B-3 If matching is not completed in the counterclockwise direction, matching is performed in the clockwise direction using the same method. At this time, by setting k=(k+N _j -1) mod N _j i=(i+N-1) mod N, primitives adjacent in the clockwise direction on the dictionary j and on the unknown object are found, respectively. The operation is performed as shown in FIGS. 8F to 8H, but since it is completely the same as in the previous section B-2 except that the direction is clockwise, the explanation will be omitted. In addition, the 8th
Figures F, 8G and 8H correspond to Figures 8C, 8D and 8E, respectively. Furthermore, when the correspondence is no longer established, kbck=k and ibck=i are set and the process proceeds to the next section B-4.

B-4 When there is a primitive that cannot be matched even after performing matching in both directions B-4- When there is only one primitive that cannot be matched among the primitives in the dictionary after matching in both directions (Figure 8I) , see Figure 8J).

kfwd primitive and kbck of dictionary j
If only one primitive remains that cannot be correlated with any other primitive, the following method is used to create a correspondence with the primitives in between.

a○ end point of kbck primitive in dictionary j
Find the coordinates (x ₁ , y ₁ ) of a point on the unknown object that is a distance L _jkbck away from the vertex on the unknown object corresponding to kbck ⁺ in the direction of the ibck primitive (see Figure 8I).

b○ Starting point of kfwd primitive of dictionary j
Find the coordinates (x ₂ , y ₂ ) of a point on the unknown object that is a distance L _jkfwd away from the vertex on the unknown object corresponding to kbck ^- in the direction of the ifwd primitive (see Figure 8I, ○).

c○ on unknown object i=ifwd modN+1 to i
= (ibck+N-1) mod Find primitives up to N that are on the straight line connecting (x ₁ , y ₁ ) and (x ₂ , y ₂ ), and if they are found, write them in the matching correspondence table (dictionary j- Add PRk, PRi, MP ⁱ _jk ). If not, set (dictionary j = PRk, 0, 0) (see Figure 8J).

B-4- If two or more primitives in the dictionary that cannot be matched remain due to matching in both directions, k = kfwd.
mod N _j +1 to k=(kbck+N _j −1)
Elements up to mod N _j (primitive vertices)
For (dictionary j-PRk, 0,0)...primitive (dictionary j-Ak ⁺ , 0,0)...add the vertex to the matching correspondence table (see Figure 8K).

The example matching set in Figure 6 (Dictionary 1-
When a matching correspondence table is created for PR1, PR7, 0.74), the 8th primitive in the unknown object matches both the radius and center coordinates, so it becomes as shown in Fig. 9.

Hidden Part Estimation Unit Here, hidden primitives and vertices in the unknown object are estimated from the matching correspondence table sent from the matching unit.

○ Primitive estimation If two or more primitives on an unknown object correspond to one dictionary primitive in the matching correspondence table, it is considered that the middle of these primitives is hidden, so that part is presume. If a circle primitive in the dictionary corresponds to a plurality of arc primitives in the unknown object, the circle in dictionary j is estimated from these arc primitives and added to the matching correspondence table. Furthermore, if a straight line primitive in the dictionary corresponds to a plurality of straight line primitives in the unknown object, the distance between the outermost end points of these straight line primitives is estimated as the primitive length. The degree of matching between the new primitive created by the estimation and the primitive k of the dictionary j is determined and added to the matching correspondence table.

○ Vertex estimation If there is no vertex on the unknown object that corresponds to the vertex in the dictionary in the matching correspondence table, the vertex is estimated by finding the intersection of the primitives at both ends of the vertex. Then, the lengths of the primitives at both ends are recalculated using the estimated vertices, and the matching degree is also estimated,
Perform new calculations on primitives.

○ Estimated figure check Check whether the estimated figure can be matched with the dictionary using data about the estimated items in the matching correspondence table. That is, in the case of a circle, if the difference in radius between the circle in the dictionary and the circle in the estimated figure is within the allowable error range, the circle in the estimated figure is output. In the case of a figure composed of straight lines, if all the primitives and vertices in the dictionary can be matched by estimation, and the degree of matching for all the correspondences is equal to or higher than the reference value, the estimated figure is output. If the estimated figure cannot be output, an instruction is issued to the matching section to send the next matching correspondence table.

Regarding the example of FIG. 6, a matching correspondence table as shown in ○A of FIG. 9 comes from the matching section. On the other hand, in the hidden part estimator, PR7 in the unknown object
Estimate the yen from PR8. Here, the radius of the estimated circle is the average of the radii of the two primitives.
Here it is 31. Since the radius of PR1 in dictionary 1 is 30, the matching degree is M ¹¹ _1-1 = {1-|R _1-1 −R ₁₁ |/R _1-1 } × {1-|L _1-1 −L ₁₁ |/L _1-1 } = {1-|30-31|/30} × {1-|180-195|/188} =0.93, and if the allowable range of matching degree is 0.9 or more, the estimated object is a dictionary 1, so dictionary 1 is output as the estimated object.

Figure 9 shows the primitive correspondence relationship including PR11 estimated from PR7 and PR8.

Unknown object update The part hidden by the estimated object is estimated based on the unknown object, and the estimated object is removed from the original unknown object and updated as a new unknown object.

1. Remove primitives and vertices on the unknown object included in the estimated object. However, even if a primitive is included in the estimated object, if the unknown object primitive data division flag is set, that primitive is partially included in the estimated object, so it should be left as is. . As a result, a cut portion is created as a contour on the unknown object.

2 We make an estimate about this cut part, but
The estimation method is basically the same as that of the hidden part estimator. In other words, if the cut point extracted by the estimated object is at one location, the primitives before and after it are determined, and if there are two or more cut points, the end point of a certain primitive extracted as the estimated object is the starting point. , for a set of two primitives whose end point is the start point of the next extracted primitive when looking in the clockwise direction. The specification method is as follows.

If both primitives are circular arcs, the two circular arcs are connected to form one circle or circular arc.

If both primitives are straight and on the same line, connect the two primitives to form one straight line primitive.

If both primitives are straight and not on the same straight line, find the intersection of the two straight line extensions and use this as the new vertex.

If one primitive is an arc and the other is a straight line, find the intersection of the extended arc and straight line and use this as the new vertex. As a result, the primitive data and vertex data of the unknown object are also changed. In the primitive where the split flag was set,
The intersection of this primitive and another primitive may appear on the primitive rather than on the extension line of the primitive. In this case, the primitives beyond the intersection can be ignored. Figure 15A, 1
Figure 5B shows such a case, where primitives PR3 and PR4 are ignored.

In the example of Figure 6, the 10th diagram shows the state in which primitives 7 and 8 on the unknown object estimated to be a circle are removed.
It is a diagram.

Now, according to the rules mentioned above, the primitive 7
When estimating the hidden part from a primitive whose starting point is the end point of , that is, a primitive whose end point is the starting point of primitive 1 and primitive 8, that is, primitive 10, this primitive is a straight line that is not on the same straight line, so the above 2. The vertex G as shown in FIG. 10 can be estimated from the - term.
Similarly, vertex F is estimated from primitive 6 and primitive 7. In this way, a new unknown object as shown in FIG. 11 is created, and this is input again as an unknown object. Since the longest primitive in this new unknown object is primitive 5 with a length of 62, the dictionary search unit searches the dictionary for a primitive that matches primitive 5, and selects primitives 1, 2, and 3 of dictionary 3 as candidates. However, since dictionary 3 is 3-fold symmetric, it is the same no matter which primitive it corresponds to. Therefore, the matching set (Dictionary 3-PR1, PR5, 0.97) is sent to the matching section. When a matching correspondence is created in the matching section, the result shown in FIG. 12A is created. From this, primitive 7 in the unknown object
A primitive that is calmer than primitive 4 is estimated, and this is newly defined as primitive 9. Since this primitive length is the distance between apex D and vertex F=59, the primitive correspondence relationship is as shown in FIG. 12B.

At this time, the average matching degree is 1+0.97+1+0.97+1+0.98/6=0.99, which satisfies the matching degree tolerance range of 0.9 or more, so dictionary 3 is output as the estimated object. FIG. 13 shows primitives 4, 5, 6, and 7 extracted from the unknown object as estimated objects.

Next, among the primitives in the remaining unknown object, vertices are estimated from primitives 3 and 8 whose connection relationship has been interrupted, and the primitive and vertex data are updated, resulting in the result shown in FIG. When this is input again as a new unknown object, a matching set of longest primitive 3 and primitive 1 of dictionary 2 (dictionary 2 - PR1, PR3, 0.98) is output from the dictionary search section. When a matching correspondence is created for this, it becomes as shown in FIG. 14C. The average matching degree is 1/8 (0.98+0.98+0.98+0.95+0.98+0.95+1+0.95)=0.97, and dictionary 2 is output as the estimated object.

As a result, there are no unmatched primitives on the unknown object, so the process ends. As a result, the unknown object is recognized as a combination of dictionaries 1, 2, and 3.

〔Effect of the invention〕

According to this invention, when recognizing an unknown object, the dictionary is searched for primitives that match primitives on the unknown object, and the dictionary is associated with the primitives and vertices of the unknown object. Since the method uses a method of estimating the primitives and vertices in the object, it is possible to recognize even superimposed objects in which multiple recognition target objects overlap each other, making it possible to expand the range of application. This brings about advantages.

[Brief explanation of the drawing]

Fig. 1 is a block diagram showing the configuration of an embodiment of the present invention, Fig. 2 is a block diagram showing a recognition device to which this invention is applied, and Fig. 3 is an explanation for explaining the circular dictionary pattern and data. Figure 4 is an explanatory diagram for explaining a rectangular dictionary pattern and data, Figure 5 is an explanatory diagram for explaining a triangular dictionary pattern and data, and Figure 6 is an example of an overlapping unknown object pattern. FIG. 6A is an explanatory diagram for explaining the primitive data shown in FIG. 6, FIG. 6B is an explanatory diagram for explaining the vertex data, and FIGS. 7A to 7E are dictionary diagrams. Flowchart for explaining the operation of the search unit, No. 8
Figures A to 8K are flowcharts for explaining the operation of the matching section, Figure 9 is an explanatory diagram for explaining the correspondence between the unknown object and the dictionary in the example of Figure 6, and Figure 10 is a diagram for explaining the correspondence between the unknown object and the dictionary in the example of Figure 6. An explanatory diagram illustrating the case where the estimated object is removed from the unknown object, FIG. 11 is a pattern diagram showing the updated unknown object, and FIG.
Figure A is an explanatory diagram for explaining the primitive data in Figure 11, Figure 11B is an explanatory diagram for explaining the vertex data, and Figure 12A is an explanatory diagram for explaining the primitive data in Figure 11.
An explanatory diagram showing the correspondence between the unknown object shown in the figure and the dictionary shown in FIG. 5, FIG. 12B is an explanatory diagram showing the correspondence when estimation is performed in FIG. 12A, and FIG. Explanatory diagram illustrating the case where an object is removed from the updated unknown object, 1st
4 is a pattern diagram showing the unknown object that has been updated again, FIG. 14A is an explanatory diagram for explaining the primitive data in FIG. 14, and FIG. 14B is an explanatory diagram for explaining the vertex data. 14th C
The figure is an explanatory diagram showing the correspondence between the unknown object shown in Fig. 14 and the dictionary shown in Fig. 4, Fig. 15A is a pattern diagram showing the case where primitives of two dictionaries overlap, and Fig. 15B is a pattern diagram showing the case where triangles are removed from FIG. 15A, and FIG. 16 is a pattern diagram showing overlapping objects. Description of symbols: 1...contour tracking unit, 2...curvature calculation unit,
3... Primitive description section, 4... Dictionary creation section, 5...
Dictionary search section, 6... Matching section (correspondence table creation section),
7... Hidden part estimation unit, 8... Unknown object update unit, 10
... Recognition target object, 11... Imaging device (TV camera), 12... Preprocessing circuit, 13... Feature extraction circuit,
14... Image memory, 15... Processor, 16... Interface circuit, 17... Terminal, M, M1, M2
...Memory, ID...Estimated object data output section.

Claims (1)

[Claims] 1. The outline of each of multiple types of objects to be recognized is traced, divided into primitives including straight line parts and circular arc parts, and the attributes of each object and the attributes of each vertex are registered in a dictionary memory as a dictionary. After performing the processing in advance, the unknown object is divided into primitives like a dictionary, the unknown object is described using the attributes of the primitives and vertices, and stored in a predetermined memory. The recognition device comprises: a dictionary search means for searching the dictionary memory for primitives that match each primitive of the unknown object, and extracting all matched primitives as a matching set; A correspondence table creation means for creating a matching correspondence table by checking the degree to which primitives and vertices in the dictionary are matched with those of the unknown object for each; hidden part estimating means for estimating primitives or vertices in the unknown object, correcting the primitive length, checking the degree of matching with a dictionary, and extracting the optimal one as the estimated object; and removing the estimated object from the unknown object and calculating the result. unknown object updating means that estimates what primitives and vertices should be added to the remaining portion and updates the unknown object based on the estimation result; An object recognition device characterized by comprising: a means for making it possible to recognize an unknown object formed by a plurality of objects touching or overlapping each other by repeating the process until the object disappears.