JPH08145738A - Active recognition device - Google Patents

Abstract

PURPOSE: To provide an active recognition device capable of high speed recognition in which the generation, integration, and transmission of the conviction of a conviction network are repeated on the basis of the characteristic quantity from a proper observing means of a plurality of observing means, and circularly performed until a target to be recognition is attained, and the conventional calculation required in the holding of a unsure conviction in the conviction network is minimized. CONSTITUTION: This device has a conviction network 1 showing the connecting relation of convictions, a plurality of observing means 6 for observing the characteristic quantity of a target in environments, a conviction generating means 3 for generating a conviction on the basis of the characteristic quantity observed by one selected observing means 6, and a conviction integrating means 5 for the conviction generated by the conviction generating means 3 with the corresponding conviction in the conviction network 1. When the integrated conviction in the conviction network 1 attains the target to be confirmed, the recognition result is outputted, and when the target is not attained, the following observing means 6 is selected and executed.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an active recognition device which holds an uncertain belief and updates the belief based on the observed feature amount of an object to recognize the object. The present invention relates to an active recognition device used for operation, navigation of a mobile robot, and the like.

A recognition device in an unknown or fluctuating environment cannot describe in advance a fixed operation procedure. The recognition device holds the information about the environmental condition obtained at that time as a belief, and the ability to automatically determine the operation procedure such as observation by the sensor based on the belief, that is, the belief maintenance ability and the observation strategy generation ability. Need to have. Generally, a recognition device having these capabilities is called an active recognition device.

Observations made by sensors can be erroneous. Further, usually, one observation cannot uniquely determine the state of the environment, and it is necessary to gradually narrow it down by a plurality of observations. Therefore, active recognizers need to be capable of handling uncertain beliefs, such as candidates for environmental conditions.

In recent years, the range of environments handled by equipment equipped with a recognition device has been expanded from known or fixed environments to unknown or variable environments. For this reason, active recognition devices that can handle uncertain beliefs have played an important role as technical elements in constructing these devices.

[0005]

2. Description of the Related Art Conventionally, Rimey and Brown shown in FIG.
(Note 1) There is an active recognition device that uses a probability distribution for the belief expression proposed in (Note 2). FIG. 9B shows the structure of the probability distribution and belief relationship data corresponding to the beliefs forming the belief network.

FIG. 10 shows a flowchart for explaining the operation of the configuration shown in FIGS. 9 (a) and 9 (b). In FIG. 10, S5
1 measures the environment by the observation means. This is shown in FIG.
The state of the object in the environment is observed by the plurality of observation means in (a).

In step S52, the belief network is operated. This operates the belief network of (b) of FIG. 9 based on the environmental information acquired by measuring the environment in S51. For example, the belief (probability distribution) is updated with environmental information.

In step S53, it is determined whether or not the recognition target for the belief (probability distribution) after the update in step S52 is achieved. In the case of YES, since the recognition target has been achieved, the process ends (END). On the other hand, in the case of NO, since it is found that the recognition target is not achieved, the process returns to S51, and the belief network is repeatedly updated based on the observed environmental information until the recognition target is achieved.

(Note 1) Rimei, Brown: "Control of Selective
Perception Using Bayes Nets and Decision Theory ", I
nt.J, of Computer Vision, Vol.12, No.2 / 3, pp.173-207 (1
994). (Note 2) Rimey, Brown: "Task-oriented Vision with Multip
le Bayes Nets ", Active Vision, pp.217-236, MIT Press
(1992).

[0010]

As described above, the conventional active recognition apparatus holds the belief as an belief network with an arbitrary probability distribution in order to enable more flexible description of environmental state candidates. Therefore, the probability distribution does not become a distribution that is easy to calculate like a normal distribution and becomes an arbitrary probability distribution, so that the spatial calculation amount and the temporal calculation amount become enormous, which is not practical. There was a problem.

The present invention solves these problems.
The belief of the belief network is repeatedly generated, integrated, and transmitted repeatedly based on the feature amount from the appropriate observation means of the plurality of observation means until the recognition target is achieved, and the conventional uncertain beliefs are maintained. The purpose is to reduce the amount of calculation required to store in the belief network and enable recognition at high speed.

[0012]

[Means for Solving the Problems] Means for solving the problems will be described with reference to FIG. In FIG. 1, a belief network 1 is a network that represents a connection of beliefs, and the belief represents candidates for the state of an object 8 existing in an environment 7.

The belief manipulating means 2 operates (updates) the belief of the belief network 1 based on the characteristic amount observed by the observing means 6, and the belief generating means 3,
It is composed of a belief transmitting means 4 and a belief integrating means 5.

The belief generating means 3 is for generating a belief based on a feature amount. The belief transmitting means 4 transmits a belief on the belief network 1. Belief integration means 5
Integrates beliefs.

The observing means 6 is for observing the characteristic amount of the object 8 in the environment 7. The environment 7 is an environment in which the recognition target object 8 exists. The target object 8 is a target object to be recognized.

The strategy generating means 11 selects the observing means 6 having the greatest effect and instructs the belief operating means 2 to operate (update) the belief network 1 based on the feature amount observed by the observing means 6. It is a thing.

[0017]

According to the present invention, as shown in FIG. 1, belief generation is performed based on the feature quantity observed by one observation means 6 selected from a plurality of observation means 6 for observing the feature quantity of the object in the environment. Means 3 generates a belief representing a candidate for the state of the object 8;
The belief integrating means 5 integrates the generated belief with the corresponding belief in the belief network 1, outputs the recognition result when the belief in the belief network 1 after integration achieves the recognition target, and achieves it. If not, the next observation means 6 is selected and executed repeatedly.

Further, the belief generating means 3 causes the belief generating means 3 to target the object 8 based on the feature quantity observed by one observation means 6 selected from a plurality of observation means 6 for observing the feature quantity of the object in the environment.
A belief representing a candidate of the state of the belief, the belief integrating means 5 integrates the generated belief with the corresponding belief in the belief network 1, and the belief transmitting means 4 converts the belief into the belief in the belief network 1. When the belief in the belief network 1 after the integration achieves the recognition target, the recognition result is output, and when it is not achieved, the next observation means 6 is selected. I try to repeat what I do.

In these cases, the degree of reinforcement of the belief network 1 is calculated based on an arbitrary feature amount within the range of the respective feature amounts of the plurality of observation means 6, and the average thereof is calculated, and this calculation is performed. The observation means 6 having the smallest difference between the average reinforcement degree and the recognition target is selected, and the belief is generated, integrated, and transmitted based on the feature amount from the selected observation means 6.

Therefore, the belief of the belief network 1, that is, the generation, integration, and transmission of the candidates of the state of the object are repeated based on the feature amount from the appropriate observing means 6 among the plurality of observing means 6, and the recognition target is set. By executing it cyclically until it is achieved, it became possible to recognize the uncertain belief (probability distribution) in the conventional belief network with a small amount of calculation and high speed recognition.

[0021]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Next, the construction and operation of an embodiment of the present invention will be described in detail with reference to FIGS.

FIG. 1 shows a block diagram of an embodiment of the present invention.
In FIG. 1, the strategy generation means 11 includes a plurality of observation means 6
The observation command is issued to one selected from among, and the belief operation command is issued based on the feature quantity at that time, and the belief network 1 is operated (belief generation, belief transmission, belief integration) and updated. However, when the updated belief reaches the recognition target 14, the recognition is achieved, and when the updated belief does not reach the recognition target 14, the next observation means 6 is cyclically repeated. The strategy generation means 11 includes a prediction unit 12 and a planning unit 1.
3, a recognition target 14, a determination unit 15, an observation strategy determination standard 16 and the like (described later).

Next, the operation of the configuration of FIG. 1 will be described in detail according to the order of the flowchart of FIG. In FIG. 2, S1 initializes the belief network. This uniformly sets the entire space as an initial value of each belief of the belief network 1, as shown in (a) of FIG. 5, which will be described later.

S2 selects appropriate observation means for strengthening the belief network. This selects one of the plurality of observation means 6 shown in FIG. 1 that seems appropriate for strengthening the belief network 1.

At S3, observation is performed. In this case, one observing means 6 selected in S2 observes the state of the object 8 to obtain a feature amount. In S4, the belief is generated based on the feature amount. This is because "A: Belief generation" is performed as described on the right side, that is, one observing means 6 selected in S3 observes the state of the target object 8 and the belief is calculated based on the feature quantity at that time. To generate. For example, as shown in the belief of the part B in FIG. 5B, which will be described later, the belief (the part B is present within the range of the angle θ from the viewpoint based on the image taken by the camera having the observing means 6). Belief).

S5 integrates the original and generated beliefs. This integrates the original belief and the belief generated in S4, for example, a common part of the original belief and the generated belief is integrated as a belief after the integration, as shown in FIG. 4 described later.

At S6, the belief is transmitted. This is S5
Communicate the beliefs integrated in to other beliefs. For example, the belief of the part B is transmitted to the belief of the part A as shown in FIG.

S7 integrates the original belief and the transmitted belief. This integrates the belief transmitted in S6 and the original belief (see (d) of FIG. 7 described later). In S8, it is determined whether the recognition target has been achieved. This is a state in which the belief generated in S5 and the belief transmitted in S7 are integrated, and the difference between the belief after the integration and the recognition target is below a predetermined threshold so that the recognition target is reached. Determine if the belief has been strengthened. In the case of YES, a belief network whose difference from the recognition target is less than or equal to a predetermined threshold is obtained, and thus the process ends (END). On the other hand, in the case of NO, S2 and subsequent steps are cyclically repeated for the next observation means 6.

As described above, one of the plurality of observation means 6 is selected to generate, transmit, and integrate the beliefs forming the belief network 1, and the belief after the integration is recognized when it reaches the recognition target. It was done. On the other hand, when the recognition target is not reached, the other observation means 6 are cyclically repeated. With these, under the configuration of FIG. 1, an appropriate observing means 6 is selected and observed, and the belief of the belief network 1 is generated, integrated, and transmitted based on the feature amount, and the changing target 8 is detected. It has become possible to perform recognition actively without holding the conventional fluctuating probability distribution in belief. This will be described in detail below.

FIG. 3 is a conceptual explanatory view (1) of the present invention. FIG. 3A shows an example of an object in which the recognition target object is composed of four parts here. Here, the four parts are the illustrated parts A, B, C, and D.
Is.

FIG. 3B shows an example of estimating the position of the part A whose recognition target is the recognition target. It is assumed that parts A, B, C, and D in FIG. 3A are in space and that part A is estimated as a recognition target.

FIG. 3C shows an example of a belief network. Since this belief network 1 has four parts A, B, C, and D from FIG. 3A, it is assumed that the beliefs are associated with each other as shown by arrows (arrows). According to the belief described below). This arrow is called, for example, the positional relationship between the parts A and D (which is a type of belief relationship data). Further, the belief of the part B is based on the feature amount observed by a certain observing means 6 as an initial value when it is assumed to exist uniformly in the entire space as shown in the enlarged view on the right side. Is generated, it becomes a region in which the illustrated part B may exist, and integrated with the original initial value, that is, a common region becomes a belief after the integration (here, the same region as the generated belief).

FIG. 3D shows an example of the feature quantity. This feature amount is obtained by observing the part A which is the object 8 by one observing means 6 in FIG. 1, for example, a line sensor, and the length L on the screen at that time and the angle from the viewpoint are θ. , The belief of part A is that the length L and angle θ
Is generated from the feature amount.

FIG. 4 (e) shows an example of belief integration.
This seeks the intersection of the original belief and the belief generated / transmitted. For example, as shown in the figure, if the original belief area is a diagonally downward-sloping area and the generated / transmitted belief area is a shaded area, the area where both of them intersect is the belief after integration. It becomes an area.

Next, the belief generation, integration, and transmission of the four parts A, B, C, and D will be specifically described with reference to FIGS. FIG. 5, FIG. 6, and FIG. 7 are views for explaining specific examples of the present invention.

In FIGS. 5, 6 and 7, (a) of FIG.
Shows an example of belief network initialization. here,
The beliefs of parts A, B, C, and D are initialized in all spaces and made uniform.

FIG. 5B shows an example of belief generation.
For example, as the feature amount of the part B, the angle θ is (b-
If obtained as shown in FIG. 1), as shown in (b-2) of FIG. 5, the belief of the part B is the angle θ from the viewpoint as shown, and the possibility that the part B exists. This is a range, and when this is integrated with the belief of the part B in FIG. 5A (initialized uniform entire space), the belief after the generation / integration is obtained by the region of the generated belief here.

FIG. 6C shows an example of belief transmission.
This is based on the belief generated by the part B in FIG. 5B based on the positional relationship of the arrows in the belief network 1.
Communicate to the belief of. Here, when the belief of the part A is transmitted, it becomes the “conveyed belief” illustrated, and when the belief of the part A of FIG. 5A (initialized uniform whole space) is integrated, here Then, it becomes the belief after the transmission / integration that the domain of the transmitted belief requires.

FIG. 7D shows an example of belief strengthening (belief generation, transmission, and integration). The belief generation, transmission, and integration of the parts A, B, C, and D are as shown below.・ Belief of part B: ・ Integrated (2) from the belief generated (1) and the original belief
It is the belief that was given.

Belief of part A: Belief integrated (4) from the belief transmitted (3) from part B and the original belief.

Belief of Part C: Belief transmitted from Part B (5) and original belief are integrated (6).

Belief of part D: Belief transmitted from part A (7) and belief transmitted from part C (8) and original belief integrated (9)
It is the belief that was given.

FIG. 7E shows a state in which the position of the part A can be estimated (the recognition target can be achieved) by repeating observation and belief strengthening. In this state, each part A, B,
The beliefs of C and D are in a state of being within the range of the recognition target, and the small area (not shown) in the entire space is the recognized range.

FIG. 8 shows a flowchart for selecting the observation means of the present invention. This is a suitable one from multiple observation means 6.
It shows a procedure for selecting one observation means 6. In FIG. 8, S11 selects one observing means candidate from the observing means set. This selects one observing means 6 from the plurality of observing means 6 in FIG.

In step S12, the range of possible feature quantities is obtained. This obtains the range of possible feature quantities when the object 8 is observed by the observation means 6 selected in S11. S
13 selects a certain feature amount from the possible range.

In step S14, the belief is strengthened (belief generation, belief transmission, belief integration) on the assumption that the selected feature amount is observed. This reinforces the belief of FIG. 7 (d) already described.

In S15, the degree of reinforcement of the belief network is calculated (for example, the degree is the belief region area). For example, as shown in (d) and (e) of FIG. 7 described above, this is the degree of the strength of each belief in the belief network 1 after the belief is strengthened, and the area of the region in which the belief exists. Calculate the smallness (entropy).

In step S16, it is determined whether or not the degree of enhancement has been calculated for a sufficient number of feature quantities to obtain an average from the range of possible feature quantities. In the case of YES, the process proceeds to S17. If NO, the process returns to S13 to select another feature amount and repeat.

In step S17, the degree of reinforcement is averaged over the range of possible feature quantities. In S18, it is determined whether or not the average degree of reinforcement has been calculated for all the observation means included in the set of observation means. In the case of YES, it progresses to S19. In the case of NO, S1 for the remaining observation means 6
Repeat 1 and later.

In S19, one observing means 6 having the smallest difference between the average degree of reinforcement and the target degree of reinforcement is selected. From the above, a sufficient number of observing means 6 can be averaged for belief strengthening (belief generation, belief transmission, belief integration) based on the feature amount within the range of each feature amount. The observation means 6 having the smallest difference from the recognition target is selected from the obtained degree of enhancement of the average, and the belief network 1 of the belief network 1 is selected on the basis of the actually observed characteristics by the observation means 6. Strengthen.

Next, the concept and operation of the present invention will be described in detail using mathematical expressions. Here, under the configuration of FIG.
An active recognition apparatus that acquires the position and orientation information of the objects by observation and performs a task in an environment where a plurality of objects exist will be described below as an example. In this case, the parameter that describes the environment is the set {X _i } of the position and orientation of each object.

X _i = (x _i , y _i , z _i ; α _i , β _i , γ _i ) εP (Formula 1) where (α _i , β _i , γ _i ) εR ³ is the object i The position of the center of gravity of (α _i , β _i , γ _i ) ∈
[0,2π] ³ represents the posture of the object i by the Euler angle. P = R ³ × [0,2π] ³ is
It represents the position and orientation space.

Assuming that the number of objects existing in the environment is N, the environment is a set of position and orientation of each object, that is, [X _1, ... X _N ].
It will be described by. N objects in the environment
Since there are N beliefs, there are N beliefs, and each belief is represented by candidate data representing state candidates for the position and orientation of each object. Since the position and orientation are 6-dimensional vectors, candidate data can be represented by a subset in 6-dimensional space. That is, the belief B _i corresponding to the object i is expressed by the following candidate data R _i .

R _i = {6-dimensional vector expressing position / orientation candidate of object i} ⊂P = R ³ × [0,2π] ³ (Equation 2) The belief network 1 is composed of the above N candidate data. . Belief Network 1 includes one or more beliefs,
It may have one or more belief-to-belief relationship data that represent interactions between beliefs.

The belief operating means 2 comprises a belief generating means 3 for handling candidate data, a belief transmitting means 4, and a belief integrating means 5. The belief generating means 3 is activated by the strategy generating means 11, and at this time, the following three arguments are passed.

Position / orientation in the working coordinate system of the visual sensor: γ Candidate data for belief to be generated: R _i Observation algorithm used for generation: A _j There may be a plurality of observation algorithms. An L-type algorithm for narrowing down candidate data is used by using the area on the image, the center of gravity, and the like when an object corresponding to the belief to be generated is photographed by a visual sensor as observation values. Each algorithm maintains the relationship between the observed area and the feature such as the center of gravity and the position and orientation of the observation target in the work coordinate system when the visual sensor is placed in the specified position and orientation in the work coordinate system. Therefore, it is possible to narrow down the candidate data by selecting only the elements suitable for the feature amount obtained as a result of the observation from the elements of the current candidate data of the belief to be generated. The candidate data resulting from the narrowing down becomes the belief that is newly generated. The following is a more rigorous formulation of the algorithm of the belief generating means 3.

For the observation algorithm A _j for the belief B _i , the belief generating means 3 arranges the visual sensor in the position / orientation rεP designated in the working coordinate system, and observes the feature amount such as the area and the center of gravity. F∈F⊂R ^d wherein belief among related data M _{(r) i} and the position and orientation P∈R ⁶ in the working coordinate system of the observation target _object, the _j, held in the form of a flame, such as a table. Position and orientation of visual sensor γ
Since the correspondence relation M _{(r) i , j in} is generally a many-to-many relation, it is defined as a subset of the product set F × P of the feature space F and the position / orientation space P. That is, M _{(r) i, j} ⊂F × P (Equation 3).

Further, the acquisition of the feature quantity f by observation has an error Δ
Assuming that f is involved, when the belief B _i has the candidate data R _i , the candidate data R _{(r) i, j, f} newly generated by the observation algorithm A _j having the correspondence M _{(r) i, j.} ^gen is ^calculated as follows.

R _{(r) i, j, f} ^gen = {pεP | (f ′, p) εM _{(r) i, j} and || f-f ′ || <Δf} (Equation 4) where , || f-f '|| represents the distance between f and f'. this
(Equation 4) is a correspondence relationship M among the candidates included in the position / orientation space P with respect to the observed feature amount f including the error Δf.
_{(r) A} candidate satisfying _{i, j} is selected, and new candidate data R
_{(r) i, j, f} ^gen is formed.

The belief integrating means 5 integrates the original candidate data R _i and the candidate data R _{(r) i, j, f} ^gen generated by the belief generating means 3 with respect to the specified belief B _i without any contradiction and newly creates the new belief data. New candidate data R _{(r) i, j, f} ^new is obtained.

R _{(r) i, j, f} ^new = R _i ∩R _{(r) i, j, f} ^gen (Equation 5) The strategy generation means 11 is a set of candidate data of the current belief of the belief network 1. R _i | i = 1 ... N}, that is,
Based on the currently obtained environmental condition candidates, the operation procedure to be executed in the future is determined. More specifically, the belief generating means 3 constituting the belief operating means 2, the belief B _j to be the subject of the belief integrating means 5, the observation algorithm A _k, and the position and orientation γ of the visual sensor are determined.

It is assumed that the strategy generating means 11 performs the belief operation based on such observations for all the observation algorithms A _k for all the beliefs R _{j, and} for all position / orientation candidates γ of the visual sensors. In that case, the effect on the task is predicted, the one with the greatest effect is selected, and the operation command is issued.

In order to quantify the effect on the task, the concept of belief entropy will be explained. In order to accomplish the task, it is necessary to obtain certain reliable information about the external environmental condition within the range necessary for accomplishing the task. In other words, it is necessary to obtain a belief with less uncertainty to the extent necessary to accomplish the task. Therefore, first, the uncertainty of belief is quantified.
A belief is a set of environmental state candidates, so the greater the number of candidates, the greater the uncertainty. Therefore, as a quantitative measure expressing the uncertainty of belief, the entropy Ent [R] of the belief R is derived from the definition of the amount of information,
It is defined as follows.

Ent [R] = logS (R) (Equation 6) where S represents deposition. Using the concept of belief entropy defined in (Equation 6) above, if the target entropy for achieving the task for each belief B _J is Ent _i ^goal , the uncertainty of belief with candidate data R _i and the expected uncertainty The squared error diff between and can be calculated as follows.

[0065] Regarding the belief obtained as a result of performing some observation, it can be considered that the observation having the smallest square error diff is the most effective observation for achieving the task.

Next, a method of predicting the effect of the belief generation on the task will be described. That is, when the set of candidate data expressing the current belief is {R _i | i = 1 ... N},
For a certain belief B _j , the effect on the task when the belief is generated by the observation by the visual sensor arranged in the position and orientation γ using the certain observation algorithm A _k is predicted. Therefore, when the range F _{(r), j, k} of the feature quantity that will be obtained when the above observation is performed, F _{(r), j, k} = {fεF | (f, p) ε M _{(r), j, k} and pεR _j } (Equation 8). Assuming that the feature quantity fεF _{(r), j, k} is obtained with an error Δf as a result of performing the observation algorithm A _k on the belief R _j , the belief R after the generation integration predicted for the belief R _i is obtained.
_{i, (r), (i, k), f} ^exp are obtained as follows.

R_{i, (r), (i, k), f} ^exp= {p ∈ R_j| (f ', p) ∈ M_{(r), j, k}And || f-f '|| <Δf} (when i = j) R_i(When i ≠ j) From the above, as a result of the observation, the feature quantity f ∈ F_{(r), j, k}Is the error Δ
Belief R when it is supposed to be obtained in f_iPredicted against
Entropy Ent of new belief_{i, (r), (j, k), f} ^expIs Ent_{i, (r), (j, k), f} ^exp= Ent [R_{i, (r), (j, k), f} ^exp] (Equation 9) is given. Furthermore, the range F of the feature amount_{(r), j, k}about
If you take the average, the observation result, belief R_iPredicted against
Expected Entropy of new beliefs Ent_{i, (r)
, (j, k)} ^expIs obtained. That is, Ent_{i, (r), (j, k)} ^exp= Aver
ge f ∈ F_{(r), j, k}[Ent_{i, (r), (j, k), f} ^exp] (Equation 10)
is there. Belief B by the expected value of this entropy_jAgainst
Observation of the position and orientation γ with a visual sensor A_kIf you do
Difference from the target entropy for task achievement dif
f_{(r), j, k}Can be calculated as follows. diff_{(r), j, k}= Σ (Ent_{i, (r), (j, k)} ^exp−Ent^goal)² (Equation 10) The strategy generating means 11 is the belief generating task described above.
The most effective observation using the method of predicting the effect on
Select and issue operation command. That is, diff _{(r), j, k}
Is the smallest (the predicted entropy value is the most
(Near) the position and orientation γ of the visual sensor and the belief B of the generation target_iWhen
Observation algorithm A_jSelect the belief generation / belief integration finger
Issue a decree.

Further, the belief transmitting means 4 is the strategy generating means 1
It is started by 1, and the following three arguments are passed at this time.・ Belief data of the belief of the transmission source: R _i・ Relationship belief data between the transmission source and the transmission destination: K _{i, j}・ Index of the belief of the transmission destination: j The two beliefs of the transmission source and the transmission destination are belief networks. It corresponds to two nodes directly connected by an arc. Therefore, the belief relationship data K _{i, j} representing the relative position / orientation relationship between the two beliefs of the transmission source and the transmission destination is held.
The belief transmitting means 4 calculates the ^unintegrated candidate data R _j ^prop that realizes the belief of the transmission destination by using the transmission source candidate data R _i and the belief relationship data K _{i, j} as follows.

R _j ^prop = {qεP | pεR _i and (p, q) εK _{i, j} } (Equation 11)

[0070]

As described above, according to the present invention,
The belief of the belief network 1, that is, the generation, integration, and transmission of candidates for the state of the object are repeated based on the feature amount from an appropriate observation means 6 of the plurality of observation means 6, and the recognition cycle is repeated until the target is achieved. Since it adopts a configuration that executes
Compared with the conventional method of holding uncertain beliefs (probability distribution) in the belief network, the amount of calculation required to hold the belief is reduced, and the amount of computation required to hold and operate the belief is significantly reduced compared to the conventional technology. Since it is possible to perform active recognition in a shorter time, it greatly contributes to the construction of a responsive system capable of responding to a complicated unknown / variable environment in real time.

[Brief description of drawings]

FIG. 1 is a configuration diagram of an embodiment of the present invention.

FIG. 2 is a flowchart explaining the operation of the present invention.

FIG. 3 is a conceptual explanatory diagram (1) of the present invention.

FIG. 4 is a conceptual explanatory view (No. 2) of the present invention.

FIG. 5 is a specific example explanatory view (1) of the present invention.

FIG. 6 is a specific example explanatory diagram (2) of the present invention.

FIG. 7 is a diagram (part 3) for explaining a specific example of the present invention.

FIG. 8 is a flowchart for selecting observation means of the present invention.

FIG. 9 is an explanatory diagram (1) of a conventional technique.

FIG. 10 is an explanatory view (No. 2) of the related art.

[Explanation of symbols]

 1: Belief network 2: Belief operating means 3: Belief generating means 4: Belief transmitting means 5: Belief integrating means 6: Observing means 7: Environment 8: Target 11: Strategy generating means

Claims (3)

[Claims] 1. A belief network (1) showing a connection of beliefs, a plurality of observation means (6) for observing the feature amount of an object in the environment, and one observation means (6) selected. A belief generating means (3) for generating a belief based on a feature amount, and a belief integrating means (3) for integrating a belief generated by the belief generating means (3) and a corresponding belief in the belief network (1). 5) and, if the belief in the belief network (1) after integration achieves the recognition target, the recognition result is output, and if not, the next observation means (6) is selected. An active recognition device characterized by executing. 2. A belief network (1) showing a connection of beliefs, a plurality of observing means (6) for observing the feature amount of an object in the environment, and one observing means (6) selected. A belief generating means (3) for generating a belief based on a feature amount, a belief generated by this belief generating means (3) and a corresponding belief in the belief network (1) are integrated and transmitted. Belief integrating means (5) for integrating the belief with the original belief, and a belief transmitting means (4) for transmitting the belief integrated by the belief integrating means (5) to the corresponding belief in the belief network (1). ), And if the belief in the belief network (1) after integration or transmission / integration achieves the recognition target, the recognition result is output, and if not, the next observation means (6)
An active recognition device that selects and executes. 3. A degree of strengthening of the belief network (1) is calculated based on an arbitrary feature amount within a range of possible feature amounts of each of the plurality of observation means (6), and an average thereof is calculated. , Selecting the observation means (6) having the smallest difference between the calculated average reinforcement degree and the recognition target, and generating, integrating, and transmitting the belief based on the feature amount from the selected observation means (6). The active recognition device according to claim 1 or 2, wherein