JP5340899B2 - Security device and sensor reaction factor estimation method - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a security device and an estimation method of a sensor reaction factor which are capable of accurately estimating a sensor reaction factor to effectively reduce false alarms caused by a false reaction of the sensor, without requiring much time and effort for setting up and without restricting types of sensors to be used. <P>SOLUTION: Using sensor reaction signals obtained in time series from a plurality of sensors each installed for each of a plurality of sensing target areas and a determination model which expresses causality of a sensor reaction as a probability model using dynamic Bayesian network, for each category of objects which may be a sensor reaction factor, a conviction degree in which an object belonging to each category may be a sensor reaction factor is computed, and a category of an object which is a sensor reaction factor is determined based on computed conviction degree. <P>COPYRIGHT: (C)2011,JPO&amp;INPIT

Description

  The present invention relates to a security device that determines an attribute of an intruding object based on a response signal of a sensor and performs a security operation when necessary, and a sensor response factor estimation method.

  Conventionally, there has been known a security device that installs a sensor outside a building, detects a suspicious person who is about to enter the inside of the building by this sensor, outputs an alarm, or reports to a security company. . Such a security device has been widely introduced in general households as the interest in security in recent years and the laying cost are reduced.

  In a security device that installs a sensor outside a building, there is a possibility that the sensor may react incorrectly due to cats, birds, fallen leaves, plants, plastic bags, etc. This leads to an increase in the burden on guards due to dispatch, and also results in mischief of residents' anxiety. For this reason, in this type of security device, it is an important issue how to reduce false alarms due to sensor false reactions.

  From the above viewpoint, for example, in the following Patent Document 1, an opposing infrared array is installed as a sensor, and the false alarm is reduced by estimating the reaction factor by the pattern of how many infrared beams are blocked. The technology is proposed.

  Moreover, in the following Patent Document 2, the reaction pattern of the human sensor is compared with a predetermined pattern, and it is determined whether or not an alarm is transmitted based on whether or not these patterns match, thereby reducing false alarms. The technique of making it be proposed is proposed.

  Moreover, in the following patent document 3, the stay state in front of the opening is estimated by a sensor, and it is determined whether or not the person is a suspicious person from the length of the stay time. Even when these sensors react, a technique has been proposed in which false alarms are reduced by estimating the degree of suspiciousness of a moving route and determining whether or not it is suspicious.

Japanese Patent Laid-Open No. 9-161163 JP-A-1-292600 JP 2007-233660 A

  However, the technique described in Patent Document 1 requires a wide space for stretching a multistage infrared beam, and such a wide space may not be secured in a general house or the like. Further, in the technique described in Patent Document 1, since the detection area of the sensor is concentrated, if a large erroneous reaction factor such as a trash bag passes, it is also considered that it is erroneously determined that a person has passed. There is a problem that the effect of reduction cannot be obtained sufficiently.

  Moreover, in the technology described in Patent Document 2, since the site configuration, possible error reaction factors, and sensor reaction patterns are various in a real environment, all patterns are specified for each environment. It is very difficult to do this, and if the sensor does not react, the pattern will be different, so there is a problem that it is difficult to ensure sufficient accuracy for complicated setting work. .

  In addition, the technique described in Patent Document 3 is a technique that focuses on determining whether or not a person is a suspicious person, and a technique that positively determines the influence of sensor false reaction caused by small animals or various natural factors. is not. That is, the technique described in Patent Document 3 is a technique in which a deviation from normal behavior by a resident is seen as a reference for a travel route and a staying time considered to be suspicious. For this reason, the reaction by small animals, which are considered to be often taken by the resident through a route that is not normally selected, is not positively determined as a reaction by small animals, etc., so it may be erroneously determined as a suspicious person. It is considered that the nature is high. In addition, it is necessary to set many parameters such as a suspicious stay time standard and a travel route considered to be suspicious.

  The present invention has been made in view of the above, and it is possible to estimate the sensor reaction factor with high accuracy without requiring much labor for setting and without restricting the type of sensor to be used. It is an object of the present invention to provide a security device and a method for estimating a sensor reaction factor that can effectively reduce false alarms resulting from a false reaction.

  In order to solve the above-described problems and achieve the object, the security device according to the present invention includes a plurality of sensors installed in each of a plurality of detection target areas that are objects of object detection, Using model storage means for storing a judgment model, which is a model expressing a causal relationship of sensor responses by a probability model using a Bayesian network, sensor response signals obtained in time series from the plurality of sensors, and the judgment model Based on the certainty factor calculated by the certainty factor calculating means and the certainty factor calculating means for calculating the certainty factor that the object belonging to each category is the sensor reaction factor Determining means for determining the category of the object that is a sensor reaction factor, and performing a guard operation according to the determination result by the determining means Characterized in that it comprises a security unit that row, the.

  The sensor reaction factor estimation method according to the present invention is a method for estimating sensor reaction factors by a plurality of sensors respectively installed in each of a plurality of detection target regions to be object detection. Using a sensor response signal obtained in time series from the plurality of sensors and a judgment model that represents a causal relationship of the sensor response by a probability model using a dynamic Bayesian network, For each category, based on the step of calculating the certainty that the object belonging to each category is a sensor reaction factor and the certainty calculated for each category of the object that can be a sensor reaction factor, Determining the category of the object that is present.

  According to the present invention, the sensor reaction factor is estimated by probabilistic inference using a dynamic Bayesian network. Therefore, the sensor reaction factor can be estimated with high accuracy, and the false alarm caused by the sensor erroneous reaction can be effectively performed. Can be reduced. In addition, since it is not necessary to predetermine the reaction pattern of each sensor, there is an advantage that setting labor is reduced compared to a method of determining all sensor reaction patterns. Furthermore, there is no particular limitation on the sensor to be used, and there is an effect that various sensors can be used.

FIG. 1 is a model diagram for explaining the outline of a Bayesian network. FIG. 2 is a diagram showing an example of a conditional probability table representing the conditional probability P (B | A) shown in FIG. FIG. 3 is a model diagram for explaining the outline of the dynamic Bayesian network. FIG. 4 is a model diagram showing a basic model in the initial state of the dynamic Bayesian network shown in FIG. FIG. 5 is a model diagram showing a state in which units are added to the basic model shown in FIG. FIG. 6 is a configuration diagram showing a schematic configuration of the monitoring apparatus according to the present embodiment. FIG. 7 is a functional block diagram showing a functional configuration inside the determination apparatus. FIG. 8 is a model diagram showing an outline of a judgment model used in the judgment apparatus. FIG. 9 is a diagram for explaining how the determination model expands as units are added to the basic model. FIG. 10 is a diagram for explaining the position model, and is a diagram showing a state where the site of a house to be guarded is divided into a plurality of areas (detection target areas). FIG. 11 is a diagram illustrating a specific example of a conditional probability table expressing the transition of the behavior of an object. FIG. 12 is a diagram illustrating a specific example of a conditional probability table expressing the transition of the position of an object. FIG. 13 is a diagram illustrating a specific example of a conditional probability table expressing the initial position of an object. FIG. 14 is a diagram illustrating a specific example of a conditional probability table expressing the initial behavior of an object. FIG. 15 is a diagram illustrating a sensor model, and is a diagram illustrating a specific example of sensor installation in a residential area to be guarded. FIG. 16 is a diagram illustrating a specific example of a conditional probability table representing a sensor model. FIG. 17 is a diagram illustrating another example of sensor installation in the site of a residence to be guarded. FIG. 18 is a diagram illustrating another example of the conditional probability table representing the sensor model. FIG. 19 is a diagram illustrating a specific example of the certainty factor calculation, and is a diagram illustrating an example of a suspicious person's behavior pattern. FIG. 20 is a diagram illustrating a specific example of the certainty factor calculation, and is a diagram illustrating an example of a behavior pattern of a small animal. FIG. 21 is a diagram illustrating a calculation result of the certainty factor in the example of FIG. FIG. 22 is a diagram illustrating a calculation result of the certainty factor in the example of FIG. FIG. 23 is a diagram in which the certainty factors calculated by the certainty factor calculation unit are plotted in time series. FIG. 24 is a diagram in which the certainty factors calculated by the certainty factor calculation unit are plotted in time series. FIG. 25 is a flowchart showing a flow of processing repeatedly executed in the determination device while the security mode is set to the “security set mode”. FIG. 26 is a flowchart showing details of the process in step S5 of FIG. FIG. 27 is a flowchart showing details of processing in the case where determination is performed by comparing the certainty factor with a threshold value in step S6 of FIG. FIG. 28 is a flowchart showing details of processing in the case where determination is performed by comparing the weighted sum of certainty factors with a threshold value in step S6 of FIG.

Exemplary embodiments of a security device and a sensor reaction factor estimation method according to the present invention will be explained below in detail with reference to the accompanying drawings.

(Overview)
The security device according to the present embodiment, for example, when an intruder attempts to enter the inside of a building in an individual house or the like, detects the intruder before entering the inside of the building, issues an alarm, etc. In particular, when multiple sensors are installed outdoors in a building, the sensor reaction signals are obtained in time series from these multiple sensors and the probability response using a dynamic Bayesian network. A sensor reaction factor is estimated using a judgment model that is a model expressing the causal relationship of the above, and when it is determined that the sensor reaction factor is a human, a security operation such as issuing an alarm is performed.

  When trying to detect an intruder based on a reaction signal from a sensor installed outside a building, there is a possibility that the sensor may react erroneously due to small animals such as cats and birds, fallen leaves, vegetation, and plastic bags. Therefore, in the security device according to the present embodiment, the outdoor site is divided into a plurality of areas using the positional relationship with the surrounding environment and the opening, etc., and each area is set as a detection target area to be an object detection target. One or more sensors are installed in each detection target area, and a factor reacting to the sensor is estimated from a time-series change in the reaction state of the sensors belonging to each detection target area and various other information.

  Specifically, an object that can be a sensor reaction factor is divided into several categories (for example, humans, small animals, inanimate objects, etc., hereinafter, the category of an object that can be a sensor reaction factor is referred to as an “object category”). From which time series of sensor response signals obtained from a plurality of sensors, the object category that includes the object generating the time series is estimated. Then, in order to estimate which object category the reaction factor is included in the time series of sensor response, prepare an action model for each object category, and how close to which action model the time series of sensor response is calculate.

  Since the above object categories are humans, small animals, etc., their actions are not uniform. That is, it is not considered that the objects belonging to these object categories take the same action every time, and the moving route, the moving speed, etc. are considered to change every time. Therefore, it is difficult to construct a definitive single model as the behavior model for each object category, and it is necessary to construct a probabilistic behavior model. Therefore, in the security device according to the present embodiment, a probability model using a dynamic Bayesian network is constructed, and a sensor reaction factor is estimated by probability inference using a dynamic Bayesian network.

  That is, in the security device according to the present embodiment, in order to execute probability inference using a dynamic Bayesian network, action selection criteria (behavior models) for various object categories are constructed using probability models, and sensors The relationship between the response of the object and the behavior of the object is used as a sensor model to construct a probability model similar to the behavior model. Then, by using a “judgment model” that combines these two probability models and the position model that represents the positions of the plurality of detection target areas described above, the sensor reaction factor is estimated from the sensor reaction signal obtained in time series. I can do it.

  In the security device according to the present embodiment, it is possible to estimate the sensor reaction factor by the probabilistic reasoning using the dynamic Bayesian network with the configuration as described above, and it is possible to reduce the false alarm due to the sensor erroneous reaction. Furthermore, since it is not a method for predetermining the response pattern of each sensor, there is an advantage that the setting effort is less compared with a method for determining all sensor reaction patterns. Furthermore, there is no particular limitation on the sensor to be used, and there is an advantage that various sensors can be used.

(Dynamic Bayesian network)
In the following, first, an outline of the dynamic Bayesian network will be described prior to specific description of the security device according to the present embodiment. A dynamic Bayesian network is a time-based expansion of a Bayesian network, which is a probabilistic model for estimating the cause from observed events.

A Bayesian network is a visual representation of a causal relationship between a plurality of elements constituting an event, and the causal relationship is expressed by a conditional probability. In other words, as shown in FIG. 1, the Bayesian network represents each element constituting an event as a node represented by a random variable, connects each node having a causal relationship with a directed side (arrow), and connects between the nodes. It is expressed as a model in which the causal relationship is defined by conditional probability. In the example shown in FIG. 1, the causal relationship between the node A and the node B constituting the event is given as the conditional probability P (B | A), and the causal relationship between the node B and the node C is the conditional probability P (C | B). Note that the leading node A has no arrow connected to this node, and the conditional probability cannot be defined, so a prior probability P (A) determined by the random variable of this node itself is given. The model of FIG. 1 expressed by such prior probabilities and conditional probabilities can be expressed by a mathematical formula as shown in the following formula (1).
P (A, B, C) = P (A) P (B | A) P (C | B) (1)

In the Bayesian network as described above, a node that is the starting point of an arrow connected to a certain node is generally called a parent node. Here, when a parent node for a certain node X is represented as Pa (X) and a model having n nodes is represented by a general formula, the following formula (2) is obtained.

  In the Bayesian network, as described above, the causal relationship between each element constituting the event is defined by the conditional probability, but the actual value of each element is generally expressed discretely. In the example shown in FIG. 1, the conditional probability P (B | A) representing the causal relationship between the node A and the node B is expressed by a conditional probability table as shown in FIG. In the conditional probability table of FIG. 2, the probability that node B is “1” when node A is “1” is 0.6, and node B is “0” when node A is “1”. The probability is 0.4, the probability that the node B is “1” when the node A is “0” is 0.1, and the probability that the node B is “0” when the node A is “0” is 0. 9 is shown. In a Bayesian network, a conditional probability table as described above is prepared as a conditional probability between nodes, and a model of the entire network is constructed.

The cause estimation using the Bayesian network is performed by fixing the elements that can be observed and considering the combination of all patterns of unknown elements. For example, in the example illustrated in FIG. 1, when it is observed that the node C is “1”, whether the result that the node C is “1” is due to the node A being “1”. Is estimated under the condition that the node A is “1” and the node B is “1” and the node C is “1”. The sum of the probabilities that node C becomes “1” and node C becomes “1” is calculated. When the above is expressed by a mathematical formula, the following formula (3) is obtained.
P (C = 1 | A = 1) = α (P (A = 1) P (B = 1 | A = 1) P (C = 1 | B = 1) + P (A = 1) P (B = 0 | A = 1) P (C = 1 | B = 0))
_{= Α (Σ B P (A} = 1) P (B | A = 1) P (C = 1 | B)) ··· (3)
Here, α is a normalization constant, and is a constant that performs normalization with the probability of A = 0 when C = 1.

  As described above, in the Bayesian network, the cause probability can be obtained by calculating the sum of the pattern probabilities for all realization values of unknown elements that cannot be observed with fixed observable element values. Estimation is possible.

  The Bayesian network described above is a static model that does not consider the time axis. Therefore, when considering application to a target where a sensor response signal can be obtained every moment as in the security device according to the present embodiment, it is necessary to construct a model corresponding to time series. Therefore, a dynamic Bayesian network, which is a model in which the Bayesian network is developed over time, is used.

  As shown in FIG. 3, the dynamic Bayesian network is a model in which one basic model is used as a unit, and units are added every time an observation result is obtained. The estimation method using such a dynamic Bayesian network is basically the same as the estimation method using the Bayesian network described above, and only the number of nodes increases each time an observation result is obtained.

In the example shown in FIG. 3, the model when the first observation result O1 is obtained is as shown in FIG. When the model of FIG. 4 is expressed as the above equation (1), the following equation (4) is obtained.
P (X0, X1, O1) = P (X0) P (X1 | X0) P (O1 | X1) (4)
Here, it is assumed that the state of X1 is estimated. O1 is fixed for the observation result, and the unknown element is X0. When the above equation (3) is used, the estimated value of the state of X1 can be obtained by calculating as the following equation (5).
P (X1 | O1) = _ΣX0P (X0) P (X1 | X0) P (O1 | X1)
= P (O1 | X1) _ΣX0P (X0) P (X1 | X0) (5)

Next, at the time of the second observation, as shown in FIG. 5, the model is obtained by adding nodes X2 and O2. When the model shown in FIG. 5 is expressed by mathematical formulas as before, the following formula (6) is obtained.
P (X0, X1, X2, O1, O2) = P (X0) P (X1 | X0) P (O1 | X1) P (X2 | X1) P (O2 | X2) (6)
Here, since the unknown elements are X0 and X1, the estimated value of the state of X2 can be obtained by calculating as in the following equation (7).
P (X2 | O1, O2) = Σ X0 Σ X1 P (X0) P (X1 | X0) P (O1 | X1) P (X2 | X1) P (O2 | X2)
= P (O2 | X2) Σ X1 P (O1 | X1) P (X2 | X1) Σ X0 P (X0) P (X1 | X0) ··· (7)

In a dynamic Bayesian network, each time a new observation result is obtained, a unit is added to the model and the calculation is performed as described above, so that the cause can be estimated from the observation result obtained in time series. However, if units are added from time to time, there is a possibility that the model becomes very large after a certain period of time, and the amount of calculation increases accordingly, which is not practical. Therefore, when causal relation estimation using a dynamic Bayesian network is performed, it is desirable to suppress an increase in calculation amount by performing recursive calculation as shown in the following formula (8).
P (Xt + 1 | O1, O2,..., Ot + 1) = P (Ot + 1 | Xt + 1) _.SIGMA.XtP (Xt + 1 | Xt, O1, O2,..., Ot) P (Xt | O1, O2, ..., Ot)
= P (Ot + 1 | Xt + 1) Σ _XtP (Xt + 1 | Xt) P (Xt | O1, O2,..., Ot) (8)
In this equation (8), the last term is the estimated value at the previous timing. Thus, by performing recursive calculation using the estimated value of the previous timing, it is possible to effectively suppress an increase in the amount of calculation.

  Next, the security device according to the present embodiment having the function of estimating the sensor reaction factor by the probability inference using the dynamic Bayesian network described above will be specifically described.

(Device configuration)
FIG. 6 is a configuration diagram showing a schematic configuration of the security device according to the present embodiment. As shown in FIG. 6, the security device according to the present embodiment includes a plurality of sensors 1 installed outside a house to be guarded, a sensor control device 2 installed in a house building, and a determination device. 3, a main body device 4, a display device 5, an operation device 6, and a monitoring device 7 installed in a security company.

  One or more sensors 1 are installed for each detection target area when the site (outdoor) of the house to be guarded is divided into a plurality of detection target areas. In other words, the security device according to the present embodiment uses a plurality of sensors 1, and the type of each sensor 1 is one suitable for object detection in the detection target area where the sensor 1 is installed. . For example, sensors that detect the opening and closing of gates, passive infrared sensors that use pyroelectric elements, sensors that detect the blocking of infrared beams, and the presence or absence of objects and the distance to objects using infrared reflection A sensor or the like can be effectively used as the sensor 1 in the present embodiment.

  The sensor control device 2 is a device for performing control of the plurality of sensors 1 and reception of reaction signals from the sensors 1. In the security device according to the present embodiment, regardless of whether or not the sensor 1 has detected an object, the sensor control device 2 periodically acquires the reaction signal of the sensor 1 at a predetermined measurement cycle, and the determination device 3 Supplied.

  The determination device 3 is a characteristic component in the security device according to the present embodiment. Based on the time series data of the response signal of the sensor 1 that is periodically acquired by the sensor control device 2, the response factor of the sensor Is a device that determines whether a human being, a small animal, or an inanimate object. The details of the determination device 3 will be described later.

  The main device 4 is a device that comprehensively controls the overall operation of the security device, such as driving of a monitoring function and holding various information by the security device according to the present embodiment. The main body device 4 has a function (security means) for issuing an alarm to a security company or a resident of a house to be guarded when the judgment device 3 determines that the reaction factor of the sensor 1 is human. Have. Specifically, when issuing a warning to the security company, the main unit 4 transmits (reports) information to the monitoring device 7 of the security company using a communication device, and warns the residents of the house. When issuing an instruction, for example, a warning message is displayed on the display device 5 or a warning sound is output from a speaker integrated with the display device 5.

  The display device 5 is a device that displays the current guard state of the guard device according to the present embodiment, the reaction state of the sensor 1, the judgment result of the judgment device 3, the judgment process by the judgment device 3, and various other information. . In the security device according to the present embodiment, the display device 5 is used when an alarm is issued to the resident of the house as described above.

  The operation device 6 is a device for a resident or the like of a house to be guarded to perform operations such as security on / off. An operation performed by a resident of the house using the operation device 6 is input to the main device 4 and reflected in the control of the security device according to the present embodiment.

  The monitoring device 7 is a device that acquires various information related to security by communication with the main body device 4 and presents the information to the security guard of the security company. In the security device according to the present embodiment, when the main device 4 issues an alarm to the security company, the monitoring device 7 receives the report from the main device 4 and transmits the alarm to the security personnel of the security company. In the security device according to the present embodiment, the determination device 3, the main device 4, the display device 5, and the operation device 6 are described as being installed in a building of a house to be guarded. It is possible to install some or all of them in a security company.

(Overview of operation)
The security device according to the present embodiment configured as described above has at least two types of security modes of “security set mode” and “security release mode”. For example, a resident of a house to be guarded By performing an operation using the operation device 6, the security mode is switched in the main body device 4. Here, the “security set mode” means that the sensor reaction factor is estimated by the determination device 3 based on the reaction signal of each sensor 1 received by the sensor control device 2, and the sensor reaction factor is determined to be human. This is a mode in which the main device 4 executes a security operation in the case of a failure. On the other hand, the “security release mode” is a mode in which the sensor reaction factor is not estimated by the determination device 3 even if the sensor 1 reacts, and the security operation by the main device 4 is not performed, that is, a security company or a house to be guarded. A mode that does not perform security operations such as issuing warnings to residents of the city.

  In the security device according to the present embodiment, when the security mode is set to the “security set mode”, the sensor control device 2 periodically receives the reaction signal of the sensor 1 at a predetermined measurement cycle, and the determination device 3 Periodically checks the response signal of the sensor 1. Then, each time the determination device 3 checks the response signal of the sensor 1, the determination device 3 performs sensor response factor estimation processing using time-series data of the sensor response signal obtained up to that point. If the sensor reaction factor is determined to be human as a result of the sensor reaction factor estimation process by the determination device 3, the main body device 4 notifies the monitoring device 7 of the security company, and the display device Security operations such as displaying a warning message to 5 and outputting a warning sound from the speaker are executed. On the other hand, if it is determined that the sensor reaction factor is a small animal other than a human being or an inanimate object, the determination device 3 resets the processing up to that point, and then the time series data of the response signal of the sensor 1 newly obtained. The sensor reaction factor estimation process is continued using. Then, the above process is repeated until the security mode is switched from the “security set mode” to the “security release mode”.

  Note that, in the security device according to the present embodiment, the security device or the security is performed as the security operation performed by the main device 4 when the determination device 3 determines that the sensor reaction factor is human during the “security set mode”. The warning is issued to the residents of the target house. In addition to the warning, for example, a threatening device that performs threatening by voice or the like is driven to perform threatening, or an electric lock, electric shutter, door You may make it perform the guard operation of operating the physical defense mechanism for raising physical strength.

(Details of judgment device)
Next, the detail of the judgment apparatus 3 which is a characteristic component in the security apparatus concerning this Embodiment is demonstrated. FIG. 7 is a functional block diagram showing a functional configuration inside the determination device 3. As shown in FIG. 7, the determination device 3 includes a model storage unit 11, a certainty factor calculation unit 12, and a determination unit 13.

  The model storage unit 11 is a database that stores a determination model that is a model that expresses the causal relationship of the sensor reaction by a probability model using a dynamic Bayesian network. As described above, the determination device 3 according to the present embodiment estimates the sensor reaction factor by the probability inference using the dynamic Bayesian network. A determination model, which is a model necessary for this estimation process, is stored in the model storage unit 11. It is remembered.

  The certainty factor calculation unit 12 is determined in advance using time series data of sensor reaction signals periodically acquired from the plurality of sensors 1 by the sensor control device 2 and the determination model stored in the model storage unit 11. For each object category that can be a sensor reaction factor, a certainty factor that an object belonging to each object category is a sensor reaction factor is calculated. Specifically, each time the sensor control signal is acquired by the sensor control device 2, the certainty factor calculation unit 12 checks the sensor response signal and buffers the sensor response signal in an internal memory or the like. Then, time-series data combining the newly obtained sensor response signal and the past sensor response signal is generated as needed, and the time-series data of the sensor response signal is stored in the model storage unit 11 By verifying, the certainty factor that is a sensor reaction factor is calculated for each object category. The certainty factor calculation unit 12 temporarily stores the calculation result in an internal memory or the like, and performs recursive calculation using the previous calculation result when calculating the certainty factor in the next processing cycle. The calculation load is reduced.

  The determination unit 13 determines an object category that is a sensor reaction factor based on the certainty factor calculated by the certainty factor calculation unit 12. Specifically, the determination unit 13 compares the certainty factor or the weighted sum of certainty factors for each object category calculated by the certainty factor calculating unit 12 with a threshold value, and calculates the certainty factor or the certainty factor weighted sum of certain object categories. When the threshold value exceeds the threshold, it is determined that the object category is an object category that is a sensor reaction factor, and the determination result is output to the main unit 4.

(Judgment model)
Next, the determination model used when the sensor reaction factor estimation process is executed in the determination device 3 will be described in more detail. FIG. 8 is a model diagram showing an outline of the judgment model. As shown in FIG. 8, the determination model is a probability model using a dynamic Bayesian network, and is roughly composed of three models: a position model, an action model, and a sensor model. Among these, the behavior model and the sensor model are probability models expressed by a conditional probability table as shown in FIG. In the determination model of FIG. 8, the subscript part of each node represents the observation time, and “0” represents the initial state. This judgment model is a time series model (dynamic Bayesian network). As shown in Fig. 9, a part with a common subscript is added as a unit at each observation timing to the basic model in the initial state. It will be done. Hereinafter, the position model, the behavior model, and the sensor model constituting the determination model will be described with specific examples.

(Position model)
The position model is a model that represents the positions of a plurality of detection target regions that are targets of object detection using the sensor 1. That is, when a private house is a security target, a general house site structure is divided into a plurality of abstract areas according to the surrounding situation and the like, and each area is set as a detection target area. The position of each area at this time is expressed as a position model. Specifically, as shown in FIG. 10, for example, a general residential site structure can be divided into a gate area, a fence area, an access area, a private area, and an opening area. The position of the area is expressed as a position model.

Here, each area of the position model illustrated in FIG. 10 has the following meaning.
Gate area: It is an area where normal access from outside is performed, the gate is normally closed, and opening and closing requires an operation such as pulling a lever.
Fence area: Represents the Fence (障) part.
Access area: This is the area through which a person coming from the outside passes to go to the entrance.
Private area: Represents the space in the back of the garden that is spatially connected to the access area but is off the route from the gate to the entrance.
Opening area: It represents the opening of the building and the space in front of it.

(Behavior model)
The behavior model is a probability model that represents the behavior style of an object belonging to each object category for each object category that can be a sensor reaction factor. That is, the behavior model is based on the conditions as shown in FIG. 2 as to what kind of behavior the object belonging to each object category moves within the building site represented by the above position model. It is a model expressed in the form of an attached probability table. As shown in FIG. 8, the determination model used in the determination apparatus 3 according to the present embodiment has two types of nodes, namely, an object action (At) and an object position (Xt). Heading to Therefore, the behavior model has two conditional probability tables: a conditional probability table that expresses the transition of the behavior of the objects belonging to each object category, and a conditional probability table that expresses the transition of the position of the object belonging to each object category. Represented by

  Here, before explaining each conditional probability table, an example of specific realization values of the object category, the object action, and the object position which are nodes of the determination model (dynamic Bayesian network) shown in FIG. 8 will be explained. .

  The object category is a classification of sensor reaction factors, and in this embodiment, there are three categories: human, animal, and inanimate. Here, inanimate is a category that collectively refers to factors that may cause sensor false alarms other than animals, such as sunlight, wind, and flying objects such as garbage bags.

  The object action is a classification of actions that can be taken by the object category. In this embodiment, the action is classified into four actions of disappearance, entry, movement, and stay. Here, “disappearance” means an operation that disappears from outside the site, and refers to a movement that leaves the outside. Moreover, approach means the operation | movement which enters the site inside from the outside. Moreover, movement means the operation | movement which moves within the site. A stay means an operation that hardly moves from the current position in the site.

  As shown in FIG. 10, the object position represents each area in the site expressed in the above position model.

  Next, two conditional probability tables expressing the behavior model will be described. FIG. 11 shows a specific example of the conditional probability table expressing the transition of the behavior of the object, and FIG. 12 shows a specific example of the conditional probability table expressing the transition of the position of the object.

  The conditional probability table that expresses the transition of an object's behavior represents the probability of the next action to be selected under the condition that the object exists at a certain position in the site by the previously selected action. Is. In other words, when the object takes a certain action and is at the current position, it represents the probability of selecting which action to continue. When the conditional probability of such a behavior transition is expressed by an equation, P (action | object, current position, previous behavior) is obtained. By expressing this conditional probability as a conditional probability table as shown in FIG. 2 using the above-mentioned classification of object behavior and the position model shown in FIG. 10, the behavior of the object as shown in FIG. A conditional probability table representing the transition is obtained.

  A conditional probability table that expresses the transition of the position of an object is a probability that expresses where the current position is, under the condition that the object has taken some action at the position where it existed at the previous timing. It is a thing. In other words, it represents the probability of where the object is after taking an action that is the position where the object was in front. When the conditional probability of such a position transition is expressed by an equation, P (position | object, previous position, previous action) is obtained. By expressing this conditional probability as a conditional probability table as shown in FIG. 2 using the above-mentioned classification of object behavior and the position model shown in FIG. 10, the position of the object as shown in FIG. A conditional probability table representing the transition is obtained.

  By the way, in the initial state determination model (the basic model shown on the left side of FIG. 9), there is no action or position node at the previous timing, so the conditional probability tables shown in FIGS. 11 and 12 are applied as they are. It is not possible. Therefore, a conditional probability table corresponding to such an initial state, that is, a conditional probability table that is suitable when the sensor 1 reacts first is prepared.

  With respect to the position in the initial state, where the object that becomes the reaction factor of the sensor 1 first exists in the building site (initial position) can be expressed as the conditional probability P (position | object), which is shown in FIG. When the conditional probability table as shown in FIG. 6 is expressed, the conditional probability table at the initial position as shown in FIG. 13 is obtained. In addition, regarding the behavior in the initial state, what kind of behavior the object that is the reaction factor of the sensor 1 takes first (initial behavior) can be expressed as a conditional probability P (action | object, current position), If this is expressed as a conditional probability table as shown in FIG. 2, a conditional probability table of initial behavior as shown in FIG. 14 is obtained.

  In the determination model shown in FIG. 8, since the nodes representing the object category (Ct) are connected by arrows, it is necessary to define a conditional probability for the causal relationship between the nodes representing the object category. However, it is considered that the object does not change while it is moving (for example, a human is not changed to a small animal), so the probability that the previous object and the current object are equal will change by 1.0. The probability should be given as 0.

(Sensor model)
The sensor model is a probabilistic model that represents the relationship between the position where the object exists and the reactions of the plurality of sensors 1. In other words, the sensor model represents how each sensor 1 reacts when an object is present at a certain position. As shown in the sensor model portion of the judgment model shown in FIG. 8, the sensor model is formed by connecting the node of the object position (Xt) and the node of the sensor response signal (Ot) as an observation result with an arrow. Defined by the conditional probability P (sensor | position). The sensor response signal takes a binary value of reaction / no reaction. Hereinafter, the sensor model will be described in more detail with specific examples.

  As shown in FIG. 10, the area of the house to be guarded is divided into five detection target areas: a gate area, an obstacle area, an access area, a private area, and an opening area. As shown in FIG. Outdoor open / close sensor 1a in the area, two opposed infrared sensors 1b and 1c in the obstacle area, outdoor space sensor 1d in the access area, outdoor space sensor 1e in the private area, and stay detection sensor 1f in the opening area. Each shall be installed. The outdoor open / close sensor 1a is a sensor that detects opening and closing of the gate, and the opposed infrared sensors 1b and 1c are sensors that detect blocking of the infrared beam. The outdoor space sensors 1d and 1e are passive infrared sensors using pyroelectric elements, and the stay detection sensor 1f is a sensor that detects the presence of an object and the distance to the object using infrared reflection. The detection areas of these sensors 1a to 1f are hatched areas in FIG.

  By expressing the conditional probability P (sensor | position) as a conditional probability table as shown in FIG. 2 under the assumption that the sensors 1a to 1f are installed in the residential premises as described above. A conditional probability table of the sensor model as shown in FIG. 16 is obtained. As described above, the sensor response signals obtained from the sensors 1a to 1f are binary values of reaction / no response, and the conditional probability table shown in FIG. The conditional probability shown in FIG. 16 expresses which sensor reacts when an object exists at a certain position in the site.

  In the conditional probability table of this sensor model, there is also a probability that an unreacted state in which none of the sensors 1a to 1f reacts even if an object exists in any area (detection target area) in the site. Giving. This is because it is difficult to completely cover each area (detection target area) defined by the position model with the detection areas of the sensors 1a to 1f, and even if an object is present in the detection area, the sensors 1a to 1f. This is because it may not respond by mistake (not reported). Thus, by giving the probability of a non-responsive state to the conditional probability table of the sensor model, it is not necessary to cover the entire area in the site defined by the position model with the detection area of the sensor 1, and the sensor It is possible to cope with the case where 1 does not react by mistake.

  As described above, the determination model used in the determination device 3 according to the present embodiment generates the position model by dividing the site of the house to be guarded into areas and generates four conditional probability tables ( 11 to 14) and a conditional probability table (FIG. 16) representing the sensor model. The above five conditional probability tables that make up the decision model are human (intruder), small animal, inanimate behavior patterns and corresponding sensor response states converted into numerical values. However, by creating these conditional probability tables using empirical know-how related to security, it is possible to construct an optimum judgment model for application to a security device. Hereinafter, an example of a specific creation criterion for the conditional probability table constituting the judgment model suitable for security will be described.

(Human behavior pattern)
First, the following assumptions are made as a basic way of thinking about human behavior patterns.
“Human moves on the ground (does not fly in the sky) and moves from outside the site to the site.
In order to make such an assumption, in the conditional probability table of the initial position shown in FIG. 13, the probability that a person exists in the area inside the site (access area, private area, opening area, etc.) in the initial state is 0. To do.

  In addition, since the definition of the object action described above defines that “entry” is to move from outside the site to inside the site, the conditional probability table for the initial action shown in FIG. 14 is other than the gate area and the obstacle area. In this area, the probability that the initial action of a human being is “entry” is set to zero. Similarly, in the conditional transition probability probability table shown in FIG. 11, it exists in an area (access area, private area, opening area, etc.) inside the site without depending on the action at the previous timing. The probability that a human selects “approach” is zero. Further, in the conditional probability table of position transition shown in FIG. 12, in the position model shown in FIG. 10, it is impossible to exist in front of the entrance immediately after entering (the distance is too far), When the previous action is “entry”, the probability that a person exists in the opening area is set to zero.

  The above is a description of an explicit action pattern regarding the physical behavior of a person based on the definition of the object action and the object position. For other parts of the conditional probability tables shown in FIGS. 11 to 14, security know-how is applied to human behavior such as intruders. Below, the example is demonstrated easily.

  Intruders tend to dislike being seen from the surroundings. Therefore, there is a tendency to prefer a place / behavior where the prospect is poor, a place / behavior that is not suspicious from the surroundings even if it is seen, and a place / behavior whose purpose is achieved in an extremely short time. Therefore, in the conditional probability table for the initial position shown in FIG. 13, “gate” is set higher than “enclosure” as the probability of the initial position of the human. This is because opening a gate is like a normal visitor and is less likely to be suspicious, and it is considered possible to enter the site in a shorter time than getting over. However, when the gate itself is strong, when the gate is firmly locked, or when the gate is high, the probability of the initial position of the person should be set higher than the gate. May be preferred.

  Similarly, staying around the gate is at risk of being found and suspicious from the surroundings, so the conditional transition probability table shown in FIG. 11 shows that in the gate area regardless of the behavior at the previous timing. The probability of selecting “move” rather than “stay” is set higher as an action taken by a human being.

  Also, the access area is an area directly connected from the gate, and when the gate is in a lattice shape, the visibility of the access area is improved. Therefore, in the conditional transition probability probability table shown in FIG. 11, the probability that a person will select “stay” is set low in the access area, and the probability that “movement” will be selected is set high. In addition, since the private area is surrounded by obstacles and the risk of being seen from the surroundings is low even when staying, the probability that a person existing in the private area will select “stay” exists in the access area. Set higher than if

  Humans go inside the building and go through the opening to enter the building. Therefore, the human heads for the opening area. Therefore, the probability that a person will select “move” is generally high in areas other than the opening area, and conversely, the probability that a person will select “stay” is high in the opening area. The conditional probability table for position transition shown in FIG. 12 represents the directionality of the movement of the object. While the intruder moves in an area where the line of sight is poor, the position changes so as to go to the area of the opening where the line of sight is poor and easy to enter. For this reason, in the conditional probability table of position transition shown in FIG. 12, the transition probability of the human position is given so as to proceed to the area preferred by the intruder.

  Conditional probabilities relating to human positions and actions are determined according to the above criteria, but these are adjusted depending on the physical structure of the house to be guarded and the surrounding environment. For example, if the fence is about 1 m, it may take a short time to get over. In such a building, the initial position of the human in the condition probability table of the initial position shown in FIG. To increase the probability of

(Small animal behavior patterns)
As for humans, as in humans, it is assumed that “move the ground (do not fly in the sky) and move from outside the site to the site. The ground is also moved within the site”. In order to make such an assumption, in the conditional probability table of the initial position shown in FIG. 13, the probability that the initial position of the small animal is “enclosure” is set high. In addition, since it is thought that a small animal does not open and close a gate, the probability that it exists in a gate area is set to 0.

  In the other conditional probability tables, items regarding small animals in each conditional probability table are filled in on the assumption that “small animals randomly select an action and move without a direction”.

(Inanimate pattern)
Inanimate objects include natural factors such as wind and sunlight, as well as flying objects such as garbage bags. Therefore, any position in the site can be an inanimate initial position, but the probability depends on the size of each area. For this reason, in the conditional probability table of the initial position shown in FIG. 13, the magnitude of the probability of the inanimate initial position is set according to the size of each area.

  Further, since it is considered that an inanimate object rarely moves, the probability of selecting “move” gives a very small probability in any state (position, behavior at the previous timing). In addition, since inanimate factors may disappear (sunlight is covered with clouds or the sun is tilted), the probability of selecting “disappear” as an action is set high. In accordance with the above concept, the inanimate items in the conditional probability tables shown in FIGS.

(Sensor response signal pattern)
The sensor response signal does not change depending on the behavior pattern of the object, but the probability of reaction changes depending on the type of sensor 1, the size of the detection area, and the installation position of sensor 1. The conditional probability table of the sensor model shown in FIG. 16 shows which sensor reacts when the object is present at a certain position X. In the example shown in FIG. The probability that a sensor other than the sensor installed in an area will react is set to zero.

  In the conditional probability table of the sensor model shown in FIG. 16, as described above, the probability that an object is present but does not react erroneously as the probability of “no reaction” depends on the type of sensor 1. Giving. In the example shown in FIG. 15, the outdoor open / close sensor 1a installed in the gate area, the opposed infrared sensors 1b and 1c installed in the enclosed area, and the stay detection sensor 1f installed in the opening area may not react by mistake. Since it is assumed that there is very little, the probability of no reaction is set to a very small value. On the contrary, the outdoor space sensors 1d and 1e installed in the access area and the private area assume that the probability of not reacting erroneously is high to some extent, so the probability of no reaction is set higher than that of other sensors.

  By the way, in the example shown in FIG. 15, it is assumed that each area (detection target area) defined by the position model is substantially covered by the detection areas of the sensors 1a to 1f. It is often difficult to completely cover all areas (detection target areas) with one detection area. Therefore, in such a case, the conditional probability table of the sensor model is also considered in consideration of the possibility that an object exists in the area defined by the position model but outside the detection area of the sensor 1. Determine the probability of “no reaction”.

  For example, as shown in FIG. 17, an outdoor space sensor 1g having a part of an access area as a detection area, a part of the access area, and a private area, as shown in FIG. It is assumed that an outdoor space sensor 1h having a part of the detection area as a detection area and an outdoor space sensor 1i having a part of the private area as a detection area are installed (the hatched areas in the figure are each sensor 1g). -1i detection areas are shown). In the example shown in FIG. 17, the area deviating from the sensor detection area is wide particularly in the private area, so even if an object exists in the private area, there is a high probability of being in an unresponsive state. Therefore, in the conditional probability table of the sensor model, when the position is a private area, the probability is set low according to the area missing from the sensor detection area. FIG. 18 shows an example of the conditional probability table of the sensor model in such a case. For simplicity, only the access area and private area are described here.

  In the example shown in FIG. 17, the outdoor space sensor 1h is installed across the access area and the private area. In this case, the outdoor space sensor 1h may react even if an object exists in either the access area or the private area. Therefore, in the conditional probability table of the sensor model shown in FIG. 18, the probability of the outdoor space sensor 1h is not set to 0 regardless of whether the position of the object is an access area or a private area. Gives a probability depending on whether or not

(Calculation of confidence)
Next, in the certainty factor calculation unit 12 of the determination device 3, a method of calculating the certainty factor that is a sensor reaction factor for each object category using the time-series data of the response signal of the sensor 1 and the determination model described above. Will be described.

The calculation formula for cause estimation using the Bayesian network is composed of a product and a sum of conditional probabilities as shown in the above formula (3). When the above equation (3) is applied to the determination model of FIG. 8 in order to calculate confidence using the determination model as shown in FIG. 8 in the sensor reaction factor estimation algorithm, the following equation (9 )become that way.
P (C0 = c | O0 = o) = α (P (C0 = c) Σ X Σ A P (A0 | C0 = c, X0) P (X0 | C0 = c) P (O0 = o | X0)) ... (9)
Moreover, the calculation formula of the certainty factor when the observed value after the second time is obtained is expressed as a calculation formula for performing recursive calculation as in the following formula (10).
P (C = c | Ot = ot, Ot-1 = ot-1, Ot-2 = ot-2,...)
_{= Α ((Σ Xt Σ At} Σ Xt-1 Σ At-1 P (At | C = c, Xt, At-1) P (Xt | C = c, At-1, Xt-1) P (Ot = ot | Xt)) P (C = c | Ot-1 = ot-1, Ot-2 = ot-2,...)) (10)
In the above formulas (9) and (10), C represents an object category, and O represents a sensor response signal (reacted sensor). A represents the action of the object, and X represents the position where the object exists. The subscript of each variable represents the observation timing, and the subscript “0” represents the initial state. Further, Σ _A (...) Means summing with respect to the action A, and Σ _X (...) Means summing with respect to the position X.

  In the above equation (9), the conditional probability at the initial position is obtained by referring to the conditional probability table as shown in FIG. 14 for P (A0 | C0 = c, X0) representing the conditional probability of the initial action. With respect to P (X0 | C0 = c) representing the value, a value is referred from the conditional probability table as shown in FIG. 13, and P (O0 = o | X0) representing the conditional probability of the sensor model is shown in FIG. By referring to the value from the conditional probability table, the certainty P (C0 = c | O0 = o) that the object category of the reaction factor is c when o is observed as the sensor response signal in the initial state. Can be calculated.

  Further, in the above equation (10), a value is referred to from the conditional probability table as shown in FIG. 11 with respect to P (At | C = c, Xt, At-1) representing the conditional probability of action transition, With respect to P (Xt | C = c, At-1, Xt-1) representing the conditional probability of the position transition, values are referred from the conditional probability table as shown in FIG. 12, and the conditional probability of the sensor model is represented. With respect to P (Ot = ot | Xt), the previous calculation result P (C = c | Ot-1 = ot-1, Ot-2 = ot- is referred to from the conditional probability table as shown in FIG. 2)), when o is observed as the second and subsequent sensor reaction signals, the certainty factor P (C = c) that the object category of the reaction factor is c | Ot = ot, Ot-1 = ot-1, Ot-2 = ot-2,.

上記の式（１１）から分かるように、知りたい物体カテゴリと観測で得たセンサ反応の部分は固定し、行動Ａと位置Ｘの組み合わせを変えた全てのパターンの確率の総和をとっている。それぞれの条件付確率は、図１１〜図１４、図１５に示したような条件付確率表に照らし合わせることで即座に具体的な数値が割り当てられ、それらの数値を使って確信度を算出する。 Here, in the example shown in FIG. 15, when the reaction of the outdoor open / close sensor 1a installed in the gate area is observed as the sensor response signal in the initial state, the certainty factor that the sensor response factor is human is calculated. If the above equation (9) is embodied with the assumption of the above, the following equation (11) is obtained.

As can be seen from the above equation (11), the object category desired to be known and the sensor response portion obtained by observation are fixed, and the sum of the probabilities of all patterns with different combinations of action A and position X is taken. Each conditional probability is immediately assigned a specific numerical value by referring to the conditional probability tables as shown in FIGS. 11 to 14 and FIG. 15, and the certainty is calculated using those numerical values. .

  In the above equations (9) to (11), α at the beginning of the right side is a normalization constant, and is calculated when c is changed in three object categories: human, small animal, inanimate. The coefficient is such that the sum of two values is 1.0. In actual calculation, the certainty factor is calculated for each of the three object categories using the above formula without considering α, and after the three values are obtained, each value is divided by the sum of the three values. To gain confidence.

(Specific example of confidence calculation)
Next, a specific example of calculating the certainty factor for each object category assuming the behavior patterns of the objects shown in FIGS. 19 and 20 will be described. FIG. 19 shows an example of a case where a suspicious person enters the access area from the obstacle area and goes directly to the opening area (entrance). It is assumed that the sensor has reacted three times with {enclosure area, access area, opening area}. On the other hand, FIG. 20 shows an example in which a small animal gets over the fence and enters the access area as in FIG. The sensor response in this case is {enclosure area, access area, enclosure area}.

  The certainty factor for each object category is calculated by using the above equation (9) and the conditional probability tables as shown in FIGS. 13, 14, and 16 for the first sensor response. From the second sensor response, calculation is performed recursively using the above equation (10) and the conditional probability tables as shown in FIGS. 11, 12, and 16 so that the calculation amount does not increase.

  FIG. 21 shows the result when the certainty factor is calculated for each object category in the example of FIG. In addition, FIG. 22 shows the result when the certainty factor is calculated for each object category in the example of FIG. In these FIG. 21 and FIG. 22, the real part represents the estimated value (confidence) of each sensor reaction factor. As shown in FIG. 21, in the case of the example of FIG. 19, it starts from the reaction of the sensor in the surrounding area facing the road, and is initially indistinguishable from the animal. However, the fact that the directionality goes straight to the opening area (entrance) has been found (that is, the sensor response O2 has been obtained), and the confidence factor that the sensor reaction factor is human is high. It becomes. On the other hand, as shown in FIG. 22, in the case of the example of FIG. 20, the reaction starts from the sensor in the surrounding area as in the example of FIG. . By obtaining a sensor response sequence that does not have a specific purpose, a certainty factor that a small animal is a sensor reaction factor as a moving object without a purpose has a high value.

(Determination of the object category that is the sensor response factor)
Next, a method for determining the object category that is a sensor reaction factor in the determination unit 13 of the determination device 3 based on the certainty factor for each object category calculated by the certainty factor calculation unit 12 will be described.

  As a method of determining the object category that is the sensor reaction factor in the determination unit 13 of the determination device 3, for example, the certainty factor for each object category calculated by the certainty factor calculation unit 12 is compared with the threshold θ1. Thus, when the certainty factor of a certain object category exceeds the threshold value θ1, a method for determining that this object category is a sensor reaction factor can be considered. FIG. 23 plots the certainty factors calculated by the certainty factor calculation unit 12 in time series. When such a series is obtained, the determination unit 13 determines the sensor reaction factor by providing the threshold θ1 with respect to the certainty factor.

That is, the determination unit 13 compares the certainty factor M ^C t for each object category (C represents the object category and t represents the observation timing) calculated by the certainty factor calculation unit 12 at each observation timing with the threshold θ1. , Θ1 <M ^C t is interrupted when an object category appears. If C (object category) satisfying θ1 <M ^C t is a human, the determination result is output to the main unit 4. Thereby, for example, a security operation such as issuing a warning to a security company or a resident of the house is executed by the main unit 4. Meanwhile, θ1 <M C ^t become C (object category) is the case were small animals and inanimate objects resets the process so far, and restarts a new determination process. For example, a value such as 0.8 or 0.9 is used as the threshold θ1.

Further, as a method of determining the object category that is a sensor reaction factor in the determination unit 13 of the determination device 3, the weighted sum of the certainty factors M ^C t for each object category calculated every moment by the certainty factor calculation unit 12. May be adopted in which the object category is determined to be a sensor reaction factor when the weighted sum of certainty factors M ^C t of a certain object category exceeds the threshold θ2.

If the time series of the calculated by the confidence factor computing unit 12 confidence M ^{C t} is a pattern as shown in FIG. 23, the confidence M ^{C t} which is calculated for each observation time by simply comparing the threshold value θ1 the determination of the sensor reaction factor can be performed appropriately, but when plotted in time series calculated by the confidence factor computing unit 12 confidence M ^{C t,} sometimes a pattern as shown in FIG. 24. One of the causes of the pattern shown in FIG. 24 is that the behavior model is not adjusted. However, since there is little information obtained in the early stage of observation, it is difficult to distinguish between humans and small animals. when is obtained, time series of such confidence M ^{C t} may be obtained. In addition to the pattern as shown in FIG. 24, there is a case where the certainty factor M ^C t for each object category shows a close value and does not appear to greatly exceed the threshold.

Therefore, in order to cope with the cases above, rather than the confidence M ^{C t} obtained for each observation time of being each time thresholding, comprehensively to determine, when confidence M calculated in sequence taking the integral of ^{C t} (summing of every confidence M ^{C t} for each object category). In this case, confidence M ^{C t} the early stages of observation is considered unreliable, and integrating over low weight in confidence M ^{C t} early stage, calculates a weighted sum IM C ^t confidence. Here, when the weight which becomes lower in the initial stage of the time series is W (t), the weighted sum IM ^C t of the certainty is IM ^C t = IM ^C t−1 + W (t) · M ^C t. .

The determination unit 13 compares the weighted sum IM ^C t of the certainty calculated as described above with the threshold θ2, and interrupts the determination process when an object category that satisfies θ2 <IM ^C t appears. If C (object category) satisfying θ2 <IM ^C t is a human, the determination result is output to the main body device 4 to execute a guard operation. On the other hand, .theta.2 <the IM C ^t C (object category) is the case were small animals and inanimate objects resets the process so far, and restarts a new determination process. For example, a value such as 1.5 or 2.0 is used as the threshold θ2. The weight W (t) is a value that increases as the observation time elapses, and various monotonically increasing functions can be used. The sigmoid function W (t) = 1.0 / (1. It is considered optimal to use 0 + exp ^−at ).

(Operation of the judgment device)
Next, a series of processing executed by the characteristic determination device 3 in the security device according to the present embodiment will be described with reference to the flowcharts of FIGS. FIG. 25 is a flowchart showing a flow of processing repeatedly executed in the determination device 3 while the security mode is set to “security set mode”. FIG. 26 shows details of the processing in step S5 of FIG. It is a flowchart to show. FIG. 27 is a flowchart showing details of processing in the case where determination is performed by comparing the certainty factor with a threshold value in step S6 of FIG. 25, and FIG. 28 is a weighting of certainty factor in step S6 of FIG. It is a flowchart which shows the detail of a process in the case of determining by comparing a sum with a threshold value.

  When the flow of FIG. 25 starts, first, the certainty factor calculation unit 12 of the determination device 3 waits until the sensor reaction signal observation timing (sensor response signal reception timing by the sensor control device 2) is reached (step S1: No). ) When the sensor response signal observation timing comes (step S1: Yes), the sensor response signal acquired by the sensor control device 2 is checked (step S2). If none of the plurality of sensors 1 is reacting as a result of the check in step S2 (step S3: No), the process returns to step S1 and the subsequent processing is repeated, and any one of the sensors 1 reacts. If it is detected (step S3: Yes), the sensor number of the sensor 1 that is responding is stored in the internal memory or the like as the initial sensor response (step S4).

  Next, using the judgment model stored in the model storage unit 11, the certainty factor calculation unit 12 performs a certainty factor calculation process that is a sensor reaction factor for each object category (step S5). And based on the calculation result of the certainty factor in this step S5, the determination part 13 performs the process which determines the object category used as the sensor reaction factor (step S6). Details of these processes will be described later.

  If the result of determination in step S6 is that the sensor reaction factor is not fixed (step S7: No), the certainty factor calculation unit 12 waits until the next sensor response signal observation timing is reached (step S8: No). When the sensor response signal observation timing is reached (step S8: Yes), the sensor response signal acquired by the sensor control device 2 is checked (step S9). If none of the plurality of sensors 1 has reacted as a result of the check in step S9 (step S10: No), a sensor number (number 0) indicating that the sensor is unresponsive is stored in the internal memory or the like. (Step S11), if any one of the sensors 1 is reacting (step S10: Yes), the sensor number of the reacting sensor 1 is stored in an internal memory or the like (step S12).

  Then, the certainty factor calculation unit 12 generates the time series data of the sensor reaction signal using the sensor number stored in the internal memory or the like in the processing so far, and stores the time series data of the sensor reaction signal and the model storage unit 11. The certainty factor calculation process is performed using the determined determination model (step S5). And the determination part 13 performs the process which determines the object category used as the sensor reaction factor based on the reliability newly calculated by the reliability calculation part 12 (step S6).

  The certainty factor calculation unit 12 and the determination unit 13 repeat the above processing until the sensor reaction factor is determined. When the sensor reaction factor is determined (step S7: Yes), the determination result that the sensor reaction factor is human is output to the main body device 4 only when the sensor reaction factor is human (step S13: Yes). (Step S14). Thereby, for example, a security operation such as issuing a warning to a security company or a resident of the house is executed by the main body device 4. Thereafter, the process up to that point in the determination device 3 is reset (step S15), and the process returns to step 1 to restart a new process. On the other hand, if the sensor reaction factor is a small animal other than a human being or an inanimate object, the process up to that point is reset without outputting the determination result to the main unit 4 (step S15), and the process returns to step 1 and a new one is returned. Processing resumes.

  In the certainty factor calculation process in step S5 of FIG. 25, as shown in the flowchart of FIG. 26, the process is switched depending on whether the process in the initial state where the sensor 1 first reacted or the second and subsequent processes. That is, in the case of processing in the initial state in which the sensor 1 has reacted first (step S101: Yes), the above conditional expression (9) and the conditional probability table of the initial position as shown in FIG. Using the conditional probability table of initial behavior as shown in FIG. 16 and the conditional probability table of sensor model as shown in FIG. 16, the certainty factor that is a sensor reaction factor is calculated for each object category (step S102). ). On the other hand, in the case of the second and subsequent processes (step S101: No), the above equation (10), the conditional probability table of action transition as shown in FIG. 11, and the position transition as shown in FIG. Using the conditional probability table of FIG. 16 and the conditional probability table of the sensor model shown in FIG. 16, a certainty factor that is a sensor reaction factor is calculated for each object category (step S103).

  The factor determination process in step S6 of FIG. 25 is a process as shown in the flowchart of FIG. 27 when determination is performed by comparing the certainty factor for each object category calculated in step S5 with the threshold value θ1. That is, P (human) is the certainty that the human is the sensor reaction factor, P (small animal) is the certainty that the small animal is the sensor reaction factor, and P ( First, it is determined whether P (human) exceeds the threshold θ1 (step S201). If P (human) exceeds the threshold θ1 (step S201: Yes), the sensor reaction factor Is determined to be a human (step S202). If P (human) is less than or equal to the threshold θ1 (step S201: No), it is then determined whether P (small animal) exceeds the threshold θ1 (step S203), and P (small animal) exceeds the threshold θ1. If so (step S203: Yes), it is determined that the sensor reaction factor is a small animal (step S204). If P (small animal) is equal to or less than the threshold θ1 (step S203: No), it is next determined whether P (inanimate) exceeds the threshold θ1 (step S205), and P (inanimate) exceeds the threshold θ1. If so (step S205: Yes), the sensor reaction factor is determined to be inanimate (step S206), and if P (inanimate) is equal to or less than the threshold θ1 (step S205: No), the sensor response factor is not determined. (Step S207).

  Further, in the factor determination process in step S6 of FIG. 25, when the weighted sum of the certainty factor for each object category calculated in step S5 is compared with the threshold value θ2, the determination is performed as shown in the flowchart of FIG. It becomes processing. That is, when the time from the start of the determination process to the present is t, the certainty of the object category x is P (x), and the weight at the time t is α (t), first, the weight of the certainty for each object category is first given. The sum αP (x, t) is calculated (steps S301 to S303). Then, it is determined whether or not the weighted sum αP (human, t) of the certainty that the human is the sensor reaction factor exceeds the threshold θ2 (step S304), and αP (human, t) exceeds the threshold θ2. If (step S304: Yes), it is determined that the sensor reaction factor is human (step S305). If αP (human, t) is equal to or smaller than the threshold θ2 (step S304: No), then, the weighted sum αP (small animal, t) of certainty that the small animal is a sensor reaction factor exceeds the threshold θ2. (Step S306), if αP (small animal, t) exceeds the threshold θ2 (step S306: Yes), it is determined that the sensor reaction factor is a small animal (step S307). If αP (small animal, t) is equal to or smaller than the threshold θ2 (step S306: No), then the weighted sum αP (inanimate, t) of certainty that inanimate is the sensor reaction factor exceeds the threshold θ2. (Step S308), if αP (inanimate, t) exceeds the threshold θ2 (step S308: Yes), it is determined that the sensor reaction factor is inanimate (step S309), and αP (inanimate, If t) is equal to or smaller than the threshold value θ2 (step S308: No), the sensor reaction factor is not determined (step S310).

  As described above in detail with reference to specific examples, in the security device according to the present embodiment, the determination device 3 that estimates the sensor reaction factor is a causal factor of the sensor response based on a probability model using a dynamic Bayesian network. Using a model storage unit 11 that stores a determination model, which is a model expressing a relationship, and sensor response signals and determination models obtained from a plurality of sensors 1 in time series, sensor response factors are provided for each object category. A dynamic Bayesian network including a certainty factor calculating unit 12 that calculates a certainty factor and a determination unit 13 that determines an object category that is a sensor reaction factor based on the certainty factor calculated by the certainty factor calculating unit 12; The sensor reaction factor is estimated by the probabilistic reasoning using. Therefore, according to the security device according to the present embodiment, it is possible to reduce false alarms due to sensor erroneous reaction. Furthermore, the security device according to the present embodiment does not require a complicated process of predetermining all the sensor response patterns in order to estimate the sensor response factor, and has an advantage that the setting effort is small. Furthermore, the security device according to the present embodiment has an advantage that various sensors can be used without any particular limitation on the sensors to be used.

  As described above, the security device and the sensor reaction factor estimation method according to the present invention are useful as a technique for accurately determining the attribute of an intruding object based on a sensor reaction signal, and in particular, a sensor erroneous reaction. It is suitable for security devices that are required to reduce the resulting false alarms.

DESCRIPTION OF SYMBOLS 1 Sensor 2 Sensor control apparatus 3 Judgment apparatus 4 Main body apparatus 5 Display apparatus 6 Operation apparatus 7 Monitoring apparatus 11 Model memory | storage part 12 Certainty factor calculation part 13 Judgment part

Claims (10)

A plurality of sensors installed one or more in each of a plurality of detection target areas to be object detection;
A model storage means for storing a judgment model, which is a model expressing the causal relationship of the sensor response by a probability model using a dynamic Bayesian network;
Using the sensor reaction signals obtained in time series from the plurality of sensors and the judgment model, for each object category that can be a sensor reaction factor, a certainty factor that an object belonging to each category is a sensor reaction factor is calculated. A certainty factor calculating means,
A determination unit that determines a category of an object that is a sensor reaction factor based on the certainty factor calculated by the certainty factor calculating unit;
A security device that performs a security operation according to a determination result by the determination device.   The determination model includes a position model representing positions of the plurality of detection target regions, a behavior model that is a probability model representing a behavior mode of an object belonging to each category for each category of the object that can be a sensor reaction factor, The security device according to claim 1, comprising a combination of a sensor model that is a probability model that represents a relationship between an existing position and a response of the plurality of sensors.   The sensor model is a probabilistic model having a probability of becoming a non-reactive state in which none of the plurality of sensors is reacting even though an object exists in any of the plurality of detection target regions. The security device according to claim 2.   The behavior model is characterized in that a conditional probability of a behavior transition of an object belonging to each category and a conditional probability of a transition of a position are determined for each category of the object that can be a sensor reaction factor. The security device according to 2 or 3.   The certainty factor calculating means uses the conditional probability of the transition of the behavior of the object and the conditional probability of the transition of the position determined by the behavior model to obtain a probability of taking an action pattern that matches the time series of the sensor response signal. Obtaining each object category and multiplying the probability by the probability of the sensor response represented by the sensor model to calculate the certainty that the object belonging to each category is a sensor reaction factor. The security device according to claim 4, wherein   The behavior model includes conditional probabilities that match when any of the plurality of sensors reacts first as conditional probabilities of behavior transition and positional transition of the object, and the plurality of sensors. 6. The security device according to claim 4 or 5, wherein a conditional probability that suits when any of the above reacts for the second time or later is determined.   The determination unit compares the certainty factor calculated by the certainty factor calculating unit with a first threshold value, and determines a category of an object having a certainty factor exceeding the first threshold value as a sensor reaction factor. The security device according to claim 1, wherein the security device is determined to be a category of an existing object.   The determination means obtains a weighted sum of certainty factors, which is a value obtained by adding the certainty factors calculated in time series by the certainty factor calculating unit by applying a weight that becomes a lower value in the initial stage of the time series. Is compared with a second threshold, and a category of an object for which the weighted sum of the certainty exceeds the second threshold is determined to be a category of an object that is a sensor reaction factor. The security device according to any one of claims 1 to 6.   The security means issues a warning to a security company or a resident of a guarded house when the category of the object that is a sensor reaction factor is determined to be human by the determination means. The security device according to any one of claims 1 to 8, characterized in that: A method of estimating a sensor reaction factor by a plurality of sensors installed in each of a plurality of detection target regions to be object detection targets,
A category of an object that can be a sensor reaction factor using a sensor reaction signal obtained in time series from the plurality of sensors and a judgment model that is a model expressing a causal relationship of the sensor reaction by a probability model using a dynamic Bayesian network For each step, calculating the certainty that an object belonging to each category is a sensor reaction factor;
A method of estimating a sensor reaction factor, comprising: determining a category of an object that is a sensor reaction factor based on the certainty factor calculated for each category of an object that can be a sensor reaction factor.

Images