WO2010032446A1 - Device for recognizing open/close operation of open/close member and method for recognizing open/close operation of open/close member - Google Patents

Abstract

A sensor microphone (10) senses a sound pressure.  An amplifier circuit (20) amplifies the sensed signal.  A filter circuit (30) filters the signal amplified through the amplifier circuit (20).  Further, an A/D converter (40) converts an analog signal obtained after passing through the filter circuit (30) to a digital signal, by use of which a signal discriminator (100) determines the presence or absence of an open/close operation of an open/close member.  The determined result is outputted through an output controller (50) to the outside.  This makes it possible to accurately detect the open/close of a door or the like without installing a special sensor at the door or the like.

Description

Opening / closing member recognition device for opening / closing member, and opening / closing operation recognition method for opening / closing member

The present invention relates to an opening / closing operation recognition device and recognition method for an opening / closing member that is used for security measures and that appropriately detects opening / closing of a door.

Conventionally, a door open / close recognition device that detects the opening / closing of doors and windows has been developed for security purposes in living spaces such as houses and buildings. A typical door open / close recognition device is a device in which a magnetic sensor or the like is installed at an open / close portion of a door or the like, and the magnetic sensor reacts in response to the opening / closing of the door to generate an alarm.

In this door open / close detection device, magnetic sensors must be installed on all doors to be detected. Accordingly, as disclosed in Japanese Patent Application Laid-Open No. 2005-44028, a speaker that emits a sound wave is arranged outdoors, and a detection device that detects the sound of the speaker using a microphone is installed indoors. A door open / close recognition device has been proposed.

According to this method, when the door or window is properly closed, the sound of the outdoor speaker is blocked by the door and the detection device cannot detect. On the other hand, when the door or the like is opened, the detection device detects the sound of the speaker. Therefore, it is not necessary to install sensors on windows and doors, and the convenience of construction can be improved.

However, the technique disclosed in Patent Document 1 has a problem in that it is necessary to install a speaker outdoors, and it must be constructed so as to withstand an environment exposed to wind and rain. In addition, there is a problem in that power consumption increases because the speaker must constantly emit sound.

Also, if an intruder notices the presence of an outdoor speaker and shuts off the electrical system of the speaker in advance, there is a problem that the opening / closing of the door cannot be detected.

The present invention has been made to solve such a problem, and has an object to provide a door opening / closing recognition device that eliminates complicated wiring and can detect only opening / closing of a door with high accuracy. To do.

The present invention solves the above problems by the following means.

The present invention for solving the above-mentioned problems is a sensor microphone for detecting sound pressure, an amplifier circuit for amplifying a signal detected by the sensor microphone, a filter circuit for selecting a signal amplified through the amplifier circuit, Based on an A / D converter that digitally converts an analog signal that has passed through a filter circuit, a signal discriminator that discriminates the opening / closing operation of an opening / closing member from the digital signal of the A / D converter, and an output of the signal discriminator And an opening / closing operation recognition device for an opening / closing member.

In the above invention, the signal discriminator of the open / close operation recognizing device that solves the above problem is converted into a frequency conversion unit that converts the digital signal of the A / D converter into a frequency component to obtain a frequency resolving variable. A feature calculation unit that calculates a feature amount using a frequency decomposition variable; and an event determination unit that recognizes and determines a sound event from the value or a combination of values based on an output of the feature calculation unit. It is characterized by that.

In the above-described invention, the open / close motion recognition device that solves the above-described problem is the conditional branching program in which the discrimination knowledge of the feature calculation unit has a tree structure, and the feature amount according to the tree structure using the frequency decomposition variable Is calculated.

The discriminating knowledge of the feature calculation unit of the open / close motion recognition device that solves the above-described problems is derived from an inductive learning program in the above invention.

The discrimination knowledge of the feature calculation unit of the open / close motion recognition device that solves the above-mentioned problems is derived from genetic programming in the above invention.

The discrimination knowledge of the feature calculation unit of the open / close motion recognition device that solves the above-mentioned problems is characterized in that, in the above-mentioned invention, includes the moving average of the frequency resolution variable.

The open / close motion recognition apparatus that solves the above-described problems uses the moving average of the frequency-resolved variables as background knowledge for generating discrimination knowledge of the feature calculation unit in the above invention.

In the above invention, the open / close motion recognition device that solves the above problem uses a moving average that uses the frequency resolution variable in a narrow band in the low frequency region as the background knowledge, and a frequency region that is higher than the low frequency region. Is characterized in that a moving average using the frequency resolution variable in a wide band is used.

The present invention for solving the above-described problems includes a sensing step of sensing sound pressure with a sensor microphone, an amplification step of amplifying a signal detected with the sensor microphone with an amplifier circuit, and a signal amplified through the amplifier circuit. A selecting step for selecting by a circuit, a converting step for digitally converting an analog signal having passed through the filter circuit by an A / D converter, and whether or not the opening / closing member is opened / closed using the digital signal of the A / D converter. An opening / closing operation recognition method for an opening / closing member, comprising: a determining step for determining; and an output step for controlling an external output based on a result of the determination.

In the above invention, the switching operation recognition method that solves the above problem is converted in the determination step by a frequency conversion unit step that converts the digital signal of the A / D converter into a frequency component to obtain a frequency decomposition variable. A feature calculation step for calculating a feature amount using the frequency decomposition variable, and an event determination step for recognizing and determining a sound event from those values or combinations of values based on the output of the feature calculation step. It is characterized by providing.

The discriminating knowledge of the feature calculation step of the open / close operation recognition method that solves the above problem is the conditional branching program having a tree structure in the above invention, and using the frequency decomposition variable, the feature quantity according to the tree structure Is calculated.

The discrimination knowledge of the feature calculation step of the open / close operation recognition method that solves the above-mentioned problems is derived from an inductive learning program in the above invention.

The discrimination knowledge of the feature calculation step of the open / close operation recognition method that solves the above problem is derived from genetic programming in the above invention.

The discrimination knowledge of the feature calculation step of the open / close operation recognition method that solves the above-mentioned problem is characterized in that, in the above-mentioned invention, includes the moving average of the frequency resolution variable.

The method for recognizing an opening / closing operation that solves the above-described problem is characterized in that, in the above-described invention, the moving average of the frequency-resolved variable is used as background knowledge for generating discrimination knowledge of the feature calculation step.

In the above invention, the open / close operation recognition method for solving the above-mentioned problem uses a moving average using the frequency resolution variable of a narrow band in the low frequency region as the background knowledge, and a frequency region higher than the low frequency region. Is characterized in that a moving average using the frequency resolution variable in a wide band is used.

According to the present invention, it is possible to detect only the opening / closing of the door appropriately by installing the door opening / closing recognition device indoors without installing any related equipment at the door or outdoors.

It is a system configuration figure showing the outline composition of the recognition device concerning the embodiment of the present invention. Similarly, it is a block diagram which shows schematic structure of the signal discriminator in the said recognition apparatus. Similarly, it is a block diagram which shows the function structure of the signal discriminator in the said recognition apparatus. Similarly, it is a flowchart which shows the flow of the door opening / closing detection of the said recognition apparatus.

Hereinafter, an opening / closing operation recognition device (hereinafter, recognition device) for an opening / closing member according to an embodiment of the present invention will be described in detail with reference to the drawings.

As shown in FIG. 1, the recognition device 1 selects a sensor microphone 10 that senses sound pressure, an amplifier circuit 20 that amplifies a signal detected by the sensor microphone 10, and a signal that is amplified through the amplifier circuit 20. Based on the output of the signal discriminator 100, the A / D converter 40 that digitally converts the analog signal that has passed through the filter circuit 30, the signal discriminator 100 that discriminates whether or not the door is open / closed from the digital signal, And an output control unit 50 for controlling the external output. The recognition device 1 is installed at a specific indoor location.

The signal discriminator 100 is a so-called MPU, and as shown in FIG. 2, a central processing unit (CPU) 150 that executes information processing by executing a program, and a RAM 152 that can temporarily hold information. A ROM 154 that stores various programs, an information storage device 156 that temporarily stores digitized sound pressure signals, and an interface 158 that inputs and outputs digital signals. In the signal discriminator 100, the CPU 150 executes a program stored in the ROM 154, thereby exhibiting a function to be described later.

As shown in FIG. 3, the functional configuration of the signal discriminator 100 includes a frequency converter 110 that converts a digital signal from a time component to a frequency component, and a feature calculator that calculates a plurality of feature amounts from the converted frequency component. 120 and an event determination unit 130 for recognizing and determining a sound event from these values or combinations of values based on the output of the feature calculation unit 120.

The frequency converter 110 converts the signal digitized through the A / D converter 40 into a frequency component. This frequency-resolved signal is input to the discrimination knowledge in the feature calculation unit 120 and calculates a door opening / closing possibility value. The conversion method to the frequency component can be selected arbitrarily, but representative methods are discrete Fourier transform (DFT), discrete cosine transform (DCT), Hadamard transform, Weblet transform, and the like.

The event determination unit 130 determines whether the possibility value of the feature calculation unit 120 exceeds a predetermined threshold value, and identifies the event. Based on the event identification result of the event determination unit 130, the output control unit 50 outputs an alarm signal.

The discrimination knowledge in the feature calculation unit 120 is a conditional branch program having a tree structure. Specifically, the tree structure is conditional branching using if then else, continuous processing of progN that executes each tree, calculation of add (addition), sub (subtraction), mul (multiplication), div (division), and constants It is configured with. The feature calculation unit 120 uses the signal sequence data (frequency decomposition variable) decomposed into frequency components, performs an operation along the tree structure as discriminating knowledge, and outputs the result as numerical data. The output result is an arbitrary value from 0 to 15, and a larger value means that the probability of door opening / closing is higher.

Note that inductive learning techniques and rules of thumb are used in combination to derive discriminant knowledge. As an inductive learning method, in this embodiment, a genetic programming method is used, and thereby discriminating knowledge is synthesized. The rule of thumb is used as background knowledge for genetic programming. A function that creates a feature value that is found when the door is opened and closed is defined, and this is used as a part of discrimination knowledge. Using heuristics as background knowledge makes it possible to improve the convergence of genetic programming.

By the way, the sound pressure signal of the door is often non-linear and cannot be separated with a simple threshold. Therefore, the signal to be recognized is converted into several variables (feature vectors) called feature quantities, and a certain threshold is applied to the feature space created by the feature vector to perform space separation.

In this embodiment, information of frequency components decomposed for each level is used as the feature amount. Note that this frequency component information also has a time-series component that is an information arrangement order. As an experience based on their own research, the present inventors have grasped in advance a frequency component of a level that is strongly involved in the opening and closing of the door in the space and a frequency component of a level that has a relatively small influence. Therefore, such experience knowledge is incorporated into the elements of the genetic program as background knowledge to improve the convergence of learning.

The background knowledge is a mathematical expression of knowledge based on the experience side (human sensuously). Therefore, in the present embodiment, the mathematical formula as the background knowledge is incorporated in the discrimination knowledge and used for calculating the feature amount. Note that discrimination using discrimination knowledge is performed by mapping a feature vector using a frequency decomposition variable onto a feature space and then performing space separation. However, if there is no background knowledge, the dimension of the feature vector will increase, and there will be a large number of straight lines to perform spatial separation (drawing). A quantity (trial and error amount, calculation amount) is required.

Therefore, in this embodiment, a basic formula for performing space separation is provided in advance by introducing a mathematical formula as background knowledge. As a result, it is possible to easily generate discrimination knowledge even when the learning amount is small.

Specifically, in the present embodiment, frequencies of 50 Hz or higher are excluded in the filter circuit 30 in advance, and 16 frequency resolution variables (freq (0 ) To freq (15)) are extracted. The present inventors have empirically found that the opening and closing of doors involves relatively low frequency components (for example, freq (0) to freq (3)). The main frequency of opening and closing the door is less than 50 Hz, which excludes frequencies above 50 Hz, and over 50 Hz, such as telephones, alarm clocks, TV / radio, falling objects, horns, sirens, etc. This is because stationary life sounds are included. Furthermore, when the present inventors define a function in which a certain frequency changes with time, it is also empirical that a characteristic of a door opening / closing event is present in a low frequency component, particularly in a band that cannot be heard by human hearing. Have found. Based on the above awareness, the following “Formula 1” is defined and given as background knowledge including the value of this parameter.

Note that ave is a moving average obtained by linearly combining frequency components, and is an element of a feature vector. w, k, m, and n are parameter constants. Freq (i) means the i-th (band level) component when the amplitude signal is frequency-converted. That is, the amplitude signal ad (= sound pressure signal ad) of the A / D converter 40 for a predetermined period is input, and discrete Fourier transform (DFT), discrete cosine transform (DCT), Hadamard transform, Weblet transform, etc. are performed. The frequency is converted, and the i-th (band level) component becomes the frequency resolution variable freq (i). The frequency decomposition variable freq (i) excludes the time element in the amplitude signal ad. However, the frequency resolution variable freq (i) itself has fluctuations (fluctuations) in the order of information arrangement within the range of the band level i. That is, when the information arrangement order element q included in freq (i) is taken into consideration, the frequency resolution variable is freq (i, q), which changes depending on the order element q. In the present embodiment, the smaller i, the lower the frequency level, and the larger i, the higher the frequency side.

When the frequency resolution variable freq (i, q) is a feature vector element, the dimension becomes i * q, the number of dimensions increases, and the derivation of discrimination knowledge becomes complicated. Therefore, in the present embodiment, the frequency resolution variable freq (i, q) is moving averaged based on the order element q, and the one-dimensionalized ave (z) is used as a feature vector. As a result, it is possible to realize a mapping that reduces the amount of combination of linear judgment expressions while keeping the number of dimensions of the feature space low. In Equation 1, one new equation is determined to determine one parameter set z (w, m, n, k). The number of dimensions is also determined by how many parameter sets z are prepared.

The reason why the moving average is used in Equation 1 is based on the experience that there is little change depending on the temporal characteristics associated with the opening and closing of the door, specifically, the information arrangement order element q. As already mentioned, if the information array sequence element q (temporal feature) is used as it is as a feature vector without introducing moving average, complex discriminating knowledge is obtained for many spectra arranged in time series. End up. This makes it difficult to mount on a microcomputer.

Especially, since the door is opened and closed by humans, it cannot be considered that the sound pressure generated by the door undergoes a violent time change. In particular, it is considered that the intruder slowly opens the door, and the temporal change of the sound pressure is less meaningful as a feature vector. Therefore, by adopting a moving average, the number of dimensions is drastically reduced and the amount of calculation is reduced.

Specifically, in the present embodiment, as background knowledge using Equation 1 above, five types of parameter sets from z = 0 to 4 are set, and the following Equation 2 group is defined and introduced.

When there are 16 frequency resolution variables (freq (0) to freq (15)) frequency-resolved to 16 levels, the present inventors have found that opening and closing of doors has a relatively low frequency component (for example, freq ( 0) ～ freq (3)) have been found empirically. Therefore, for low frequency components, a moving vector is calculated using freq (i) alone or using a total sum component of 2 to 3, and a feature vector is set (see A to C in Equation 2). On the other hand, for high frequency components, a moving average is calculated using a plurality of sum components of freq (i), desirably 3 or more, more preferably 4 or more (see D and E in Equation 2). ). As a result, while reducing the dimension, it is possible to extract and distinguish the characteristics of the event in detail for the low frequency component portion.

教師 Use supervised learning data to generate discriminative knowledge recursively based on background knowledge. In order to prepare this learning data in advance, a recording device called a data logger is installed in a condominium or a detached room, and sound information about many door opening / closing patterns is recorded. On the other hand, it is also necessary to record data to be rejected. For example, sound information such as a windy day and a thunderstorm is collected. In particular, wind sounds, construction sounds, car passing sounds, air conditioner operations, wind noises, and the like have relatively low frequencies, so it is difficult to determine whether the door is open or closed with a simple sound pressure. Therefore, in this embodiment, learning is performed using genetic programming based on various door opening / closing pattern sound information and the rejection sound information, and the discrimination knowledge is synthesized. The discriminating knowledge to be synthesized is incorporated in the feature calculation unit 120.

Next, the door opening / closing determination procedure in the recognition device 1 will be described.

In step 200, sound pressure information is first sensed by the sensor microphone 10, and in step 202, this information is amplified by the amplifier circuit 20. Further, in step 204, the amplified information is input to the filter circuit 30 in order to obtain only the frequency band to be discriminated (in this embodiment, less than 50 Hz). Note that the role of the filter circuit 30 also has the meaning of suppressing aliasing noise when digitally converted thereafter.

The signal that has passed through the filter circuit 30 is input to the A / D converter 40 and digitally converted in step 206, and then input to the information storage device 156 of the signal discriminator 100 in step 208. In this embodiment, the information storage device 156 is a FIFO (First in first で は out) type memory. The FIFO type memory means a memory structure in which a first input signal is first extracted. In this information storage device 156, digital signals for a certain period of time are accumulated.

The time-series discrete signals stored in the FIFO memory are extracted in units of 2 n units in step 210, input to the frequency conversion unit 110, and converted into frequency components. In step 212, the frequency-converted signal is stored again in the FIFO memory of the information storage device 156 for a predetermined time together with time-series data.

Thereafter, in step 214, a frequency-converted signal sequence for a predetermined time is taken out from the information storage device 156 and input to the feature calculation unit 120. Specifically, the feature calculation unit 120 inputs this signal sequence to the discrimination knowledge, and numerically converts the possibility of opening and closing the door. In step 216, based on the output value of the feature calculation unit 120, the event determination unit 130 determines whether the event is opening / closing of the door based on a threshold value. In step 218, a predetermined alarm is issued based on the result of the event determination unit 120.

As described above, according to the recognition apparatus 1 of the present embodiment, the signal discriminator 100 can detect the sound pressure of only the door with high accuracy. Therefore, it is possible to detect the opening / closing of the door only by placing the recognition device 1 in the space without installing a sensor or the like outside the door or the door.

Further, the present inventors have clarified that the opening / closing sound of the door includes an extremely low frequency that cannot be heard by human ears. Therefore, in the signal discriminator 100, the frequency conversion unit 110 performs frequency decomposition, and the feature calculation unit 120 calculates the feature amount using the frequency decomposition variable. In this way, it is possible to properly detect only the opening and closing of the door while actively removing the frequency decomposition variable of the low frequency level related to the opening and closing of the door and removing the noise component unrelated to the opening and closing of the door. I have to. In particular, since the discrimination knowledge of the feature calculation unit 120 is a conditional branch program having a tree structure, it is possible to calculate feature quantities easily and at high speed.

Furthermore, in this embodiment, the discrimination knowledge of the feature calculation unit 120 is derived by an operation using an inductive learning program, particularly genetic programming. Accordingly, by learning in advance a large number of pattern sounds for actual door opening and closing, it becomes possible to acquire highly accurate discrimination knowledge, and it is possible to increase the detection accuracy of door opening and closing. In addition, by learning in advance a large number of pattern sounds to be rejected such as wind and cars, it is possible to appropriately exclude noise with high occurrence frequency by the conditional branching program having a tree structure.

Furthermore, in this embodiment, the discrimination knowledge of the feature calculation unit 120 includes a feature vector based on a moving average using a frequency decomposition variable. As a result, it is possible to obtain an effect of filtering the fluctuation of the frequency, which is included in the frequency decomposition variable, which is not related to the opening / closing of the door, and it is possible to greatly reduce the number of dimensions of the feature vector.

It should be noted that this feature vector is appropriately incorporated into the discrimination knowledge by defining it in the feature calculation unit as background knowledge in advance. That is, the discrimination knowledge of the feature calculation unit 120 is generated using the moving average of the frequency resolution variable.

Note that the mathematical formula introduced as background knowledge uses a moving average using a frequency resolution variable of a narrow band (for example, freq (0), freq (1), freq (2), etc.) in a low frequency region. However, in a frequency range higher than the low frequency range, a moving average using a wide-band frequency resolution variable (for example, the sum of freq (4 to 7), the sum of freq (8 to 15), etc.) is used. . As a result, it is possible to perform a fine analysis in a low frequency band in which the door opening / closing characteristics tend to appear.

Note that the recognition apparatus 1 according to the present embodiment exemplifies discrete Fourier transform (DFT), discrete cosine transform (DCT), Hadamard transform, Weblet transform, and the like as a frequency transforming method. However, the present invention is not limited to this, and other methods can be used.

Further, the feature calculation unit 120 in the recognition apparatus 1 shows a case where a moving average of frequency decomposition variables is used in background knowledge and discriminating knowledge, but the present invention is not limited to this, and smoothing by other methods is possible. Is also possible.

The present invention can be used for opening / closing detection of opening / closing members such as doors, doors, and lids.

Claims (16)

A sensor microphone that senses sound pressure;
An amplifier circuit for amplifying a signal detected by the sensor microphone;
A filter circuit for selecting a signal amplified through the amplifier circuit;
An A / D converter for digitally converting the analog signal passed through the filter circuit;
A signal discriminator for discriminating the presence / absence of an opening / closing operation of the opening / closing member from the digital signal of the A / D converter;
An output control unit for controlling an external output based on the output of the signal discriminator;
An opening / closing operation recognition device for an opening / closing member. The signal discriminator is
A frequency converter that converts the digital signal of the A / D converter into a frequency component to obtain a frequency decomposition variable;
A feature calculation unit for calculating a feature amount using the converted frequency decomposition variable;
An event determination unit for recognizing and determining a sound event from those values or combinations of values based on the output of the feature calculation unit;
The opening / closing operation recognition device for an opening / closing member according to claim 1, comprising: The discrimination knowledge of the feature calculation unit is a conditional branch program having a tree structure, and the feature quantity is calculated according to the tree structure using the frequency decomposition variable. Opening / closing operation recognition device for opening / closing member. 4. The opening / closing operation recognition device for an opening / closing member according to claim 2, wherein the discrimination knowledge of the feature calculation unit is derived by an inductive learning program. The open / close motion recognition device for an open / close member according to claim 2, 3 or 4, wherein the discrimination knowledge of the feature calculation unit is derived by genetic programming. 6. The opening / closing operation recognition device for an opening / closing member according to any one of claims 2 to 5, wherein the discrimination knowledge of the feature calculation unit includes a moving average of the frequency resolution variable. The opening / closing operation recognition device for an opening / closing member according to any one of claims 2 to 6, wherein a moving average of the frequency decomposition variable is used as background knowledge for generating discrimination knowledge of the feature calculation unit. As the background knowledge,
In the low frequency region, using a moving average using the frequency resolution variable in a narrow band,
In a frequency region higher than the low frequency region, a moving average using the frequency resolution variable in a wide band is used.
The opening / closing operation recognition device for an opening / closing member according to claim 7. A sensing step of sensing sound pressure with a sensor microphone;
An amplification step of amplifying a signal detected by the sensor microphone with an amplifier circuit;
A selection step of selecting a signal amplified through the amplifier circuit with a filter circuit;
A conversion step of digitally converting the analog signal that has passed through the filter circuit with an A / D converter;
A determination step of determining presence / absence of an opening / closing operation of the opening / closing member using a digital signal of the A / D converter;
An output step for controlling an external output based on the result of the determination;
A method for recognizing the opening / closing operation of the opening / closing member. In the determination step,
A frequency converter step of converting the digital signal of the A / D converter into a frequency component to obtain a frequency resolving variable;
A feature calculation step of calculating a feature amount using the transformed frequency decomposition variable;
An event determination step for recognizing and determining a sound event from the value or combination of values based on the output of the feature calculation step;
The method for recognizing an opening / closing operation of an opening / closing member according to claim 9. The discrimination knowledge of the feature calculation step is a conditional branch program having a tree structure, and the feature quantity is calculated according to the tree structure using the frequency decomposition variable. Of recognizing opening / closing operation of the opening / closing member 12. The opening / closing operation recognition device for an opening / closing member according to claim 10 or 11, wherein the discrimination knowledge of the feature calculation step is derived by an inductive learning program. The method for recognizing an opening / closing operation of an opening / closing member according to claim 10, 11 or 12, wherein the discrimination knowledge of the feature calculation step is derived by genetic programming. The method for recognizing an opening / closing operation of an opening / closing member according to any one of claims 10 to 13, wherein the discrimination knowledge of the feature calculation step includes a moving average of the frequency resolution variable. The method for recognizing an opening / closing operation of an opening / closing member according to any one of claims 10 to 14, wherein a moving average of the frequency decomposition variable is used as background knowledge for generating discrimination knowledge of the feature calculation step. As the background knowledge,
In the low frequency region, using a moving average using the frequency resolution variable in a narrow band,
The method for recognizing an opening / closing operation of an opening / closing member according to claim 15, wherein a moving average using the frequency resolution variable in a wide band is used in a frequency region higher than the low frequency region.

Images