JP4252028B2 - Traffic sound identification device, traffic sound determination program for causing computer to function as traffic sound identification device, recording medium, and traffic sound determination method - Google Patents

Description

  The present invention relates to a traffic sound identification device, a traffic sound determination program for causing a computer to function as a traffic sound identification device, a recording medium, and a traffic sound determination method, and more particularly to a traffic sound identification device and a computer for determining the type of traffic sound. The present invention relates to a traffic sound determination program, a recording medium, and a traffic sound determination method for functioning as a traffic sound identification device.

  In order to ensure traffic safety and reduce accidents, prevent accidents by using a system that detects and reports accidents at an early stage, seeks early resolution of accident post-processing, and analyzes the mechanism of accidents. It is essential.

  In order to analyze the mechanism of accident occurrence, it is necessary to detect accidents and near miss events, and accidents are detected based on traffic sounds such as collision sounds and sudden braking sounds due to accidents. Yes.

  However, many traffic sounds have characteristics similar to those of accidents when an accident occurs. The sound includes, for example, noise of a large vehicle carrier or metal sound generated during construction work. Therefore, it is very difficult to distinguish between accident sounds and sounds other than accident sounds (hereinafter also referred to as non-accident sounds).

Therefore, techniques for accurately identifying accident sounds using a neural network are disclosed in Japanese Patent Application Laid-Open No. 2000-275096 (Patent Document 1) and Japanese Patent Application Laid-Open No. 2001-33304 (Patent Document 2).
JP 2000-275096 A JP 2001-33304 A

  However, in the technique disclosed in Japanese Patent Application Laid-Open No. 2000-275096 (Patent Document 1) and Japanese Patent Application Laid-Open No. 2001-33304 (Patent Document 2), an output value of a neural network for identifying a traffic sound is “ The traffic sound cannot be accurately identified unless it is a combination of integers such as “0” and “1”.

  In general, the output value from the output layer of a neural network to which traffic sound data that is very difficult to identify is input is often not an integer but a decimal, even if the neural network is sufficiently learned. . Therefore, in the technique disclosed in Japanese Patent Laid-Open No. 2000-275096 (Patent Document 1) and Japanese Patent Laid-Open No. 2001-33304 (Patent Document 2), if the output value from the output layer of the neural network is a decimal number, The possibility of accurately identifying traffic sounds is low.

  Further, since the technique disclosed in Japanese Patent Application Laid-Open No. 2000-275096 (Patent Document 1) and Japanese Patent Application Laid-Open No. 2001-33304 (Patent Document 2) uses only one neural network, If the output value from the output layer is an intermediate decimal number between “0” and “1” (for example, 0.4), no result is obtained, and traffic sound identification becomes more difficult.

  The present invention has been made to solve the above-described problems, and one of the objects of the present invention is to provide a traffic sound identification device capable of identifying traffic sounds with very high accuracy. .

  Another object of the present invention is to provide a traffic sound determination program for causing a computer to function as a traffic sound identification device capable of identifying traffic sounds with very high accuracy.

  Still another object of the present invention is to provide a recording medium in which a traffic sound determination program for causing a computer to function is recorded as a traffic sound identification device capable of identifying traffic sounds with very high accuracy. .

  Still another object of the present invention is to provide a traffic sound determination method capable of identifying traffic sounds with very high accuracy.

In order to solve the above-described problem, according to one aspect of the present invention, there is provided a traffic sound identification device that determines which of a plurality of predetermined types of traffic sounds is collected. The identification device calculates a power spectrum of traffic sound, divides the power spectrum calculated by the arithmetic circuit into a plurality of subbands, and based on the power spectrum divided into the plurality of subbands, I (natural number) ) Pre-ordered from the first to the Lth (natural number) for determining the type of traffic sound based on the subband circuit for generating a plurality of input data and the power spectrum divided into a plurality of subbands L neural networks and a learning circuit for learning the L neural networks are provided, and the learning circuit has N (natural number) traffic for learning. A process of selecting N traffic sound data at random without interfering with the data from the data is performed L times, so that L sets of learning data are sequentially generated, and a learning process of sequentially learning with L neural networks is performed. Each of the learned L neural networks outputs a judgment result of the type of traffic sound based on the I input data , and the traffic sound identification device is output from each of the learned L neural networks. And a determination circuit for determining the type of traffic sound based on the majority decision of the L determination results.

  According to another aspect of the present invention, there is provided a traffic sound identification device for determining which collected traffic sound is any of a plurality of predetermined traffic sounds, wherein the traffic sound identification device is a power of traffic sound. In order to determine the type of traffic sound based on an arithmetic circuit that calculates a spectrum, a subband circuit that divides the power spectrum calculated by the arithmetic circuit into a plurality of subbands, and a power spectrum that is divided into a plurality of subbands 1 to L (natural number) in advance, and N learning networks for learning the L neural networks, and the learning circuit has N (natural number) for learning. By repeating the data selection process that selects N traffic sound data randomly without disturbing the overlap from each traffic sound data, the L sets of learning data are ordered. A learning process of generating and sequentially learning with the L neural networks is performed, and the learning circuit learns the (k + 1) th (k + 1) (k: natural number smaller than L) of the L neural networks in the data selection process. When the process is performed, traffic sound data that is difficult to learn in the learning process of the kth neural network among the L pieces is preferentially selected, and each of the learned L number of neural networks includes a plurality of pieces. Based on the power spectrum divided into sub-bands, the traffic sound type determination result is output, and the traffic sound identification device weights the L determination results output from the learned L neural networks, respectively. And a determination circuit for determining the type of traffic sound.

  Preferably, when the learning circuit causes L neural networks to learn, the type of traffic sound used for learning determination is M (natural number), and the M traffic sounds are Q (natural numbers smaller than M). When the determination circuit determines the traffic sound, it is determined to which of the Q groups the determined traffic sound belongs.

According to another aspect of the present invention, a traffic sound determination program executed by a computer including a microphone for collecting traffic sound, the traffic sound determination program calculating a power spectrum of the traffic sound; Dividing the power spectrum into a plurality of subbands , generating I (natural number) input data based on the power spectrum divided into the plurality of subbands , and learning N (natural number) Data selection processing for selecting N traffic sound data at random without interfering with traffic sound data is performed L times, so that L sets of learning data are generated sequentially, and the learning data generated sequentially are learned sequentially Each of the learned L neural networks is sequentially configured, and each of the learned L neural networks is converted into I input data . A step of outputting a judgment result of the type of traffic sound, and a step of judging the type of traffic sound based on the majority decision of the L judgment results outputted from the L neural networks learned. Is executed on the computer.

  According to another aspect of the present invention, a traffic sound determination program executed by a computer including a microphone for collecting traffic sound, the traffic sound determination program calculating a power spectrum of the traffic sound; A step of dividing the power spectrum into a plurality of subbands, and a data selection process for selecting N traffic sound data randomly from N (natural number) traffic sound data for learning without preventing overlap. By sequentially generating L sets of learning data, and sequentially configuring L neural networks in which the sequentially generated learning data is sequentially learned, and in each of the learned L neural networks, Based on the power spectrum divided into multiple subbands, the step of outputting the judgment result of the type of traffic sound and the learning The computer executes a step of determining the type of traffic sound based on the weighting calculation of the L determination results respectively output from the L neural networks. When learning processing is performed on the (k + 1) th (k: natural number smaller than L) of the neural networks, priority is given to traffic sound data that is difficult to learn with the kth neural network of L. The computer further performs the step of selecting automatically.

  According to still another aspect of the present invention, the recording medium is a medium on which a traffic sound determination program is recorded.

According to another aspect of the present invention, there is provided a traffic sound determination method for determining which collected traffic sound is any of a plurality of predetermined types of traffic sounds, the traffic sound determination method comprising: A step of calculating a spectrum, a step of dividing the calculated power spectrum into a plurality of subbands , generating I (natural number) input data based on the power spectrum divided into the plurality of subbands , and learning By performing a data selection process for selecting N traffic sound data randomly from N (natural number) traffic sound data for L times without hindering duplication, L sets of learning data are sequentially generated and sequentially Based on the I input data , each of the L neural networks trained sequentially comprises L neural networks that sequentially train the generated learning data, and A step of outputting a determination result of the type of traffic sound, and a step of determining the type of traffic sound based on the majority of the L determination results output from the learned L neural networks.

  According to another aspect of the present invention, there is provided a traffic sound determination method for determining which collected traffic sound is any of a plurality of predetermined types of traffic sounds, the traffic sound determination method comprising: A step of calculating a spectrum, a step of dividing the calculated power spectrum into a plurality of subbands, and N traffic sound data randomly from N (natural number) traffic sound data for learning without hindering duplication. By performing the data selection process to be selected L times, L sets of learning data are sequentially generated, L neural networks sequentially learning the sequentially generated learning data, and learning L A step of causing each of the neural networks to output a judgment result of the type of traffic sound based on the power spectrum divided into a plurality of subbands; And a step of determining the type of traffic sound based on the weighted calculation of L determination results respectively output from the network, and in the data selection process, the Lth (k + 1) (k: L) When the learning process of the (smaller natural number) th neural network is performed, traffic sound data that is difficult to learn with the kth neural network out of L is preferentially selected.

  The traffic sound identification device according to the present invention includes L neural networks for determining the type of traffic sound and a learning circuit for learning the L neural networks, and the learned L neural networks. The type of traffic sound is determined on the basis of the majority decision of the L determination results respectively output from.

  Therefore, since the type of traffic sound is determined based on the majority of the plurality of determination results output from the plurality of neural networks, a traffic sound identification device capable of identifying traffic sound with very high accuracy is provided. It becomes possible to provide.

  The traffic sound identification device according to the present invention includes L neural networks for determining the type of traffic sound, and a learning circuit for learning the L neural networks, and has learned L pieces of learning networks. The type of traffic sound is determined based on the weighting calculation of the L determination results output from the neural network.

  Accordingly, since the type of traffic sound is determined based on the weighting calculation of the plurality of determination results respectively output from the plurality of neural networks, the traffic sound identification device capable of identifying the traffic sound with very high accuracy Can be provided.

  The traffic sound determination program according to the present invention sequentially generates L sets of learning data, sequentially configures L neural networks in which the sequentially generated learning data is sequentially learned, and learns the L neural networks. The type of traffic sound is determined based on the majority decision of the L determination results output from each.

  Therefore, since the type of traffic sound is determined based on the majority of the plurality of determination results output from the plurality of learned neural networks, the traffic sound can be identified with very high accuracy. As the identification device, it is possible to provide a traffic sound determination program for causing a computer to function.

  The traffic sound determination program according to the present invention sequentially generates L sets of learning data, sequentially configures L neural networks in which the sequentially generated learning data is sequentially learned, and learns the L neural networks. The type of traffic sound is determined on the basis of the weighting calculation of the L determination results respectively output from.

  Therefore, since the type of traffic sound is determined based on the weighted calculation of the plurality of determination results respectively output from the plurality of learned neural networks, the traffic sound can be identified with very high accuracy. It is possible to provide a traffic sound determination program for causing a computer to function as the sound identification device.

  The recording medium according to the present invention records a traffic sound determination program. The traffic sound determination program sequentially generates L sets of learning data, sequentially configures L neural networks that have sequentially learned the generated learning data, and outputs each of the learned L neural networks. The type of traffic sound is determined based on the majority decision of the L determination results.

  Therefore, since the type of traffic sound is determined based on the majority of the plurality of determination results output from the plurality of learned neural networks, the traffic sound can be identified with very high accuracy. As an identification device, it is possible to provide a recording medium for causing a computer to function.

  The recording medium according to the present invention records a traffic sound determination program. The traffic sound determination program sequentially generates L sets of learning data, sequentially configures L neural networks that have sequentially learned the generated learning data, and outputs each of the learned L neural networks. The type of traffic sound is determined based on the weighted calculation of the L determination results.

  Therefore, since the type of traffic sound is determined based on the weighted calculation of the plurality of determination results respectively output from the plurality of learned neural networks, the traffic sound can be identified with very high accuracy. As a sound identification device, it is possible to provide a recording medium for causing a computer to function.

  The traffic sound determination method according to the present invention sequentially generates L sets of learning data, sequentially configures L neural networks that have sequentially learned the learning data, and learns the L neural networks that have been learned. The type of traffic sound is determined based on the majority decision of the L determination results output from each.

  Therefore, since the type of traffic sound is determined based on the majority of the plurality of determination results output from the plurality of learned neural networks, the traffic sound can be identified with very high accuracy. A determination method can be provided.

  In addition, the traffic sound determination method according to the present invention sequentially generates L sets of learning data, sequentially configures L neural networks that sequentially learn the generated learning data, and learns the L pieces of learned data. The type of traffic sound is determined based on the weighting calculation of the L determination results output from the neural network.

  Therefore, since the type of traffic sound is determined based on the weighted calculation of the plurality of determination results respectively output from the plurality of learned neural networks, the traffic sound can be identified with very high accuracy. It is possible to provide a sound determination method.

  Hereinafter, embodiments of the present invention will be described with reference to the drawings. In the following description, the same parts are denoted by the same reference numerals. Their names and functions are also the same. Therefore, detailed description thereof will not be repeated.

<First Embodiment>
FIG. 1 is a block diagram showing a configuration of a traffic sound identification apparatus 1000 according to the present embodiment.

  Referring to FIG. 1, traffic sound identification apparatus 1000 includes a microphone 100, a sound pressure calculation circuit 110, a power spectrum calculation circuit 120, a subband calculation circuit 130, and a neural network unit 200.

  The microphone 100 has a function of collecting traffic sounds near a road (for example, near an intersection). The sound pressure calculation circuit 110 measures the sound pressure value of the traffic sound collected by the microphone 100 and performs a predetermined calculation process. The power spectrum calculation circuit 120 calculates a power spectrum based on the sound pressure value calculated by the sound pressure calculation circuit 110. The subband calculation circuit 130 divides the power spectrum calculated by the power spectrum calculation circuit 120 into a plurality of frequency bands (subbands). The neural network unit 200 determines the type of traffic sound based on the power spectrum data divided into a plurality of subbands by the subband arithmetic circuit 130, and outputs the determination result.

  FIG. 2 is a diagram showing an internal configuration of the neural network unit 200 in the present embodiment.

  Referring to FIG. 2, the neural network unit 200 includes neural networks 200.1, 200.2,..., 200. L is included. In the following, the neural network is also simply referred to as “NN”.

  FIG. 3 is a diagram showing a detailed internal configuration of the neural network 200.1 in the present embodiment.

  Referring to FIG. 3, neural network 200.1 includes an input layer 200.1A, a hidden layer 200.1B, and an output layer 200.1C.

  The input layer 200.1A has a plurality of input units. The “unit” is a model of a “neuron” that is a neuron element of the human brain. A plurality of data are respectively input to the plurality of input units. In the present embodiment, the number of input units is, for example, 32. The number of input units is not limited to 32 and may be an arbitrary number.

  The hidden layer 200.1B has a plurality of hidden units.

  The output layer 200.1C has a plurality of output units.

  The plurality of input units and the plurality of hidden units are coupled to each other. The plurality of hidden units and the plurality of output units are coupled to each other. A plurality of result data OT1, OT2,..., OTc by the neural network 200.1 are output from the plurality of output units, respectively. The plurality of result data OT1, OT2,..., OTc are pre-ordered from the first to the c-th. The plurality of result data OT1, OT2,... OTc are also collectively referred to as determination results.

  The result data OT1, OT2,..., OTc are associated with a plurality of types of traffic sounds. The result data OT1 is associated with, for example, “collision sound”.

  2, 200... 200. Each of L has the same configuration and function as the neural network 200.1 of FIG.

  Again referring to FIG. 1, traffic sound identification apparatus 1000 further includes a learning circuit 210 and a determination circuit 250.

  Based on a plurality of determination results output from the neural network unit 200, the learning circuit 210 performs neural networks 200.1, 200.2,. , TDL is input with teacher data TD1, TD2,..., TDL, respectively, so that the neural networks 200.1, 200.2,. Let L learn. In addition, the process which makes a some neural network learn is also called group learning process. The group learning process is performed by the learning circuit 210.

  The determination circuit 250 determines what kind of traffic sound is based on the plurality of determination results output from the neural network unit 200, and outputs result data RT for specifying the determined traffic sound. .

  Next, a method of learning a plurality of neural networks and determining input data using the learned neural networks will be described. Typical examples of the method include bagging and Adaboost.

  In the present embodiment, a bagging group learning process (hereinafter also referred to as a bagging group learning process) will be described.

  In the bagging group learning process, for example, a learning set S composed of N pieces of traffic sound data is used. Hereinafter, the learning set S is also referred to as learning data. The learning set S is represented by the following equation (1).

In the equation (1), x ₁ , x ₂ ,..., X _N represent input vectors. y _i represents the category label of the input vector x _i . In this embodiment, the input vector x _i indicates the 32 data. Note that the number of data indicated by the input vector x _i is not limited to 32, and may be arbitrary. In the present embodiment, the category label y _i is an integer. For example, if the category label y _i is “2”, the neural network is trained so that the second result data OT2 is “1”. C represents the number of category labels or groups to be described later.

  FIG. 4 is a flowchart showing the flow of the bagging group learning process.

  Referring to FIG. 4, in step S100, an initialization process is performed. In the initialization process, the number of neural networks to be learned and other parameters are set. The initial value of k is set to “1”. In the present embodiment, the number of neural networks is assumed to be L. Then, it progresses to step S110.

In step S110, a k-th learning set _Sk generation process (hereinafter also referred to as a learning set generation process) is performed.

  FIG. 5 is a diagram for explaining learning set generation processing.

  FIG. 5A shows the above-described equation (1).

FIGS. 5B, 5C, and 5D illustrate learning sets S ₁ , S ₂ , and S _L generated by the learning set generation process. The learning sets S ₁ , S ₂ , and S _L are randomly selected from the learning set S (learning data) consisting of N traffic sound data with a certain probability (for example, 1 / N) without obstructing duplication. Generated by selecting traffic sound data.

In other words, there may be a plurality of the same traffic sound data in the generated learning set. Moreover, all the traffic sound data included in the generated learning set may be the same. In addition, at least two of the generated learning sets S ₁ , S ₂ , and S _L may be the same.

Referring again to FIG. 4, in step S110, the generation processing of the k-th training set _{S k} is terminated, the process proceeds to step S120.

In step S120, the learning sets _{S k} a k-th neural networks (e.g., the neural network 200.1) NN learning process of learning is carried out.

Since k = 1 at this time, in the learning process, the values of the _first input vector x ₁ (see FIG. 5B) of the learning set S ₁ are stored in the plurality of input units of the neural network 200.1 in FIG. Entered.

The neural network 200.1 outputs result data OT1, OT2,... OTc to be output with respect to the input values, the first category label y ₁ (see FIG. 5B) of the learning set S ₁ , and Based on the above, a process of changing the weight of connection between units of the neural network 200.1 (hereinafter also referred to as a connection weight change process) is performed.

Specifically, the connection weights changing process, and category label corresponding to the input vector x _1, the error value of the data neural network is output by the input vector x _1, relative to the weight of the interconnection between the units Give feedback. Thereby, the weight of the coupling between the units can be changed. This rate of change is referred to as “learning rate”.

Then, also for each input vector from the second training set S ₁ to N-th, by repeating the same processing as described above, to train the neural network 200.1. Thereafter, the process proceeds to step S130.

  In step S130, the kth neural network (neural network 200.1) is added to the group of classifiers. Here, the discriminator is a neural network. The group of classifiers is a group of learned neural networks used in processing for determining the type of traffic sound described later. For the first time, nothing is included in the group of classifiers before the process of step S130 is performed. Thereafter, the process proceeds to step S140.

  In step S140, k is incremented by one. Thereafter, the process of step S150 is performed.

  In step S150, it is determined whether k is greater than L. If NO in step S150, the process proceeds to step S110 again.

  By repeating the processes of steps S110, S120, S130, and S140 described above until the condition of step S150 is satisfied, neural networks 200.1, 200.2,. All of L is already learned. At this time, the trained neural networks 200.1, 200.2,..., 200. L is included.

  If YES in step S150, the bagging group learning process ends.

  Next, processing for determining the type of collected traffic sound in bagging (hereinafter also referred to as bagging determination processing) will be described.

  FIG. 6 is a flowchart showing the flow of the bagging determination process.

  FIG. 7 is a diagram for explaining the bagging determination process.

  FIG. 7A is a graph showing traffic sound input from the microphone 100. In FIG. 7A, the horizontal axis represents the elapsed time, and the vertical axis represents the sound pressure value.

  Next, the bagging determination process will be described with reference to FIG. 1, FIG. 6, and FIG.

  In step S200, traffic sound collection processing is performed. In the traffic sound collection process, traffic sound is collected from the microphone 100. The sound pressure calculation circuit 110 samples the collected traffic sound as analog data at 48 kHz. Note that the sampling frequency is not limited to 48 kHz, and may be another value.

  Then, the sound pressure calculation circuit 110 sets a certain section of 1024 data continuously sampled at 48 kHz as one section, and obtains the total value of the sound pressure values in the one section. Hereinafter, the total value is also referred to as a section sound pressure value. In FIG. 7A, each of the sections J, K, L, M, and N corresponds to the above-described one section.

  Next, the sound pressure calculation circuit 110 sets two adjacent sections (for example, the section J and the section K) as the first section and the second section. Here, the first section is a section (section J) of a time before the second section. The second section is a section (section K) of a time later than the first section.

  The sound pressure calculation circuit 110 performs a process for obtaining a difference between the section sound pressure values of the first and second sections by subtracting the section sound pressure value of the first section from the section sound pressure value of the second section. Do. This process is performed by the sound pressure calculation circuit 110 every time a predetermined time elapses. That is, the sound pressure value of the traffic sound (the waveform in FIG. 7A) is always monitored by the sound pressure calculation circuit 110. Thereafter, the process proceeds to step S210.

  In step S210, a process for determining whether or not a traffic sound of a predetermined level or higher has occurred (hereinafter also referred to as a heavy traffic sound determination process) is performed.

  In the heavy traffic sound determination process, the sound pressure calculation circuit 110 determines whether or not the difference between the section sound pressure values of the first and second sections is larger than a predetermined threshold value.

  If NO in step S210, the process proceeds to step S200 again. On the other hand, if YES at step S210, the process proceeds to step S220.

  In step S220, power spectrum calculation processing is performed.

  In the power spectrum calculation process, the power spectrum calculation circuit 120 includes a second section set by the sound pressure calculation circuit 110 and three sections (sections L, M, and N) that are continuous from the second section (section K). Then, FFT (Fast Fourier Transform) calculation processing is performed.

  7B, FIG. 7C, FIG. 7D, and FIG. 7E show the waveforms of the power spectrum calculation circuit 120 in section K, section L, section M, section M, and section N, respectively. Is a power spectrum obtained by performing an FFT calculation process.

  Then, the power spectrum calculation circuit 120 synthesizes the graphs obtained by performing the FFT processing on the section K, the section L, the section M, and the section N, respectively.

  FIG. 7F shows a waveform synthesized by the power spectrum calculation circuit 120. The power spectrum calculation circuit 120 generates a synthesized waveform based on the maximum value at each frequency of the synthesized waveform. The combined waveform is a thick line waveform in FIG. Thereafter, the process proceeds to step S230.

  In step S230, subband calculation processing is performed.

  In the subband calculation processing, the subband calculation circuit 130 calculates an average value of power spectra included in each of a plurality of continuous frequency bands (for example, 1 kHz to 10 kHz) with respect to the generated composite waveform. Is done. As a result, the power spectrum is divided into a plurality of frequency bands (subbands). In the following, the average value corresponding to each subband is referred to as feature data. It is assumed that the number of feature data in this embodiment is 512. The number of feature data is not limited to 512. Thereafter, the process proceeds to step S240.

  In step S240, an averaging calculation process is performed. If the number of feature data is 512, the number of input data of the neural network is large. Therefore, in the averaging calculation process, the subband calculation circuit 130 performs every 16 consecutive feature data out of 512 feature data. Find the average value. Further, the subband arithmetic circuit 130 normalizes the average value so that it falls within the range of “0” to “1”. In the present embodiment, the normalized average value (hereinafter also referred to as normalized average value) is used as input data for the neural network. The number of the input data is 32 from 512/16.

  In addition, when calculating | requiring the average value for every feature data, it is not limited to the above-mentioned 16 pieces. Further, the number of input data of the neural network may remain 512. That is, the process of step S240 may not be performed. Thereafter, the process proceeds to step S250.

  In step S250, the neural networks 200.1, 200.2,. The 32 input data obtained in step S240 are respectively input to the 32 input units of each L. Neural networks 200.1, 200.2,. L outputs output data OUT1, OUT2,..., OUTL, respectively, according to the input data. The output data OUT1, OUT2,..., OUTL are input to the determination circuit 250.

  Here, the output data output from the neural network will be described in detail. As an example, the neural network 200.1 will be described. In the neural network 200.1, when the 32 input data obtained in step S240 are input to 32 input units, respectively, a plurality of result data OT1, OT2,. OTc is output respectively. The result data OT1, OT2,..., OTc are decimal numbers between “0” and “1” (for example, 0.2, 0.8, etc.). Further, at least two of the result data OT1, OT2,..., OTc may have the same value.

  Data composed of result data OT1, OT2,..., OTc output from the neural network 200.1 is output data OUT1.

  Neural network 200.2,. Similarly to the neural network 200.1, when each of the 32 input data obtained in step S240 is input to each of the 32 input units, each of L is output from the plurality of output units to the plurality of result data OT1. , OT2,..., OTc. Data composed of result data OT1, OT2,..., OTc output from the neural network 200.2 is output data OUT2. The neural network 200. Data composed of result data OT1, OT2,..., OTc output from L is output data OUTL. Thereafter, the process proceeds to step S260.

  In step S260, a traffic sound determination process is performed. The traffic sound determination process is a process for determining which of the plurality of predetermined types of traffic sounds collected in S200.

  FIG. 8 is a flowchart showing the flow of traffic sound determination processing.

  Referring to FIG. 8, in step S262, determination circuit 250 performs data processing on input output data OUT1, OUT2,..., OUTL.

  As an example, data processing of output data OUT1 output from the neural network 200.1 will be described. The determination circuit 250 validates the result data indicating a predetermined threshold value (for example, 0.5) or more among the result data OT1, OT2,... OTc constituting the output data OUT1. Thus, there may be more than one valid result data.

  The determination circuit 250 determines a traffic sound associated with valid result data. In addition, when there are two or more effective result data among the plurality of result data output from the neural network 200.1, two traffic sounds (for example, “collision sound”) respectively corresponding to the two effective result data. , “Sudden brake sound”) is determined.

  As for the output data OUT2,..., OUTL output from the neural network 200.2,..., L, respectively, the determination circuit 250 performs data processing in the same manner as the above-described neural network 200.1. Do not repeat. Thereafter, the process proceeds to step S264.

In step S264, the neural network 200.2,. The majority of each L determination result is taken. Specifically, the determination circuit 250 includes neural networks 200.1,. The traffic sounds determined for each of L are tabulated. For example, if “collision sound” is determined by two neural networks, two “collision sounds” are set. For example, if “sudden brake sound” is determined by three neural networks, “sudden brake sound” is set to three. Such aggregation is performed, and finally, the largest number of traffic sounds are set as the determination results of the neural network unit 200. In the present embodiment, when there are a plurality of traffic sounds, the category label y corresponding to the input vector x _i as the traffic sound collected from the microphone 100 among the plurality of traffic sounds. If the traffic sound corresponding to _i is included, it is assumed that the neural network unit 200 performs a correct determination. And this traffic sound determination process is complete | finished and it returns to a bagging determination process.

  Referring again to FIG. 5, after the process of step S260, the bagging determination process ends.

  In the bagging determination process in the present embodiment, the result data indicating that the result data output from each of the plurality of neural networks is equal to or greater than a predetermined threshold is processed as valid data. The output values of the result data output by may be summed up for each corresponding traffic sound, and the traffic sound corresponding to the largest value may be determined as the determined traffic sound.

  Next, description will be given of the above-described processing by bagging and the results of processing when specific data is actually used.

  FIG. 9 is a diagram showing a data table T100 used for determination of traffic sound.

  Referring to FIG. 9, 119 traffic sound data are used in the present embodiment. The 119 traffic sound data belong to categories N1 to N11, that is, any of 11 categories. Further, with 119/11 = 10.8, approximately 11 traffic sound data per category are obtained. The number of traffic sound data to be used is not limited to 119, and may be any number. Also, the number of categories is not limited to 11 and may be any number.

  As an example, the category N1 will be described. The category N1 is a category of “collision sound”. The category N1 includes 10 pieces of traffic sound data.

  The categories N1 to N11 are divided into a plurality of groups. In this case, it was divided into two groups (Group A and Group B). Group A was “accident sound” and Group B was “non-accident sound” other than accident sound. Note that the number of groups is not limited to two and may be an arbitrary number.

  The experimental conditions are described below. Of the 119 traffic sound data, 100 were used for learning and the remaining 19 were used for determination. The number of neural networks was 10. The number of input units of the neural network is 32. The number of hidden layers in the neural network is one. The number of hidden units in the neural network was 20. The learning rate of the neural network was set to “0.2”. The range of the value of the result data output from the neural network is set to be “0” to “1”.

  In the present embodiment, the experiment was performed under three conditions of condition A, condition B, and condition C. When the condition A causes the neural network to learn, the learning circuit 210 determines which of the 11 categories the traffic sound data belongs to and learns the neural network. When determining the traffic sound using the learned neural network, the determination circuit 250 determines which category of the 11 categories the traffic sound data belongs to.

  In condition B, when the neural network learns, the learning circuit 210 determines which of the two groups the traffic sound data belongs to and learns the neural network. When determining the traffic sound using the learned neural network, the determination circuit 250 determines which group of the two groups the traffic sound data belongs to.

  Condition C is the same as condition A when the neural network is trained. When the traffic sound is determined by the learned neural network, it is the same as the condition B.

  First, the bagging collective learning process based on the data table T100 under the bagging condition A will be described.

  Referring to FIG. 4 again, in step S100, the number of neural networks is set to ten. Further, the number of input units of each neural network is set to 32. The number of hidden layers of the neural network is set to one. The number of hidden units in the neural network is set to 20. The learning rate of the neural network is set to “0.2”. The range of the value of the result data output from the neural network is set to be “0” to “1”.

Further, according to the condition A, the number of output units of the neural network is set to 11. Therefore, the data indicated by the category labels y ₁ ,..., Y ₁₀₀ corresponding to the input vectors x ₁ ,..., X ₁₀₀ of ₁₀₀ traffic sound data for learning are “1” to “11”, respectively. Is set to one of the following. Note that the values “1” to “11” indicated by the category labels y _i correspond to the categories N1 to N11, respectively. Then, it progresses to step S110.

  In step S110, the learning set generation process described above is performed. Thereafter, the process proceeds to step S120.

  In step S120, the above-described neural network learning process is performed. Thereafter, the processes of steps S130 and S140 are performed in order.

  Since the processes in steps S130 and S140 are the same as the processes in steps S130 and S140 described above, detailed description will not be repeated. Thereafter, the process proceeds to step S150.

  In step S150, it is determined whether k is larger than L (10). If NO in step S150, the process proceeds to step S110 again.

  All of the neural networks 200.1, 200.2,..., 200.10 have been learned by repeating the processing of steps S110, S120, S130, and S140 described above until the condition of step S150 is satisfied. It becomes. At this time, the group of classifiers includes learned neural networks 200.1, 200.2, ..., 200.10.

Next, the bagging determination process based on the data table T100 under the bagging condition A will be described. Here, 19 of traffic sound data are used for determining the input vector _becomes x 101 _{~x 119.}

  Referring to FIG. 6 again, steps S200, S210, S220, S230, and S240 are performed in order. Since the processes in steps S200, S210, S220, S230, and S240 are respectively similar to the processes in steps S200, S210, S220, S230, and S240 described above, detailed description thereof will not be repeated. Thereafter, the process proceeds to step S250.

  In step S250, the 32 input data obtained in step S240 are input to the 32 input units included in each of the neural networks 200.1, 200.2,. The neural networks 200.1, 200.2,..., 200.10 output output data OUT1, OUT2,. The output data OUT1, OUT2,..., OUT10 are input to the determination circuit 250.

  Since the number of categories is 11, each of OUT1, OUT2,..., OUT10 is composed of result data OT1, OT2,. Thereafter, the process proceeds to step S260.

  In step S260, a traffic sound determination process is performed.

  Since the traffic sound determination process is the same as the process when L = 10 and c = 11 in steps S262 and S264 described above, detailed description will not be repeated. Then, the bagging determination process ends.

  Then, random selection processing of 100 traffic sound data used for learning out of 119 traffic sound data, random selection processing of 19 traffic sound data used for determination out of 119 traffic sound data, 100 The learning process for each piece of traffic sound data and the judgment process for the 19 pieces of traffic sound data are repeated 20 times, and the probability that correct judgment is made (hereinafter also referred to as “correct answer rate”) is obtained.

  FIG. 10 is a diagram showing a data table T200 showing experimental results.

  Referring to FIG. 10, the column of NN (neural network) indicates a correct answer rate when learning and determination processing are performed by one conventional neural network instead of a plurality of neural networks. The accuracy rate of the bagging determination process based on the data table T100 under the above-described bagging condition A was 0.555. Under condition A, the correct answer rate of bagging was almost equal to the correct answer rate (0.554) of one neural network.

  Next, the bagging group learning process based on the data table T100 under the bagging condition B will be described.

Referring again to FIG. 4, step S100 is different from condition A only in that the number of output units of the neural network is set to two. Therefore, the data indicated by the category labels y ₁ ,..., Y ₁₀₀ corresponding to the input vectors x ₁ ,..., X ₁₀₀ of the ₁₀₀ traffic sound data for learning are “1” and “2”, respectively. Is set to one of the following. The values “1” and “2” indicated by the category label y _i correspond to the groups A and B, respectively. Other than that, it is the same as the process of step S100 in the case of condition A, and therefore detailed description will not be repeated. Thereafter, the processes of steps S110, S120, S130, S140, and S150 are performed in order.

  Since the processes in steps S110, S120, S130, S140, and S150 are the same as the processes in steps S110, S120, S130, S140, and S150 for the condition A described above, detailed description will not be repeated. Then, the bagging group learning process ends.

Next, the bagging determination process based on the data table T100 under the bagging condition B will be described. Here, 19 traffic sound data are used for determination, and the input vectors are x ₁₀₁ to x ₁₁₉ .

  Referring to FIG. 5 again, steps S200, S210, S220, S230, and S240 are performed in order. Since the processes in steps S200, S210, S220, S230, and S240 are the same as the processes in steps S200, S210, S220, S230, and S240 for condition A, detailed description thereof will not be repeated. Thereafter, the process proceeds to step S250.

  In step S250, a process similar to the process of S250 for condition A is performed. The output data OUT1, OUT2,..., OUT10 are input to the determination circuit 250.

  Since the number of groups is 2, each of OUT1, OUT2,..., OUT10 includes result data OT1 and OT2. Thereafter, the process proceeds to step S260.

  In step S260, a traffic sound determination process is performed.

  Since the traffic sound determination process is the same as the process when L = 10 and c = 2 in steps S262 and S264 described above, detailed description will not be repeated. The bagging determination process ends with the above-described process.

  Then, random selection processing of 100 traffic sound data used for learning out of 119 traffic sound data, random selection processing of 19 traffic sound data used for determination out of 119 traffic sound data, 100 The learning process for each piece of traffic sound data and the determination process for the 19 pieces of traffic sound data are repeated 20 times to obtain the above-mentioned correct answer rate.

  Referring to FIG. 10 again, the accuracy rate of the bagging determination process based on the data table T100 in the above-described bagging condition B is 0.824. Under condition B, the accuracy rate of the bagging determination process is improved by about 6% from the accuracy rate of one neural network (0.766). Therefore, it can be said that the traffic sound identification device according to the present invention was able to identify traffic sounds with very high accuracy.

  Next, the bagging collective learning process based on the data table T100 under the bagging condition C will be described.

  Referring to FIG. 4 again, in condition C, the processes of steps S100, S110, S120, S130, S140, and S150 for condition A are performed in order.

  Since the processes in steps S100, S110, S130, S120, S140, and S150 are the same as the processes in steps S100, S110, S130, S120, S140, and S150 when the condition A is described above, detailed description will be given. Will not repeat.

Next, the bagging determination process based on the data table T100 under the bagging condition C will be described. Here, 19 traffic sound data are used for determination, and the input vectors are x ₁₀₁ to x ₁₁₉ .

  Referring to FIG. 5 again, under condition C, the processes of steps S200, S210, S220, S230, S240, S250, and S260 for condition B are performed in order. The processes in steps S200, S210, S220, S230, S240, S250, and S260 are the same as the processes in steps S200, S210, S220, S230, S240, S250, and S260 in the case of condition B. The explanation is not repeated. Then, the bagging determination process ends.

  Then, random selection processing of 100 traffic sound data used for learning out of 119 traffic sound data, random selection processing of 19 traffic sound data used for determination out of 119 traffic sound data, 100 The learning process for each piece of traffic sound data and the determination process for the 19 pieces of traffic sound data are repeated 20 times to obtain the above-mentioned correct answer rate.

  Referring to FIG. 10 again, the accuracy rate of the bagging determination process based on the data table T100 under the above-described bagging condition C is 0.843. Under condition C, the accuracy rate of the bagging determination process is improved by about 4% from the accuracy rate of one neural network (0.800).

  In addition, the correct answer rate of condition C in which learning is performed at the category level and determination is performed at the group level is 0, compared to the correct answer rate (0.824) of condition B in which learning is performed at the group level and determination is performed at the group level. .843, an improvement of about 2%. Therefore, it can be said that the bagging of the condition C is very effective for identifying traffic sounds. Therefore, it can be said that the traffic sound identification device according to the present invention was able to identify traffic sounds with very high accuracy.

<Second Embodiment>
In the first embodiment, bagging has been described, but in this embodiment, Adaboost will be described. In the Adaboost, it will be described in detail later, how to create a training set S _k is different from the bagging.

  Since the traffic sound identification device used in the present embodiment is the same as traffic sound identification device 1000 of the first embodiment, detailed description of the configuration and functions will not be repeated.

  Next, the Adaboost collective learning process (hereinafter also referred to as Adaboost collective learning process) will be described.

  In the AdaBoost group learning process, the learning set S expressed by the above-described equation (1) is used as in the bagging group learning process.

  FIG. 11 is a flowchart showing the flow of the AdaBoost group learning process.

Referring to FIG. 11, in step S300, initialization processing similar to that in step S100 is performed. In step S300, the distribution w _k used for creating the learning set S _k is further initialized. The distribution w _k is updated every time the learning set S _k is created. More specifically, the first distribution W ₁ is required. Distribution _{W 1} is determined by the following equation (2).

B is a set of all mislabels. Here, the mislabeling, if the case i is an index of the input vector x _i, is that of incorrect label y of cases i. That is, (i, y) indicates an incorrect label pair. B is expressed by the following equation (3).

  | B | in the equation (2) is the number of elements of the mislabel set, and is represented by the following equation (4).

  Here, c indicates the number of the category labels or groups described above.

The distribution W ₁ is obtained from the equations (2), (3), and (4). Thereafter, the process proceeds to step S310.

At step S310, the generation processing of the k-th training set _{S k} is performed by using the distribution _{w k.} At this time, the initial value of k is set to “1”.

Referring to FIG. 5 again, learning sets S ₁ , S ₂ , and S _L are based on W ₁ , W ₂ , and W _L from learning set S (learning data) including N pieces of traffic sound data, respectively. , By randomly selecting N traffic sound data without preventing duplication.

Since W ₁ is a constant, only the _first learning set S ₁ is generated by randomly selecting N traffic sound data from the learning set S with a certain probability without preventing duplication.

Referring again to FIG. 11, in step S310, the when generation processing of the k-th training set _{S k} is terminated, the process proceeds to step S320.

In step S320, the learning sets _{S k} a k-th neural networks (e.g., the neural network 200.1) NN learning process of learning is carried out. Since the NN learning process is similar to the NN learning process in step S120 described above, detailed description will not be repeated. Thereafter, the process proceeds to step S330.

In step S330, the k-th neural network pseudo loss ε _k is calculated. The pseudo loss ε _k is _obtained by the following equation (5).

h _k (x _i , y _i ) is a value (for example, 0.8) of the output OT _yi of y _i when x _i is input to the k th neural network, and h _k (x _i , y) Is the sum (for example, 0.3) of the output values of category labels other than y _i when x _i is input to the k-th neural network. The larger the pseudo loss ε _k is, the more difficult it is to learn traffic sound data. The smaller the value of the pseudo loss ε _k (closer to “0”), the easier it is to learn traffic sound data. Thereafter, the process proceeds to step S340.

In step S340, the distribution w _k is updated. The distribution w _k is updated by the following equation (6).

Z _k is a normalization constant for making w _{k + 1 a} probability part distribution. β _k is _obtained by the following equation (7).

(5), (6) and (7), the distribution _{w k + 1} updating the distribution _{w k} is determined. Thereafter, the process proceeds to step S350.

  In step S350, the k-th learned neural network (neural network 200.1) is added to the group of classifiers described above. Note that for the first time, nothing is included in the group of classifiers before the process of step S350 is performed. Thereafter, the process proceeds to step S360.

  In step S360, k is incremented by one. Thereafter, the process of step S370 is performed.

  In step S370, it is determined whether k is greater than L. If NO in step S370, the process proceeds to step S310 again.

  The processes of steps S310, S320, S330, S340, S350, and S360 described above are repeated until the condition of step S370 is satisfied, so that the neural networks 200.1, 200.2,. All of L is already learned. At this time, the trained neural networks 200.1, 200.2,..., 200. L is included.

By the above processing, in S310, each time increasing the value of k, to change the distribution _{w k,} to generate the training set _{S k.} The learning set _Sk includes traffic sound data that is difficult to learn as the value of k increases. That is, the neural network 200 as input data the training set S _k value is large k. k can determine the traffic sound more accurately.

  If YES in step S370, this Adaboost group learning process ends.

  Next, a process for determining the type of collected traffic sound in the AdaBoost (hereinafter also referred to as an AdaBoost determination process) will be described.

  FIG. 12 is a flowchart showing the flow of the AdaBoost determination process.

  Next, the AdaBoost determination process will be described with reference to FIGS.

  In step S400, the same processing as in step S200 is performed, and thus detailed description will not be repeated. Thereafter, the process proceeds to step S410.

  In step S410, processing similar to that in step S210 is performed, and thus detailed description will not be repeated. If NO in step S410, the process proceeds to step S400 again. On the other hand, if YES in step S410, the processes in steps S420, S430, S440, and S450 are performed in order.

  Since the processes of steps S420, S430, S440, and S450 are the same as the processes of steps S220, S230, S240, and S250, detailed description will not be repeated. Thereafter, the process proceeds to step S460.

In step S460, traffic sound determination processing is performed. In the traffic sound determination process, the determination circuit 250 includes the neural networks 200.1, 200.2,. The traffic sound is determined by weighting calculation of the output data OUT1, OUT2,. Specifically, the determination circuit 250 determines the traffic sound based on the category label h _fin (x) obtained by the following equation (8) of the weighting calculation.

For example, if the category label h _fin (x) is “N2”, the traffic sound to be determined is determined to be the sudden brake sound of the category N2.

  Thereafter, this Adaboost determination process ends.

  Next, description will be given of the above-described processing by Adaboost and the processing results when specific data is actually used.

  Since the experimental conditions in the present embodiment are the same as the experimental conditions in the first embodiment, detailed description will not be repeated.

  First, the AdaBoost group learning process based on the data table T100 under the Aboost condition A will be described.

  Referring to FIG. 11 again, in step S300, the same processing as step S100 in condition A is performed, and therefore detailed description will not be repeated. Thereafter, steps S310, S320, S330, S340, S350, and S360 are sequentially performed.

  Since the processes in steps S310, S320, S330, S340, S350, and S360 are the same as those in steps S310, S320, S330, S340, S350, and S360, detailed description will not be repeated. Thereafter, the process proceeds to step S370.

  In step S370, it is determined whether k is larger than L (10). If NO in step S370, the process proceeds to step S310 again.

  The processes of steps S310, S320, S330, S340, S350, and S360 described above are repeated until the condition of step S370 is satisfied, whereby the neural networks 200.1, 200.2,. Everything has been learned. At this time, the group of classifiers includes learned neural networks 200.1, 200.2, ..., 200.10.

  Next, the AdaBoost determination process based on the data table T100 under the AdaBoost condition A will be described.

  Referring to FIG. 12 again, steps S400, S410, S420, S430, and S440 are performed in order. Since the processes in steps S400, S410, S420, S430, and S440 are the same as the processes in steps S400, S410, S420, S430, and S440 described above, detailed description will not be repeated. Thereafter, the process proceeds to step S450.

  In step S450, the 32 input data obtained in step S240 are input to the 32 input units of each of the neural networks 200.1, 200.2,. The neural networks 200.1, 200.2,..., 200.10 output output data OUT1, OUT2,. The output data OUT1, OUT2,..., OUT10 are input to the determination circuit 250.

  Since the number of categories is 11, each of OUT1, OUT2,..., OUT10 is composed of result data OT1, OT2,. Thereafter, the process proceeds to step S460.

  In step S460, traffic sound determination processing is performed.

  Since the traffic sound determination process is the same as the process when L = 10 and c = 11 in step S460 described above, detailed description will not be repeated. Then, the Adaboost determination process ends.

  Then, as in the case of the above-described bagging, the learning process and the determination process are repeated 20 times to obtain the above-described accuracy rate.

  Referring to FIG. 10 again, the accuracy rate of the AdaBoost determination process based on the data table T100 under the above-described Aboost condition A is 0.556. Under the condition A, the accuracy rate of the AdaBoost determination process was almost equal to the accuracy rate of one neural network (0.554) and the accuracy rate of bagging (0.555).

  Next, the AdaBoost group learning process based on the data table T100 under the AdaBoost condition B will be described.

  Referring to FIG. 11 again, in step S300, the difference is that the number of output units of the neural network is set to two as compared with the case of condition A. Otherwise, the above-described case of bagging is performed. Detailed description will not be repeated. Thereafter, steps S310, S320, S330, S340, S350, S360, and S370 are performed in order.

  The processes in steps S310, S320, S330, S340, S350, S360, and S370 are the same as the processes in steps S310, S320, S330, S340, S350, S360, and S370, respectively, under the condition A described above. Therefore, detailed description will not be repeated. Then, the AdaBoost group learning process ends.

  Next, the AdaBoost determination process based on the data table T100 under the AdaBoost condition B will be described.

  Referring to FIG. 12 again, steps S400, S410, S420, S430, and S440 are performed in order. Since the processes in steps S400, S410, S420, S430, and S440 are the same as the processes in steps S400, S410, S420, S430, and S440 for condition A, detailed description will not be repeated. Thereafter, the process proceeds to step S450.

  In step S450, a process similar to the process of S450 in the case of condition A is performed. The output data OUT1, OUT2,..., OUT10 are input to the determination circuit 250.

  Since the number of groups is 2, each of OUT1, OUT2,..., OUT10 includes result data OT1 and OT2. Thereafter, the process proceeds to step S460.

  In step S460, traffic sound determination processing is performed.

  The traffic sound determination process is the same as the process when L = 10 and c = 2 in step S460 for the condition A described above, and thus detailed description will not be repeated. Then, the Adaboost determination process ends.

  Then, as in the case of bagging described above, the learning process and the determination process are repeated 20 times to obtain the correct answer rate.

  Referring to FIG. 10 again, the accuracy rate of the AdaBoost determination process based on the data table T100 under the above-described Condition B under AdaBoost is 0.793. Under condition B, the accuracy rate of the AdaBoost determination process is about 3% worse than the accuracy rate of bagging (0.824), but is about 3% higher than the accuracy rate of one neural network (0.766). ing. Therefore, it can be said that the traffic sound identification device according to the present invention was able to identify traffic sounds with very high accuracy.

  Next, the Adaboost collective learning process based on the data table T100 under the Adaboost condition C will be described.

  Referring to FIG. 11 again, under condition C, steps S300, S310, S320, S330, S340, S350, S360, and S370 for condition A are sequentially performed.

  The processes of steps S300, S310, S320, S330, S340, S350, S360, and S370 are the same as the processes of steps S300, S310, S320, S330, S340, S350, S360, and S370, respectively, under the above-described condition A. Since the process is performed, detailed description will not be repeated.

  Next, the AdaBoost determination process based on the data table T100 under the AdaBoost condition C will be described.

  Referring to FIG. 12 again, under condition C, steps S400, S410, S420, S430, S440, S450, and S460 in the case of condition B are sequentially performed. Since the processes of steps S400, S410, S420, S430, S440, S450, and S460 are the same as the processes of steps S400, S410, S420, S430, S440, S450, and S460 for the condition B, respectively. Detailed description will not be repeated. Then, the Adaboost determination process ends.

  Then, as in the case of the above-described bagging, the learning process and the determination process are repeated 20 times to obtain the above-described accuracy rate.

  Referring to FIG. 10 again, the correct rate of Adaboost based on the data table T100 under the above-mentioned Condition C in Adaboost is 0.805. Under the condition C, the correct answer rate of AdaBoost was almost equal to the correct answer rate (0.800) of one neural network.

  In addition, the correct answer rate of the condition C in which the learning is performed at the category level and the determination at the group level is 0 than the correct answer rate (0.793) of the condition B in which the learning is performed at the group level and the determination is performed at the group level. .805, an improvement of about 1%. Therefore, it can be said that the Adaboost of the condition C is very effective for identifying traffic sounds. Therefore, it can be said that the traffic sound identification device according to the present invention was able to identify traffic sounds with very high accuracy.

<Third Embodiment>
The present invention can also be applied to a personal computer (hereinafter also referred to as a PC (Personal Computer)) to which a microphone is connected. In that case, the communication program may be read from the recording medium on which the traffic sound determination program 180A is recorded, installed on the PC, and the PC may be operated based on the traffic sound determination program 180A. Detailed description will be given below.

  FIG. 13 is a block diagram showing an internal configuration of PC 500 in the present embodiment. Note that FIG. 13 also shows a communication unit 570 and a recording medium 555 for explanation. FIG. 13 shows a configuration when the traffic sound determination program 180A is installed in the PC 500. In this case, the PC 500 operates as the traffic sound identification device 1000.

  Referring to FIG. 13, communication unit 570 exchanges data with a network in a wired or wireless manner. Communication unit 570 exchanges data with the network. The communication unit 570 is a communication interface (for example, a router) using Ethernet (registered trademark).

  Further, the communication unit 570 may be any of a communication interface for performing data communication using a wireless technology such as IEEE802.11a, IEEE802.11b, IEEE802.11g, which are wireless LAN standards.

  A display unit 530, a mouse 542, a keyboard 544, and a microphone 546 are connected to the PC 500.

  Display unit 530 displays an image based on the image data output from PC 500. Display unit 530 displays an image in accordance with an instruction from control unit 510. The display unit 530 is a liquid crystal display (LCD (Liquid Crystal Display)), a CRT (Cathode Ray Tube), an FED (Field Emission Display), a PDP (Plasma Display Panel), an organic EL display (Organic Electroluminescence Display), a dot matrix, or the like. Any other display device of an image display method may be used.

  The mouse 542 is an interface for the user to operate the PC 500. The keyboard 544 is an interface for the user to operate the PC 500. The microphone 546 has a function of collecting traffic sound, similar to the microphone 100 described above.

  The PC 500 includes a control unit 510, a data temporary storage unit 522, a storage unit 520, a communication unit 560, a VDP (Video Display Processor) 532, a CGROM (Character Graphic Read Only Memory) 534, and a VRAM (Video Random Access). Memory) 536, an input unit 540, and a recording medium access unit 550.

  The CGROM 534 stores image data for generating an image to be displayed on the display unit 530 such as font data and graphic data.

  The storage unit 520 stores a communication program 180A for causing the control unit 510 to perform predetermined processing, and other various data. The storage unit 520 is accessed by the control unit 510.

  The storage unit 520 is accessed by the control unit 510. The storage unit 520 is a hard disk capable of storing a large amount of data. Note that the storage unit 520 is not limited to a hard disk, and may be any medium (for example, a flash memory) that can hold data without being supplied with power.

  That is, the storage unit 520 uses an EEPROM (Erasable Programmable Read Only Memory) capable of erasing and writing the memory any number of times, an EEPROM (Electronically Erasable and Programmable Read Only Memory) capable of being electrically rewritten, and ultraviolet rays. Any of a UV-EPROM (Ultra-Violet Erasable Programmable Read Only Memory) capable of erasing and rewriting stored contents any number of times and a circuit having a configuration capable of storing and storing data in a nonvolatile manner may be used.

  Control unit 510 has a function of performing various types of processing, arithmetic processing, and the like for each device inside PC 500 in accordance with traffic sound determination program 180A stored in storage unit 520. The control unit 510 includes a microprocessor, an FPGA (Field Programmable Gate Array), which is an LSI (Large Scale Integration) that can be programmed, and an ASIC (Application, which is an integrated circuit designed and manufactured for a specific application. Specific Integrated Circuit) or any other circuit having an arithmetic function may be used.

  The control unit 510 also instructs the VDP 532 to generate an image and display the image on the display unit 530 according to the traffic sound determination program 180A stored in the storage unit 520 (hereinafter referred to as a “drawing instruction”). Call out).

  The VDP 532 is connected to the display unit 530. The VDP 532 reads necessary image data from the CGROM 534 in accordance with a drawing instruction from the control unit 510 and generates an image using the VRAM 536. The VDP 532 reads out the image data stored in the VRAM 536 and causes the display unit 530 to display an image based on the image data.

  The VRAM 536 has a function of temporarily storing an image generated by the VDP 532.

  The data temporary storage unit 522 is accessed as data by the control unit 510 and used as a work memory for temporarily storing data.

  The data temporary storage unit 522 is a high-speed RAM (Random Access Memory), SRAM (Static Random Access Memory), DRAM (Dynamic Random Access Memory), SDRAM (Synchronous DRAM), or double data rate mode capable of temporarily storing data. DDR-SDRAM (Double Data Rate SDRAM), which is an SDRAM with an excellent data transfer function, RDRAM (Rambus Dynamic Random Access Memory), DRAM that adopts high-speed interface technology developed by Rambus, and Direct-RDRAM (Direct Rambus Dynamic) Random Access Memory) or any other circuit having a configuration capable of storing and storing data in a volatile manner.

  A mouse 542, a keyboard 544, and a microphone 546 are connected to the input unit 540. An input instruction from the mouse 542 or the keyboard 544 is transmitted to the control unit 510 via the input unit 540. Control unit 510 performs predetermined processing based on an input instruction from input unit 540. The traffic sound collected from the microphone 546 is transmitted to the control unit 510 via the input unit 540.

  The control unit 510 uses the sound pressure calculation circuit 110, the power spectrum calculation circuit 120, the subband calculation circuit 130, and the neural network described in the first and second embodiments for the traffic sound input from the input unit 540. 200.1, 200.2, ..., 200. L, processing similar to that of the learning circuit 210 and the determination circuit 250 is performed.

  The recording medium access unit 550 has a function of reading the traffic sound determination program 180A from the recording medium 555 on which the traffic sound determination program 180A is recorded. The traffic sound determination program 180A stored in the recording medium 555 is read from the recording medium access unit 550 and stored in the storage unit 520 by the operation (installation process) of the control unit 510.

  The recording medium 555 includes a DVD-ROM (Digital Versatile Disk Read Only Memory), a CD-ROM (Compact Disk Read Only Memory), an MO (Magneto Optical Disk), a floppy (registered trademark) disk, a CF (Compact Flash) card, and an SM. (Smart Media (registered trademark)), MMC (Multi Media Card), SD (Secure Digital) memory card, Memory Stick (registered trademark), xD picture card and USB memory, magnetic tape, and other non-volatile memory Good.

  Communication unit 560 exchanges data with control unit 510. Communication unit 560 exchanges data with communication unit 135 in a wired or wireless manner. The communication unit 560 is an interface similar to the communication unit 570 described above.

  The communication unit 560 may be any one of USB (Universal Serial Bus) 1.1, USB 2.0, and other communication interfaces for performing serial transfer. The communication unit 560 may be any of a Centronics specification, IEEE 1284 (Institute of Electrical and Electronic Engineers 1284), and other communication interfaces that perform parallel transfer. The communication unit 560 may be any one of communication interfaces using IEEE 1394 or other SCSI standards.

  As described above, even with a PC on which the traffic sound determination program 180A is installed, the same processing as in the first and second embodiments can be performed, and the present invention can be applied.

  Therefore, also in this embodiment, the same effect as in the first and second embodiments can be obtained.

  The device on which the traffic sound determination program 180A is recorded is not limited to a PC, but a device capable of collecting traffic sounds is connected, the traffic sound determination program 180A can be installed, and the device is based on the traffic sound determination program 180A. Any device can be used.

  As described above, in the present invention, the bagging or Adaboost technique is performed using a neural network, but the present invention is not limited to this, and a technique similar to Bagging or Adaboost may be used.

  The experimental results described in the first and second embodiments are numerical values that do not have a very high accuracy rate because only 100 traffic sound data are used for learning. Increasing the number of sound data is considered to improve the accuracy rate.

  The embodiment disclosed this time should be considered as illustrative in all points and not restrictive. The scope of the present invention is defined by the terms of the claims, rather than the description above, and is intended to include any modifications within the scope and meaning equivalent to the terms of the claims.

It is a block diagram which shows the structure of the traffic sound identification apparatus in this Embodiment. It is the figure which showed the internal structure of the neural network part in this Embodiment. It is the figure which showed the detailed internal structure of the neural network in this Embodiment. It is a flowchart which shows the flow of a bagging group learning process. It is a figure for demonstrating a learning set production | generation process. It is a flowchart which shows the flow of a bagging determination process. It is a figure for demonstrating a bagging determination process. It is a flowchart which shows the flow of a traffic sound determination process. It is a figure which shows the data table used for determination of a traffic sound. It is a figure which shows the data table which shows an experimental result. It is a flowchart which shows the flow of an Ada boost group learning process. It is a flowchart which shows the flow of an Ada boost determination process. It is a block diagram which shows the internal structure of PC in this Embodiment.

Explanation of symbols

  100 microphone, 110 sound pressure calculation circuit, 120 power spectrum calculation circuit, 130 subband calculation circuit, 180A traffic sound determination program, 200 neural network unit, 200.1, 200.2,. L neural network, 210 learning circuit, 250 determination circuit, 500 PC, 555 recording medium, 1000 traffic sound identification device.

Claims (8)

A traffic sound identification device that determines which of a plurality of predetermined types of traffic sounds is collected,
An arithmetic circuit for calculating a power spectrum of the traffic sound;
A subband circuit that divides the power spectrum calculated by the arithmetic circuit into a plurality of subbands, and generates I (natural number) input data based on the power spectrum divided into the plurality of subbands ;
L neural networks pre-ordered from first to L (natural number) for determining the type of traffic sound based on the power spectrum divided into the plurality of subbands;
A learning circuit for learning the L neural networks,
The learning circuit sequentially generates L sets of learning data by performing a process of selecting N traffic sound data randomly from N (natural number) traffic sound data for learning without interfering with duplication. Then, a learning process for sequentially learning with the L neural networks is performed,
Each of the learned L neural networks outputs a judgment result of the traffic sound type based on the I input data ,
A traffic sound identification device further comprising a determination circuit for determining the type of traffic sound based on a majority decision of L determination results respectively output from the learned L neural networks. A traffic sound identification device that determines which of a plurality of predetermined types of traffic sounds is collected,
An arithmetic circuit for calculating a power spectrum of the traffic sound;
A subband circuit that divides the power spectrum calculated by the arithmetic circuit into a plurality of subbands;
L neural networks pre-ordered from first to L (natural number) for determining the type of traffic sound based on the power spectrum divided into the plurality of subbands;
A learning circuit for learning the L neural networks,
The learning circuit performs L data selection processing to select N traffic sound data randomly from N (natural number) traffic sound data for learning without preventing duplication, thereby obtaining L sets of learning data. A learning process that sequentially generates and sequentially learns with the L neural networks is performed,
In the data selection process, the learning circuit performs the learning process of the (k + 1) th (k + 1) (k: a natural number smaller than L) of the L pieces of neural networks, and the kth of the L pieces of learning circuits. Select traffic sound data that was difficult to learn in the neural network learning process,
Each of the learned L neural networks outputs a determination result of the traffic sound type based on the power spectrum divided into the plurality of subbands,
A traffic sound identification device further comprising a determination circuit for determining the type of the traffic sound based on a weighting calculation of L determination results respectively output from the learned L neural networks. When the learning circuit causes the L neural networks to learn, the type of traffic sound used for learning determination is M (natural number),
The M traffic sounds are divided into Q groups (natural numbers smaller than M),
3. The traffic sound identification device according to claim 1, wherein when the determination circuit determines the traffic sound, the traffic sound to be determined belongs to which group of the Q groups. A traffic sound determination program executed by a computer having a microphone for collecting traffic sound,
Calculating a power spectrum of the traffic sound;
Dividing the calculated power spectrum into a plurality of subbands , and generating I (natural number) input data based on the power spectrum divided into the plurality of subbands ;
By performing a data selection process for selecting N traffic sound data randomly from N (natural number) traffic sound data for learning without interfering with duplication, L sets of learning data are sequentially generated, Sequentially configuring L neural networks in which the learning data generated sequentially is trained;
Causing each of the learned L neural networks to output a determination result of the traffic sound type based on the I input data ;
A traffic sound determination program for causing a computer to execute a step of determining the type of traffic sound based on a majority decision of L determination results respectively output from the L neural networks learned. A traffic sound determination program executed by a computer having a microphone for collecting traffic sound,
Calculating a power spectrum of the traffic sound;
Dividing the computed power spectrum into a plurality of subbands;
By performing a data selection process for selecting N traffic sound data randomly from N (natural number) traffic sound data for learning without interfering with duplication, L sets of learning data are sequentially generated, Sequentially configuring L neural networks in which the learning data generated sequentially is trained;
Causing each of the learned L neural networks to output a judgment result of the traffic sound type based on the power spectrum divided into the plurality of subbands;
Making the computer perform the step of determining the type of the traffic sound based on the weighting calculation of the L determination results output from the learned L neural networks,
In the data selection process, when learning processing of the (k + 1) th (k + 1) (k: natural number less than L) of the L pieces of neural networks is performed, learning is performed with the kth neural network of the L pieces. A traffic sound determination program for causing a computer to further execute a step of preferentially selecting traffic sound data that has been difficult. The recording medium which recorded the traffic sound determination program of Claim 4 or Claim 5 . A traffic sound determination method for determining whether the collected traffic sound is any of a plurality of predetermined traffic sounds,
Calculating a power spectrum of the traffic sound;
Dividing the calculated power spectrum into a plurality of subbands , and generating I (natural number) input data based on the power spectrum divided into the plurality of subbands ;
By performing a data selection process for selecting N traffic sound data randomly from N (natural number) traffic sound data for learning without interfering with duplication, L sets of learning data are sequentially generated, Sequentially configuring L neural networks in which the learning data generated in sequence is sequentially learned;
Outputting each of the learned L neural networks based on the I pieces of input data to output a judgment result of the traffic sound type;
And a step of determining the type of the traffic sound based on a majority decision of the L determination results output from the learned L neural networks. A traffic sound determination method for determining whether the collected traffic sound is any of a plurality of predetermined traffic sounds,
Calculating a power spectrum of the traffic sound;
Dividing the computed power spectrum into a plurality of subbands;
By performing a data selection process for selecting N traffic sound data randomly from N (natural number) traffic sound data for learning without interfering with duplication, L sets of learning data are sequentially generated, Sequentially configuring L neural networks in which the learning data generated in sequence is sequentially learned;
Outputting each of the learned neural networks based on the power spectrum divided into the plurality of sub-bands based on the traffic sound type determination results;
Determining the type of the traffic sound based on a weighting calculation of L determination results respectively output from the L neural networks learned.
In the data selection process, when learning processing of the (k + 1) th (k + 1) (k: natural number less than L) of the L pieces of neural networks is performed, learning is performed with the kth neural network of the L pieces. A traffic sound determination method that preferentially selects difficult traffic sound data.

Images