Detection device, detection method, and detection program

Abstract

To provide a detection device capable of suppressing a decline in detection performance even in the case of a plurality of sound sources, a detection method, and a detection program.SOLUTION: A detection device comprises: a likelihood generation unit which generates a likelihood from observation information related to sound; a Bayesian estimation unit which calculates a posterior probability distribution via Bayesian estimation using the likelihood; a sorting unit which resorts probability values in the posterior probability distribution; a number estimation unit which estimates the number of probability values corresponding to sound sources from the resorted probability values; a calculation unit which calculates a test statistic related to the presence or absence of sound sources on the basis of the number of probability values; and a determination unit which determines the presence or absence of sound sources on the basis of the test statistic.SELECTED DRAWING: Figure 14

Description

[Technical Field]

[0001] The present invention relates to a detection device, a detection method, and a detection program for detecting a sound source based on observation information. [Background technology]

[0002] Conventionally, tracking devices are known that automatically detect and track moving objects based on observation information regarding the direction, frequency, speed, distance, etc. of the moving object observed by sonar, radar, lidar, etc. The tracking device acquires acoustic power for each direction using, for example, a microphone array or a hydrophone array. The acquired acoustic power information for each direction is then regarded as the likelihood of the moving object's direction, and a posterior probability is calculated using Bayesian estimation, and the moving object is detected based on the posterior probability (see, for example, Non-Patent Document 1). [Prior art documents] [Non-patent literature]

[0003] [Non-Patent Document 1] REBethel and GJParas, “A PDF Multitarget Tracker”, IEEE Transactions on Aerospace and Electronic Systems, vol. 30, NO. 2, April 1994, p. 386-403 Summary of the Invention [Problem to be solved by the invention]

[0004] In conventional tracking devices, the search unit detects sound sources one by one and passes the information to the tracking unit to begin tracking. However, there are cases where the search unit has not yet detected multiple sound sources. In such cases, the posterior probability distribution in conventional tracking devices is normalized so that it becomes 1 when integrated over the entire space. Therefore, compared to the posterior probability when there is one sound source, the posterior probability for the same sound source when multiple sound sources are present is relatively lower. On the other hand, since the detection threshold is determined from the false alarm probability and is constant regardless of the number of sound sources, the detection ability of moving objects is reduced as a result.

[0005] The present invention has been made in light of the above-mentioned problems, and aims to provide a detection device, a detection method, and a detection program that can suppress a decrease in detection capability even in the case of multiple sound sources. [Means for solving the problem]

[0006] The detection device according to the present invention includes a likelihood generation unit that generates likelihoods from observation information related to sound, a Bayesian estimation unit that uses the likelihoods to determine a posterior probability distribution through Bayesian estimation, a sorting unit that rearranges probability values ​​in the posterior probability distribution, a number estimation unit that estimates the number of probability values ​​corresponding to a sound source from the rearranged probability values, a calculation unit that calculates a validation quantity related to the presence or absence of a sound source based on the number of probability values, and a determination unit that determines the presence or absence of a sound source based on the validation quantity.

[0007] The detection method according to the present invention includes the steps of generating a likelihood from observational information about a sound, determining a posterior probability distribution by Bayesian estimation using the likelihood, rearranging the probability values ​​in the posterior probability distribution, estimating the number of probability values ​​corresponding to a sound source from the rearranged probability values, calculating a validation metric regarding the presence or absence of a sound source based on the number of probability values, and determining the presence or absence of a sound source based on the validation metric.

[0008] A detection program according to the present invention is a program that causes a processor of a detection device to execute the above detection method. [Effects of the Invention]

[0009] According to the present invention, the number of probability values ​​corresponding to a sound source is estimated from the probability values ​​obtained by rearranging the posterior probability distribution, and the presence or absence of a sound source is determined using a calibration quantity calculated based on the number of probability values, thereby making it possible to suppress a decrease in detection ability even in the case of multiple sound sources. [Brief explanation of the drawings]

[0010] [Figure 1] FIG. 1 is a schematic diagram illustrating the configuration of a tracking device according to a conventional technique. [Figure 2] FIG. 1 is a schematic diagram illustrating the configuration of a search unit in a tracking device according to a conventional technique. [Figure 3] 10A and 10B are diagrams illustrating a mask process performed by a mask processing unit of a searching unit according to the prior art. [Figure 4] FIG. 10 is a schematic diagram illustrating the configuration of a detection processing unit of a searching unit according to the prior art. [Figure 5] FIG. 1 is a schematic diagram illustrating the configuration of a tracking unit in a tracking device according to a conventional technique. [Figure 6] 10A and 10B are diagrams illustrating mask processing by a mask processing unit of a tracking unit according to the prior art. [Figure 7] 10A and 10B are diagrams illustrating an example of mask generation in a mask generation unit of a tracking unit according to the prior art; [Figure 8] FIG. 10 illustrates a situation where there is one sound source. [Figure 9] This is an image of the posterior probability distribution when there is one sound source. [Figure 10] FIG. 1 illustrates a situation where there are two sound sources. [Figure 11] This is an image of the posterior probability distribution when there are two sound sources. [Figure 12] 1 is a schematic configuration diagram of a tracking device according to a first embodiment. [Figure 13] FIG. 2 is a schematic configuration diagram of a searching unit according to the first embodiment. [Figure 14] 3 is a schematic configuration diagram of a detection processing unit of a searching unit according to the first embodiment. FIG. [Figure 15] 5 is a diagram illustrating a detection process in a detection processing unit according to the first embodiment. FIG. [Figure 16] FIG. 10 is a diagram for explaining a problem solved in the second embodiment. [Figure 17] FIG. 10 is a schematic configuration diagram of a tracking device according to a second embodiment. [Figure 18] FIG. 10 is a schematic configuration diagram of a searching unit according to a second embodiment. [Figure 19] FIG. 10 is a schematic configuration diagram of a detection processing unit of a searching unit according to a second embodiment. [Figure 20] 10 is a diagram illustrating a detection process in a detection processing unit according to the second embodiment. FIG. DETAILED DESCRIPTION OF THE INVENTION

[0011] First, a tracking device 500 according to the prior art will be described. The following description will be given taking as an example a case where observation information relating to direction obtained by sonar 200 is used. FIG. 1 is a schematic configuration diagram of a tracking device 500 according to the prior art. As shown in FIG. 1, the tracking device 500 includes a searching unit 51 and a plurality of tracking units 52a to 52c. Although three tracking units 52a to 52c are shown in FIG. 1, the number of tracking units may be two or less, or four or more.

[0012] When tracking processing is started in the tracking device 500, observation information from the sonar 200 is input to the search unit 51. The observation information is a sensor output, and is amplitude information of sound received by a microphone array or hydrophone array composed of multiple receivers provided in the sonar 200. When the observation information is input, the search unit 51 calculates a likelihood related to the direction and searches for the sound source based on the likelihood.

[0013] When the searching unit 51 detects a sound source, it outputs information about the detected sound source to the tracking units 52a to 52c. The information about the detected sound source is the estimated direction, ID number, and prior probability distribution of the sound source detected by the searching unit 51. The tracking units 52a to 52c each track a sound source and output information about the sound source being tracked. The information about the sound source being tracked is the estimated direction and ID number of the sound source being tracked by each of the tracking units 52a to 52c.

[0014] At this time, in order to prevent the sound source being tracked by each of the tracking units 52a to 52c from affecting the searching unit 51 and the other tracking units, each of the tracking units 52a to 52c generates a mask that suppresses the influence of the sound source being tracked and outputs it to the searching unit 51 and the other tracking units. The searching unit 51 and the tracking units 52a to 52c apply the mask generated by the other tracking units to the likelihood, thereby suppressing the influence of the sound source being tracked by the other tracking units.

[0015] Next, the operation of the searching unit 51 will be described. The searching unit 51 receives as input observation information and masks output from all of the active (sound source tracking) tracking units 52a to 52c, and searches for sound sources that have not yet been detected. When the searching unit 51 detects a sound source, it outputs the estimated direction, ID number, and prior probability distribution of the sound source detected by the searching unit 51 to any of the inactive tracking units 52a to 52c, and activates the inactive tracking units 52a to 52c.

[0016] For example, a case will be described in which the tracking unit 52a is not active, the searching unit 51 detects a sound source, and the tracking unit 52a is activated. The estimated direction, ID number, and prior probability distribution of the sound source detected by the searching unit 51 are output from the searching unit 51 and input to the tracking unit 52a, activating the tracking unit 52a. The active tracking unit 52a only generates a mask and outputs the generated mask to the searching unit 51, as well as to the active tracking units 52b and 52c other than the tracking unit 52a.

[0017] Furthermore, in a processing cycle in which the tracker 52a becomes active, the tracker 52a outputs the estimated direction, ID number, and prior probability distribution of the sound source detected by the searcher 51, which are input from the searcher 51, as the estimated direction and ID number of the sound source tracked by the tracker 52a. In subsequent processing cycles, the tracker 52a receives input of observation information and masks output from the active trackers 52b and 52c other than the tracker 52a. The tracker 52a then calculates the estimated direction of the sound source being tracked using the input mask and observation information, and generates a mask corresponding to the sound source being tracked by the tracker 52a based on the estimated direction. The tracker 52a then outputs the generated mask and the estimated direction and ID number of the sound source tracked by the tracker 52a.

[0018] Next, detailed operations of the searching unit 51 and the tracking units 52a to 52c in Fig. 1 will be described. Fig. 2 is a schematic configuration diagram of the searching unit 51 in a tracking device 500 according to the prior art. The searching unit 51 includes a likelihood generating unit 511, a mask processing unit 512, a Bayesian estimation unit 513, a detection processing unit 514, a Markov updating unit 515, a direction estimating unit 516, and an ID assigning unit 517. Each unit of the searching unit 51 is implemented by a dedicated device or a dedicated processing circuit. Alternatively, the tracking device 500 may include a processor such as a CPU and a memory, and the processor may execute a program stored in the memory to implement each unit of the searching unit 51. Alternatively, each unit of the searching unit 51 may be implemented by a combination of dedicated device or dedicated circuitry and software.

[0019] The search unit 51 receives the observation information and the masks a to c output from the active tracking units 52a to 52c. The likelihood generation unit 511 receives the observation information and calculates a likelihood function L t(θ), where t represents time. The likelihood generation unit 511 generates a Bearing Level indicating the power level for each direction by, for example, performing beamforming, and outputs the Bearing Level as a likelihood. Here, likelihood generation using beamforming is described as an example, but the likelihood generation unit 511 can use various techniques for generating likelihood, such as calculating likelihood from a direction estimated by MUSIC (Multiple Signal Classification) or the like. The likelihood generation unit 511 calculates the likelihood function L t (θ) is output to the mask processing unit 512.

[0020] The mask processing unit 512 calculates the likelihood function L output from the likelihood generation unit 511. t (θ) and the masks a to c output from the active tracking units 52a to 52c are input, and the likelihood function L t (θ), the influence of sound sources already tracked by the active trackers 52a to 52c is suppressed. FIG. 3 is a diagram illustrating the masking process by the masking processor 512 of the searcher 51 according to the prior art. FIG. 3 shows the masking process when only the trackers 52a and 52b are active. FIG. 3(a) shows the likelihood function L before the masking process. t (θ), and FIG. 3(b) shows the mask a generated by the tracking unit 52a, FIG. 3(c) shows the mask b generated by the tracking unit 52b, and FIG. 3(d) shows the masked likelihood function L mask t (θ).

[0021] The mask processing unit 512 uses the likelihood function L shown in FIG. t A weighting coefficient is applied to (θ) by multiplying the mask a generated by the tracking unit 52a shown in FIG. 3(b) and the mask b generated by the tracking unit 52b shown in FIG. 3(c). This results in a masked likelihood function L mask t(θ) can be obtained. In this way, by suppressing the sound sources being tracked by the tracker 52a and the tracker 52b, sound sources that have not yet been tracked remain as peaks. The mask processing unit 512 calculates the masked likelihood function L mask t (θ) is output to the Bayesian estimation unit 513.

[0022] The Bayesian estimation unit 513 calculates the masked likelihood function L output from the mask processing unit 512. mask t (θ) and the prior probability distribution P calculated one processing cycle ago prior t|t-1 (θ) and are input, and the posterior probability distribution P posterior t Then, the Bayesian estimation unit 513 calculates the calculated posterior probability distribution P posterior t (θ) is output to the detection processing unit 514, the Markov update unit 515, and the direction estimation unit 516. Here, "'" in equation (1) indicates a variable when calculating the summation.

number

[0023] The detection processing unit 514 uses the posterior probability distribution P calculated by the Bayesian estimation unit 513 posterior t (θ) is used as input, and the presence or absence of the sound source is determined. detect t is output to the Markov update unit 515, the direction estimation unit 516, and the ID assignment unit 517. detect t =1 indicates that a new sound source has been detected, and flag detect t =0 indicates that no new sound source has been detected.

[0024] The Markov update unit 515 receives the posterior probability distribution P posterior t (θ) and the determination result flag output from the detection processing unit 514 detectt The Markov update unit 515 receives the result flag detect t If is 0, the posterior probability distribution P posterior t Based on (θ), the prior probability distribution P at time t+1 prior t+1|t (θ) and outputs it to the Bayesian estimation unit 513 after one processing cycle. prior t+1|t (θ) is the prior probability distribution P prior t|t-1 (θ) is used as the judgment result flag detect t If is 1, the Markov update unit 515 updates the prior probability distribution P prior t+1|t (θ) is output from the search unit 51.

[0025] The direction estimation unit 516 receives the posterior probability distribution P posterior t (θ) and the determination result flag output from the detection processing unit 514 detect t The result of the judgment is input as flag. detect t If is 1, the direction estimation unit 516 calculates the posterior probability distribution P posterior t Based on (θ), the estimated direction θ^ is calculated using techniques such as MAP (Maximum A Posteriori) estimation or weighted average expected value calculation. t is calculated and output as the estimated direction. In this specification, the symbol "^" is a hat that indicates an estimated value and is written above the symbol on the right. detect t If is 0, the direction estimation unit 516 does not perform processing.

[0026] The ID assigning unit 517 receives the determination result flag output from the detection processing unit 514. detect t If the value is 1, the ID number is output. detectt If the value is 0, the ID assigning unit 517 does not perform the process.

[0027] Next, the detection processing unit 514 of the search unit 51 in the conventional technology will be described. FIG. 4 is a schematic configuration diagram of the detection processing unit 514 of the search unit 51 according to the conventional technology. As shown in FIG. 4, the detection processing unit 514 has a maximum value search unit 541 and a threshold determination unit 542. The posterior probability distribution P posterior t When (θ) is input to the detection processing unit 514, the maximum value search unit 541 calculates the posterior probability distribution P posterior t The maximum value of the posterior probability at (θ) is p max t The maximum value p detected by the maximum value search unit 541 is max t is output to the threshold determination unit 542.

[0028] The threshold determination unit 542 determines the maximum value p output from the maximum value search unit 541. max t The threshold determination unit 542 performs threshold determination as shown in the following formula (2), and outputs the determination result flag detect t Output.

number

[0029] where threshold detect is the detection threshold, and flag detect t =0 indicates that no new sound source has been detected, and flag detect t =1 indicates that a new sound source has been detected. detect is typically preset to a desired false alarm probability in the absence of a sound source.

[0030] Next, the operation of the trackers 52a to 52c in the conventional technology will be described. Since the basic operation is the same for each of the trackers 52a to 52c, the operation of the tracker 52a will be described as an example. The operation of the tracker 52a is triggered when the searcher 51 inputs the estimated direction, ID number, and prior probability distribution of the sound source detected by the searcher 51 to the tracker 52a. The tracker 52a becomes active upon receiving this input and starts operating. The active tracker 52a only performs mask generation. Furthermore, of the estimated direction and ID number of the sound source detected by the searcher 51 and the prior probability distribution, the estimated direction and ID number in the searcher 51 are output from the tracker 52a in the processing cycle in which the tracker 52a becomes active.

[0031] FIG. 5 is a schematic configuration diagram of a tracking unit 52a in a tracking device 500 according to the related art. As shown in FIG. 5, the tracking unit 52a includes a likelihood generation unit 521, a mask processing unit 522, a Bayesian estimation unit 523, a Markov update unit 524, a tracking determination unit 525, a direction estimation unit 526, a mask generation unit 527, and an output control unit 528. Each unit of the tracking unit 52a is implemented by a dedicated device or a dedicated processing circuit. Alternatively, the tracking device 500 may include a processor such as a CPU and a memory, and the processor may execute a tracking program stored in the memory to implement each unit of the tracking unit 52a. Alternatively, each unit of the tracking unit 52a may be implemented by a combination of dedicated device or dedicated circuitry and software.

[0032] The tracker 52a receives as input the observation information and a mask output from an active tracker other than the tracker 52a. For example, when only the tracker 52a and the tracker 52b are active, the input to the tracker 52a is the observation information and the mask b generated by the tracker 52b.

[0033] The likelihood generating unit 521, like the likelihood generating unit 511 of the searching unit 51, generates a likelihood function L tracker,1 t(θ) is generated and output to the mask processing unit 522. t represents time, and the superscript "1" represents the number of the tracking unit. Here, since "tracking unit 52a" is used as an example, "1" is written as the number of "tracking unit 52a."

[0034] The mask processing unit 522 calculates the likelihood function L output from the likelihood generation unit 521. tracker,1 t (θ) and the mask output from the active tracking unit other than the tracking unit 52a are input, and the likelihood function L tracker,1 t (θ) is suppressed by suppressing the influence of sound sources already tracked by active trackers other than the tracker 52a. FIG. 6 is a diagram illustrating masking processing by the mask processing unit 522 of the tracker 52a according to the prior art. FIG. 6 shows masking processing when only the trackers 52a and 52b are active. FIG. 6(a) shows the likelihood function L before masking processing. tracker,1 t (θ), and FIG. 6(b) shows the mask b generated by the tracking unit 52b, and FIG. 6(c) shows the masked likelihood function L mask,tracker,1 t (θ).

[0035] In the example of FIG. 6, in order to suppress the influence of sound sources already tracked by active trackers other than the tracker 52a, the likelihood function L tracker,1 t By multiplying (θ) by the mask b generated by the tracking unit 52b, a weighting coefficient is applied to the masked likelihood function L, which is the likelihood that the influence of the sound source being tracked by the tracking unit 52b is suppressed. mask,tracker,1 t The mask processing unit 522 obtains the masked likelihood function L mask,tracker,1 t (θ) is output to the Bayesian estimation unit 523.

[0036] The Bayesian estimation unit 523 calculates the masked likelihood function L output from the mask processing unit 522. mask,tracker,1 t (θ) and the prior probability distribution P calculated one processing cycle ago prior,tracker,1t|t-1 (θ) is input, and the posterior probability distribution P posterior,tracker,1 t (θ) is calculated. The posterior probability distribution P posterior,tracker,1 t (θ) is output to the Markov update unit 524 , the tracking determination unit 525 , and the direction estimation unit 526 .

[0037] The Markov update unit 524 updates the posterior probability distribution P posterior,tracker,1 t Based on (θ), the prior probability distribution P at time t+1 prior,tracker,1 t+1|t (θ) is calculated. prior,tracker,1 t+1|t (θ) is output to the Bayesian estimation unit 523 after one processing cycle.

[0038] The tracking determination unit 525 calculates the posterior probability distribution P posterior,tracker,1 t Based on (θ), it is determined whether the sound source is still being tracked or has been lost, and the result flag continuous,tracker,1 t to the output control unit 528. In detail, the tracking determination unit 525 outputs the posterior probability distribution P posterior,tracker,1 t By adding the probability values ​​around the direction where (θ) is maximum, the calibration quantity TS tracker,1 t is calculated and a threshold determination is made as in the following equation (3), thereby determining whether or not the sound source is being tracked at time t.

number

[0039] where threshold continuous is the threshold for determining whether tracking continues, and flag continuous,tracker,1 t =0 indicates that detection has ceased, and flag continuous,tracker,1 t =1 indicates that detection is ongoing. Detection threshold continuousis set in advance by considering the trade-off between the continuity of the sound source and the continuity of the false alarm.

[0040] The direction estimation unit 526, like the direction estimation unit 516 of the search unit 51, calculates the posterior probability distribution P posterior,tracker,1 t Based on (θ), the estimated direction θ^ tracker,1 t and outputs it to the output control unit 528 and the mask generation unit 527.

[0041] The mask generation unit 527 generates the estimated azimuth θ̂ output from the azimuth estimation unit 526. tracker,1 t Based on this, the azimuth width θ mask , and a certain value (for example, 1) for the rest. Fig. 7 is a diagram showing an example of mask generation in the mask generation unit 527 of the tracking unit 52a according to the conventional technology. Fig. 7(a) shows the posterior probability distribution P posterior,tracker,1 t 7(b) shows a mask a generated by the mask generation unit 527. As shown in FIG. 7, the mask generation unit 527 generates a mask a having an azimuth width θ mask The weighting coefficients other than are 1, and the azimuth width θ mask A weighting factor of less than 1 will produce a mask.

[0042] Although Fig. 7 shows an example of generating a rectangular mask, any shape can be used, not just a rectangle, as long as it is a method of masking the likelihood of the sound source being tracked, such as multiplying by a small value to suppress the likelihood of only the direction close to the predicted direction being tracked. mask,tracker,1 t (θ) is output to the output control unit 528.

[0043] The output control unit 528 receives the determination result flag, which is the output of the tracking determination unit 525. continuous,tracker,1 t , the estimated direction θ^ output from the direction estimation unit 526 tracker,1 t , and the mask w output from the mask generation unit 527 mask,tracker,1t The output control unit 528 receives the determination result flag (θ). continuous,tracker,1 t When the value of θ^ is 1 (when the sound source is tracked by the tracking unit 52a), the estimated direction θ^ output from the direction estimation unit 526 is tracker,1 t , the ID number input from the search unit 51, and the w output from the mask generation unit 527. mask,tracker,1 t (θ) is output as the estimated direction, ID number, and mask a. The output control unit 528 outputs the determination result flag continuous,tracker,1 t is 0 (when the tracking unit 52a no longer detects the sound source, that is, when it is determined that detection has ceased), the estimated direction θ^ output from the direction estimation unit 526 tracker,1 t , the ID number output from the search unit 51, and the mask w output from the mask generation unit 527. mask,tracker,1 t (θ) is not output. That is, the judgment result flag continuous,tracker,1 t If is 0, there is no output from the output control unit 528.

[0044] Here, the problem to be solved by the present invention relates to the search unit 51. Here, for simplicity of explanation, an example will be explained in which all of the tracking units are not activated. Fig. 8 is a diagram showing a situation in which there is one sound source, and Fig. 9 is an image of the posterior probability distribution in the case in which there is one sound source. As shown in Fig. 8, consider a situation in which there is one sound source, and this sound source has not yet been detected. In this case, the posterior probability distribution P posterior t In (θ), as shown in FIG. 9, the maximum value p max t is the detection threshold detect When the amplitude exceeds 1 / 2, the sound source can be detected.

[0045] In the conventional tracking device 500, the search unit 51 detects sound sources one by one and passes the information to the trackers 52a to 52c to start tracking, but there is a possibility that there are multiple sound sources that have not yet been detected at the stage of the search unit 51. Fig. 10 is a diagram showing a situation when there are two sound sources, and Fig. 11 is an image of the posterior probability distribution when there are two sound sources. For example, as shown in Fig. 10, consider a situation where there are two sound sources and neither of these two sound sources has yet been detected. In such a situation, the posterior probability distribution P posterior t As shown in equation (1), (θ) is normalized to be 1 when integrated over the entire space. Therefore, as shown in Figure 11, the posterior probability for the same sound source is relatively lower than the posterior probability for a single sound source. On the other hand, the detection threshold detect Since is determined from the false alarm probability, it is constant regardless of the number of sound sources. Therefore, as shown in Figure 11, the maximum value p max t is the detection threshold detect As a result, the detection capability is reduced. A tracking device 100 according to an embodiment of the present invention that solves this problem will be described below.

[0046] Embodiment 1 Fig. 12 is a schematic configuration diagram of a tracking device 100 according to the first embodiment. The tracking device 100 is an example of a detection device of the present invention. As shown in Fig. 12, the tracking device 100 of this embodiment includes a searching unit 1 and a plurality of tracking units 2a to 2c. Although three tracking units 2a to 2c are shown in Fig. 12, the number of tracking units may be two or less, or four or more.

[0047] When tracking processing is started in the tracking device 100, observation information about sound from the sonar 200 is input to the search unit 1. The observation information is a sensor output, and is amplitude information of sound received by a microphone array or hydrophone array made up of multiple receivers provided in the sonar 200. When the observation information is input, the search unit 1 calculates a likelihood related to the direction and searches for the sound source based on the likelihood.

[0048] When the search unit 1 detects a sound source, it outputs information about the detected sound source to the trackers 2a to 2c. The information about the detected sound source is the estimated direction, ID number, and prior probability distribution of the sound source detected by the search unit 1. The trackers 2a to 2c each track a sound source and output information about the sound source being tracked. The information about the sound source being tracked is the estimated direction and ID number of the sound source being tracked by each of the trackers 2a to 2c.

[0049] At this time, in order to prevent the sound source being tracked by each of the tracking units 2a to 2c from affecting the search unit 1 and the other tracking units, each of the tracking units 2a to 2c generates a mask that suppresses the influence of the sound source being tracked and outputs it to the search unit 1 and the other tracking units. The search unit 1 and the tracking units 2a to 2c suppress the influence of the sound source being tracked by the other tracking units by assigning the mask generated by the other tracking units to the likelihood.

[0050] Next, the configuration of the searching unit 1 in this embodiment will be described. FIG. 13 is a schematic configuration diagram of the searching unit 1 according to the first embodiment. As shown in FIG. 13, the searching unit 1 includes a likelihood generating unit 11, a mask processing unit 12, a Bayesian estimation unit 13, a detection processing unit 14, a Markov updating unit 15, a direction estimation unit 16, and an ID assigning unit 17. Each unit of the searching unit 1 is realized by a dedicated device or a dedicated processing circuit. Alternatively, the tracking device 100 may include a processor such as a CPU and a memory, and the processor may execute a tracking program stored in the memory to realize each unit of the searching unit 1. Alternatively, each unit of the searching unit 1 may be realized by a combination of dedicated device or dedicated circuit and software.

[0051] The search unit 1 receives the observation information and the masks a to c output from the active tracking units 2 a to 2 c. The likelihood generation unit 11 receives the observation information as an input, similar to the likelihood generation unit 511 of the prior art, and calculates a likelihood function L t The likelihood generating unit 11 generates the likelihood function L t (θ) is output to the mask processing unit 12.

[0052] The mask processing unit 12, like the mask processing unit 512 of the prior art, calculates the likelihood function L output from the likelihood generating unit 11. t (θ) and the masks a to c output from the active tracking units 2a to 2c are input, and the likelihood function L t The mask processing unit 12 reduces the influence of the sound source already tracked by the active trackers 2a to 2c on (θ). mask t (θ) is output to the Bayesian estimation unit 13.

[0053] The Bayesian estimation unit 13, like the Bayesian estimation unit 513 of the prior art, calculates the masked likelihood function L output from the mask processing unit 12. mask t (θ) and the prior probability distribution P calculated one processing cycle ago prior t|t-1 (θ) and are input, and the posterior probability distribution P posterior t Then, the Bayesian estimation unit 13 calculates the calculated posterior probability distribution P posterior t (θ) is output to the detection processing unit 14, the Markov update unit 15, and the direction estimation unit 16.

[0054] The detection processing unit 14 calculates the posterior probability distribution P posterior t (θ) is used as input, and the presence or absence of the sound source is determined. detect t is output to the Markov update unit 15, the direction estimation unit 16, and the ID assignment unit 17. detect t =1 indicates that a new sound source has been detected, and flag detect t =0 indicates that no new sound source has been detected.

[0055] The Markov update unit 15, like the Markov update unit 515 of the prior art, detect t If is 0, the posterior probability distribution P posteriort Based on (θ), the prior probability distribution P at time t+1 prior t+1|t (θ) is calculated and output to the Bayesian estimation unit 13 after one processing cycle. detect t If is 1, the Markov update unit 15 updates the prior probability distribution P prior t+1|t (θ) is output from search unit 1.

[0056] The direction estimation unit 16 receives the determination result flag detect t When θ is 1, the estimated direction θ̂ is calculated as in the direction estimation unit 516 of the conventional art. t is calculated and output as the estimated direction. detect t If is 0, the direction estimation unit 16 does not perform the processing.

[0057] The ID assigning unit 17, like the ID assigning unit 517 of the prior art, receives the determination result flag detect t If the value is 1, the ID number is output. detect t If the value is 0, the ID assigning unit 17 does not perform the process.

[0058] Next, the detection processing unit 14 of the search unit 1 in this embodiment will be described. Fig. 14 is a schematic configuration diagram of the detection processing unit 14 of the search unit 1 according to embodiment 1. As shown in Fig. 14, the detection processing unit 14 of this embodiment has a sorting unit 141, an addend estimation unit 142, a calibration amount calculation unit 143, and a threshold determination unit 144.

[0059] The sorting unit 141 sorts the input posterior probability distribution P posterior t The sorting unit 141 sorts the probability values ​​(hereinafter also referred to as "posterior probabilities") in (θ) in ascending or descending order. posterior,sort t (i) is output to the addend estimation unit 142 and the calibration amount calculation unit 143.

[0060] The addend estimation unit 142 estimates the sequence P of the posterior probabilities sorted by the sorting unit 141. posterior,sort t Based on (i), the number of additions N^, which is the number of probability values ​​corresponding to the sound source, is estimated. The number of additions estimation unit 142 outputs the estimated number of additions N^ to the calibration amount calculation unit 143.

[0061] The calibration quantity calculation unit 143 calculates the sequence P of posterior probabilities output from the sorting unit 141. posterior,sort t (i) and the addend N^ output from the addend estimation unit 142 are input, and a sequence P of posterior probabilities rearranged according to the addend N^ is obtained. posterior,sort t By adding (i), the calibration quantity TS t The calibration quantity calculation unit 143 calculates the calculated calibration quantity TS t to the threshold determination unit 144.

[0062] The threshold determination unit 144 calculates the calibration quantity TS output from the calibration quantity calculation unit 143. t is used as input, and by threshold judgment, it is determined whether or not a sound source exists at time t, and the judgment result flag detect t The threshold value determining unit 144 of this embodiment outputs the threshold detect The object to be compared with is p max t Instead, the test quantity TS t Other than that, it is the same as the threshold value determination unit 542 of the prior art.

[0063] Next, the operation of the detection processing unit 14 will be described. First, the sorting unit 141 sorts the posterior probability distribution P posterior t Specifically, the sorting unit 141 sorts the probability values ​​of (θ) according to the magnitude relationship. sort_descend Using (x), the posterior probability distribution P posterior t The probability values ​​of (θ) are sorted in descending order, and the posterior probability sequence P posterior,sort t (i) is obtained.

number

[0064] The sorting unit 141 sorts the sequence P of posterior probabilities in descending order. posterior,sort t (i) is output to the addend estimation unit 142. i is an index indicating the sample number after rearranging in descending order. In the case of descending order, P posterior,sort t (1) has the highest posterior probability value. In addition, although the order is descending in the explanation here, it may be rearranged in ascending order if the correspondence is appropriate in the subsequent processing.

[0065] The addend estimation unit 142 estimates the sequence P of posterior probabilities output from the sorting unit 141. posterior,sort t Based on (i), the addition number N^, which is the number of probability values ​​corresponding to the sound source, is calculated. Specifically, the addition number estimation unit 142 calculates a function function est_sourceNumber The summation number N^ is calculated using (x). Here, the number of posterior probabilities corresponding to the sound source (summation number N^) is not the number of maximum values ​​of the probability value corresponding to the sound source, but the number of probability values ​​including the maximum value and its surroundings, and is calculated using a method such as AIC (Akaike's information criterion).

number

[0066] The addend estimation unit 142 outputs the estimated addend N^ to the test quantity calculation unit 143. Here, the addend estimation unit 142 uses AIC as an example for estimating the addend N^, but it is also possible to use other model selection criteria such as BIC (Bayesian Information Criterion).

[0067] The calibration quantity calculation unit 143 calculates the sequence P of posterior probabilities output from the sorting unit 141. posterior,sort t(i) and the number of additions N^ estimated by the addition number estimation unit 142 are input, and the test quantity TS t Calculate the calculated calibration metric TS t is output to the threshold value determination unit 144.

number

[0068] By operating as described above, the detection processing unit 14 of this embodiment can determine whether or not a sound source exists in a situation where multiple sound sources exist. When a sound source is detected, the direction of the sound source is estimated by the direction estimation unit 16. However, particularly when MAP estimation is applied in the direction estimation unit 16, there is a high possibility that the direction of the sound source with a higher probability value will be estimated. For such a sound source with a higher probability value, information on the estimated direction, ID number, and prior probability distribution is passed from the search unit 1 to the trackers 2a to 2c. For a sound source with a lower probability value, a mask related to the direction of the sound source with a higher probability value generated by the trackers 2a to 2c is input in the next processing cycle or later, and the search unit 1 performs processing to detect the sound source.

[0069] 15 is a diagram illustrating the detection processing in the detection processing unit 14 according to the first embodiment. As described above, first, the sorting unit 141 sorts the posterior probability distribution P posterior t (θ) is the sequence of posterior probabilities P shown in FIG. 15(b) depending on the magnitude of the probability value. posterior,sort t 15(b), the summation number estimation unit 142 estimates the summation number N^, which is the number of probability values ​​corresponding to the sound source. The verification quantity calculation unit 143 then adds up the top N^ probability values ​​among the posterior probabilities rearranged as in FIG. 15(b), and calculates the verification quantity TS t It is said that.

[0070] As described above, in the tracking device 100 of this embodiment, even when there are multiple sound sources, by estimating the summation number N^, which is the number of posterior probabilities corresponding to the sound sources, and adding the probability values, it is possible to obtain a verification amount close to that when there is only one sound source. As a result, even when there are multiple sound sources, it is possible to obtain a verification amount close to that when there is only one sound source. detect Therefore, the detection capability can be improved compared to the conventional technology.

[0071] Embodiment 2 Next, a second embodiment will be described. Fig. 16 is a diagram for explaining the problem solved by the second embodiment. The posterior probability distribution P posterior t (θ) is the sequence of posterior probabilities P posterior,sort t When rearranged as in (i), the probability distribution corresponding to each sound source has a spread, and as a result, the sorted posterior probability may change gradually, as shown in Figure 16(b). In such a situation, when the addend number estimation unit 142 separates the probability values ​​corresponding to the sound source from the probability values ​​corresponding to the noise, there is a possibility that an erroneous estimation result will be obtained for the number of probability values ​​corresponding to the sound source. If the number of probability values ​​corresponding to the sound source is estimated to be low and underestimated, signals will be missed, and the probability values ​​corresponding to the sound source cannot be sufficiently added, resulting in a decrease in detection ability. On the other hand, if the number of probability values ​​corresponding to the sound source is estimated to be high and overestimated even when a sound source does not exist, the number of false alarms will increase.

[0072] The tracking device 100 of the second embodiment solves the above problems in addition to the problems with the conventional technology. Fig. 17 is a schematic configuration diagram of a tracking device 100A according to the second embodiment. The tracking device 100A is an example of a detection device of the present invention. As shown in Fig. 17, the tracking device 100A of this embodiment includes a searching unit 1A and a plurality of tracking units 2a to 2c. Although three tracking units 2a to 2c are shown in Fig. 17, the number of tracking units may be two or less, or four or more. The configuration and operation of the tracking units 2a to 2c of this embodiment are the same as those of the tracking units 52a to 52c in the conventional technology.

[0073] The operation of the tracking device 100A of this embodiment will be described. In the tracking device 100A of this embodiment, the searching unit 1A outputs estimated directions, ID numbers, and prior probability distributions corresponding to multiple sound sources. As in the first embodiment, the searching unit 1A receives as input observation information and masks output from all active (sound source tracking) tracking units, and searches for sound sources that have not yet been detected. In conventional technology, the estimated directions, ID numbers, and prior probability distributions in the searching unit 1A always correspond to one sound source, but the searching unit 1A of this embodiment outputs ID numbers and prior probability distributions in the same number as the number of detected sound sources, corresponding to one or more sound sources detected simultaneously.

[0074] When the search unit 1A detects a sound source, it outputs the estimated direction, ID number, and prior probability distribution for each detected sound source to inactive trackers, and activates trackers equal to the number of detected sound sources. For example, we will explain a case where the search unit 1A detects two sound sources simultaneously in a situation where there is no active tracker, and activates trackers 2a and 2b.

[0075] The estimated direction, ID number, and prior probability distribution output from the search unit 1A correspond to each of the simultaneously detected sound sources. The estimated direction, ID number, and prior probability distribution corresponding to each of the simultaneously detected sound sources are input to the tracker 2a and the tracker 2b, activating the tracker 2a and the tracker 2b. The activated tracker 2a and the tracker 2b perform the same operation as the tracker described in the prior art.

[0076] Next, the configuration of the searching unit 1A in this embodiment will be described. FIG. 18 is a schematic configuration diagram of the searching unit 1A according to the second embodiment. As shown in FIG. 18, the searching unit 1A includes a likelihood generating unit 11, a mask processing unit 12, a Bayesian estimation unit 13, a detection processing unit 14A, a Markov updating unit 15, a direction estimating unit 16, and an ID assigning unit 17. Each unit of the searching unit 1A is realized by a dedicated device or a dedicated processing circuit. Alternatively, the tracking device 100A may include a processor such as a CPU and a memory, and the processor may execute a tracking program stored in the memory to realize each unit of the searching unit 1A. Alternatively, each unit of the searching unit 1A may be realized by a combination of dedicated device or dedicated circuit and software.

[0077] The operations of the likelihood generating unit 11, the mask processing unit 12, and the Bayesian estimation unit 13 in this embodiment are the same as those in the first embodiment. The detection processing unit 14A of the searching unit 1A calculates the posterior probability distribution P posterior t (θ) is used as input, and the presence or absence of the sound source is determined. detect t and the estimated number of sound sources N^ obtained in the process of determining whether a sound source exists or not. source and are output to the Markov update unit 15, the direction estimation unit 16, and the ID assignment unit 17. The detection processing unit 14A also outputs sound source peak information θ obtained in the process of determining whether or not a sound source exists. source,peak is output to the direction estimation unit 16.

[0078] The Markov update unit 15 updates the posterior probability distribution P posterior t (θ), the determination result flag which is the output of the detection processing unit 14A detect t , and the estimated number of sound sources N^ source is input, and the prior probability distribution P prior,tracker,1 t+1|t (θ) is output to the Bayesian estimation unit 13 after one processing cycle, and the estimated number of sound sources N^ sourceIn this case, all the prior probability distributions output simultaneously are the same.

[0079] The direction estimation unit 16 receives the determination result flag output from the detection processing unit 14A. detect t , the estimated number of sound sources N^ source , and sound source peak information θ source,peak is used as input, and the estimated number of sound sources is N^ source The number of sound source peak information θ source,peak Based only on the surrounding probabilities, the direction of each sound source is estimated within the same processing cycle using techniques such as MAP estimation or weighted average expected value calculation, and output as the estimated direction.

[0080] The ID assigning unit 17 receives the determination result flag detect t and the estimated number of sound sources N^ source is used as input, and the estimated number of sound sources is N^ source Output the ID numbers.

[0081] Next, the detection processing unit 14A of the search unit 1A in this embodiment will be described. Fig. 19 is a schematic configuration diagram of the detection processing unit 14A of the search unit 1A according to Embodiment 2. As shown in Fig. 19, the detection processing unit 14A of this embodiment has a peak detection unit 411, a peak sorting unit 412, a sound source number estimation unit 413, a calibration amount calculation unit 414, a sound source peak extraction unit 415, and a threshold determination unit 416.

[0082] The peak detector 411 detects the input posterior probability distribution P posterior t (θ) posterior,peak t (θ peak ) and the orientation θ corresponding to each maximum value peak and detect the maximum value P posterior,peak t (θ peak ) is the posterior probability distribution P posterior t The peak detector 411 detects the local maximum value P posterior,peakt (θ peak ) and the orientation θ corresponding to each maximum value peak is output to the peak sort unit 412.

[0083] The peak sorting unit 412 sorts the maximum value P of the posterior probability output from the peak detection unit 411. posterior,peak t (θ peak ) are sorted according to their magnitude relationship, and the maximum value of the sorted posterior probability P posterior,peak,sort t (i) is output to the sound source number estimation unit 413 and the verification amount calculation unit 414. In addition, the rearranged maximum value P posterior,peak,sort t (i) The direction θ corresponding to each maximum value peak and outputs the results to the calibration quantity calculation unit 414 and the sound source peak extraction unit 415.

[0084] The sound source number estimation unit 413 corresponds to the addition number estimation unit 142 in the detection processing unit 14 of the first embodiment. The sound source number estimation unit 413 receives a sequence P posterior,sort t Instead of (i), we use the posterior probability P posterior,peak,sort t The sound source number estimation unit 413 receives the input posterior probability P posterior,peak,sort t Based on (i), the number of sound sources N^ source is estimated and output to the sound source peak extraction unit 415, and is also output as an output of the detection processing unit 14A.

[0085] The calibration quantity calculation unit 414 calculates the posterior probability distribution P posterior t (θ) and the maximum value of the sorted posterior probability P posterior,peak,sort t (i) and the orientation θ corresponding to the rearranged maximum peak,sort (i) and the estimated number of sound sources N^ output from the sound source number estimation unit 413 source and are input, and the calibration metric TS t The calibration quantity calculation unit 414 calculates the calculated calibration quantity TSt is output to the threshold determination unit 416.

[0086] The sound source peak extraction unit 415 extracts the azimuths θ that have been sorted to correspond to the maximum values ​​output from the peak sort unit 412. peak,sort (i) and the estimated number of sound sources N^ output from the sound source number estimation unit 413 source and the flag output from the threshold determination unit 416 detect t and the sound source peak information θ source,peak Output.

[0087] The threshold determination unit 416 calculates the calibration quantity TS output from the calibration quantity calculation unit 414. t is used as input, and by threshold judgment, it is determined whether or not a sound source exists at time t, and the judgment result flag detect t The threshold value determining unit 416 of this embodiment outputs the flag detect t to the sound source peak extraction unit 415.

[0088] Next, the operation of the detection processing unit 14A of this embodiment will be described. The peak detection unit 411 of the detection processing unit 14A detects the posterior probability distribution P posterior t (θ) is input, and the function function find_peak Using (x), the maximum value of the posterior probability P is calculated as shown in the following equation (7). posterior,peak t (θ peak ) to detect.

number

[0089] θ peak is the direction corresponding to each maximum value. The peak detection unit 411 detects the maximum value P posterior,peak t (θ peak ), and the orientation θ corresponding to each maximum value peak is output to the peak sorting unit 412.

[0090] The peak sorting unit 412 uses a function function to sort the posterior probabilities in descending order according to the magnitude relationship. sort_descend (x) and find the maximum posterior probability P posterior,peak t (θ peak ) and the orientation θ corresponding to each maximum value peak Specifically, the peak sorting unit 412 sorts the maximum value P of the posterior probability as shown in the following equation (8). posterior,peak t (θ peak ) in descending order, and the sequence of maximum posterior probabilities P posterior,peak,sort t (i) and the direction θ corresponding to each maximum value peak are rearranged to correspond to the maximum value, and θ peak,sort (i)

number

[0091] Here, i is an index indicating the sample number after rearranging the maximum values ​​of the posterior probability in descending order. In the explanation here, the order is descending, but in the subsequent processing, the order may be ascending if the correspondence is appropriate. The peak sorting unit 412 sorts the sequence P of the rearranged maximum values ​​of the posterior probability. posterior,peak,sort t (i) is output to the sound source number estimation unit 413 and the calibration amount calculation unit 414, and the direction θ corresponding to each maximum value is calculated. peak,sort (i) is output to the calibration quantity calculation unit 414 and the sound source peak extraction unit 415 .

[0092] The calibration quantity calculation unit 414 calculates the sequence P of the maximum values ​​of the posterior probability output from the peak sorting unit 412. posterior,peak,sort t (i), P posterior,peak,sort t (i) corresponding direction θ peak,sort (i), and the estimated number of sound sources N^ output from the sound source number estimation unit 413 source is used as input, and the summation range parameter θ sum Using the following equation (9), the calibration metric TS t Calculate.

number

[0093] The calibration metric TS calculated using equation (9) t is output to the threshold determination unit 416. The summation range parameter θ sum is a parameter for adding each peak of the posterior probability distribution taking into account the spread in the azimuth direction, and can be set, for example, based on the Markov update in the Markov update unit 15 and a parameter for spreading the probability distribution (a parameter for adjusting how much the probability distribution spreads after one processing cycle).

[0094] The sound source peak extraction unit 415 extracts the azimuths θ that have been sorted to correspond to the maximum values ​​output from the peak sort unit 412. peak,sort (i) The estimated number of sound sources N^ output from the sound source number estimation unit 413 source , and the judgment result flag output from the threshold judgment unit 416 detect t The input is the orientation θ, which is rearranged to correspond to the maximum value. peak,sort Of (i), the first to N^ source The corresponding direction information θ peak,sort (1) to θ peak,sort (N^ source ) is the sound source peak information θ source,peak and output it as

[0095] 20 is a diagram illustrating the detection processing in the detection processing unit 14A according to the second embodiment. As described above, first, the peak detection unit 411 detects the posterior probability distribution P posterior tAfter detecting the maximum value of (θ), the peak sorting unit 412 sorts the maximum values ​​of the probability values ​​as shown in Fig. 20(b), and the sound source number estimation unit 413 estimates the number of sound sources from the sorted results. By sorting after detecting peaks in the posterior probability in this way, it is possible to clearly distinguish the maximum values ​​corresponding to sound sources from other small maximum values, and it is possible to estimate the number of sound sources without missing any. As a result, as shown in Fig. 16, the posterior probability series P posterior,sort t When rearranged as in (i), it is possible to suppress a decrease in detection ability due to the spread of the probability distribution corresponding to each sound source.

[0096] In addition, by utilizing the outputs of the peak sorting unit 412 and the sound source number estimation unit 413, the sound source peak extraction unit 415 can detect multiple sound sources in the same processing cycle and multiple tracking units can be activated simultaneously, so that sound sources can be detected earlier than with conventional technology that detects only one sound source in one processing cycle.

[0097] The above is a description of the embodiments of the present invention, but the present invention is not limited to the configurations of the above embodiments, and various modifications or combinations are possible within the scope of the technical concept. For example, in the above embodiments, the tracking device 100 and the tracking device 100A are described as examples of the detection device of the present invention, but the searching unit 1 and the searching unit 1A may be configured as independent searchers, and this may be used as the detection device. In addition, the tracking units 2a to 2b may also be configured as independent trackers.

[0098] Furthermore, in the above embodiment, an example of a sound source detection device using observation information from a sonar 200 equipped with a microphone array or a hydrophone array has been described, but it is also possible to calculate a posterior probability distribution using observation information regarding direction, frequency, speed, and distance obtained not only by sonar 200 but also by radar or lidar, and input this posterior probability distribution to the detection processing units 14 and 14A of the above embodiment.

[0099] Furthermore, in the above embodiment, it has been described that a model selection criterion is used to estimate the number of posterior probabilities corresponding to a sound source. However, for example, for a series of probability values ​​arranged in descending or ascending order, an index of change in probability values, such as a difference or a ratio, may be calculated, and a threshold may be set for the calculated index, thereby estimating the number of posterior probabilities corresponding to a sound source.

[0100] Furthermore, in the above-mentioned second embodiment, a technique for detecting a maximum value of the posterior probability and estimating the posterior probability around the maximum value was described, but the same effect can be obtained by detecting a minimum value by calculating the inverse of the posterior probability or by inverting the sign of the posterior probability. [Explanation of symbols]

[0101] 1, 1A Search unit, 2a to 2c Tracking unit, 11 Likelihood generation unit, 12 Mask processing unit, 13 Bayesian estimation unit, 14, 14A Detection processing unit, 15 Markov update unit, 16 Direction estimation unit, 17 ID assignment unit, 51 Search unit, 52a to 52c Tracking unit, 100, 100A Tracking device, 141 Sorting unit, 142 Addition number estimation unit, 143 Validation amount calculation unit, 144 Threshold judgment unit, 200 Sonar, 411 Peak detection unit, 412 Peak sorting unit, 413 Sound source number estimation unit, 414 Validation amount calculation unit, 415 Sound source peak extraction unit, 416 Threshold judgment unit, 500 Tracking device, 511 Likelihood generation unit, 512 Mask processing unit, 513 Bayesian estimation unit, 514 Detection processing unit, 515 Markov update unit, 516 direction estimation unit, 517 ID assignment unit, 521 likelihood generation unit, 522 mask processing unit, 523 Bayesian estimation unit, 524 Markov update unit, 525 tracking determination unit, 526 direction estimation unit, 527 mask generation unit, 528 output control unit, 541 maximum value search unit, 542 threshold determination unit.

Claims

1. a likelihood generation unit that generates likelihoods from observation information related to sounds; a Bayesian estimation unit that calculates a posterior probability distribution by Bayesian estimation using the likelihood; a sorting unit that sorts the probability values ​​in the posterior probability distribution; a number estimation unit that estimates the number of probability values ​​corresponding to sound sources from the sorted probability values; a calculation unit that calculates a verification amount regarding the presence or absence of the sound source based on the number of the probability values; a determination unit that determines the presence or absence of the sound source based on the calibration amount.

2. The detection device according to claim 1 , wherein the sorting unit sorts the probability values ​​in the entire space of the posterior probability distribution in descending order or ascending order.

3. a peak detection unit that detects a maximum value in the posterior probability distribution, the sorting unit sorts the maximum values ​​in descending order or ascending order, The detection device according to claim 1 , wherein the number estimation unit estimates the number of the probability values ​​corresponding to the sound source from the rearranged maximum values.

4. the calculation unit calculates the calibration amount using an addition range parameter; The detection device according to claim 3 , wherein the addition range parameter is a parameter for providing a spread to the maximum value.

5. the sorting unit sorts the maximum values ​​and the orientations corresponding to the maximum values, The detection device according to claim 3 , further comprising a sound source peak extraction unit that extracts sound source peak information including direction information of the plurality of sound sources based on the rearranged directions, the number of the probability values, and the determination result of the determination unit.

6. 6. The detection device according to claim 1, wherein the observation information is information relating to at least one of direction, frequency, distance, and speed.

7. The detection device according to any one of claims 1 to 5, wherein the number estimation unit estimates the number of the probability values ​​corresponding to the sound source using a model selection criterion including AIC or BIC.

8. a searching unit having the likelihood generating unit, the Bayesian estimating unit, the sorting unit, the number estimating unit, the calculating unit, and the determining unit; a plurality of tracking units; When the determination unit determines that the sound source is present, the search unit outputs the estimated direction, ID number, and prior probability distribution of the sound source to an inactive tracking unit among the plurality of tracking units; The detection device according to any one of claims 1 to 5, wherein the tracking unit becomes active and tracks the sound source when the estimated direction of the sound source, the ID number, and the prior probability distribution are input from the search unit.

9. generating likelihoods from observations about the sounds; a step of obtaining a posterior probability distribution by Bayesian estimation using the likelihood; reordering the probability values ​​in the posterior probability distribution; estimating the number of probability values ​​corresponding to sound sources from the sorted probability values; calculating a validation measure for the presence or absence of the sound source based on the number of probability values; and determining the presence or absence of the sound source based on the calibration quantity.

10. A detection program that causes a processor of a detection device to execute the detection method according to claim 9.

Images