JP7725938B2 - Tracking device, tracking system, tracking method, and tracking method program - Google Patents

Description

This technology relates to a tracking device, tracking system, tracking method, and tracking method program that tracks a tracking target based on a signal. In particular, it relates to tracking multiple tracking targets.

Tracking devices are known that, in a noisy environment, estimate the state quantity indicating the presence of a tracked object in a state quantity space based on signals emitted by the object, and then track the object based on the estimate. For example, tracking devices use sound waves emitted by the tracked object as signals and perform tracking by estimating the direction of arrival of the sound waves and the time-varying frequency of the sound waves. The state quantity indicating the presence of a tracked object corresponds to the direction of arrival of the sound waves and the frequency of the sound waves (see, for example, Non-Patent Document 1).

R. E. Bethel and G. J. Paras, “A PDF Multitarget Tracker”, IEEE Transactions on Aerospace and Electronic Systems, vol. 30, NO. 2, April 1994, p. 386-403

When there are multiple tracking targets, multiple trackers track each target. However, if the orientations of the multiple tracking targets are close to each other, each tracker cannot automatically determine which tracking target the estimated orientation obtained at a given time corresponds to. As a result, multiple trackers tracking different tracking targets may end up estimating the orientation of the same tracking target and tracking it after the multiple tracking targets cross each other in terms of orientation.

Therefore, there was a desire to develop a tracking device that could accurately track multiple tracking targets.

The tracking device disclosed herein comprises a data processing device having a likelihood generation unit that calculates likelihoods in a plurality of directions from observation data obtained by observing signals related to a plurality of tracking targets and performs processing to generate a likelihood function with the direction as a parameter variable , and a plurality of tracking units that determine the presence or absence and direction of a tracking target, respectively, for one of the plurality of tracking targets based on the likelihood function, and each tracking unit includes an orientation change rate estimation unit that performs orientation change rate estimation processing to estimate an orientation change rate in the direction of the corresponding tracking target and calculates an orientation change rate estimated value, and a tracking unit that calculates an orientation change rate estimated value based on the orientation change rate estimated value. a weighting unit that generates a weighted likelihood function by superimposing the orientation change rate weight on a likelihood function; a Bayesian estimation calculation unit that performs processing to generate a posterior probability distribution from the weighted likelihood function; a Markov update unit that performs processing to generate a priori probability distribution; a detection and determination unit that performs processing to determine the presence or absence of a tracking target based on the posterior probability distribution and detect the tracking target; an ID assignment unit that assigns an ID number to the tracking target detected by the detection and determination unit; and an orientation estimation unit that estimates the orientation of the tracking target based on the posterior probability distribution and calculates the estimated orientation.

The tracking system disclosed herein also includes a receiver array having multiple receivers and a signal conversion device that converts signals received by the receivers into digital data to generate observation value data, and the tracking device described above.

Furthermore, the tracking method disclosed herein is a tracking method for tracking each tracking target based on a likelihood function with orientation as a parameter variable, which is generated from likelihoods related to a plurality of orientations calculated from observation value data obtained by observing signals related to a plurality of tracking targets, and includes the steps of: calculating an orientation change rate estimate value by performing an orientation change rate estimation process to estimate an orientation change rate in the orientation of the corresponding tracking target; generating an orientation change rate weight that performs weighting based on the orientation change rate based on the orientation change rate estimate value; generating a weighted likelihood function by superimposing the orientation change rate weight on the likelihood function; generating a posterior probability distribution from the weighted likelihood function; generating a priori probability distribution; determining the presence or absence of the tracking target based on the posterior probability distribution and detecting the tracking target; assigning an ID number to the detected tracking target; and estimating the orientation of the tracking target based on the posterior probability distribution to calculate the estimated orientation.

Furthermore, the tracking method program disclosed herein is a tracking method program that tracks each tracking target based on a likelihood function with orientation as a parameter variable, generated from likelihoods related to multiple orientations calculated from observation value data obtained by observing signals related to multiple tracking targets, and causes a computer to perform the following steps: calculating an orientation change rate estimate value by performing orientation change rate estimation processing to estimate an orientation change rate in the orientation of the corresponding tracking target; generating an orientation change rate weight that performs weighting based on the orientation change rate based on the orientation change rate estimate value; generating a weighted likelihood function by superimposing the orientation change rate weight on the likelihood function; generating a posterior probability distribution from the weighted likelihood function; generating a priori probability distribution; determining the presence or absence of a tracking target and detecting the tracking target based on the posterior probability distribution; assigning an ID number to the detected tracking target; and estimating the orientation of the tracking target based on the posterior probability distribution to calculate the estimated orientation.

According to this disclosure, the weight generation unit generates an orientation change rate weight based on the orientation change rate estimated value calculated by the orientation change rate estimation unit. The weighting unit then superimposes the orientation change rate weight on a likelihood function, and the Bayesian estimation calculation unit calculates a posterior probability distribution based on the weighted likelihood function weighted based on the orientation change rate related to the orientation change, and the detection and determination unit and orientation estimation unit detect the tracking target and estimate the orientation. This makes it possible to calculate a probability distribution that includes orientation changes, enabling multiple tracking targets to be accurately tracked.

1 is a diagram showing the configuration of a tracking system centered around a tracking device 100 according to a first embodiment. FIG. 2 is a diagram showing the configuration of a tracking unit 500 according to the first embodiment. 5A and 5B are diagrams showing examples of masks used in mask processing by a mask processing unit 501 according to the first embodiment. 10 is a diagram illustrating weight processing by a weight generating unit 510 according to the first embodiment. FIG. FIG. 2 is a diagram illustrating an example of a processing flow in the data processing device 110 according to the first embodiment. FIG. 1 illustrates a situation where the azimuths of two targets intersect. FIG. 10 is a diagram showing an example of time transition of the azimuth of a target tracked by each tracking unit 500 when multiple targets intersect. FIG. 10 is a diagram showing the mask likelihood function l ^m,i _t (θ) in each tracking unit 500 when erroneous tracking occurs. 10A to 10C are diagrams illustrating the effects according to the first embodiment. FIG. 10 is a diagram showing the configuration of a tracking unit 500 according to a second embodiment. FIG. 10 is a diagram showing the configuration of a tracking unit 500 according to a third embodiment. FIG. 11 is a diagram illustrating the processing of a peak likelihood extractor 512 in the third embodiment. FIG. 11 is a diagram illustrating the processing of the applied likelihood recalculation unit 513 in the third embodiment. FIG. 11 is a diagram illustrating a regenerated likelihood function processed by a likelihood function regeneration unit 514 in the third embodiment. 10A and 10B are diagrams illustrating the effects of the tracking device 100 according to the third embodiment. FIG. 10 is a diagram illustrating a configuration of a tracking unit 500 according to a fourth embodiment. 10A and 10B are diagrams illustrating the effects of the tracking device 100 according to the fourth embodiment.

The following describes tracking devices and other devices according to the embodiments, with reference to the drawings. In the following drawings, items with the same reference numerals are the same or equivalent, and this applies throughout the entire description of the embodiments described below. Furthermore, the sizing relationships between components in the drawings may differ from those in reality. Furthermore, the configurations of components shown throughout the specification are merely examples and are not limited to the configurations described in the specification. Not all of the devices described in the specification may be included. In particular, the combinations of components are not limited to the combinations in each embodiment, and components described in other embodiments may be applied to other embodiments. Furthermore, when multiple similar devices are distinguished by subscripts, the reference numerals, subscripts, etc. may be omitted if there is no need to distinguish or identify them.

Embodiment 1.
FIG. 1 is a diagram showing the configuration of a tracking system centered on a tracking device 100 according to a first embodiment. In explaining the tracking system according to the first embodiment, a sound wave signal generated by sound waves will be used as an example of a signal used to track a target to be tracked. Hereinafter, the tracking target that serves as a sound source will be referred to as a target. The sound wave signal includes sound waves that serve as acoustic signals obtained from the target as well as sound waves that serve as noise. Here, the tracking system according to the first embodiment will be explained assuming a case in which two targets A and B that serve as sound sources are tracked.

The tracking system in embodiment 1 includes a tracking device 100, a microphone array 200, and a communication line 300. As shown in FIG. 1, the microphone array 200 is a receiver array in which multiple microphones 210-1 to 210-n, which convert received sound signals into electrical signals, are arranged in any shape, such as a line. The microphone array 200 also includes a signal conversion device 220 and a signal transmission device 230. The signal conversion device 220 periodically performs sampling and encoding processes on the sound signals received by each microphone 210, converting them into observation value data, which is digitalized data to be processed, and performs data generation processing. The signal transmission device 230 transmits a signal including the observation value data.

The communication line 300 is a line that connects the microphone array 200 and the tracking device 100 for communication. Here, the communication line 300 is assumed to be a dedicated line for communication, but is not limited to this. For example, signals may be sent from the microphone array 200 to the tracking device 100 via a commonly used telecommunications line. The communication line 300 may also be wired or wireless.

The tracking device 100 is a device that performs processing related to target tracking based on observation data. The tracking device 100 in embodiment 1 has a data processing device 110, a storage device 120, a signal input device 130, and a signal output device 140. The signal input device 130 converts the data in the signal sent from the microphone array 200 via the communication line 300 into observation data that can be processed by the data processing device 110. The signal output device 140 generates a signal including data from the data processing and detection judgment results, and transmits the signal to an external device (not shown). Here, in embodiment 1, the tracking device 100 has the signal input device 130 and the signal output device 140, but they may also be included in a device other than the tracking device 100.

Furthermore, the storage device 120 provided in the tracking device 100 in embodiment 1 temporarily or long-term stores data used for processing by the data processing device 110. Here, as will be described later, the data processing device 110 has multiple tracking units 500. Here, the storage device 120 collectively stores data, parameters, etc. for the multiple tracking units 500. However, this is not limited to this. An independent storage device 120 may be provided for each tracking unit 500. Here, the values related to the data, parameters, etc. required for each tracking unit 500 to perform its respective processing may differ.

The data processing device 110 included in the tracking device 100 of embodiment 1 generates data on estimated azimuths θ ^* of multiple targets and ID numbers assigned to each target based on observation value data related to acoustic signals. The data processing device 110 has a likelihood generation unit 400 and multiple tracking units 500. In FIG. 1, the data processing device 110 has two tracking units 500-1 and 500-2. Each tracking unit 500 basically has the same configuration and performs the same processing. Each unit within the data processing device 110 performs processing such as calculation and judgment, and generates various data for processing by the data processing device 110.

Here, the data processing device 110 of the tracking device 100 is typically configured as a device that performs control and arithmetic processing, such as a computer centered around a CPU (Central Processing Unit). The data processing device 110 typically executes pre-programmed procedures for the signal processing methods performed by each unit to realize the processing of each unit. Here, for example, the storage device 120 stores the program data. However, this is not limited to this, and each unit may also be configured as a separate dedicated device (hardware).

The storage device 120 also has a volatile storage device (not shown) such as random access memory (RAM) that can temporarily store data, and a non-volatile auxiliary storage device (not shown) such as a hard disk drive. The non-volatile auxiliary storage device may be a solid state drive, a flash memory that can store data long-term, or the like. Here, the tracking device 100 may be configured such that the hardware that executes the processing of each section of the data processing device 110 and the storage device 120 is located in different locations, and the sections communicate with each other to perform the processing operations of the tracking device 100. The tracking unit 500 may also be an independent device that functions as a tracker.

The likelihood generation unit 400 generates likelihoods for multiple directions based on observation value data related to the sound wave signal received by the microphone array 200, and calculates a likelihood function l _t (θ) with the direction θ as a parameter variable. Here, t represents time (t = 1, 2, ...), and the period from time t to time t + 1 is the processing cycle of the data processing device 110. The tracking device 100 of embodiment 1 will be described for processing observation value data at a certain time t. The likelihood generation unit 400 forms directivity for multiple directions simultaneously, for example, by beamforming. The likelihood generation unit 400 then arranges and processes the sound power values in each direction, and defines the obtained data as the likelihood function l _t (θ). Here, the likelihood generation unit 400 generates data of the likelihood function l _t (θ) using beamforming, but this is not limited to this. Other techniques for generating the likelihood function l _t (θ) can be used, for example, by assuming a distribution of target azimuths that are sound sources and calculating the likelihood function l _t (θ) around the target's estimated instantaneous azimuth θ ^* .

2 is a diagram showing the configuration of tracking units 500 according to embodiment 1. Each tracking unit 500 generates an estimated azimuth θ ^* and an ID number of the corresponding target as data based on the likelihood function l _t (θ) generated by likelihood generation unit 400. Each tracking unit 500 according to embodiment 1 includes a mask processing unit 501, a weighting unit 502, a Bayesian estimation calculation unit 503, a Markov update unit 507, a mask generation unit 508, a detection determination unit 504, an ID assignment unit 505, an azimuth change rate estimation unit 509, a weight generation unit 510, and an azimuth estimation unit 506.

FIG. 3 is a diagram showing an example of a mask used in mask processing by the mask processing unit 501 in embodiment 1. The mask processing unit 501 performs mask processing to superimpose a mask such as that shown in FIG. 3( a) generated by a mask generation unit 508 (described later) on the likelihood function l _t (θ) generated by the likelihood generation unit 400, thereby generating a masked likelihood function l ^m,i _t (θ). Here, i is a number that distinguishes the tracking unit 500. The mask processing is processing to suppress likelihoods other than those in a specific range (width θmask) of directions for the likelihood function l _t (θ) such as that shown in FIG. 3( b), as shown in FIG. 3( c). Here, the mask processing unit 501 performs processing using a mask that is stored in a mask storage unit 522 (described later) and that is generated by the mask generation unit 508 (described later) at time t-1, which is one period before time t. If the initial value of the mask used by the mask processing unit 501 during processing at time t=1 is not set in advance, the same value, such as 1, is superimposed on all directions so that the relationship between the direction and the likelihood does not change due to the mask. On the other hand, if the direction, etc. are set in advance, the initial value of the mask is set to a value based on the mask set in advance. The mask likelihood function l ^m,i _t (θ) is used as data in the processing performed by the weighting unit 502.

The weighting unit 502 performs weighting processing to superimpose an orientation change rate weight _wt (θ') generated by a weight generation unit 510 (described later) on the masked likelihood function lm ^,it ₍ θ) generated by the mask processing unit 501 through masking processing, thereby generating a weighted likelihood function Lw ^, _it (θ,θ'). The weighted likelihood function Lw ^, _it (θ,θ') is used as data in the processing performed by the Bayesian estimation calculation unit 503. Here, θ' represents the orientation change rate. The weighted likelihood function Lw ^, _it (θ,θ') is a vector with the orientation θ and the orientation change rate θ' as parameters (variables).

The Bayesian inference calculation unit 503 performs arithmetic processing based on the weighted likelihood function L ^w,i _t (θ, θ') related to the processing by the weighting unit 502 and the prior probability distribution P ^prior,i _t|t-1 (θ, θ') related to the processing by the Markov updating unit 507 described later, to calculate a posterior probability distribution P ^posterior,i _t (θ, θ'). Furthermore, the Bayesian inference calculation unit 503 calculates an orientation posterior probability distribution P ^posterior,i _t (θ) by adding the posterior probability distribution P ^posterior,i _t (θ, θ') to the orientation change rate direction for each orientation θ, for target determination and orientation estimation. The posterior probability distribution P ^posterior,i _t (θ, θ') generated by the Bayesian inference calculation unit 503 is used as data in the processing performed by the Markov updating unit 507. Furthermore, the posterior probability distribution P ^posterior,i _t (θ) of the direction generated by the Bayesian estimation calculation unit 503 is used as data in the processing performed by the direction estimation unit 506 and the detection determination unit 504 .

The prior probability distribution P ^prior,i _t|t-1 (θ, θ') used by the Bayesian estimation calculation unit 503 during calculation processing is the prior probability distribution P ^prior,i _t|t-1 (θ, θ') at one period before time t (time t-1) stored in the prior probability buffer unit 521 described later. Here, the initial value of the prior probability distribution P prior ^,i _1|0 (θ) used by the Bayesian estimation calculation unit 503 during processing at time t=1 is generally set to a value that results in equal probability, 1/Ng, for all target orientations θ and orientation change rates θ', when Ng is the number of intervals in the state space in the prior probability distribution. On the other hand, if set in advance, the initial value of the prior probability distribution P ^prior,i _t|t-1 (θ, θ') is set to a value that results in an arbitrary distribution based on the prior setting.

The detection determination unit 504 performs a determination process to determine the presence or absence of a target that is a sound source based on the posterior probability distribution P ^posterior,i _t (θ, θ′) calculated by the Bayesian estimation calculation unit 503 .

The ID assignment unit 505 performs a process of assigning an ID number to a target determined by the detection and determination unit 504. Here, while the detection and determination unit 504 continuously determines that a target is present, the ID assignment unit 505 continues to assign the same ID number. Furthermore, if the detection and determination unit 504, which had previously determined that no target was present, determines that a target is present, the ID assignment unit 505 assigns a new ID number. The ID number may be any identifiable number, such as a number, letter, or symbol.

The Markov update unit 507 calculates a prior probability distribution P ^prior,i _{t+1|t(θ, θ') at time t based on the posterior probability distribution P posterior,i t(θ, θ') calculated by the Bayesian estimation calculation unit 503. The prior probability distribution P prior,i t+1 |} ^{t (} _θ , θ') calculated by the Markov update unit 507 is used in the processing of the mask generation unit 508 and the orientation change rate estimation unit 509. In addition, the prior probability distribution P ^prior,i _t+1|t (θ, θ') is stored in a prior probability buffer unit 521, which will be described later. The prior probability distribution P ^prior,i _t+1|t (θ, θ') stored in the prior probability buffer unit 521 is used in the Bayesian estimation calculation unit 503 as the prior probability distribution P ^prior,i _t|t-1 (θ, θ') one period before (time t) in processing one period after time t (time t+1).

The mask generation unit 508 estimates the target's azimuth θ at time t+1 by calculating the expected value of the azimuth θ using a weighted average and estimating the azimuth θ _t+1|t that maximizes the prior probability. Based on the estimated azimuth θ _t+1|t , the mask generation unit 508 then generates a mask in which a certain value (e.g., 1) is set for a width θmask based on the azimuth at time t+1, and a value lower than the certain value (e.g., 0) is set for other azimuths. The mask data generated by the mask generation unit 508 is stored in a mask storage unit 522 (described later) and is used for mask processing performed by the mask processing unit 501 in the next cycle.

The azimuth change rate estimation unit 509 performs azimuth change rate estimation processing to calculate an azimuth change rate estimated value θ' ^* by estimating the azimuth change rate θ' for the target of interest.

The weight generation unit 510 calculates the heading change rate weight w t (θ') to be used in the processing of the next period (time t+1) based on the heading change rate estimated value _θ ' ^* processed by the heading change rate estimation unit 509 and the set adjustment parameter σ ² _θset stored in the adjustment parameter storage unit 524. The data of the heading change rate weight w _t (θ') generated by the weight generation unit 510 is stored in the weight storage unit 523 and is used in the weighting processing performed by the weight assignment unit 502 in the next period.

Then, the direction estimation unit 506 performs direction estimation processing to calculate an estimated direction θ* that estimates the direction of the sound source (target) from which the acoustic signal is emitted, based on the posterior probability distribution P ^posterior,i _t (θ, θ ^' ) generated by the Bayesian estimation calculation unit 503. Here, the direction estimation processing performed by the direction estimation unit 506 is processing performed using techniques such as MAP estimation and weighted averaging.

In addition, data required for processing by the tracking unit 500 is stored in the storage device 120. Here, it is assumed that the corresponding data is stored in the prior probability buffer unit 521, mask storage unit 522, weight storage unit 523, and adjustment parameter storage unit 524 in particular. Here, the prior probability buffer unit 521, mask storage unit 522, weight storage unit 523, and adjustment parameter storage unit 524 are included in the storage device 120, but for convenience of explanation, they will be described below as part of the tracking unit 500.

The prior probability buffer unit 521 stores one cycle of the prior probability distribution P ^prior,i _t+1|t (θ) at time t calculated by the Markov update unit 507. As described above, the prior probability distribution P ^prior,i _t+1|t (θ, θ') stored in the prior probability buffer unit 521 is used as the prior probability distribution P ^prior,i _t|t-1 (θ, θ') in the processing of the tracking device 100 at time t+1.

The mask memory unit 522 stores the mask generated by the mask generation unit 508 for one cycle. The mask stored in the mask memory unit 522 is used for mask processing by the mask processing unit 501 at time t+1.

The weight storage unit 523 stores the heading change rate weight w _t (θ′) for one period generated by the weight generation unit 510. The heading change rate weight w _t (θ′) stored in the weight storage unit 523 is used in the processing of the weight assignment unit 502 at time t+1.

The adjustment parameter storage unit 524 also stores a preset setting adjustment parameter σ ² _θset as data. The setting adjustment parameter σ ² _θset is a parameter that adjusts the spread in the direction of the azimuth change rate θ' of the azimuth change rate weight w _t (θ') obtained by the weight generation unit 510's processing one cycle before. Here, if the setting adjustment parameter σ ² _θset is set to a small value, it becomes easier to distinguish between the azimuth change rates θ' in the azimuth where two targets intersect. However, the estimation accuracy of the azimuth change rate θ' of each target decreases. Therefore, the value of the setting adjustment parameter σ ² _θset is set in advance, taking into account the trade-off between the azimuth change rate θ' and the estimation accuracy of the azimuth change rate θ'. For example, if the ability to distinguish between the azimuth change rate θ' of two targets is emphasized, the setting adjustment parameter σ ² θset is set to the step size of the azimuth change rate θ' in the prior probability distribution P ^prior,i _t|t-1 (θ, θ') and the posterior probability distribution P ^posterior,i _t (θ, _θ') .

Next, the processing of the data processing device 110 included in the tracking device 100 in embodiment 1 will be described. Here, the processing of the tracking unit 500 at time t will be mainly described. Here, the initial values of the mask used by the mask processing unit 501 are different to track different targets, but the same processing is performed in each tracking unit 500. First, the processing in the azimuth change rate estimation unit 509 and weight generation unit 510 at time t-1 used at time t will be described.

The orientation change rate estimation unit 509 calculates an orientation change rate prior probability distribution P ^prior,i _t|t-1 (θ') related to the orientation change rate θ' by adding, in the orientation direction, the prior probability distribution P ^prior,i t|t _- 1 (θ, θ') obtained by the processing of the Markov update unit 507. Then, the orientation change rate estimation unit 509 performs estimation processing by calculating the expected value of the orientation change rate θ' using a weighted average for the orientation change rate prior probability distribution P ^prior,i _t|t-1 (θ') or estimation processing of the orientation change rate θ' that maximizes the prior probability, thereby calculating an orientation change rate estimated value θ' ^* at time t. Here, the orientation change rate estimation unit 509 calculated the orientation change rate prior probability distribution P ^prior,i _t|t-1 (θ') to calculate the orientation change rate estimated value θ' ^* , but the present invention is not limited to this. With respect to the prior probability distribution P ^prior,i _t|t-1 (θ, θ'), the orientation change rate estimated value θ'* may be calculated by performing an estimation process in the space of the orientation θ and the orientation change rate θ' by calculating the expected value of the orientation change rate θ' using a weighted average, or by performing an estimation process for the orientation change rate θ' ^that maximizes the prior probability.

4 is a diagram illustrating weighting processing by the weight generation unit 510 according to embodiment 1. The weight generation unit 510 performs weighting processing to calculate the orientation change rate weight w t (θ') expressed by equation (1) based on the orientation change rate θ' processed by the orientation change rate estimation unit 509 and the set adjustment parameter σ ² _θset stored in the adjustment parameter storage unit 524. By performing calculations based on equation (1), the weight generation unit 510 generates orientation change rate weights _{w t} (θ') whose values vary depending on the orientation change rate θ', as shown in FIG.

5 is a diagram illustrating an example of the processing flow in the data processing device 110 according to embodiment 1. The following mainly describes the processing of each tracking unit 500 based on the likelihood function l _t (θ) generated by the likelihood generating unit 400. As described above, the likelihood generating unit 400 generates likelihoods in multiple directions based on the observation value data related to the acoustic signal received by the microphone array 200, and calculates and generates the likelihood function l _t (θ) (step S0).

The mask processing unit 501 generates a masked likelihood function l ^m,i _t (θ) based on the likelihood function l _t (θ) generated by the likelihood generating unit 400 and the mask stored in the mask storage unit 522 (step S1).

The weighting unit 502 performs a calculation based on equation (2) using the masked likelihood function l ^m,i _t (θ) generated by the mask processing unit 501 through masking processing and the orientation change rate weight w _t (θ') generated by the weight generation unit 510, and generates a weighted likelihood function L ^w,i _t (θ, θ') (step S2).

Here, the orientation change rate weight w _t (θ') is a vector related to the orientation change rate θ'. Also, the mask likelihood function l ^m,i _t (θ) is a vector related to the orientation θ. In equation (2), the result of multiplying the combinations of all vector elements related to the orientation change rate θ' and the orientation θ is the weight likelihood function L ^w,i _t (θ, θ'), which is a matrix.

The Bayesian inference calculation unit 503 calculates the posterior probability distribution P ^{posterior, i} _t (θ, θ') according to equation (3) using the weighted likelihood function L w,i t (θ, θ') related to the processing by the weighting unit 502 and the prior probability distribution P ^prior ,i _{t |t-1} (θ, θ') stored in the prior probability buffer unit 521. Furthermore, the Bayesian inference calculation unit 503 calculates the orientation posterior probability distribution P ^posterior,i _t (θ) by adding the posterior probability distribution P ^posterior,i _t (θ, θ') to the orientation change rate direction for each orientation θ according to equation (4) (step S3).

The detection determination unit 504 performs a determination process to determine whether or not a target that is a sound source exists (step S4) based on the directional posterior probability distribution P ^posterior,i _t (θ) obtained by the processing of the Bayesian estimation calculation unit 503. Then, the ID assignment unit 505 performs a process to assign an ID number to the target determined by the detection determination unit 504 (step S5).

Furthermore, the direction estimation unit 506 generates an estimated direction θ* that estimates the direction of the sound source (target) from which the acoustic signal is emitted, based on the direction posterior probability distribution P ^posterior,i _t (θ) obtained by processing by the Bayesian estimation calculation unit ⁵⁰³ (step S6).

Here, the detection/determination unit 504 and the orientation estimation unit 506 performed processing based on the orientation posterior probability distribution P ^posterior,i _t (θ) calculated by the Bayesian estimation calculation unit 503 from the posterior probability distribution P posterior, ⁱ t(θ,θ'), but this is not limiting. The detection/determination unit 504 and the orientation estimation unit 506 may perform processing by detecting a peak in the space of the orientation θ and the orientation change rate θ' for the posterior probability distribution P ^posterior,i _t (θ,θ') calculated by the Bayesian estimation calculation unit 503.

Next, the processing of the Markov update unit 507 will be described. The Markov update unit 507 calculates and generates a prior probability distribution P ^prior ,i _{t +1|t} (θ, θ′) at time t based on the posterior probability distribution P ^posterior,i t(θ, θ′) obtained by the processing of the Bayesian estimation calculation unit 503 (step S7). The prior probability distribution P ^prior,i _t+1|t (θ, θ′) is stored in the prior probability buffer unit 521.

Furthermore, as described above, the orientation change rate estimation unit 509 adds the prior probability distribution P ^prior,i _t+1|t (θ, θ') obtained by processing of the Markov update unit 507 in the orientation direction to calculate the orientation change rate prior probability distribution P ^prior,i _t+1|t (θ') related to the orientation change rate θ' (step S8).

Then, the weight generation unit 510 performs weighting processing based on the orientation change rate θ′ processed by the orientation change rate estimation unit 509 and the set adjustment parameter σ ² _θset stored in the adjustment parameter storage unit 524, and generates the orientation change rate weight w _t (θ′) (step S9).

Figure 6 shows a situation where the orientations of two targets intersect. As shown in Figure 6, when viewed from the position of the microphone array 200, the two targets that serve as sound sources are moving in opposite directions.

7 is a diagram showing an example of the transition over time of the azimuth of a target tracked by each tracking unit 500 when multiple targets intersect. FIG. 7(a) shows the estimated azimuth θ ^* of the target by each tracking unit 500 when the target is being correctly tracked. FIG. 7(b) shows the estimated azimuth θ ^* of the target by each tracking unit 500 when erroneous tracking occurs. In FIG. 7, (a) shows a situation in which targets A and B, which are separated in different azimuths, gradually approach each other. (b) shows a situation in which the two targets pass each other in terms of azimuth, causing the azimuths of the two targets to overlap. (c) shows a situation in which the two targets gradually move apart after passing each other.

FIG. 8 is a diagram showing the mask likelihood function l ^m,i _t (θ) in each tracker 500 when erroneous tracking occurs. FIG. 8(a) shows the mask likelihood function l ^m,i _t (θ) in tracker 500-1 and tracker 500-2 in situation a in FIG. 7(a). FIG. 8(b) shows the mask likelihood function l ^m,i _t (θ) in tracker 500-1 and tracker 500-2 in situation b in FIG. 7(a). FIG. 8(c) shows the mask likelihood function l ^m,i _t (θ) in tracker 500-1 and tracker 500-2 in situation c in FIG. 7(a). As shown in FIG. 8(a), when the targets are far apart, the mask likelihood function l ^m,i _t (θ) has a peak likelihood orientation for each target that is different. On the other hand, the mask likelihood function l ^m,i _t (θ) in FIG. 8(b) contains the likelihoods of two targets, and the likelihood peak cannot be reduced to one.

When Bayesian estimation calculation unit 503 performs processing based on mask likelihood function l ^m,i _t (θ) in FIG. 8B, the peak in the posterior probability distribution in tracking unit 500-2 will be at the position of target A. The prior probability distribution calculated by Markov update unit 507, which performs processing using the posterior probability distribution, will also have a peak at target A. As a result, as shown in FIG. 7B, after the targets cross, tracking unit 500-2 will begin tracking target A, the same as tracking unit 500-1, rather than target B.

9 is a diagram illustrating the effect according to embodiment 1. According to tracking device 100 of embodiment 1, weight generation unit 510 calculates orientation change rate weight w t (θ') from orientation change rate estimated value θ' ^* obtained by orientation change rate estimation unit 509. Then, weighting unit 502 generates weighted likelihood function L ^w,i _t (θ, θ') weighted based on orientation change rate weight _{w t} (θ').

As shown in Figure 9(a), in the situation a shown in Figure 7, target A changes its azimuth in the positive direction. Therefore, the weighted likelihood function L ^w,i _t (θ, θ') in tracking unit 500-1, which is tracking target A, has a high likelihood for a positive azimuth change rate θ'. On the other hand, target B changes its azimuth in the negative direction. Therefore, the weighted likelihood function L ^w,i _t (θ, θ') in tracking unit 500-2, which is tracking target B, has a high likelihood for a negative azimuth change rate θ'.

9(b), in the situation shown in b in FIG. 7, both the tracking units 500-1 and 500-2 generate weighted likelihood functions L ^{w,i t (θ, θ') in which the acoustic signals of targets A and B are combined with respect to the direction θ for each direction change rate θ'. As a result, the posterior probability distributions P posterior,i} _t ⁽ _θ , θ') processed by the Bayesian estimation calculation unit 503 in each tracking unit 500 also reflect the direction change rates θ' for targets A and B.

7, the azimuths θ of targets A and B become increasingly separated over time. Therefore, as shown in FIG. 9C, the weighted likelihood function L ^w,i _t (θ, θ') in each tracking unit 500 reflects the azimuth θ and azimuth change rate θ' of the target being tracked. Therefore, tracking device 100 of the tracking system in embodiment 1 can accurately track multiple targets even when the multiple targets intersect.

Embodiment 2.
Fig. 10 is a diagram showing the configuration of tracking unit 500 according to embodiment 2. In Fig. 10, mask processing unit 501, weighting unit 502, Bayesian estimation calculation unit 503, Markov update unit 507, mask generation unit 508, detection determination unit 504, ID assignment unit 505, and orientation change rate estimation unit 509 perform the same processing as that described in embodiment 1.

Here, in the tracking device 100 of the second embodiment, the estimated direction θ ^{* obtained by the direction estimation unit 506 through estimation processing is not only output as a signal but also stored as data in the estimated direction buffer unit 526 described later. Furthermore, the weight generation unit 510 calculates the direction change rate weight w t (θ')} to be used in processing in the next period (time _t +1) based on the error standard deviation σθ obtained through processing by the error standard deviation calculation unit 511 described later and the adjustment parameter k stored in the adjustment parameter storage unit 524. In the second embodiment, the adjustment parameter k stored in the adjustment parameter storage unit 524 is a parameter that adjusts the degree to which the error standard deviation σθ related to the direction change rate θ' is reflected in the spread of the direction change rate weight w _t (θ').

Also, as shown in FIG. 10 , in embodiment 2, each tracking unit 500 of the tracking device 100 has an error standard deviation calculation unit 511. Furthermore, the tracking device 100 of embodiment 2 has an estimated azimuth change rate buffer unit 525 and an estimated azimuth buffer unit 526 in the storage device 120.

The estimated heading change rate buffer unit 525 stores, as data, one cycle of the heading change rate θ' obtained through processing by the heading change rate estimation unit 509. Here, the estimated heading change rate buffer unit 525 stores, as data, the heading change rate θ' for T ^θ cycles from time t-T ^θ+1 to time t.

The estimated direction buffer unit 526 also stores as data the estimated direction θ ^* obtained by the direction estimation process of the direction estimation unit 506. Here, the estimated direction buffer unit 526 stores as data the estimated direction θ ^* for T ^θ+1 periods from time t-T ^θ to time t.

The error standard deviation calculation unit 511 performs calculation processing based on equations (5) and (6) using the estimated heading θ ^* stored in the estimated heading buffer unit 526 and the heading change rate θ' stored in the estimated heading change rate buffer unit 525 to calculate the error standard deviation σθ' ^* related to the heading change rate θ'. Here, ΔT _proc in equation (6) is the time period for calculating the estimated heading θ ^* . The error standard deviation calculation unit 511 calculates the heading change rate θ', which is the magnitude of the heading change per unit time, using equation (6).

The weight generation unit 510 performs calculation processing based on equation (7) to generate the heading change rate weight w _t (θ'). The data of the heading change rate weight w _t (θ') is stored in the weight storage unit 523 and is used as the heading change rate weight w _t (θ') in the weighting processing performed by the weighting unit 502 in the next period.

For example, in the tracking device 100 of the first embodiment, the weight generation unit 510 determines the spread of the orientation change rate weight w _t (θ') related to the orientation change rate θ' using the set adjustment parameter σ ² _θset stored in the adjustment parameter storage unit 524. Here, if the set adjustment parameter σ ² _θset is set to reduce the spread of the orientation change rate weight w _t (θ'), and if the orientation change rate estimated value θ' ^* differs significantly from the actual orientation change rate θ' of the target during processing by the tracking device 100 at a certain time, the orientation change rate θ' in subsequent calculations may become incorrect. On the other hand, if the spread of the orientation change rate weight w _t (θ') is set to be large, it becomes difficult to distinguish between multiple targets using the orientation change rate θ', and there is a possibility that accurate tracking will not be possible when multiple targets intersect.

On the other hand, in the tracking device 100 of embodiment 2, the azimuth change rate weight w t (θ') for the azimuth change rate θ' can be automatically adjusted in accordance with the product of the difference between the estimated value of the azimuth change rate θ' and the actual azimuth change rate θ' of the target that is the sound source and the adjustment parameter k. Specifically, the azimuth change rate weight w _t (θ') for the azimuth change rate θ' can be automatically adjusted so that the larger the difference between the estimated value of the azimuth change rate θ' and the actual azimuth change rate θ' of the target, the wider the azimuth change rate weight w _t (θ') for the azimuth change rate θ' becomes, and the smaller the difference between the estimated value of the azimuth change rate θ' and the actual azimuth change rate θ' of the target, the narrower the azimuth change rate weight w _t (θ') for the azimuth change rate θ' becomes. As a result, the azimuth change rate θ' can be estimated with high accuracy and multiple targets can be accurately tracked.

Embodiment 3.
Fig. 11 is a diagram showing the configuration of tracking unit 500 according to embodiment 3. In Fig. 11, the components denoted by the same reference numerals as in Fig. 10 perform the same processing as described in embodiments 1 and 2.

As shown in FIG. 11, the tracking unit 500 of the tracking device 100 in embodiment 3 has a peak likelihood extraction unit 512, an applied likelihood recalculation unit 513, and a likelihood function regeneration unit 514. Furthermore, the tracking unit 500 in embodiment 3 has a maximum likelihood buffer unit 527 and an orientation likelihood width parameter storage unit 528 in the storage device 120.

The maximum likelihood buffer unit 527 stores, as data, a set number of maximum likelihoods l ^peak extracted by the peak likelihood extraction unit 512 (described later) (for T ^lmax from time t-T ^lmax to time t). The orientation likelihood width parameter storage unit 528 stores, as data, an orientation likelihood width parameter. The orientation likelihood width parameter is a parameter for adjusting the spread of the likelihood in the orientation direction used by the Bayesian estimation calculation unit 503.

The peak likelihood extraction unit 512 extracts the maximum likelihood l ^peak , which is the maximum likelihood, from the masked likelihood functions l ^m,i _t (θ) generated by the mask processing unit 501, and performs processing to set the orientation θ from which the peak likelihood was extracted as the maximum orientation θ ^peak . The maximum likelihood l ^peak extracted by the peak likelihood extraction unit 512 is used as data in processing performed by the applied likelihood recalculation unit 513. In addition, the maximum likelihood l ^peak extracted by the peak likelihood extraction unit 512 is stored as data in the maximum likelihood buffer unit 527.

The applicable likelihood recalculation unit 513 performs processing based on the set maximum likelihood l ^peak stored in the maximum likelihood buffer unit 527 and the maximum likelihood l ^peak and maximum orientation θ ^peak obtained by processing of the peak likelihood extraction unit 512, and generates an applicable likelihood to be used in the Bayesian estimation calculation unit 503 and an orientation orientation corresponding to the applicable likelihood.

The likelihood function regeneration unit 514 recalculates the likelihood function based on the applied likelihood obtained by processing the applied likelihood recalculation unit 513, the applied orientation, and the orientation likelihood width parameter stored in the orientation likelihood width parameter storage unit 528, to generate a regenerated likelihood function.

Next, we will further explain the processing of the tracking unit 500 in the tracking device 100 of embodiment 3. Here, we will particularly explain the processing of the peak likelihood extraction unit 512, applied likelihood recalculation unit 513, and likelihood function regeneration unit 514.

The peak likelihood extraction unit 512 generates the maximum likelihood l ^peak and the maximum orientation θ ^peak based on equation (8). Here, the function findpeaks is a function that detects the maximum peak (maximum value). [l ^peak,i _t , θ ^peak,i _t ] stores the maximum likelihood l ^peak and the maximum orientation θ ^peak corresponding to the maximum likelihood l ^peak in descending order. The peak likelihood extraction unit 512 then stores the maximum likelihood l ^peak that is the peak among [l ^peak,i _t , θ ^peak,i _t ] in the maximum likelihood buffer unit 527 as data.

12 is a diagram illustrating the processing of the peak likelihood extraction unit 512 in embodiment 3. FIG. 12(a) shows the relationship between the mask likelihood function l ^m,i _t (θ) and the maximum likelihood l peak in the tracker 500 tracking each of multiple targets (two targets in this example) when the targets are in the state shown in FIG. 7(a). As shown in FIG. 12(a), if the azimuths of the two targets are apart, the peak likelihood extraction unit 512 of the tracker 500-1 extracts the maximum likelihood l ^{peak ,1} _t (A), which is the peak associated with the acoustic signal from target A. Furthermore, the peak likelihood extraction unit 512 of the tracker 500-2 extracts the maximum likelihood l ^peak,2 _t (B), which is the peak associated with the acoustic signal from target B. The maximum likelihood l ^peak obtained by extraction by each tracker 500 is stored as data in the maximum likelihood buffer unit 527.

On the other hand, Figure 12(b) shows the relationship between the mask likelihood function l ^m,i _t (θ) and the maximum likelihood l ^peak in the tracker 500 tracking each target when multiple targets are in the state shown in Figure 7(b). As shown in Figure 12(b), if the azimuths of two targets are close to each other, the tracker 500-1 extracts the maximum likelihood l ^peak,1 _t (A) and the maximum likelihood l ^peak,1 _t (B) related to the acoustic signals from the two targets. The tracker 500-2 also extracts the maximum likelihood l ^peak,2 _t (A) and the maximum likelihood l ^peak,2 _t (B).

Next, the applied likelihood recalculator 513 performs processing based on the maximum likelihood l ^peak for the set amount stored in the maximum likelihood buffer 527 and the maximum likelihood l ^peak and maximum direction θ ^peak obtained by processing by the peak likelihood extractor 512. First, the applied likelihood recalculator 513 calculates the average value μ ^l,i _t of the maximum likelihood for T ^lmax , which is the set number of cycles, based on equation (9).

The application likelihood recalculator 513 further calculates the application likelihood l ^re_apply , _it based on equation (10) and the application orientation ^θ re_apply,it based on equation (11) using the maximum likelihood l ^peak , the maximum orientation θ ^{peak ,} and the average value μ _l , _it . Here, i ^peak.apply,i in equations (10) and (11) is expressed by equation (12).

13 is a diagram illustrating the processing of applied likelihood recalculator 513 in embodiment 3. As shown in the upper part of Fig. 13, for example, applied likelihood recalculator 513 of tracking unit 500-1 can generate maximum orientation θ peak corresponding to maximum likelihood l ^peak ,1 t (A) and maximum likelihood l peak,1 _t (A) out of maximum likelihood l ^peak ,1 _t (A) and maximum likelihood l ^{peak ,1} _t (B), based on maximum likelihood l ^peak _in a time series for a set time stored in maximum likelihood ^{buffer unit} 527.

On the other hand, as shown in the lower part of Figure 13, the applied likelihood recalculation unit 513 of the tracking unit 500-2 can generate the maximum likelihood l ^peak,2 _t (B) and the maximum orientation θ ^peak corresponding to the maximum likelihood l ^peak,2 _t (B) based on the maximum likelihood for the set time stored in the maximum likelihood buffer unit 527.

In this way, the applied likelihood recalculator 513 does not generate the magnitude of the maximum likelihood in the masked likelihood function l ^m,i _t (θ), but generates a maximum likelihood l ^peak and a maximum orientation θ ^peak that are closer to the past maximum likelihoods stored in the maximum likelihood buffer unit 527. As a result, in Fig. 13(b), the applied likelihood recalculator 513 of the tracker 500-1 recalculates [l ^peak,1 _t , θ ^peak,1 _t ], and the applied likelihood recalculator 513 of the tracker 500-2 recalculates [l ^peak,2 _t , θ ^peak,2 _t ].

14A and 14B are diagrams illustrating the regenerated likelihood function processed by the likelihood function regenerator 514 in embodiment 3. Fig. 14A shows a masked likelihood function. Fig. 14B shows the regenerated likelihood function. The likelihood function regenerator 514 uses the applied likelihood obtained by processing in the applied likelihood recalculator 513, the applied orientation corresponding to the applied likelihood, and the orientation likelihood width parameter stored in the orientation likelihood width parameter storage unit 528 to perform calculation based on equation (13) to generate the regenerated likelihood function l ^re,i _t (θ).

The weighting unit 502 performs weighting processing based on equation (14) using the regenerated likelihood function obtained by processing in the applied likelihood recalculation unit 513 and the orientation change rate weight w _t (θ') obtained by processing in the weight generation unit 510 to generate a weighted likelihood function L ^re_w,i _t (θ, θ'). Here, w _t (θ') is a vector related to the orientation change rate θ'. Furthermore, l ^re,i _t (θ) and l ^re,i _t (θ) are vectors related to the orientation θ. The weighting unit 502 multiplies all combinations of θ and θ' based on equation (14) and arranges the results as a matrix with the elements being the result.

As described above, according to the tracking device 100 of the third embodiment, the maximum likelihood buffer unit 527 stores a set amount of data on the maximum likelihood l ^peak extracted by the peak likelihood extraction unit 512. Furthermore, the applied likelihood recalculation unit 513 generates an applied likelihood. Then, the likelihood function regeneration unit 514 generates a regenerated likelihood function based on the applied likelihood. Therefore, each tracker 500 does not perform Bayesian estimation using the instantaneous likelihood based on the masked likelihood function l ^m,i _t (θ), but instead uses the peak of likelihood that is most likely as the likelihood of the target to be tracked from the distribution of the past maximum likelihood l ^peak as the applied likelihood, and performs Bayesian estimation using the regenerated likelihood function. As a result, in a situation where the orientations of the targets are close to each other, such as immediately after the intersection of multiple targets shown in Figure 7(b), the tracking unit 500-1 can generate a likelihood function for target ^A , _the original target to be tracked, from the maximum likelihood l ^peak of multiple targets that cannot be distinguished and are combined using the mask likelihood function l m,i t (θ), as shown in Figure 14.

Figure 15 is a diagram illustrating the effects of the tracking device 100 of embodiment 3. In the tracking device 100 of embodiment 3, as shown in Figure 15, each tracking unit 500 can perform processing based on a likelihood function corresponding to the respective azimuth and azimuth change rate of the target to be tracked. As a result, it is possible to improve tracking capability in tracking situations that include multiple intersecting targets.

Embodiment 4.
Fig. 16 is a diagram showing the configuration of a tracking unit 500 according to embodiment 4. In Fig. 16, the components denoted by the same reference numerals as in Fig. 11 and the like perform the same processing as that explained in embodiments 1 to 3.

16 , tracking unit 500 of tracking device 100 according to the fourth embodiment has an applied likelihood buffer unit 529 in storage device 120 instead of maximum likelihood buffer unit 527 according to the third embodiment. Applied likelihood buffer unit 529 stores the applied likelihood obtained by the processing of applied likelihood recalculation unit 513 as data for a set amount (for T ^lapply from time t−T ^lapply to time t). Here, applied likelihood buffer unit 529 stores, as data, maximum likelihood l ^peak extracted and obtained by peak likelihood extraction unit 512 for a certain period of time after target tracking begins, and then stores the applied likelihood as data after the certain period of time has elapsed.

Next, the processing of the tracking unit 500 according to the third embodiment will be further described. Here, the processing of the applied likelihood recalculator 513 will be particularly described. The applied likelihood recalculator 513 performs processing based on the maximum likelihood l ^peak for a set amount stored in the applied likelihood buffer unit 529 and the maximum likelihood l ^peak and maximum direction θ ^peak obtained by processing by the peak likelihood extractor 512. First, the applied likelihood recalculator 513 calculates the average value μ ^lapply,i _t of the maximum likelihood l ^peak for T ^lapply based on equation (15).

The applied likelihood recalculator 513 further calculates the applied likelihood based on the above-mentioned formula (10) and the applied orientation based on the formula (11) using the maximum likelihood l ^peak , the maximum orientation θ ^peak , and the average value μ t. Here, in the fourth embodiment, i ^peak.apply,i in the formulas (10) and (11) is expressed by the formula (16).

The applied likelihood recalculation unit 513 stores the processed applied likelihood as data in the applied likelihood buffer unit 529 after a certain time has elapsed since target tracking began.

17 is a diagram illustrating the effect of the tracking device 100 of the fourth embodiment. First, a case will be described in which the maximum likelihood l ^peak is stored in the maximum likelihood buffer unit 527, as in the third embodiment. When the azimuths θ of multiple targets are close to each other, as shown in FIG. 17( a), the maximum likelihood l ^peak,2 _t (A) related to target A is also stored in the maximum likelihood buffer unit 527 of the tracking unit 500-2 that is performing processing related to target B. As a result, if the time period in which the two targets are close to each other becomes long, after the two targets cross, the likelihood related to target A is used in the processing of the tracking unit 500-2, which may hinder accurate estimation of the azimuth of target B.

In view of this, the tracking device 100 of embodiment 4 is configured to store the applied likelihood as data in the applied likelihood buffer unit 529. Therefore, by storing the applied likelihood obtained by recalculation in the applied likelihood recalculation unit 513 in the applied likelihood buffer unit 529, the data used in the calculations by the applied likelihood recalculation unit 513 does not include likelihood data relating to different targets. Therefore, as shown in FIG. 17(b), for example, it is possible to prevent the applied likelihood recalculation unit 513 from recalculating the applied likelihood based on the likelihood relating to target A that is not being tracked by the tracking unit 500-2. As a result, it is possible to further improve tracking capability in tracking situations that include multiple intersecting targets.

Embodiment 5.
In the tracking device 100 of the first and second embodiments described above, the orientation change rate estimation unit 509 performs processing based on the prior probability distribution P ^prior,i t+1|t(θ) calculated by the Markov update unit 507. However, this is not limiting. The orientation change rate estimation unit 509 may also perform processing based on the posterior probability distribution P ^posteriori t(θ) calculated by the Bayesian estimation calculation unit 503.

Regarding the weight generation unit 510 described in the first and second embodiments, equation (2) in the first embodiment and equation (8) in the second embodiment use a shape according to a Gaussian distribution, but this is not limited to this. Any distribution that can be predicted in advance can be applied, not limited to a Gaussian distribution. Furthermore, the distribution of the orientation change rate can be estimated and applied based on the orientation change rate obtained by the time difference of the estimated orientation θ ^* stored in the estimated orientation buffer unit 526.

In addition, although the orientation likelihood width parameter of the orientation likelihood width parameter storage unit 528 in the third embodiment is a parameter that is set in advance, this is not limiting. For example, the orientation likelihood width parameter may be set based on the variation in the maximum orientation θ ^peak obtained by the processing of the peak likelihood extraction unit 512, and used as the orientation likelihood width parameter.

In the above-described first to fourth embodiments, the data processing device 110 has two tracking units 500 and tracks two targets, but this is not limited to this. By increasing the number of tracking units 500, it is possible to track three or more targets.

The tracking device 100 in the above-described first to fourth embodiments tracks a target based on the sound wave signals received by the microphones 210 of the microphone array 200. Here, the microphone array 200 is assumed to be mounted on a passive sonar, but this is not limited to this. It can also be applied to an active sonar, which emits sound waves and converts the sound waves reflected from the target into acoustic signals. It can also be applied to signals obtained by RADER (Radio Detecting and Ranging), LIDAR (Light Detection and Ranging), etc.

100 Tracking device 110 Data processing device 120 Storage device 130 Signal input device 140 Signal output device 200 Microphone array 210, 210-1 to 210-n Microphones 220 Signal conversion device 230 Signal transmission device 300 Communication line 400 Likelihood generation unit 500, 500-1, 500-2 Tracking unit 501 Mask processing unit 502 Weighting unit 503 Bayesian estimation calculation unit 504 Detection determination unit 505 ID assignment unit 506 Orientation estimation unit 507 Markov update unit 508 Mask generation unit 509 Orientation change rate estimation unit 510 Weight generation unit 511 Error standard deviation calculation unit 512 Peak likelihood extraction unit 513 Applied likelihood recalculation unit 514 Likelihood function regeneration unit 521 Prior probability buffer unit 522 Mask storage unit 523 Weight storage unit 524 Adjustment parameter storage unit 525 Estimated heading change rate buffer unit 526 Estimated heading buffer unit 527 Maximum likelihood buffer unit 528 Heading likelihood width parameter storage unit 529 Applied likelihood buffer unit

Claims (9)

a likelihood generation unit that calculates likelihoods in a plurality of directions from observation data obtained by observing signals related to a plurality of tracking targets and generates a likelihood function with the directions as parameter variables ;
a data processing device having a plurality of tracking units that determine the presence and direction of a tracking target corresponding to one of the plurality of tracking targets based on the likelihood function;
Each of the tracking units comprises:
an azimuth change rate estimation unit that performs an azimuth change rate estimation process to estimate an azimuth change rate in the corresponding azimuth of the tracking target and calculates an azimuth change rate estimated value;
a weight generation unit that generates a heading change rate weight based on the heading change rate estimation value, and
a weighting unit that generates a weighted likelihood function by superimposing the orientation change rate weight on the likelihood function;
a Bayesian estimation calculation unit that performs processing to generate a posterior probability distribution from the weighted likelihood function;
a Markov update unit that performs processing to generate a prior probability distribution;
a detection determination unit that performs processing to determine the presence or absence of the tracking target based on the posterior probability distribution and detect the tracking target;
an ID assigning unit that assigns an ID number to the tracking target detected by the detection and determination unit;
and an azimuth estimation unit that estimates the azimuth of the tracking target based on the posterior probability distribution and calculates an estimated azimuth. The tracking device described in claim 1, wherein the azimuth change rate estimation unit performs the azimuth change rate estimation process based on the prior probability distribution generated by the Markov update unit. The tracking device described in claim 1, wherein the azimuth change rate estimation unit performs the azimuth change rate estimation process based on the posterior probability distribution generated by the Bayesian estimation calculation unit. A tracking device according to any one of claims 1 to 3, wherein the weight generation unit generates the azimuth change rate weight from data relating to a time series of the estimated azimuth estimated by the azimuth estimation unit and the azimuth change rate estimated value estimated by the azimuth change rate estimation unit. Each of the tracking units comprises:
a peak likelihood extracting unit that extracts a maximum likelihood from the likelihood function generated by the likelihood generating unit as a maximum likelihood;
an applied likelihood recalculation unit that calculates an average of the maximum likelihoods for a preset period and calculates an applied likelihood and an applied orientation corresponding to the applied likelihood to be used in processing by the Bayesian estimation calculation unit;
5. The tracking device according to claim 1, further comprising: a likelihood function regeneration unit that regenerates the likelihood function based on the applied likelihood and the applied direction. The tracking device described in claim 5, wherein the applied likelihood recalculation unit calculates the average of the applied likelihood for the set period after a certain time has elapsed since the start of tracking, and calculates the applied likelihood and the applied direction corresponding to the applied likelihood to be used in processing by the Bayesian estimation calculation unit. a receiver array having a plurality of receivers and a signal converter that converts signals received by the receivers into digital data and generates observation value data;
A tracking system comprising the tracking device according to any one of claims 1 to 6. A tracking method for tracking each of a plurality of tracking targets based on a likelihood function with a direction as a parameter variable, the likelihood function being generated by likelihoods related to a plurality of directions calculated from observation value data obtained by observing signals related to the plurality of tracking targets, the method comprising:
a step of calculating an azimuth change rate estimated value by performing an azimuth change rate estimation process for estimating an azimuth change rate in the corresponding azimuth of the tracking target;
generating a heading rate weight based on the heading rate estimate; and
generating a weighted likelihood function by superimposing the heading change rate weight on the likelihood function;
generating a posterior probability distribution from the weighted likelihood function;
performing a process to generate a prior probability distribution;
performing a process of determining the presence or absence of the tracking target based on the posterior probability distribution and detecting the tracking target;
assigning an ID number to the detected tracking target;
and estimating the direction of the tracking target based on the posterior probability distribution to calculate an estimated direction. A tracking method program for tracking each of a plurality of tracking targets based on a likelihood function with a direction as a parameter variable, the likelihood function being generated by likelihoods related to a plurality of directions calculated from observation value data obtained by observing signals related to the plurality of tracking targets,
a step of calculating an azimuth change rate estimated value by performing an azimuth change rate estimation process for estimating an azimuth change rate in the corresponding azimuth of the tracking target;
generating a heading rate weight based on the heading rate estimate; and
generating a weighted likelihood function by superimposing the heading change rate weight on the likelihood function;
generating a posterior probability distribution from the weighted likelihood function;
performing a process to generate a prior probability distribution;
performing a process of determining the presence or absence of the tracking target based on the posterior probability distribution and detecting the tracking target;
assigning an ID number to the detected tracking target;
and estimating the direction of the tracking target based on the posterior probability distribution and calculating the estimated direction.