IROS2019國際學(xué)術(shù)會議論文集 1857

Development of a Location estimation System for Minute Sound Source by Using Human Acoustic System with Stochastic Resonance Katsuyoshi Tsujita Abstract The purpose of this study is to develop a system that enables location estimation of a small sound source The location estimation of a small sound source has some diffi culties such as high computational costs or disturbances from the ambient noises and refl ected waves The proposed system is composed of a biologically inspired system which uses a hearing mechanism based on the human auditory model and a mechanism for perceiving weak sound signals that use stochastic resonance The location estimation mechanism in the proposed system is based on the time lag detecting architecture On the other hand the stochastic resonance mechanism can pick up the small sound signals among the ambient noises Using this proposed system we implemented the location estimation of the small sound source through hardware experiments Good results were obtained for the small sound source location estimation I INTRODUCTION In recent years large progress has been made in the technology of location estimation of sound source and it has become possible to perform high precision presumption of location by using methods such as microphone array system the MUSIC multiple signal classifi cation method the ESPRIT estimation of signal parameters via rotational invariance technique method etc 1 3 These systems have very accurate location performance of the sound source But there are some issues to deal with in these methods That is the system has high computational cost complexity in system architecture or insuffi cient performance for location estimation of a small sound source Besides the location estimation system of the sound source has problems with application in the outdoor envi ronment The location estimation of sound sources in the outdoor environment is hampered by the infl uence of the refl ected waves Countermeasures law of the refl ected waves fi t the location estimation system of sound sources with the echo suppression mechanism 4 Since the accuracy of the location estimation is deteriorated by ambient noise in cases where it is especially aimed at small sound sources the location estimation of a sound source is diffi cult Therefore a new location estimation system for such sound source and the environment is expected to be de veloped especially for the application of rescue inspecting devices at the disaster site or industrial mechanical inspection systems and so on Among the problems of the sound source estimation we aim at the development of an effective sound estimation system for small sound sources Additionally we assume the use of a sound source estimation system in an environment with little infl uence of refl ection but with ambient noises Performing location estimation of a sound source using the time lag detecting and sound pressure difference of sounds to both ears is known 5 Human auditory mechanism of location estimation of sound sources has a dextrous property that one can hear and divide various types of sounds even in environments where the target sound sources are buried in the surrounding noise like the cocktail party effect 6 Moreover it is widely said that for pure sounds in the low frequency the region human location estimation of a sound source is performed by using the time lag between sound signals 7 On the other hand it utilizes the sound pressure difference of sound signals in a high frequency region When the candidate for detection is considerably low frequency wave such as human voice a sound pressure difference is not needed It is thought that an auditory model is effective in the location estimation of a small sound by using a time lag detecting method To detect the small sound source using the stochastic resonance phenomenon can be effective way 8 Stochastic resonance is known as a phenomenon in which a response is enhanced by adding a suitable quantity of noise to a weak signal 9 12 Stochastic resonance is also used as a nonlinear receiver digital signal processing system of the baseband binary 13 In this research we built a location estimation system of a sound source combining two auditory systems One is an auditory method that technologically imitates the human auditory system which uses time lag detection between both ears The other system is a small sound detecting system which uses the stochastic resonance phenomenon to detect sounds buried in noise 14 In this paper the system s validity is verifi ed through hardware experiments II SYSTEM ARCHITECTURE The architecture of the proposed location estimation sys tem of a sound source is summarized in Fig 1 The system has two sound signal fl ows the left and the right Both signals are processed in two subsystems One is a location estimation system using a time lag detection mechanism The other is a small sound detecting system using stochastic resonance phenomena In the system a sound signal is an input into microphones on both the right and the left In the small sound detection system the input signals on both sides are enhanced and detected by using stochastic resonance phenomena with a controlled gaussian noise The detected signals are succeeding input to the location estimation system of a sound source and are 2019 IEEE RSJ International Conference on Intelligent Robots and Systems IROS Macau China November 4 8 2019 978 1 7281 4003 2 19 31 00 2019 IEEE7302 Fig 1 System architecture processed by the auditory periphery model of the human auditory system Now let us summarize the human auditory model com paring artifi cial model Human auditory mechanism of lo cation estimation system of a sound source is processed in the cochlea an auditory organ in the inner ear The cochlea receives the vibrations from sound conducted by the tympanum and ossicles which consist of three small bones to amplify the vibrations of the tympanum that is a mechanical signal amplifi er The cochlea is fi lled with watery liquid lymph The amplifi ed vibration runs through the lymph and is detected by hair cells whose hairs can detect a very small vibration Each hair cell responds to a specifi c frequency of the vibration The detected vibrations are separated in various frequency component signals which can be artifi cially constructed by a set of bandpass fi lters The signals are transformed into nerve signals and are sent to the higher center through cochlear nerves that is the location estimation system The proposed system uses a model of the human auditory system to detect the time lag Stochastic resonance is a phenomenon whose response to a system improves it by adding a proper quantity of noise to a faint signal Moreover it is understood that stochastic resonance will occur and be observed in a system if three elements are present the threshold value characteristic a noise source and a weak input signal 15 When a system has a threshold value to respond to the sound signals the system will not respond to faint signals which are below the threshold level However when an optimum amount of noise is added to this system it will respond to the faint signals when a tuned amount of artifi cial noise added to the signals exceeds the threshold level Weak signals are detectable via this response Furthermore Collins et al 16 devised how to parallelize the threshold value element to the input signal and reproduce not only the periodic response but the input waveform itself Noise independent to each input is added and the reconstructed waveform of a faint signal is obtained by adding the output from each element In this paper the model of Collins et al 16 was adopted as a small sound detection system and it was built as a system on a computer Noise is added to the input signal and input to the parallelized threshold value element independently of each other White noise generated by the Box Muller method 17 of Eq 1 from the uniform random number was used for the added noise X1 2logeU1 1 2cos2 U2 X2 2logeU1 1 2sin2 U2 1 For the parallelized threshold value element Schmidt Trig ger type devices are used which have hysteresis properties to the input signals 18 The number of the parallelized thresh old value elements is 100 By summing up the output pulse from each threshold value element a faint signal is detected The intensity of the added noise and the threshold value of the element must be optimized to make the appropriate stochastic resonance phenomena to an input signal In this paper a value higher than the voltage obtained from the input is fi xed as the threshold value and the intensity of the added noise is set up by hand tuning according to measurement data In the location estimation system of a sound source an input on either side is fi rst processed by a cochlear model on both the right and left sides respectively The cochlear organ in the human auditory system decomposes the sound signals according to frequency 19 and it can be modeled as parallelized artifi cial bandpass fi lter groups The fi lter group consisted of Butterworth fi lters of low pass side 6 dB oct and high region side 18 dB oct in this research The bandwidth of the fi lter group uses the integer values in the Bark scale which are equivalent to the frequency resolution on a human cochlear nerve model The relation between a Bark value and frequency is denoted in Eq 2 20 B 13arctan 0 76f 0 35arctan f 7 5 2 2 Next the output of the cochlear model performs nonlinear conversion rectifi ed using Eq 3 with the hair cells model f x x 1 3 x 0 1 4x 1 3 x 0 3 After attenuating a negative signal it changes into a pulse like signal using the cochlear nerve model In this case the ignition rate increases in proportion to a rise of the signal from the hair cells The output of the cochlear nerve model is performed as follows using a random number sequence n t 1 h t r t 0 h t r t 4 A pulse signal on either side is input into a time lag detection model The time lag detection system was designed based on Jeffress s model 21 shown in Fig 2 An input pulse on the left side is transmitted sequentially at every time interval to the end of the right side in the time lag detection model and the input pulse on the right side is processed 7303 in opposite direction The transmitted pulse signals coming from both sides coincide with each other at a certain point of the line of ignition elements and ignite the element at the meeting point When the sound source is located on one side the sound signals which reach both sides of the microphone shave a time lag Therefore the meeting point of the ignition element shifts to the corresponding side on the line of the ignition elements The location estimation of the sound source can be performed by investigating the position where the element was ignited Fig 2 Jeffress model III HARDWARE EXPERIMENTS Figure 3 shows the experimental setup which is composed of microphones amplifi ers a sound source speaker system a computer and an FFT Fast Fourier Transform analyzer In this experiment omni directional back electret capacitor type microphones EB 1461 Four Leaf Ltd were used The sensitivity of each microphone is 44 3 dB 0 dB 1 V Pa with a signal to noise ratio of 69 dB and the frequency characteristic of 50 Hz 20 kHz In the experiment these two microphones have been arranged at an interval of 0 2 m The input signal was recorded using an FFT analyzer CF3600 Ono SokkiCo Ltd after amplifying with the gain of 60 dB The sampling frequency is 44 1 kHz The measured signal performs location estimation only by using a time lag A pure sound of 500 Hz which is a frequency band which can perform location estimation only for time lag information was used After generating the pure sound on the computer it was given to the system as a sound source using the amplifi er DRA 201SA DENON Co Ltd and a speaker system CM 5 KENWOOD Corp The sound source position was arranged horizontally on semicircles of radius 1 0 m or 2 0 m from the center of the two microphones like Fig 4 It may be pointed out that if we investigate the performance of the proposed system aiming at the application in the outside environment the conditions of the radius of 1 0 m or 2 0 m are too small But in this study the purpose of the experiment is to verify the validity of the time lag detection system and the signal enhancement performance of the stochastic resonance fi rst The sound source orientation was set from the middle point of the both microphones The front was at 0 deg and there were a total of fi ve points 30 deg 80 deg of right and left intervals as the sound source positions Moreover Fig 3 Experimental setup this measurement was executed in a soundproof room Two conditions of the sound pressure level were et as 50dB and 40dB The approximate index of sound pressure level is as follows 60 dB almost same as the human ordinary conversation 50 dB almost same as the driving sound of air conditioner in a room small voice conversation etc 40 dB almost same as the loudness of human whis pering At fi rst the sound pressure was set signifi cantly high at 50 dB Then the location estimation the performance of the system was measured by decreasing the sound pressure level step by step When the sound pressure level became small we investigated the performances of the proposed system with the small sound detecting system Fig 4 Location of the sound source in the experiments IV EXPERIMENTAL RESULTS A Case of Enough Sound Pressure The location estimation processing was performed using several averages of measurement data When the sound pressure level was about 50 dB loud enough for the location estimation was possible The results are shown in Figs 5 6 In the fi gure the horizontal axis is the orientation of the estimated location of the sound source The vertical axis is the fi red frequency of ignition element in the time lag detection system From the fi gures the peaks of the 7304 frequency distribution of fi red ignition elements were located at the correct orientations of the sound source in all cases The results showed that it is possible to estimate the sound source position and the location estimation system itself was well functioning and effective for sound source location even if the distance of the sound source changes a 0 deg 1 0 m b 30 deg 1 0 m c 80 deg 1 0 m Fig 5 Result of location estimation of the sound source 500 Hz 50dB 1 0m B Cases of Small Sound Source As a result of performing location estimation using a sound weaker than the above mentioned sound source 40 dB a presumption of the sound source position was diffi cult due to the infl uence of the background noise By using the location estimation system of the sound source and the small sound detection system in combination the performance of the system was evaluated with a small sound signal whose source position was diffi cult to locate using only the location estimation system of the sound source The results are shown in Fig 7 a 0 deg 2 0 m b 80 deg 2 0 m Fig 6 Result of location estimation of the sound source 500 Hz 50dB 2 0m Even when using the same measurement data which the location estimation system by itself had diffi culty in detecting the position of the sound source the positioning accuracy of the integrated small sound detection system improved greatly The location estimation of this system was even able to presume the location of a small sound whose position was diffi cult to estimate using only the location estimation system It was shown that the small sound detection system using stochastic resonance is effective in the location esti mation system of small sounds C Performance under the noisy environment In the previous section the results of the sound source estimation experiment with the state removed as much as possible are shown In this section we show the results of experiments in consideration of environmental noise Consider the difference in performance as compared with the experiment that removes ambient noise as much as possible For this time artifi cial pseudo white noise was prepared separately from the sound source and it was set as background noise by generating a sound pressure that was almost the same as the microsound source Figure 8 shows the power spectrum of the environmental noise of this experimental case Figure 9 shows the performance of the proposed system under environmental noise In the condition that the level of ambient noise was almost the same as the level of the signal the frequency of fi ring overall increased and a sharp peak did not appear at the position of the sound source However using the proposed system it was verifi ed the validity of the system 7305 a 0 deg b 30 deg c 80 deg Fig 7 Result of location estimation of the sound source 500 Hz 40dB Fig 8 Power spectrum of environmental noise D Comparison with Other Sound estimation Method In this section the performance of the proposed system is compared with one of the usual sound estimation system a 0 deg b 30 deg c 80 deg Fig 9 Result of location estimation of the sound source within the environmental noise 500 Hz 50dB Noise level 50dB KINECT system 22 The orientation of the sound source is given as 30 deg Fig 11 shows an example result of KINECT system and the proposed system for the sound pressure level is 50 dB In this case both systems worked well KINECT system has better performance in the precision of location than the proposed system at this sound pressure level But the case of sound pressure of the sound source is less than 40 dB that is almost same as the loudness of human whispering KINECT system did not detect the sound source while the proposed system was able to detect and almost localized the small sound source V CONCLUSIONS In this research a location estimation system which can be used with the small sound source was developed This system is composed of a location estimation system of the sound source using a time lag detection system and a small sound detection syste