An Automated Acoustic System to Monitor and Classify Birds

Transcription

1 University f Nebraska - Lincln DigitalCmmns@University f Nebraska - Lincln 2003 Bird Strike Cmmittee USA/Canada, 5th Jint Annual Meeting, Trnt, ONT Bird Strike Cmmittee Prceedings August 2003 An Autmated Acustic System t Mnitr and Classify Birds C. Kwan Intelligent Autmatin, Inc., Rckville, MD K. H U. f Missuri at Clumbia G. Mei Intelligent Autmatin, Inc., Rckville, MD Y. Li U. f Missuri at Clumbia Z. Ren Intelligent Autmatin, Inc. See next page fr additinal authrs Fllw this and additinal wrks at: Part f the Envirnmental Health and Prtectin Cmmns Kwan, C.; H, K.; Mei, G.; Li, Y.; Ren, Z.; Xu, R.; Zha, G.; Stevensn, M.; Stanfrd, V.; and Rchet, C., "An Autmated Acustic System t Mnitr and Classify Birds" (2003) Bird Strike Cmmittee USA/Canada, 5th Jint Annual Meeting, Trnt, ONT This Article is brught t yu fr free and pen access by the Bird Strike Cmmittee Prceedings at DigitalCmmns@University f Nebraska - Lincln. It has been accepted fr inclusin in 2003 Bird Strike Cmmittee USA/Canada, 5th Jint Annual Meeting, Trnt, ONT by an authrized administratr f DigitalCmmns@University f Nebraska - Lincln.

2 Authrs C. Kwan, K. H, G. Mei, Y. Li, Z. Ren, R. Xu, G. Zha, M. Stevensn, V. Stanfrd, and C. Rchet This article is available at f Nebraska - Lincln:

3 Presentatins f Bird Strike 2003 Page 1 f 11 An Autmated Acustic System t Mnitr and Classify Birds C. Kwan +, K. H *, G. Mei +, Y. Li *, Z. Ren +, R. Xu +, G. Zha +, M. Stevensn +, V. Stanfrd &, and C. Rchet & + Intelligent Autmatin, Inc. * Dept. f EE & NIST 7519 Standish Place U. f Missuri at Clumbia Technlgy Building (225) Suite Engineering Building West Rm A216 Rckville, MD Clumbia, MO65211 Gaithersburg, MD USA USA USA Abstract Cllisins between aircraft and birds have becme an increasing cncern fr human health and safety. Mre than fur hundred peple and ver fur hundred aircraft have been lst glbally in recent years, accrding t a FAA reprt. T minimize the number f birdstikes, micrphne arrays have been used t mnitr birds near the airprt r sme critical lcatins in the airspace. Hwever, the range f existing arrays is nly limited t a few hundred meters. Mrever, the identificatin perfrmance in lw signal-t-nise envirnment is nt satisfactry. Under the supprt f the US Air Frce, Intelligent Autmatin, Incrprated (IAI) and the University f Missuri at Clumbia, prpse a nvel system t imprve bird mnitring and recgnitin system in nisy envirnments. First, a micrphne dish cncept (micrphne array with many cncentric rings) is prpsed that prvides very directinal and lng range (a few thusand meters) acquisitin f bird sunds, can simultaneusly pick up and track sund frm different directins, and the cst f the dish will be a few hundred dllars. Secnd, an efficient recgnitin algrithm is prpsed which uses Hidden Markv Mdel (HMM) and Gaussian Mixture Mdels (GMM). The verall system is suitable fr real-time mnitring and recgnitin fr a large number f birds. Here we will summarize sme preliminary results f ur prpsed methd. First, we will give a brief verview f the prpsed system, which cnsists f several majr parts: micrphne dish and data acquisitin system, Directin f Arrival (DOA) estimatin, beamfrmer t eliminate interferences, and bird classifier. Secnd, we will describe a new wideband DOA estimatin algrithm and prvide a cmparative study between estimatin results using linear array and ur circular array. Third, beamfrming algrithm will be intrduced thrugh a cmparative study between the linear array and ur circular array. A new beamfrming algrithm fr dish array has been develped. It was fund that the dish array has several key advantages ver the linear array, including less ambiguity angles, mre cnsistent perfrmance, etc. Furth, bird classificatin results using GMM methd will be presented. Fifth, the develpment f a prttype micrphne dish will be included. A dish array cnsisting f 64 micrphne elements has been develped and used t cllect sund data in labratry and in an pen space. Sixth, experimental results will be described t shw the perfrmance f the sftware and hardware. 1. Overall System Descriptin Main Results Figure 1 shws the prpsed system which cnsists a micrphne dish, a data acquisitin system, and sftware prcessing algrithms such as directin finder, beamfrmer, and bird sund classificatin. A/D cnversin Directin finder Beamfrmer Bird sund segmentatin Bird verificatin Micrphne Array

4 Presentatins f Bird Strike 2003 Page 2 f 11 Figure 1 Prpsed autmated bird mnitring and recgnitin system. 2. DOA Estimatin Algrithm 2.1 Estimatin Algrithm fr Circular Array A beamfrmer requires the directin f arrival (DOA) f the surce signal fr beamfrming t enhance the desired signal. The signal DOA is nt knwn in practice and needs t be estimated. This sectin presents the DOA estimatin f a wideband surce signal, based n the MUSIC algrithm fr narrwband signal. MUltiple Signal Classificatin(MUSIC) algrithm is a DOA estimatin algrithm fr narrwband signal. Fr wideband signal DOA estimatin, a simple technique exists by dividing the wideband signal int many narrwbands and then applying MUSIC n thse narrwbands. The DOA estimatin fr the wideband signal is generated by cmbining estimated results frm all the narrwband cmpnents. The prcess is shwn in Figure 2. Other wideband DOA estimatin techniques can be fund in [1,2], which are cmputatinally intensive. At present, we implemented the narrwband cmbining technique. Array Signal FFT Narrwband MUSIC : : Cmbining DOA estimates Wideband DOA estimate Narrwband MUSIC Figure 2 Blck diagram f the DOA estimatin f a wideband surce. As shwn in Fig. 2, the DOA estimatin algrithm cnsists f the narrwband MUSIC algrithm, which is fllwed by peak search technique t btain the DOA estimate fr each frequency band, and the cmbinatin f the DOAs frm different frequency bands t frm the final estimate. Figure 3 shws the MUSIC spatial spectrum btained frm 4 cncentric circular arrays with a ttal f 30 elements. The surce signal used are tw randm amplitude narrwband signals at 500Hz, cming frm ( θ = 90, φ = 70 ) and ( θ = 45, φ = 60 ) respectively, where the angle θ is with respect t the x-axis in the x-y plane and φ is the angle with respect t the z-axis. As shwn in the figure, the MUSIC spectrum cntains 2 peaks suggesting 2 DOAs. Figure 3 Narrwband MUSIC spectrum fr 2 DOAs. Figure 4 Narrwband MUSIC spectrum with nly lcal maxima. After the MUSIC spatial spectrum is btained, the remaining task is t identify the lcatin f thse peaks in the spectrum which crrespnd t the DOAs. We use the MUSIC spectrum in Figure 3 as an example t illustrate the 2-D peak searching algrithm described in [3]. Figure 4 shws such a prcessed MUSIC spectrum. Finally, it shuld be nted that the narrwband DOA estimatin results have a bias, especially in φ directin. We

5 Presentatins f Bird Strike 2003 Page 3 f 11 fund ut that when we use windwing t cmpute FFT f the array signal, the spectrum smearing f windwing will intrduce a bias in the result. Thus lnger windw is preferred. And this als suggests that larger number f spectral cmpnents generally gives smaller bias in estimatin result. Based n this understanding, the estimated results frm narrwband MUSIC are cmbined in a way taking their spectrum energy int cnsideratin. The peak value in the MUSIC spectrum will be assciated with an estimated DOA as its cnfidence value. A histgram is generated t cmbine the narrwband DOA estimates using the cnfidence values f the estimated narrwband DOAs as shwn in Figure 5. After btaining the histgram f DOA estimates frm different frequency cmpnents, the 2-D peak searching algrithm described in [3] is used again t yield the final wideband DOA estimate. Figure 5 Cmbined narrwband DOA estimates. Figure 6 MUSIC spectrum fr a linear array. 2.2 Cmparisn with DOA Estimatin Using Linear Array The DOA ambiguity set f a linear array is a cne arund the linear array. Thus it cannt be used t estimate the directin f a cming signal in 3-D space. T illustrate the advantage f using a circular array instead f a linear array in DOA estimatin, the MUSIC spectrum generated by an 11 element with half wavelength spacing linear array is shwn in Figure 6. There is nly ne narrwband signal at 500 Hz cming frm ( θ = 45, φ = 60 ). The SNR is 3dB. Althugh there is nly ne signal, there are tw stripes f spectrum peaks, crrespnding t the ambiguity set f a cne arund the linear array. It is clear that fr linear array it is nt pssible t yield an accurate DOA estimate withut ambiguity. 3. Beamfrming Algrithm The prpsed beamfrmer design fllws the technique develped by the authrs [4]. We shall nly summarize the array structure and sme cmparative results here. The details f the derivatin f the array are given in [4]. The circular array has 7 rings and 102 elements. The radius f the array is abut 0.5m which is very cmpact. Figure 7 shws the array structure. One nvelty f the prpsed design is that the circular array can perfrm wideband beamfrming, thrugh the cmpund ring apprach. In the cmpund ring structure, sme rings are shared by several frequency bands and therefre resulting in savings in array elements. The prpsed cmpund ring structure has 4 perating frequency bands as listed in the secnd clumn f Table 1. The third clumn in the table shws the number f rings in each band and the furth clumn is the number f elements in each ring fr the frequency band cnsidered. The gruping f the rings fr different bands is shwn in Figure 8. The minimum separatin between tw array elements is 1 δλ 4 2kHz d = =0.0402m (1) 4 4π The largest radius, and hence the size f the array, is 12d = m (2)

6 Presentatins f Bird Strike 2003 Page 4 f 11 The details in deriving (1) and (2) is skipped here and will be prvided upn request. Because f reusing array elements in different subarrays, the ttal number f elements is =102. Ring array cnfiguratin Subarray 4 (3.5 8k Hz) Subarray 2 ( Hz) 0 0 Subarray 3 ( Hz) Subarray 1 ( Hz) Figure 7 The prpsed circular array cnfiguratin. Figure 8 Gruping f the rings in the fur subarrays. Table 1: Gruping f rings int different subarrays fr brad-band beamfrming Apprx. Operating Frequency Range Number f Rings Number f Elements in Each Ring Subarray Hz 700 Hz 3 [6, 10, 14] Subarray Hz 1.5k Hz 4 [6, 10, 14, 18] Subarray 3 1.5k Hz 3.5k Hz 4 [6, 10, 14, 18] Subarray khz 8k Hz 2 [10, 14] In general, the larger the number f rings in a subarray, the larger will be the attenuatin in the ambient nise level. The pwer spectral density f birds have higher energy frm 700 4k Hz. That is why subarrays 2 and 3 have 4 rings t prvide larger attenuatin t the nise. Figure 9 shws a typical beam pattern f the prpsed circular array at 1 khz. A main advantage f the prpsed design is that it prvides clse t a fixed level f residue side lbes. Figure 9 The beam pattern f the prpsed circular array at 1 khz. Figure 10 The crdinate system used. The crdinate system shwn belw in Fig. 10, where θ crrespnds t the angle with respect t the x-axis in the x- y plane, and φ is the angle frm the z-axis with respect t the x-y plane. Fr cmparisn purpse, a cmpund linear array that has the same number array elements as the prpsed circular array (102 elements) is used. The cmpund linear array cmpses f 5 subarrays perating at frequency ranges arund 500Hz, 1kHz, 2kHz, 4kHz and 8kHz respectively. Each subarray cntains 34 elements. Half f the elements frm a subarray f higher frequency will be reused in the fllwing lwer frequency subarray. Thus ttal number f elements is: 34+17*4=102. Figure 11 shws a cmpund linear array with 5 subarrays and 4 elements within each subarray. (Subarray with as much as 34 elements are difficult t shw.)

7 Presentatins f Bird Strike 2003 Page 5 f 11 Figure 11 Cnfiguratin f cmpund linear array The smallest distance between tw array elements is λ8khz d = =0.0214m 2 The size f the 102 elements cmpund linear array is: λ500hz (34 1) = 11.32m 2 which is very large. Because f the cmpund array structure, the beampattern fr different center frequency is same. A 3-D beampattern fr ne f the subarray is shwn in Fig. 12, the DOA in assumed t be ( θ = 45, φ = 45 ). A linear array has an ambiguity regin that appears as a cne. Figure 12 Three-D beampattern f cmpund linear array. Figure 13 Three-D beampattern fr ring array The cmpund ring array used is the ne described earlier. It has 7 rings and cntains 102 elements. The array diameter is abut 1 m. The 3-D beampattern fr ne f the subarray is shwn in Fig. 13, the DOA in assumed t be ( θ = 45, φ = 45 ) Besides the abve beampattern cmparisn, a series f ther cmparisns between a linear array and the prpsed circular array have been carried ut. We have cncluded that: Circular array has an ambiguity set f Directin f Arrival (DOA) f nly 2 directins, while linear array has a larger ambiguity set f (DOA) which is cne. The beampattern f circular array can be rtated t arbitrary directin in the x-y plane withut suffering great fluctuatin. This is nt the case fr linear array. The beampattern f circular array varies with the φ much mre than the linear array des. Cmpund linear array utperfrms circular array in terms f ambient nise level mst f the time but it requires very large array size. Cmpund linear array is incapable f attenuating directinal interference in the DOA ambiguity set, circular array has much less ambiguity set thus it can remve the directinal interference in mst cases linear array fails. Cmprmise between several factrs (array size, nise level, ambiguity set f DOA etc) is necessary. Overall, circular array gives better perfrmance.

8 Presentatins f Bird Strike 2003 Page 6 f Bird Classificatin Algrithm We have develped tw different bird classificatin algrithms. One is based n Hidden Markv Mdel (HMM) and the ther ne is based n Gaussian Mixture Mdel (GMM). Based n ur wn evalutins [3], GMM perfrms better than HMM. Due t page limitatins, here we nly fcus n GMM apprach. Accrding t the evaluatins dne by Natinal Institute f Standards and Technlgy (NIST) engineers, GMM has been prven t be quite useful in speaker verificatin applicatins. The birds have similar spectrum as humans. The individual cmpnent densities f a multi-mdal density may mdel s many underlying set f acustic classes. A linear cmbinatin f Gaussian basis functins is capable f representing a large class f sample distributin. The bird classificatin cnsists f tw majr steps: 1) preprcessing the extract features; 2) applying GMM mdels t classify different birds. 4.1 Preprcessing t Extract Features f Birds T identify the bird species, the algrithm we have been using is t first extract the feature vectrs frm the bird sund data, then match these feature vectrs with Gaussian Mixture Mdels, each trained specifically fr each bird class. The difference between the prbabilities is cmpared t a pre-set threshld t decide if a given bird sund belngs t a specific bird class. The feature extractin subsystem can be best described by Fig. 14. Preemphasis Frame blcking Windwing FFT Mel-scale Filter bank Cepstral cefficient Make Feature vectr Mean nrmalizatin Figure 14 Preprcessing steps in the feature extractin subsystem. The purpse f feature extractin is t cnvert each frame f bird sund int a sequence f feature vectrs. In ur system, we use cepstral cefficients derived frm a Mel-frequency filter bank t represent a shrt-term bird speech spectra. The digital bird sund data is first preprcessed (pre-emphasized, set t verlapped frames and windwed) and then Mel Frequency Cepstral Cefficient Analysis is applied. Typically feature extractin prcess cmpresses arund 256 samples f bird sund data dwn t between 13 t 40 features. 4.2 Gaussian Mixture Mdel fr Birds The Gaussian mixture mdel fr birds is a prbabilistic mdel by which the distributin f data is mdeled as a linear cmbinatin f several multivariate Gaussian densities. There are tw mtivatins fr using Gaussian Mixture Densities as a representatin f bird identity [5]. The first is the intuitive ntin that the individual cmpnents densities f a multi-mdal density, like the GMM, may mdel sme underlying set f acustic classes. The secnd mtivatin is the empirical bservatin that a linear cmbinatin f Gaussian basis functin is capable f representing a large class f sample distributin. The GMM is usually trained with the Expectatin-Maximizatin (EM) algrithm t maximize the likelihd f the bservatin data frm an individual class.

9 Presentatins f Bird Strike 2003 Page 7 f 11 z SIR= 18dB (90 O, 30 O ) Helicpter (45 O, 45 O ) Bird (cangse) (135 O, 90 O ) y Jeep SIR= 15dB SNR=0dB x Figure 15 (a) Simulatin Scenari 1; (b) Beamfrming results frm Scenari 1. Our results f bird classificatin by using GMM perfrmed very well [3]. Figure 15 shws ne simulatin scenari where the desired bird signal is CanGse. The tw interferences are helicpter nise at 18dB SIR and jeep nise at 15dB SIR. The backgrund nise is the Hth nise at 0 db SNR. Hth nise, rughly speaking, is a lwpassed Gaussian nise with spectrum similar t vice. The beamfrming and classificatin prblem is mre challenging if the backgrund nise is Hth. This is because the nise and signal spectra verlaps extensively in the frequency dmain. As a matter f fact, the prblem is easier if the nise is white. Als in Fig. 15 is the received signal befre and after beamfrming, and the errr between the true and the beamfrmed signal. Befre beamfrming, the nise and interference dminates. The bird surce signal becmes apparent after beamfrming. Table 2 shws the classificatin results. Table 2: Classificatin Accuracy f CanGse Percentage f crrect classificatin Befre Beamfrming 13.11% After Beamfrming 100% Befre beamfrming, the classificatin results were nt satisfactry, due t the large amunt f interference and backgrund nise. It is clear that after beamfrming, the classificatin results imprve significantly. 5. Micrphne Dish Design and Hardware Prttype 5.1 Array Cnfiguratin The cmpund ring array is cmpsed f 10 rings: Subarray 1 is cmpsed f the th rings, with elements distributed as Subarray 2 is cmpsed f the th rings, with elements distributed as (4) (8) 6 8. Subarray 3 is cmpsed f the th rings, with elements distributed as (8) (8) 6 8. Subarray 4 is cmpsed f the th rings, with elements distributed as (8) (8) 6 8. The numbers in brackets are elements reused frm a previus subarray. Ttal number f elements: 4+4+(6+8)*4=64. The tp view f the ring array is shwn in Fig. 15.

10 Presentatins f Bird Strike 2003 Page 8 f 11 θ Figure 15 Circular array cnfiguratin. Figure 16 Definitin f the angle. Table 3 summarizes the ring radii. The largest radius is δλ 4 500Hz R10 = = m 4π where δ 4 is the 4 th rt f the 0 th rder Bessel functin f the 1 st kind and δ 4 = The smallerst radius is 1 R1 = R10 = m. 32 Array elements n each ring are equally distrubuted. Instead f starting all the 1 st element f each ring at θ = 0, we use the fllwing scheme t frm a mre symmetric structure fr the ring array. This will reduce the fluctuatin in beampattern when signal s DOA is rtated in the plane where the ring array lies. Table 3: Summary f ring radii Ring Radius R 1 2 R1 3 R1 4 R1 6 R1 8 R1 12 R1 16 R1 24 R1 32 Number f Elements The 1 st element f the 1 st ring is placed at θ = 0 and the remaining elements n that ring are anti-clckwise, equally placed alng the circle. Fr the 2 nd ring, its 1 st π element is placed at θ = and the remaining elements n that ring 4 are als anti-clckwise, equally placed alng the circle. And s n fr ther rings. Table 4 belw lists the lcatins (expressd by degree value f θ ) f each element n the rings. Table 4: Angular distributin f array elements Ring Ring Ring Ring Ring Ring Ring Ring Ring Ring R 1

11 Presentatins f Bird Strike 2003 Page 9 f Cnstructin f the Circular Array Based n the circular array cnfiguratin described in Sectin 5.1, we cnstructed a micrphne array as shwn in Fig. 17. The array bard was spliced by tw pieces f wd. Hles were drilled and micrphnes were put int them. Wd bard Micrphne elements Micrphne amplificatin bards. (a) Frnt view f the circular array; (b) Rear view f the array Figure 17 Finished circular micrphne array. It has a dimensin f 1.2 meter in diameter. 5.3 Data Acquisitin Electrnics Assciated With the Circular Array We built data acquisitin hardware fr the micrphne array. Figure 18 shws hw we implement the data acquisitin system. Basically, there are tw main parts in this system. One is the micrphne data acquisitin bard which is analg and perfrms A/D, the ther ne is the mther bard which is digital and sends digital data t cmputer thrugh Ethernet. Micrphne data acquisitin bard (a) Data acquisitin schematics (b) Hardware relatinships. Fig. 18 Data acquisitin system fr the circular micrphne array. Figure 19 is an verview f the micrphne array analg bard. There are 8 micrphnes n this bard. We need t have at least 64 micrphnes in ur applicatin, s there will be 8 analg bards in this system. There are three circuit parts n this bard: The micrphne amplificatin stage. The digitalizatin stage. The mtherbard cnnectin stage.

12 Presentatins f Bird Strike 2003 Page 10 f 11 Figure 19 A finished micrphne bard. Figure 20 The verview f the mtherbard Mtherbard In Fig. 20, the FPGA is the gathering part f the data acquisitin system. The cables at the tp f the picture are the data cllectin cables cnnected t the 8 analg bards. The red DIP is here t fix the MAC address f the micrphne array. The LEDs are t give the status f the micrphne array. 6. Experimental Results 6.1 DOA Estimatin Results The micrphne dish system was fully tested by taking many measurements in ur labratry as well as in an pen field (parking lt). Table 4 summarizes ne experiment in the pen field. The experimental setup invlved micrphne dish and the data acquisitin system t cllect the data, tw laptps displaying the sund surces. One surce emulated the bird and the ther ne emulated an aircraft. The distance between the surces and the micrphne dish was abut 40 ft. It can be seen that there are less than 10 % f estimatin errr in the θ directin and 2% f errr in the φ directin. There are three majr surces f errrs. One is that the distance between the surces and the micrphne dish is still nt lng enugh. Hence the micrphne dish can nt be treated as a pint. The secnd reasn is that distance measurements were dne manually and may have sme inherent errrs. The third reasn is that the number f elements in the dish is nly 64. The angular reslutin will be much better if mre than 100 elements are used in the array. Table 4: Summary f DOA estimatin and the expected DOA values DOA Estimatin Expected DOA 6.2 Beamfrmer Outputs θ φ θ φ Interference Bird Once the DOAs are estimated, beamfrming algrithm eliminated the effects f interference and cllected a clean bird signal. Figure 21 shws signals befre and after the beamfrming. It can be seen the befre beamfrming, the signals were very nisy. Hwever, after beamfrming, the signals were very clean. We culd nt hear any interference and backgrund nise after the beamfrming. Figure 21 Signals befre beamfrming and after beamfrming.

13 Presentatins f Bird Strike 2003 Page 11 f 11 Cnclusins and Future Research Directins There are three majr research directins: Reduce the beamwidth by increasing the number f micrphnes In this prttype system, we have nly 64 elements. In rder t reduce the beamwidth, we need t increase the number f elements t abut 200. This is nt easy as each micrphne is digitized at a sampling rate f 22 khz. Als the memry requirement will be huge in real-time system Integrate hardware and sftware t prduce a real-time bird mnitring system As can be seen frm the previus sectins, there are many sftware and hardware elements. The integratin f these cmpnents is very challenging. Field tests Once the integratin is dne, we will bring the system t the airprts fr tests and evaluatins. Acknwledgements This research was supprted by the Air Frce Office f Scientific Research under cntract F C References [1] H.Wang and M. Kaveh, Cherent signal-subspace prcessing fr the detectin and estimatin f angles f arrival f multiple wide-band surces, IEEE trans. Acust., Speech, Signal Prcessing, vl. ASSO-33, pp , Oct [2] K M. Burckley and L. Griffiths, Brad-band signal-subspace spatial-spectrum(bass-ale) estimatin, IEEE trans. Acust., Speech, Signal Prcessing, vl. 36, pp , July [3] C. Kwan, K. H, et al., Phase 1 Prgress Reprt 3 t the Air Frce, April, [4] Y. Li, K.C. H, and C. Kwan, Design f Brad-band Circular Ring Micrphne Array fr Speech Acquisitin in 3-D, accepted Int. Cnf. n Acustics, Speech, and Signal Prcessing, [5] D. A. Reynlds and R. C. Rse, Rbust Text-Independent Speaker Verificatin Using Gaussian Mixture Speaker Mdels, IEEE Trans. Speech and Audi Prcessing, vl. 3, n. 1, 1995.