extended kalman filter with probabilistic data …€¦ · soumitro chakrabarty1, konrad...

1
1. Introduction 2. Problem Formulation 3. Candidate Selection 4. Extended Kalman Filter with PDA 5. Experimental Results 6. Conclusion EXTENDED KALMAN FILTER WITH PROBABILISTIC DATA ASSOCIATION FOR MULTIPLE NON-CONCURRENT SPEAKER LOCALIZATION IN REVERBERANT ENVIRONMENTS E-Mail: [email protected] 1 International Audio Laboratories, Erlangen, Germany 2 Fraunhofer Institute for Integrated Circuits (IIS), Erlangen, Germany Soumitro Chakrabarty 1 , Konrad Kowlaczyk 2 , Maja Taseska 1 and Emanuël A.P. Habets 1,2 Aim - Estimate the source position at each time step within an acoustic environment using signals acquired by distributed microphones Conventional Method - Use broadband time-difference-of-arrival (TDOA) estimates as measurements within a state-space model based probabilistic framework Proposed Method - Use narrowband direction-of-arrival (DOA) estimates selected based on magnitude squared coherence between microphone signals as measurements within an extended kalman filter (EKF) framework with probabilistic data association (PDA) direct sound + early reflections late reverberation noise signal array index frequency index time index reflection order microphone index Signal model in TF domain state variable process noise measurement noise measurements State-space model array center Measurement function • Given the model, compute the state estimate Measurements Multiple narrowband DOA estimates per array • Measurement for m-th array - number of candidates ¯ θ (m,2) n ¯ θ (m,3) n ¯ θ (m,1) n ¯ θ (m,4) n ˜ θ (m,3) n ˜ θ (m,2) n ˜ θ (m,4) n ˜ θ (m,1) n Number of Occurrences Direction of Arrival Candidates selected based on magnitude squared coherence (MSC) • Search for local maxima in the histogram of all narrowband DOA estimates: • Frequency indices of the candidates obtained by: • Selected candidates: Extended Kalman Filter • Kalman Gain: • Measurement prediction covariance: • Jacobian Matrix: • Innovation vector: Probabilistic Data Association Modifications Innovation vector for multiple candidates: a posteriori probability Computed: State estimate: State covariance: event: i-th measurement is from true source position accounts for measurements not corresponding to true source position each element 0 0.5 1 1.5 2 2.5 3 3.5 4 -0.02 -0.01 0 0.01 0.02 Actual Path PDA with ESPRIT ESPRIT SRP-PHAT Amplitude 0 0.5 1 1.5 2 2.5 3 3.5 4 3 4 5 6 7 8 X [m ] 0 0.5 1 1.5 2 2.5 3 3.5 4 3 4 5 6 7 8 Y[ m] Time [s] 0 0.1 0.2 0.3 0.4 0.5 0.6 0.2 0.3 0.4 0.5 0.6 0.7 Reverberation Time[s] RMSE[m] 0 10 20 30 40 50 60 0.2 0.4 0.6 0.8 1 SNR[dB] RMSE[m] PDA with ESPRIT ESPRIT SRP- PHA T Simulation setup STFT: 32 ms, 50% DOA estimation method: ESPRIT Figure: Microphone signal (top), and tracked position along X and Y dimensions (middle and bottom, respectively) for RT 60 = 0.2s and SNR = 20dB. Figure: Simulation setup Figure: Position RMSE for varying reverberation times and constant SNR = 30dB (top) and varying SNR levels with a constant RT 60 = 0.3s (bottom). broadband speech presence probability process noise covariance measurement noise covariance • A method for acoustic source tracking using EKF with PDA was proposed • Multiple narrowband DOA estimates selected based on the MSC were used as measurement candidates • The proposed method yields an improved tracking performance over single-candidate based methods estimated narrowband DOAs Length 9.0 m Width 9.0 m [7.5m, 4m] [5.2m, 5.4m] Source 1 Trajectory [3m, 3.5m] [4.4m, 5.5m] Source 2 Trajectory Mic.Array 1 Mic.Array2 6cm 6cm 2m 2m 4.5m 4.5m

Upload: others

Post on 28-Jun-2020

0 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: EXTENDED KALMAN FILTER WITH PROBABILISTIC DATA …€¦ · Soumitro Chakrabarty1, Konrad Kowlaczyk2, Maja Taseska1 and Emanuël A.P. Habets1,2 • Aim - Estimate the source position

1. Introduction

2. Problem Formulation

3. Candidate Selection

4. Extended Kalman Filter with PDA

5. Experimental Results

6. Conclusion

EXTENDED KALMAN FILTER WITH PROBABILISTIC DATA ASSOCIATION FOR MULTIPLE NON-CONCURRENT SPEAKER LOCALIZATION IN REVERBERANT ENVIRONMENTS

E-Mail: [email protected]

1International Audio Laboratories, Erlangen, Germany 2Fraunhofer Institute for Integrated Circuits (IIS), Erlangen, Germany Soumitro Chakrabarty1, Konrad Kowlaczyk2, Maja Taseska1 and Emanuël A.P. Habets1,2

• Aim - Estimate the source position at each time step within an acoustic environment using signals acquired by distributed microphones• Conventional Method - Use broadband time-di�erence-of-arrival (TDOA) estimates as measurements within a state-space model based probabilistic framework

• Proposed Method - Use narrowband direction-of-arrival (DOA) estimates selected based on magnitude squared coherence between microphone signals as measurements within an extended kalman �lter (EKF) framework with probabilistic data association (PDA)

direct sound + early re�ections late reverberation noise signal array index

frequency indextime index re�ection order microphone index

Signal model in TF domain

state variable process noise

measurement noisemeasurements

State-space model

array center

Measurement function

• Given the model, compute the state estimate

Measurements• Multiple narrowband DOA estimates per array• Measurement for m-th array

- number of candidates

θ̄(m,2)nθ̄(m,3)

n θ̄(m,1)n θ̄(m,4)

n

θ̃(m,3)n

θ̃(m,2)n

θ̃(m,4)n

θ̃(m,1)n

Number

ofOccurrences

Direction of Arrival

• Candidates selected based on magnitude squared coherence (MSC)• Search for local maxima in the histogram of all narrowband DOA estimates:• Frequency indices of the candidates obtained by:

• Selected candidates:

Extended Kalman Filter• Kalman Gain:

• Measurement prediction covariance:

• Jacobian Matrix:

• Innovation vector:

Probabilistic Data Association Modi�cations• Innovation vector for multiple candidates: a posteriori probability

• Computed:

• State estimate:

• State covariance:

event: i-th measurement is from true source position

accounts for measurements not corresponding to true source position

each element

0 0.5 1 1.5 2 2.5 3 3.5 4

−0.02

−0.01

0

0.01

0.02

Actual Path PDA with ESPRIT ESPRIT SRP−PHAT

Am

plit

ud

e

0 0.5 1 1.5 2 2.5 3 3.5 4

3

4

5

6

7

8

X [

m]

0 0.5 1 1.5 2 2.5 3 3.5 4

3

4

5

6

7

8

Y[m

]

Time [s]

0 0.1 0.2 0.3 0.4 0.5 0.6

0.2

0.3

0.4

0.5

0.6

0.7

Reverberation Time[s]

RM

SE

[m]

0 10 20 30 40 50 600.2

0.4

0.6

0.8

1

SNR[dB]

RM

SE

[m]

PDA with ESPRIT ESPRIT SRP−PHATSimulation setup•

• STFT: 32 ms, 50% •

• DOA estimation method: ESPRIT

Figure: Microphone signal (top), and tracked position along X and Y dimensions (middle and bottom, respectively) for RT60 = 0.2s and SNR = 20dB.

Figure: Simulation setup

Figure: Position RMSE for varying reverberation times and constant SNR = 30dB (top) and varying SNR levels with a constant RT60 = 0.3s (bottom).

broadband speech presence probability

process noise covariance

measurement noise covariance

• A method for acoustic source tracking using EKF with PDA was proposed

• Multiple narrowband DOA estimates selected based on the MSC were used as measurement candidates• The proposed method yields an improved tracking performance over single-candidate based methods

estimated narrowband DOAs

Length 9.0 m

Width 9.0 m

[7.5m, 4m]

[5.2m, 5.4m]

Source 1 Trajectory

[3m, 3.5m]

[4.4m, 5.5m]Source 2 Trajectory

Mic.Array 1

Mic

.Arr

ay2

6cm

6cm

2m

2m

4.5m

4.5m