SLIDE 1 Localization of Simultaneous Moving Sound Sources for Mobile Robot Using a Frequency-Domain Steered Beamformer Approach
Jean-Marc Valin, François Michaud, Brahim Hadjou, Jean Rouat Department of Electrical Engineering and Computer Engineering Université de Sherbrooke, Québec, Canada Jean-Marc.Valin@USherbrooke.ca
SLIDE 2
Approaches to Sound Source Localization
Binaural audition
T wo microphones Interaural phase difgerence Interaural intensity difgerence Imitate human auditory system
Microphone array audition
Larger number of microphones Phase difgerence only Increased redundancy compensating for high complexity of human audition
SLIDE 3
Approach Overview
Sounds arrive at microphones with difgerent delays (depending on distance)
Hypothesis: point sound sources
Steered beamformer: scans all directions for energy peaks Probabilistic post-processing: applies Bayesian inference
SLIDE 4
Steered Beamformer
Delay-and-sum beamformer Beamformer energy
SLIDE 5
Frequency Domain Computation
SLIDE 6
Spectral Weighting
Cross-correlation peaks are very wide
Poor angular accuracy Overlap between close sources
Solution: spectral weighting
Whiten spectrum Give less weight to noisy regions of spectrum
SLIDE 7
Search
Set of possible directions of arrival represented as sphere Defjning a homogeneous grid
Recursive subdivision of icosahedron Resulting grid with 2562 points
SLIDE 8
Search
Find directions with highest energy
SLIDE 9
Bayesian Post-fjlter
Data from beamformer is noisy Express localization in terms of source probability of presence Probability computed for each grid point Use Bayes' rule to compute probability using past and present observations
SLIDE 10
Bayesian Post-fjlter
beamformer probability a priori probability combined probability
SLIDE 11
Estimator Combination
All previous steps computed twice
Short frames (~40 ms) Medium frames (~200 ms)
Need to combine both estimators
Estimators are not independent
Weighted geometric average of the dependent case and the independent case:
SLIDE 12
Results
Detection accuracy over distance
Difgerent sounds Rate of detection(#detections / #occurences)
SLIDE 13
Results (2 moving speakers)
time azimuth
SLIDE 14
Results (2 moving speakers)
time azimuth
SLIDE 15
Results (4 moving speakers)
time azimuth
SLIDE 16
Results (moving robot)
Localization in 3D
time azimuth elevation
SLIDE 17 Conclusion
Robust localization of sound sources
Moving sources or robot Up to 4 simultaneous sources reliably Reliable detection up to 5 meters
T wo-step method
Steered beamformer Bayesian post-fjlter
Related work
T racking sources over time Separating sound
separated
SLIDE 18
Questions?
SLIDE 19
Search (cont.)
1) Steered beamformer direction search
Finding the direction with highest energy
SLIDE 20
Bayesian Post-fjlter (cont.)
Beamformer assigns instantaneous probability for each grid point A priori probability assuming a Markov process Current probability
SLIDE 21
Results (7 sources)
SLIDE 22
Search (cont.)
2) Complete search
Finding all sources
