Method estimates relative position of two microphone arrays with one source
Self-Localization of Distributed Microphone Arrays Using One Source and Acoustic Reflections
Sound
Summary
Determining where multiple microphones are in a room can be tricky, especially when they are in different places and angles. This paper presents a way to figure out the positions and orientations of two separate microphone groups using sound from just one source and the way the sound echoes in the room. The authors introduce a new technique called ARCTIC that matches the reflected sound points captured by each microphone group, even when the data is noisy or incomplete. Their tests show this method works well in real rooms, accurately locating the microphones within a few centimeters and degrees.
What this means in practice
- •For audio engineers: Determine precise placement and orientation of distributed microphone arrays using a single sound source and room reflections without prior layout knowledge.
- •For acoustic system installers: Align multiple microphone arrays more efficiently in complex indoor environments by automatically estimating their relative positions from acoustic signals.
Authors
Thomas Guffroy, Jean-Daniel Pascal Prieto, Cédric Foy, Antoine Deleforge
Abstract
We present a method to estimate the relative position and orientation of two non-coplanar microphone arrays using a room impulse response (RIR) from a single source. The approach exploits early acoustic reflections, retrieved by each array in the form of a 3D image-source point cloud, using a recently-proposed image-source localization technique. The task is then reduced to aligning the two unlabeled point clouds, which is made difficult by their inherent symmetries and by spurious, missing and noisy estimates. To this aim we introduce ARCTIC, a point-cloud registration algorithm that performs a ranked search over congruent triangles anchored on the true source. Simulated and real experiments show that ARCTIC strongly outperforms two baselines, achieving mean angular and radial self-localization errors below 5 and 10 cm across 36 real RIRs measured in 6 different rooms.