Integrating Smooth Motion Assumptions with Ransac-Based Sound Source Localization

The momentary localization of a single sound source, in an environment with microphones distributed at known positions, can be done with multilateration, using time-difference-of-arrival (TDOA) estimates. However, TDOA estimates from cross-correlation are noisy in real environments, and a robust multilateration method must handle outliers. Assuming constraints on the smoothness of movement over time, the location estimate can be improved for both stationary and moving sources. In this work, the smooth motion assumption is explored in different stages of a RANSAC-based (Random Sample Consensus) implementation. The evaluation is done on real recordings from the public LuViRA dataset, giving the first 3D baseline result on the dataset. Each of the proposed steps is shown to reduce the localization error compared to the benchmark method.