Confidence measure based on higher-order statistics for spatial feature in particle filtering

This paper proposes to estimate the reliability of each direction-of-arrival (DOA) estimate explicitly. Observed spatial features, which give DOA estimates, are usually distorted by acoustical noise and reverberation. Then, distorted spatial features are filtered out by particle filters with a system model, which describes sound source dynamics on spatial state space. After particle filtering is carried out, set of the weighted particles can be considered as the discrete approximation of the true spatial feature. It is reasonable that particles with large weights concentrate around the true DOA on spatial state space. In this paper, crest factor, skewness, and kurtosis are calculated based on the second, third, and forth order moments from the particle set for representing how particles concentrate simply. It is confirmed that the skewness confidence measure is the most suitable for estimating DOA reliability under both directional and distributed noise conditions.