Automated extraction of dolphin whistles—A sequential Monte Carlo probability hypothesis density approach

Abstract

The need for automated methods to detect and extract marine mammal vocalizations from acoustic data has increased in the last few decades due to the increased availability of long-term recording systems. Automated dolphin whistle extraction represents a challenging problem due to the time-varying number of overlapping whistles present in, potentially, noisy recordings. Typical methods utilize image processing techniques or single target tracking, but often result in fragmentation of whistle contours and/or partial whistle detection. This study casts the problem into a more general statistical multi-target tracking framework, and uses the probability hypothesis density (PHD) filter as a practical approximation to the optimal Bayesian multi-target filter. In particular, a particle version, referred to as a Sequential Monte Carlo PHD (SMC-PHD) filter, is adapted for frequency tracking and specific models are developed for this application. Based on these models, two versions of the SMC-PHD filter are proposed and their performance is investigated on an extensive real-world dataset of dolphin acoustic recordings. The proposed filters are shown to be efficient tools for automated extraction of whistles, suitable for real-time implementation.

Tracking time differences of arrivals of multiple sound sources in the presence of clutter and missed detections

“…Moreover, it is often useful to assume that not every measurement is equally likely to initialize a newborn target. 32,33 In this study, the measurements that are more likely to initialize a newborn target are the ones with a higher amplitude.…”

“…The target amplitude likelihood g a ðajdÞ is defined in Sec. III B 4, and g k ðzjxÞ is g k ðzjxÞ ¼ N ðz; ẑ; SÞ; (7) where N ðÁ; ẑ; SÞ denotes a Gaussian density function with mean ẑ and covariance S. Further, measurements are obtained given a certain probability of detection, p D , which can be trained from the data 32,33 when no prior knowledge is available or it can be approximated analytically. In sonar and radar, 35,49 the probability of detection given a threshold k can be modeled as the function of the SNR only, p k D ðdÞ.…”

“…54. These are typically used to evaluate the whistle tracking performance 32,33,40,54 and measure the quantity and quality of the detected tracks. Briefly, precision measures the percentage of the detections that are correct, recall measures the percentage of the expected detections that are retrieved, coverage measures the average percentage of a ground truth track that is detected, fragmentation measures the average number of detections per ground truth track, and mean deviation measures the average deviation between the ground truth track and its corresponding detection(s).…”

“…24 and 27 The PHD filter has been previously used to track multiple targets in sonar 28,29 and video, 30 multiple speakers in reverberant environments, 25,31 and multiple overlapping dolphin whistles in spectrograms. 32,33 Although the problems of whistle tracking 32,33 and TDOA tracking appear similar, there are key differences that require the filters to be derived and defined differently. These differences pertain to what is measured: a whistle is produced by only one animal, whereas a TDOA track is produced by a group of likely multiple animals, resulting in a much higher track variance in the TDOA domain.…”

“…Meanwhile, missed detections in TDOA tracking can also be the result of animals not vocalizing during certain periods of time, resulting in increased track fragmentation due to multiple, prolonged periods of missed detections. Thus, although whistle tracking filters can extract meaningful tracks with less informative measurements based on only one feature (in this case, frequency), 32,33 TDOA tracking filters require more informative measurements with additional features for successful tracking.…”