A Sound Processing Pipeline for Robust Feature Extraction to Detect Elephant Rumbles

Abstract

Abstract A signi cant number of human and elephant lives have been lost due to the humanelephant con ict in Sri Lanka. To save lives of humans and elephants, it is therefore essential to minimize encounters between them. An early warning system, which detects and localize the presence of elephants through their infrasonic emissions is a viable solution to mitigate such con icts. The high cost of infrasonic detectors is an important challenge to the realworld deployment of such localization systems, in particular in developing countries where the human-elephant con ict occurs. ElOC is a system developed as a part of inventing a lowcost automatic elephant detection and localization system. Which is capable of localizing the infrasonic source within a ten-meter accuracy. In this dissertation, a novel approach is proposed to extend the ELOC to identify the elephant infrasound automatically. The novelty of this approach is the capability of distinguishing the infrasonic emissions from the elephant on top of the low-cost, resource-limited hardware platform of the ELOC. The approach rst applies a sequence of operations to reduce the e ect of noise contained in the infrasonic signal captured by the ELOC node. Then the spectral features of the infrasonic signal are extracted with wavelet-based signal reconstruction to analyze the signal more precisely. Finally, the extracted features are feed to the pre-trained classi er to distinguish the infrasound emissions from the elephants. This study is able to classify elephant rumbles with an accuracy of 82%. Thereby the proposed approach exhibits promising results in elephant detection and capable of operating on the resource-limited hardware platform of the ELOC. This study also contributes to the domain of digital signal processing since the study is the rst attempt of wavelet-based feature extraction in the domain of infrasound elephant rumble detection.