Automatic Analysis of Elephant Vocalizations

The decline of habitat for elephants as a result of expanding human activity combined with increasing elephant numbers in spatially separated clusters of conservation areas is a serious conservation problem in Africa. Nearly 80% of the distributional range of elephants in Southern Africa stretches beyond the borders of officially protected areas. This fact leads to deadly conflicts between humans and elephants. We are a team of biologists and computer scientists who want to alleviate these conflicts by the development of early warning and information systems for humans living near the corridors who regularly get into serious conflict with traveling elephants. Such systems require the robust recognition and categorization of elephant vocalizations. Today, no system exists that fulfills the requirements for automatic recognition of elephant vocalizations under natural conditions. The goals of the proposed project are (i) to investigate the complex vocal communication system of elephants and (ii) to develop automatic techniques for the analysis of elephant calls. The acquired knowledge and the automatic techniques will form the basis for a future autonomous warning and information system. Elephants make extensive use of powerful infrasonic calls which travel distances of up to several kilometers. This qualifies the elephant as a perfect model species for acoustic observation since it is possible to detect elephants even if they are out of sight. We will incorporate visual information into the automatic analyses where improvements can be expected (e.g. at open habitats such as waterholes). We will investigate and evaluate among others automatic statistical methods to detect elephants, estimate group-size from vocalization rates, and classify call types automatically. Another focus will be the manual and automatic analysis of age and gender of the caller due to characteristics in call structure. This will allow us to determine the demography of families and populations. We additionally aim at defining family and bond group specific acoustic signatures and vocal dialects within the study population. The project is proposed for an initial duration of three years and will be conducted by the Department for Cognitive Biology of the University of Vienna in collaboration with the Interactive Media Systems Group, Institute of Software Technology and Interactive Systems, Vienna University of Technology. One 50% post-doc and two 60% doctoral students shall be employed by the project partners. There is a clear definition of responsibilities and exchange of results between the project partners. At the end of the project, a large set of fully annotated audio and video material will be available together with statistical evaluations of the material. Additionally, novel analysis techniques will be available that are able to robustly recognize and classify elephant calls from data obtained under natural conditions.

The decline of habitat for elephants as a result of expanding human activity is a serious conservation problem in Africa and Asia. This fact leads to an increase of the human-elephant conflict that seriously threatens the long-term survival of elephants. The project team of biologists and computer scientists wants to alleviate these conflicts by the development of early warning and information systems for humans who regularly get into serious conflict with elephants. Such systems require the robust recognition and categorization of elephant vocalizations. The main goals of the project entitled Automatic Analysis of Elephant Vocalizations were (1) to investigate the complex vocalizations of elephants and (2) to develop automatic techniques for the analysis of these elephant calls. Elephants make extensive use of powerful infrasonic calls (rumbles) that travel distances of up to several kilometers. This qualifies the elephant as a perfect model species for acoustic observation since it is possible to detect elephants even if they are out of sight. As the main overall project result we present a method for the automated detection of elephant vocalizations (focusing on the low-frequency rumbles) that is robust to the diverse noise sources present in the field. We evaluated the method on data recorded under natural field conditions. The proposed method outperformed existing approaches and accurately detected elephant rumbles. We also investigated natural rumble variability. We found that rumbles vary considerably in acoustic structure due, among others, to differences in sound production mechanisms (oral versus nasal sound emission) and maturational processes (characteristics for age groups). Automatic classification distinguished these structural rumble variants with high accuracy. During our project we further investigated the capabilities of visual elephant detection and developed a first method for visually detecting and tracking elephants in wildlife videos. Our experiments on visual detection confirmed that tracking elephants in wildlife videos is feasible and particularly robust at near distances. We were able to detect and track elephants of different sizes and postures and in different activities. We conclude that a combined acoustic and visual elephant detector is a promising solution because both techniques potentially compensate their weaknesses, which might lead to a more reliable detection in future. We clearly identified the most critical challenges and necessary improvements in order to transfer the promising fundamental results into practical implementation, and conclude that our findings can form the basis for a future automated early warning system for elephants.