Wild elephants and villagers near nature reserves are often caught up in the human-wildlife conflict. Even in countries where the elephant is sacred, this may lead to elephants being killed by villagers. Out of protection, by accident (as on the photo), or out of revanche.
Elephant rumbles may be detected at distances from 5 to 10km, depending amongst other things on vegetation. Being able to detect them long before the enter the village gives both the villagers and the elephants a better chance of survival.
To this end we are developing an acoustic elephant detection and monitoring system. It requires knowledge of elephant behaviour, knowledge of acoustic signal processing and analysis, and tough engineering skills to get a fully functional system that survives the forces of nature. To complement our team we sought collaboration with various esteemed researchers, specialised in elephant vocalisations and machine listening:
- Angela Stoeger, from the Department of Cognitive Biology, University of Vienna, Austria,
- Shermin de Silva, from Trunks & Leaves Inc,
- Matthias Zeppelzauer, from the St. Pölten University of Applied Sciences,
- Karol Piczak, from the Warsaw University of Technology, and
- Ard Kuijpers, from M+P
The Wild Elephant Approaching Alerting system is based or our acoustic SERVAL sensor. The SERVAL sensor is connected to IoT sensor platform SCCSS, allowing us to further analyse the signals and send out alerts to all rangers and/or community members involved in a project.
We managed to develop the SERVAL-sensor. That is, a sensor system which can be trained to distinguish different types of sounds. To this end we use a combination of known physical sound characteristics and deep learning. The trained neural networks are being deployed on the sensor, which sends out alerts to the SCCSS-platform, from where it can be dispatched to early responders.
We have collected terabytes of elephant sounds samples. During this stage we received addition field data from:
- Peter Wrege, Director of the Elephant Listening Project at the Cornell Lab of Ornithology
- Blaise Droz, Documentaliste indépendant et Journaliste chez Journal du Jura
Ongoing: preparing the data for automated processing; experimenting with different models and configurations.
June 2017: With the help of Karol Piczak we managed to deploy the trained convolutional (neural) network on a Raspberry Pi and are now working on power optimization.
July 2017: with the help of M+P we have selected an affordable yet powerful off-the-shelf measurement microphone, which can be deployed in combination with our Raspberry Pi-solution.
September 2017: The Dutch branch of the World Wildlife Fund (WNF) is sponsoring part of the field-testing, which will be conducted in Africa. First, however, we have to further develop the algorithms.
June 2018: First live tests of SERVAL in Amsterdam. The sounds of the city jungle are quite different from the natural jungle, but close to home and easier to verify. The first results are very encouraging! Tested city sounds that can also be used to detect people in nature reserves, include engines, talking people, dog barks, and radio sounds (talkshows, music). In the mean time, we keep working on the elephant sounds recognition algorithm..