EleVoc

The Project

Asian elephants are an endangered species and human-elephant conflict poses a grave threat to their existence. Human-elephant conflict refers to the negative interactions between humans and elephants such as in electrocutions and crop-raiding. It has undesirable consequences for people and their resources as well as elephants and their habitats. Every year, more than 500 humans and 100 elephants are killed due to human-elephant conflict. A device using bio-acoustics and machine-learning is proposed to build an early warning system and determine the proximity and behavior of elephants by classifying elephant vocalizations. This early warning device indicates the presence of elephants in the proximity as well as whether they are likely to raid even when elephants are not visible due to darkness or thick foliage. This would help village administration and villagers to take necessary actions to curtail human-elephant conflict and prevent casualties. This device uses machine learning to detect when an elephant vocalizes and identify the type of vocalization - Chirp, Roar, Rumble, or Trumpet. Data from recordings of 147 vocalizations were annotated and pre-processed. A unique approach was taken to train machine-learning models to classify this data. Two levels of machine-learning models were trained hierarchically. The first level contained one machine learning model that classifies vocalizations into two categories - high frequency and low frequency. The second level contains two models that further sub-classify the vocalizations. Uniquely modified mel-scale filter banks were extracted from the vocalizations and used to train the multiple models. This two-level hierarchical-model approach achieved an accuracy of 96.88% for the first level and 98.00% and 75.13% for the second level models. The models run live on a Raspberry Pi along with a microphone and an alarm system. This early warning device raises an alarm and sends a message with location information when elephants are identified to be in the surroundings. This message is received by the village's emergency response team and they can give the appropriate guidance to villagers.

View this project on YouTube

Team Comments

As wildlife enthusiasts, my family and I visit wildlife sanctuaries and involve in wildlife conservation. In one such visit, we were charged by elephants even though we were only passing by in a jeep. This was my first introduction to conflict and I wanted to find a way to solve it.

The most difficult part was to build the machine learning model with good accuracy. I had to do multiple iterations to find a ML model with more than 80% accuracy. Later, I needed to find how to deploy them. Since RaspberryPi can run TFLite models, I then learnt how to deploy ML models on RPi.

If I had more time, I would add a location sensor to detect the exact location of the device and send location information as latitude and longitude along with the message sent to the emergency response team. Currently, I am only sending the device id mapped to the installation location.