By recognizing the movements of animals in the wild, using a sensor attached to their body, it may well be possible to detect if poachers are nearby. These ‘animal activity recognition’ sensors can also help in biodiversity research or cattle management. Researcher Jacob Kamminga of the University of Twente developed a motion sensor with built-in intelligence for recognizing motion patterns of a wide range of animals. The sensor consumes very little energy and it is prepared for harsh conditions.
To this day, many elephants are killed for their ivory and rhinos for the alleged healing properties of their horn. Although stricter rules cause some improvement, far too many wild animals are victims of poaching. If you would be able to recognize the movements of animals, you might be able to detect their response to the presence of humans. Satellite, GPS data and remote sensing already proof to be valuable in this. Data coming from sensors that are directly connected to the animal’s body may have substantial added value.
Kamminga did research on the type of measurements needed for this type of recognition, as well as the built-in intelligence. A remarkable conclusion of his work is that in most cases, a single sensor is sufficient, an accelerometer. “I added a gyroscope as well, that measures rotation. This can make it some more accurate, but this comes with a prize. It consumes 100 times more energy than the accelerometer. In most cases, just the accelerometer is accurate enough”, says Kamminga. Replacing a battery every once in a while, is not an option, so energy efficiency is one of the top priorities.
The movements of the sensor will be recognized by the in-built intelligence. An option would be to train the system with many possible movements, but this is very labour-intensive and you would need to do it for every animal species. This is called ‘labeled data’, Kamminga shows that the sensor intelligence can operate using mostly ‘unlabeled data’, using just a small set of labeled data as a basis. The actual recognition could be done using relatively simple decision trees, but nowadays it is also possible to include a ‘deep-learning’ neural network in the sensor. This improves the flexibility of the system. Kamminga already analyzed the movement patterns of goats, sheep and horses.
After measuring and classification, the data has to be sent using a mobile network or a satellite connection. To avoid using too much energy for this, the sensor only transmits data when there is a change. Rough natural circumstances may form another challenge: if the sensor band moves, the data should still be accurate. Kamminga developed a solution for that as well.
This type of animal activity data can also be used for analyzing biodiversity in a certain area. Do animals in this have enough food and freedom of movement, also when the ? These are typically questions for the UT-Faculty ITC – for Geoinformation Science and Earth Observation. Professor Andrew Skidmore was involved in Kamminga’s work: ‘Linking wild animal movement recorded using sensors with remotely sensed imagery and GIS models is promising technology to better understand the ecological requirements of species, as well as inform management and policy decisions with conservation outcomes and biodiversity.”
Kamminga did his resrarch in the Pervasive Systems group of Professor Paul Havinga. This group is specialized in small, energy efficient, often autonomously communicating sensors for a wide range of applications, including monitoring water quality or the maintenance status of infrastructural works like bridges and tunnels.
Jacob Kamminga recently successfully defended his PhD thesis ‘Hiding in the deep – online animal recognition using motion sensors and machine learning’.