Internet of Wings
Artificial intelligence is extending the battery life of bio-loggers and collecting more useful observations of animal behavior.
Understanding the behavior, physiology, and social interactions of many animals poses significant challenges. In some cases, the animals may be a bit on the shy side, and the act of observing changes their behavior. Or it may be that the behaviors that we wish to observe happen in remote, or otherwise inaccessible, locations that make observation impractical.
These issues have been addressed technologically with a technique called bio-logging, in which cameras or other sensors are mounted on animals. The animals are then freed so that they can go about their normal business while the sensors report data back to the researchers. Bio-logging is not without its own challenges, however. Many animals we wish to observe are quite small, which means the corresponding electronics must also be small — and for power hungry sensors like cameras, that means a short battery life.
A team of Japanese researchers recognized that bio-logging would be able to capture much more interesting data if the power hungry sensors were only turned on when they were needed. To this end, they have recently published a paper proposing a method that makes use of low-power sensors and artificial intelligence to strategically determine when to turn on the higher power draw sensors.
A machine learning model is pre-trained using data from accelerometers and GPS receivers. The model is deployed on a device powered by an ATmega328 microcontroller. When an activity of interest is recognized, using accelerometer and GPS data, the camera, or any other high power consumption devices are turned on. Device battery life was extended tenfold using this new method, as compared with traditional always-on bio-loggers.
The researchers collected 95 videos of black-tailed gulls using traditional bio-logging devices and found that only two contained any of the desired target behaviors. Another set of 184 videos were collected using the new method. In this case, 55 contained target behavior. While there is still room for improvement, this new method is already showing itself to be a useful path forward.
