On the Development of a Wearable Animal Monitor

Abstract

Animal monitoring is a task traditionally performed by pastoralists, as a way of ensuring the safety and well-being of animals; a tremendously arduous and lonely task, it requires long walks and extended periods of contact with the animals. The Internet of Things and the possibility of applying sensors to different kinds of devices, in particular the use of wearable sensors, has proven not only to be less invasive to the animals, but also to have a low cost and to be quite efficient. The present work analyses the most impactful monitored features in the behavior learning process and their learning results. It especially addresses the impact of a gyroscope, which heavily influences the cost of the collar. Based on the chosen set of sensors, a learning model is subsequently established, and the learning outcomes are analyzed. Finally, the animal behavior prediction capability of the learning model (which was based on the sensed data of adult animals) is additionally subjected and evaluated in a scenario featuring younger animals. Results suggest that not only is it possible to accurately classify these behaviors (with a balanced accuracy around 91%), but that removing the gyroscope can be advantageous. Results additionally show a positive contribution of the thermometer in behavior identification but evidences the need for further confirmation in future work, considering different seasons of different years and scenarios including more diverse animals' behavior.

Keywords: animal monitoring; behavior prediction; machine learning; wearable sensors.

Figure 1

Monitoring platform: (a) monitoring collar, (b) android app that gathers collar data and video records animal behavior.

Figure 2

Model’s performance as function of depth: (a) tree depth influence in balanced accuracy, and (b) tree depth influence in number of leaves.

Figure 3

Confusion matrix of best model (AllFeatures). The different cell colors represent the classification accuracy; the more accurate the classification of a state, the darker the cell is represented.

Figure 4

Balanced accuracy evolution with tree depth.

Figure 5

Tree leaf count evolution with tree depth.

Figure 6

Confusion matrix of the best model (InertFeatures + TMP). The different cell colors represent the classification accuracy; the more accurate the classification of a state, the darker the cell is represented.

Figure 7

Performance of Best Model (InertFeatures + TMP) against Lamb. The different cell colors represent the classification accuracy; the more accurate the classification of a state, the darker the cell is represented.