Cow behavioural activities classification by convolutional neural networks

As has been shown in several studies, behavioural activities of animals provide im- portant parameters for the evaluation of their health and welfare. In recent years the use of wearable sensors to record animal activity has become an important practice especially in extensive farms, where there is an infrequent farmer-to-animal contact. Accelerometers allow the measurements of movements of a body in space and are currently very popular in the zootechnical field for monitoring livestock, as they can be worn without being invasive for animals. The objective of this work was to address the task of classifying cow behavioural activities using a Convolutional Neural Net- work (CNN) to discriminate five classes: feeding in standing position, feeding while walking, walking, lying and rumination in lying position. To carry out this study, ac- celerometer data were acquired at 4 Hz by customized devices attached to cow collars, containing triaxial accelerometers. The acquired samples were previously labelled by using video-labelling, and then grouped in windows and pre-processed. The developed model is a CNN with 1D convolutions, which receives as input a 3-channel batch of windows, where channels are the three axes. Firstly, the model processes the data in parallel branches, which analyse different combination of channels. Features maps ob- tained from each branch are concatenated and provided as input to another cascade of convolutional layers. The model finally returns the prediction of the behavioural class. Our approach classified the five behavioural classes with an average F1 score of 81.51%. When merging the feeding in standing position and feeding while walking classes, F1 score reached 90.01%.