On the Benefits of Deep Convolutional Neural Networks on Animal Activity Recognition

Monitoring the behavior of animals (e.g., eating habits) can lead to conclusions regarding animal's welfare. To achieve this, remote monitoring of animal activity with the aid of inertial sensors and use of machine learning algorithms over the collected data can be used. However, these algorithms rely on handcrafted features extracted by statistical or heuristic functions over raw motion data. To this purpose, we employ deep Convolutional Neural Networks (ConvNets) for activity recognition of livestock animals, such as goats and horses. We investigate the potential gains of ConvNets compared with other machine learning algorithms of the literature, which are about 12.5% greater accuracy and more than 7% higher F1-score. Moreover, we designate the advantages of late sensor fusion (2D convolution) and also show that an increase on the number of filters on each layer does not necessarily lead to a greater classification accuracy. To this end, we benchmark various ConvNet architectures and demonstrate the role of hyperparameter tuning to optimize the overall accuracy. To the best of our knowledge, ConvNets are employed for animal activity recognition here for the first time.