A framework for energy-efficient equine activity recognition with leg accelerometers

Abstract

Automated behavioral detection and classification through sensors can enhance the horses’ health and welfare. Since monitoring needs to be carried out continuously, an energy-efficient method is needed. The number of logging axes, sampling rate, and selected features of accelerometer data not only have a significant impact on classification accuracy in activity recognition but also on the sensors’ energy needs. Three models are designed for detecting horses’ activities namely, a Random Forest classifier (RF), a Convolutional Neural Network (CNN) and a hybrid CNN, i.e. a CNN fused with statistical features that retain knowledge about the global time series form. The models are validated using an experimental dataset obtained from six different horses each performing seven different activities. The results indicate that using one leg accelerometer data is sufficient for high classification accuracies (>98.6%) for the three models. The hybrid CNN substantially improves over the RF and CNN at a sampling rate of 5 Hz with an increase in accuracy of 1.88% and 2.79%, respectively. The hybrid CNN is capable of excellent performance, detecting nearly 99.59% of the behaviours at 10 Hz. The experiments show that the CNN and hybrid CNN use as much as 17.2 and 13.5 times less energy respectively, than the RF-based method. The experimental results showed that, although the recognition rate of the proposed optimized hybrid CNN model is similar to that of the original hybrid CNN model, the former requires only 6% of the Multiply-Accumulate (MAC) operations. For automatic detection of the horses’ behavior those results suggest using one leg accelerometer data sampled at 10 Hz classified by an optimized hybrid CNN.