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

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