Current demographic trends push livestock production toward more intensive farming while reducing environmental impact and maintaining a high grade of animal welfare. Under this new scenario, farmers need to be provided with tools to manage their farm more efficiently. Currently the health status of small ruminants is monitored by farmers using mostly visual inspection however this become ineffective for larger stock numbers. This piece of work presents a new machine learning algorithm based on autoregressive hidden Markov model to analyse triaxial accelerometer data to classify sheep behaviouralpatterns. The algorithm creates a model of animal’s activity directly from raw datafrom the sensors and was specifically design for running on embedded devices.This approach allows the model, when used on the initial healthy animal, to be automaticallytuned for each individual which subsequently provides a normal reference for abnormaldetection.Finally, a set of experiments were performed where a sheep was fitted with the collarprototype over two days. The animal under study was maintained in a shed with twoother sheep. The results of this experiment shows that algorithm achieved an average of93.40±0.09% accuracy when determining the animals’ activities using only the accelerometerdata which is comparable to more computing expensive classifying techniques.
|Number of pages||8|
|Publication status||Accepted/In press - 2017|
|Event||8th European Conference on Precision Livestock Farming 2017 - Nantes, France|
Duration: 12 Sep 2017 → 14 Sep 2017
|Conference||8th European Conference on Precision Livestock Farming 2017|
|Period||12/09/17 → 14/09/17|
- Smart collar
- activity monitoring
- hidden Markov model
- variational Bayesian inference
Record, P. M. (Accepted/In press). Novel activity monitoring system based on smart collar and variational Bayesian learning of multivariate autoregressive hidden Markov models. Paper presented at 8th European Conference on Precision Livestock Farming 2017, Nantes, France.