Classification of sows’ activity types from acceleration patterns using univariate and multivariate models

Automatic monitoring of animal behavior in livestock production opens up possibilities for on-line monitoring of, among others, oestrus, health disorders, and animal welfare in general. The aim of this study is to use time series of acceleration measurements in order to automatically classify activity types performed by group-housed sows. Extracts of series collected for 11 sows are associated with 5 activity types: feeding (FE), rooting (RO), walking (WA), lying sternally (LS) and lying laterally (LL). A total of 24 h of three-dimensional series is used. One univariate model and four multivariate models are used to describe all five activity types. Three multivariate models differ in their variance/covariance structure; a fourth alternative multivariate model (MU) simply combines the 3-axes of the univariate model, assuming independence. For each model, the activity-specific parameters are estimated using the EM algorithm. The classification method, based on a Multi-Process Kalman Filter provides posterior probabilities for each of the 5 activities, for a given series. For the univariate model, LL is the activity which is best recognized by the 3-axes; FE, RO and WA are best recognized by one particular axis; LS is poorest recognized. The average results are improved by using all four types of multivariate models. The percentages of activity recognition are similar among the multivariate models. By grouping the activity types into active (FE, RO, WA) vs. passive (LS, LL) categories, the method allows to correctly classify 96% of the active category and 94% of the passive category.