Recognition of Activity States in Dairy Cows with SVMs and Graphical Models

Abstract

Activity patterns of dairy cattle have received increasing interest in recent years because they promise insights into health state and well-being. The fusion with data from additional sensor signals promises a comprehensive monitoring of activity patterns composed of sequences of single activity states. We used a combination of a Support Vector Machine (SVM), a state of the art classification method, and a Conditional Random Field (CRF). SVMs distinguish single states, whereas CRFs label state sequences under consideration of specified constraints. In a preliminary experiment, a Local Positioning System was combined with a heart rate sensor in order to estimate seven spatiotemporal activity states. The application of the CRF to the SVM result caused a slight increase in accuracy (5%) but a major improvement at the correct determination of long sequences (increasing length of the longest common subsequence from 3481 to 6207 periods). This robust detection of long lying sequences allowed for the unaffected extraction of the resting pulse.