Acceleration sensors in a Wireless Sensor Network (WSN) were used to monitor the behaviour of dairy cows. The data processing from 3D acceleration into behaviour classification (lying, standing or walking) was based on a two-steps method: first the distinction between lying and standing/walking was based on the computed average values during standing and the calculated distance from the actual values and the standing average. Secondly the distinction between standing and walking was based on the variance in 10 successive measurements. With this method both on-node data processing and state-based triggering are possible. This classification was tested during one day with three cows were video recordings were available as reference data. The calculated behaviour corresponds highly with the observed behaviour, the distinction between lying and standing/walking is the same during 99.15% of the time. 99.5% Of the observed walking periods have a matching calculated walking period (if measurements are available).