Manual observations either directly or by analysis of video recordings of dairy cow behaviour in loose housing systems are costly. Therefore progress could be made if reliable estimates of duration of walking and standing could be based on automatic recordings.
In this study we developed algorithms for the detection of walking and standing in dairy cows based on the output from an electronic device quantifying acceleration in three dimensions.
Ten cows were equipped with one movement sensor on each hind leg. The cows were then walked one by one in the alleys of the barn and encouraged to stand and walk in sequences of approximately 20s for the period of 10min. Afterwards the cows were stimulated to move/lift the legs while standing in a cubicle. The behaviour was video recorded, and the recordings were analysed second by second for walking and standing behaviour as well as the number of steps taken.
Various algorithms for predicting walking/standing status were compared. The algorithms were all based on a limit of a moving average calculated by using one of two outputs of the accelerometer, either a motion index or a step count, and applied over periods of 3 or 5s. Furthermore, we investigated the effect of additionally applying the rule: a walking period must last at least 5s.
The results indicate that the lowest misclassification rate (10%) of walking and standing was obtained based on the step count with a moving average of 3s and with the rule applied. However, the rate of misclassification given walking and standing differed between algorithms, thus the choice of algorithm should relate to the specific question under consideration.
In conclusion, the results suggest that the number of steps taken per time unit as well as the frequency and duration of walking and standing can be estimated with a reasonable accuracy.