Animal behavior classification performance in a case deflection of an attached device

Cattle behavior data is core to the assessment of cow health. Behavior recognition systems using accelerometers are an appropriate solution. The accelerometer or accelerometer sensor is affordable and simple to set up, which collects accelerometer data mounted on cattle’s leg and collar, using machine learning algorithms to provide high classification performance. In our previous work, a cow behavior classifier was proposed based on synchronized collar and leg-mounted accelerometer data to increase classification performance. In the actual implementation context, we found that cow activity can affect the position of the collar mounted sensor; specifically, this sensor’s position can be deflected from the original position, resulting in changes in data obtained from the accelerometer. In this study, we simulated data from the sensor attached to the collar in the case of deflection to evaluate the influence on the results of behavioral classification: lying, walking, feeding, and standing. Using features (mean, RMS, range, and median) with a data window of 16 seconds (one sample/second) and the Random Forest (RF) algorithm, the evaluation results in our test cases with indicators (accuracy, precision, and sensitivity) have a similarity. These results show that the deflection of the collar mounted sensor does not significantly impact the behavioral classification results when applying the proposed classification system.