Purpose <p>Neonatal activity is an important physiological parameter in the neonatal intensive care unit (NICU). The degree of neonatal activity is associated with under and over-sedation and may also indicate the onset of disease. Activity may also cause motion noise on physiological signals leading to false readings of important parameters such as heart rate, respiratory rate or oxygen saturation or, in extreme cases, a failure to calculate the parameter at all. Here we report on a novel neonatal activity monitoring technology we have developed using a Random Forest machine learning algorithm trained on features extracted from a depth video stream from a commercially available depth sensing camera.</p> Methods <p>A cohort of twenty neonates took part in the study where depth information was acquired from various camera locations above and to the side of each neonate. Depth data were processed to provide features indicating changes corresponding to the activity of the neonate and then input into a Random Forest model which was trained and tested using a leave-one-out cross validation paradigm.</p> Results <p>Applying the thresholds found in training the Random Forest model during testing with leave-one-out cross validation, the mean (standard deviation) of the sensitivity and specificity of the optimal points and the corresponding area under the receiver operator curve (ROC-AUC) were 92.0% (8.8%), 93.2% (11.1%) and 97.7% (2.5%) respectively. The activity identified by the model also appeared to match well with noisy segments on the corresponding respiratory flow signal.</p> Conclusions <p>The results reported here indicate the viability of continuous non-contact monitoring of neonatal activity using a depth sensing camera system.</p>