Abstract/Summary

dentification of individual humans from RGB image data is well-established. However, in many domains, such as in healthcare or applications involving children, ethical issues have been raised around using traditional RGB image data because individuals can be identified from these data. The widespread availability of reliable depth data, and the associated human skeleton data derived from these data, presents an opportunity to differentiate between individuals while potentially avoiding individually identifiable features. Using skeleton data only, we developed a unique 20-dimensional bone segment length feature vector for 1,761 trials (1,759,980 image frames) of data, captured from 14 participants who engaged in a one-hour group intervention playing Xbox One Kinect Bowling twice-weekly for 24 weeks. We then evaluated our novel feature using representative batch processing (k-nearest neighbour) and real-time (multi-layer perceptron) models, validated against manually-labelled ground-truth data. Our results suggest that our skeleton feature can differentiate between instances (i.e., individuals) with an accuracy over all participants of 100% for batch processing and 96.57% in real-time, and deals well with class imbalances. Our results suggest that we can reliably differentiate between individual persons using only skeleton data derived from depth image data in medical research.