Machine learning based posture estimation for a wireless canine machine interface

Effective training and accurate interpretation of canine behaviors are essential for dog welfare and to obtain the maximum benefits provided by working dogs. We are developing a canine body area network based interface to incorporate electronic sensing and computational behavior modeling into canine training, where computers will be able to provide real time feedback to trainers about canine behavior. In this study, we investigated the accuracy of machine learning algorithms in identifying canine posture through wireless inertial sensing with 3-axis accelerometers and 3-axis gyroscopes. Data was collected from two dogs performing a sequence of 5 postures (sit, stand, lie, stand on two legs, and eat off the ground). A two-stage cascade learning technique was used: one for differentiating samples of behaviors of interest from transitions between behaviors, and one for posture classification of the behaviors. The algorithms achieved high posture classification accuracies demonstrating potential to enable a real time canine computer interface.