A general descriptor for detecting abnormal action performance from skeletal data

We propose an action-independent descriptor for detecting abnormality in motion, based on medically-inspired skeletal features. The descriptor is tested on four actions/motions captured using a single depth camera: sit-to-stand, stand-to-sit, flat-walk, and climbing-stairs. For each action, a Gaussian Mixture Model (GMM) trained on normal motions is built using the action-independent feature descriptor. Test sequences are evaluated based on their fitness to the normal motion models, with a threshold over the likelihood, to assess abnormality. Results show that the descriptor is able to detect abnormality with accuracy ranging from 0.97 to 1 for the various motions.