Personalized Gestures Through Motion Transfer

Abstract

With the growing ubiquitousness of pervasive sensing and toward ambient intelligence, pervasive surveillance becomes a very real privacy threat, where private gesture interaction is likely to be observed and automatically interpreted by other (even benign) pervasive intelligence tools. We propose motion transfer, the example-guided modification of motion to translate from default motion and gesture interaction alphabets to personal ones. Apart from privacy, incentive to use personalized gesture interaction alphabets include convenience as well as physical handicaps (i.e., inability to conduct certain movements). We demonstrate the concept using motion transfer in RGB-video data. We further show that the approach is feasible also for point-cloud-based gesture recognition methods. In particular, we implemented an end-to-end model for human motion transfer with 3D (<italic>x</italic>-<italic>y</italic>-time) or 4D (<italic>x</italic>-<italic>y</italic>-<italic>z</italic>-time) point-cloud datasets. Point-cloud-based motion transfer is a privacy protecting way of customizing gestures to control devices, hence lowering the risk of disclosing the nature of interaction to surrounding pervasive surveillance installations.