Abstract

Enabling artificial systems to recognize human actions
is a requisite to develop intelligent assistance systems that
are able to instruct and supervise users in accomplishing
tasks. In order to enable an assistance system to be wearable,
head-mounted cameras allow to perceive a scene visually
from a user’s perspective. But realizing action recognition
without any static sensors causes special challenges.
The movement of the camera is directly related to the user’s
head motion and not controlled by the system. In this paper
we present how a trajectory-based action recognition can
be combined with object recognition, visual tracking, and a
background motion compensation to be applicable in such
a wearable assistance system. The suitability of our approach
is proved by user studies in an object manipulation
scenario.