Motion model enabled appearance prediction for partial human body tracking in robot follower

Abstract

Robot follower, a robot following its human operator, has found its application in many areas such as senior care, manufacturing, transportation, and etc. Tracking the target person is a key technique for the follower. In this paper, we present a new method for partial human body tracking, namely human feet tracking. Human feet tracking suffers from weak visual features and appearance variations, making it more critical to continuously update the foot appearance model. We propose to utilize the human motion model to predict foot appearance. It is achieved by first defining a motion phase to each human foot appearance. Due to the fact that the foot appearance across different motion cycles with the same motion phase is similar, we can predict the target appearance using the current motion phase and the target images stored from previous walking cycles. A phase labeled exemplar pool is built to serve the motion phase indexed appearance searching. We combine this phase labeled exemplar pool into particle filtering and have achieved robust human feet tracking.