A method for predicting personalized pelvic motion based on body meta-features for gait rehabilitation robot

Training for balancing, which is governed by the motion of pelvis and thorax, is a key for gait rehabilitation. COWALK, which is a gait rehabilitation robot under development in our institute, is capable of pelvic motion training. In this paper, we describe a statistical method to generate pelvic motion which is considered to fit each person, i.e., personalized pelvic motion. We measured 14 anthropometric features of human and captured gait motion using an optical motion capture system from 113 healthy subjects. We setup a database of gait motion and body measurements; we define a 4 dimensional compact vector representation of pelvic motion, and body meta-feature, which is a weighted linear combination of the anthropometric measurements, to maximize statistical correlation between the former and the latter. To synthesize a personalized pelvic motion for a new subject, we search for k nearest neighbors in the space of body meta-feature (k-NN algorithm), and average the pelvic motions of them. We validate the algorithm using the database of 113 subjects by excluding each person, synthesizing a personalized pelvic motion for the subject, and comparing it with actual motion of the subject.