Motion Recognition Employing Multiple Kernel Learning of Fisher Vectors Using Local Skeleton Features

Yusuke Goutsu, Wataru Takano, Yoshihiko Nakamura; Proceedings of the IEEE International Conference on Computer Vision (ICCV) Workshops, 2015, pp. 79-86

Abstract


We propose a skeleton-based motion recognition system focusing on local parts of the human body closely related to a target motion. In this system, a skeleton feature is com- posed of a sequence of relative positions between paired joints calculated by Inverse Kinematics. Several joints of skeleton model are connected as a Local Skeleton Feature. The temporal sequence is modeled as human motion model by using Hidden Markov Model. Motion features are rep- resented as Fisher vectors parameterized by the human mo- tion models, and weighted and integrated by using Multiple Kernel Learning. This system makes it possible for robots to recognize human actions in our daily life. The experimental results based on two datasets show an improvement in per- formance of classification rate, which shows that the design of motion feature is effective for motion recognition.

Related Material


[pdf]
[bibtex]
@InProceedings{Goutsu_2015_ICCV_Workshops,
author = {Goutsu, Yusuke and Takano, Wataru and Nakamura, Yoshihiko},
title = {Motion Recognition Employing Multiple Kernel Learning of Fisher Vectors Using Local Skeleton Features},
booktitle = {Proceedings of the IEEE International Conference on Computer Vision (ICCV) Workshops},
month = {December},
year = {2015}
}